Professional Documents
Culture Documents
Probability and Statistics in Engineering
Probability and Statistics in Engineering
in Engineering
Probability and Statistics
in Engineering
Fourth Edition
William W. Hines
Prqjessor Emeritus
School of Industrial and Syste.rns Engineen"n.g
Georgia fnstit,ge of Technology
Douglas C. Montgomery
Professor of Engineering and Statistics
Department of Industrial Engineen"ng
Arizona State University
David M. Goldsman
Professor
School of Industrial and Sys!erns Engineering
Georgia Institute of Technology
~
WILEY
John Wiley & Sons, Inc.
Acquisitions Edi!or Wayne Anderson
Associate Editor Jenny Welter
Marketi.'1g Manager Kathedne Hepburn
Se::ior Production Editor valerie A. mrgas
Senior Designer Dawn Stamey
Cover Image Alfredo PasiekalPhoto Researchers
Production Management Services Argosy Publishing
This book was set in 10/12 Tunes: Roman by Argosy Publishing and printed and bound by Hamil.:on
Printing. The cover was printed by Phoenix Color.
Copyright 2oo3© John WIley &: Sons, me, All rights reserved.
No p31t of this publication may be rcproduced. stored in a retrieval system. or transmitted in any form or by any
means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except ;as permitted under Sec·
tions 107 or 108 of the 1976 United St:ates CopyrightAct. witb.outeither the poor written permission of the Pub-
lisher or au:horiz.ation through payn::tent of the appropriate per-eopy fee to the Copyright Clearance Center. 222
Rosewood Drive. Danvers, MA OHi23. (978) 750-S400. fux (973) 750-44-70. Requests to the Publisher for per-
mission should be nddressed to the PermissionS D<?artment, John Wuey & Sons, Inc., 111 River Street, Hobok-en.
1') 07030, (201)748-{)OCQ, mx
(201)748-6088, E-mail: PERMJffiQ@W!LEYCOM,
10 order books or fur custOtr..er service please call1(800}-CALL vm.EY (225-5945),
This book is written for a first course in applied probability and statistics for undergradu-
ate smdents in engineering, physical sciences, and management science curricula. We have
found that the text can be used effectivelY as a two-semester sophomore~ or junior-level
course sequence, as well as a one~semester refresher course in probability and statistics for
first-year graduate students.
The text has undergone a major overhaul for ille fourth edition, especially wiill regard
to many of the statistics chapters. The idea has been to make the book more accessible to a
wide audience by including more motivational examples, rea1~world applications. and use-
ful computer exercises. With the aim of making the course material easier to learn and eas-
ier to teach, we have also provided a convenient set of cOurse notes, available on the Web
site www.wiley.comico.11ege/hines. For instructors adopting the text, the complete solu-
tions are also available ou a password~protected portion of this Web site.
Structurally speaking, we start ille book off with prObability illeory (Chapter I) and
progress through random variables (Chapter 2), functions of random variables (Cbapter3),
joint random variables (Chapter 4), discrete and continuous distributions (Chapters 5 and
6), and normal distribution (Chapter 7). Then we introduce statistics and data description
techniques (Chapter 8). The statistics chapters follow the Same rough outline as in the pre-
vious edition, namely, sampling distributions (Chapter 9), parameter estimation (Chapter
10), hypothesi, testing (Chapter 11), single- and multifactor design of experiments (Chap-
ters 12 and 13), and simple and multiple regression (ChapterS 14 and 15). Subsequent spe-
cial-topics chapters include nonparametric statistics (Chapter 16), qUality control and
reliability engineering (Chapter 17), and stochastic processes and queueing theory (Chap-
ter 18), Frnally, there is an entirely new Chapter, on statistical techniques for computer sim~
ulation (Chapter 19)-.--perhaps ille first of its kind in this type of statistics text.
The chapters that have seen the most substantial evolution are Chapters 8-14. The dis-
cussion in Chapter 8 on descriptive data analysis is greatly enhanced over that of the previ-
ous edition~s. We also expanded the discussion on different types of interval estimation:in
Chapter 10. In addition. an emphasis has been placed On real-life computer data analysis
examples, Throughout ille book, we incorporated other SlIUctural changes. In all chapters,
we included new examples and exercises; including numerous computerflbased exercises.
A few words on Chapters 18 and 19. Stochastic processes and queueing theory arise
naturally out of probability, and wefeelillat Chapter 18 serves as a good introduction to the
subject-nonnally taught in operations research. management science, and certain engi-
neering disciplines. Queueing theory bas garnered a great deal of use in such diverse fields
as telecommunications, manufacturing, and production planning, Computer simulation. the
topic of Chapter 19, is perhaps the most widely used tool in operations research and man~
agement science, as well as in a number of physical sciences. Simulation :ma..'7ies all th.e
tools of probability and statistics and is used in everyJring from financial analysLs to factory
control and planning, Our text provides what amounts to a simulation mL.1icourse, covering
the areas of Monte Carlo experimentation. random number and variate generation, and
simulation output data analysis.
v
vi Preface
We are grateful to the following individuals for their help during the process of com-
pleting the current revision of the text. Christos Alexopoulos (Georgia fustitute of Tech-
nology), Michael Cararnanis (Boston University), David R. Clark (Kettering University),
J, N. Hool (Auburn University). Johu S. Ramberg (University of Arizona), and Edward J.
Williarus (Gniversity of )":lichigan - Dearborn), served as reviewers and provided a great
deal of valuable feedback. Beamz Valdes (Argosy Publishing) did a wor.derful job super-
vising the typesetting and page proofing of the text, and Jennifer Welter at Wiley provided
great leadership at every tum. Everyone was certainly a pleasure to work with. Of course,
we thank our families for their infinite patier:.ce and support throughout the endeavor.
2-1 Introduction 33
5, Some Important Discrete
2-2 The Distribution Function 36
Distributions 106
2-3 Discrete Random Variables 38
2-4 Continuous Random Variables 41 5-1 Introduction 106
2-5 Some Characteristics ofPi;.}tnOUtiODS 44 5w2 Bernoulli Trials a.."1d the Bernoulli
2-6 Chebyshev', Ioequality 48 Dis:ributior.. 106
2-7 SuInmary 49 5~3 The Binomial Dis:ri::outio:l. 108
2-8 Exercises 50 5-3.1 Mean and Variance of the Binomial
Distribution 109
3. Functions of One Random Variable and 5~3.2
The Cumulative Binomial
Exp~tion 52 Distribution - ltO
5~3.3 An Application of the Binomial
3-1 Introduction 52 Distribut:!on 111
3·2 Equivalent Events 52 54 The Geometric Distribution 112
3·3 Functions of a Discrete Random 5-4,1 Mean and Variance of the Geometric
Variable 54 Distribution 113
3-4 Continuous Functions of a Continuous 5~5 The Pascal Distribution 115
Random Variable 55 5-5.1 Mear. a::.d Variance 0: the Pascal
3-5 Ex,pectation 58 Distric:.r:iolJ 115
3-6 Approximations to E[H(Xl] and I1H(X)] 62 5-6 The Multinomial DistnlJotioa 116
3-7 The Moment-Gene.'1Iting Function 65 5-7 The Hypergeometric Distribution 117
3-8 Sum:na:y 67 5-7_1 Mean ar.d Variance ofthc Hyper-
3-9 E--:ercises 68 geometric Distribution 1 I8
viii Contents
9-2.2 Finitc Populations and Enumerative 10"8 Othe:: Interval EstiJ:nation Procedures 255
Studies 204- 10-8.1 P:ediction In:eIYals 255
9-3 The Chi-Square Distribution 205 10~8,2 Tolerance Ir.tervals 257
9-4 The t Distribution 208 ;0-9 Su:n:nru:y 258
9~5 The F Distribution 211 10-10 Exercises 260
9-6 Summary 214
9-7 Exercises 214 11. Tests of Hypotheses 266
r ~,,
216 11-1 Introduction 266
(~: ,Parameter Estimation
!l·Ll Statistk:al Hypotheses 266
10-1 Point Estimation 216 11-1.2 Type! and Type II Errors 267
10-1.1 Properties of Estimators 217 !l-13 One-Sided and Two-Sided
10-1.2 The Method of Maximum Hypotheses 269
Likelihood 221 11-2 Tests of Hypotheses on a Single
10-1.3 The Metbod of Moments 224 Sample 271
10-1.4 BaycsianInferencc 226 11 ~2.1 Tests of Hypotheses on the Mean
10-1.5 Applications to Estimation 227 of a Normal Distribution. Variance
10·1.6 Precision of Estimation: KnoVitn 271
The Standard:Error 230 11~2.2 Tests of Hypotheses on the Mean
10-2 Single-Sample Confidence Interval of a Kormal Disttfoution. Varianee
Estimation 232 Unkao~n 27&
10-2.1 Confidence Llterval on the Mean 11-2.3 Tests of Hypotheses on the
of a Normal Distribution, Variance Variance of a Normal
Known 233 Distribution 281
10-2.2 Confidence Interval on the Mean 11-2.4 Tests of Hypotheses 0:.1 a
of a Nonnal Distribution, Variance Proportion 2&3
Unknown 236 1l~3 Tests of Hypotheses on Two Samples 286
10-2.3 Confidence Interval on the 11-3.1 Tests of Hypotheses on the Means
Variance of a Non:u.ll of Two Normal Distributions,
Distribution 238 Variances KnO\\T. 2&6
10-2A Confidence Interval on a 11~3.2 TesLS of Hypotheses on the Means
P;oportion 239 of1\vo Normal Distributions,
10-3 Two-Sampl. Confidence Interval Variances Unknovro 288
Es~on 242 11-33 The Paired t- Test 292
10-3_1 Confidence Interval on the 11 ~3.4 Tests for the Equality ofIwo
Difference between Means of Two Variances 294
Normal Distr1outions, Variances 11 ~3.5 Tests of Hypotheses on 1\vo
Known 242 Proportions 296
10-3,2 Confidence Interval on the 11-4 Testing for Goodrless of Fit 300
Difference between Mea:n:s of'IWo 11-5 Contingency Table Tests 307
Normal Distnoutious. Va.."'iatices 11-6 Sample Computer Output 309
Unknown 244 11-7 Summary 312
10-3.3 Confider.ce Interval an IJ.: - J4 for 11-8 Exercises 312
Pl'rired Observations 247
10-3.4 Confidence Interval 0:.1 the Ratio 12_ Design and Analysis of Single.
ofVariauces of1'v,"o Normal Factor Experiments: The Analysis
IAstributions 248 of Variance 321
10-3.5 Confidence bterval on the
Difference between Two 12-1 The Completely Randomized Single-Factor
Propor::ions 250 Experiment 321
10-4 Approximate Confidence Intervals in 12-:_1 AD Example 321
Maximum Likelihood Esti:trtation 251 12-1.2 The Analysis of
10-5 Simultaneous Confidence Intervals 252 Variance 323
10-6 Bayesian Confidence Intervals 252 12-1.3 Esti.mation of the Model
10-7 Bootstrap Confidence Intervals 253 Parameters 327
X Contents
12-1.4 Residual Analysis and Model 14. Simple Linear Regression and
Checking 330
Correlation 409
12-1.5 An linbaLmcedDe,ign 33\
12-2 Tests on Individual Treaonen~ 14-1 Simple Linear Regression 409
Means 332 14-2 Hypot..~esis Tes!ing in Simple Linear
12-2.1 Orthogonal Contrasts 332 Regression 414
12-2.2 'f\Jkey's Test 335 14-3 lnterval Estimation in Simple Linear
12-3 The Random-Effects Model 337 Regression 417
12-4 The Randomized Block Design 341 14-4 Prediction of New Observations 420
12-4.1 Design and Statistical 14-5 Measuring the Adequacy of the Regression
Analysis 341 Model 421
12-4.2 Tests on Individual Treatment 14-5.1 Residual Analysis 421
Mea:lS 344 14-5.2 The Lack-of·Fit Test 422
12-4,3 Residual Analysis and Model 14-5.3 The Coefficient of
Checking 345 Dete:rmination 426
12-5 Deterrr.:n:.ng Sample Size in Si."l.gle-Factor 14-6 Transformations to a Straight
Experiments 347 Line 426
12~6 Sample Co:rnputer Ou~ut 348 14w-7 Correlation 427
12·7 Summary 350 14~8 SampleComputerOu~ut 43:
12-8 Exercises 350 14·9 Summary 431
14-10 Exercises 432
13. Design of Experiments with Several
Factors 353 15. Multiple Regressiol' __43_7_ _ __
--~~--~------------
13-1 Examples of Experimental Design 15~ 1 Multiple Regression Models 437
Applications 353 15~2 Estimation of the Parameters 438
13-2 Factorial Experi.-uents 355 15-3 Confidence Intervals in Multiple Linear
13-3 Tvio-Factor Factorial Regression 444
Experiments 359 15-4 Prediction of New Observations 446
13-3.1 Statistical Analysis of the Fixed~ 15~ 5 HypOthesis T~'ting in Multiple Linear
Effects Model 359 Regression 447
13-3.2 Model Adequacy 15-5,1 Test for Significance of
Checking 364 Regression 447
13-3.3 One Observation per 15-5.2 Tests on Individual Regression
Cell 364 Coefficients 450
13-3.4 The Random·Effects 15·6 MeasuresofMcdelAdequacy 452
Model 366 15-6.1 The Coefficient of Multiple
13-3.5 The Mixed Model 367 Detennination 453
13-4 General Factorial Experiments 369 15-6.2 ResidualAnalysis 454
13·5 The 2' FaC'.orial Design 373 15-7 Polynomial Re:gression 456
13·5.1 The2'Design 373 15-8 Indicator Variables 458
13·5.2 The 2' Design for k ~ 3 15-9 The Correlation Matrix 461
r"'actors 379 15-10 Problems in Multiple Regression 464
13-5.3 A Single Replicate of the 2~ 15-1O.! Multicollinearity 464
Design 386 15-10.2 Influential Observations m
13·6 Confounding in the 2' Design 390 Regression 470
13~7 Fractional Replication of the 2k 15-10.3 Autocorrel1.tion 472
Design 394 15,11 Selection of Var'.ab1e.s in Multiple
13-7.1 The One-Half Fracuon of the 2' RegressioQ 474
Design 394 15-11.1 The Model·Buililing
13M7.2 Smaller Fractions: The 2k -p Problem 474
Fractional Factorial 400 15-11.2 Computational Procedures for
13·8 Sample Computer Output 4D3 Variable Selection 474
13·9 SlUlllllary 404 15-12 SUlIlIrulJ:)' 486
13·10 Exercises 404 15-13 Exercises 4&6
Contents xi
19-4.2 Common Random Chart VIII Operating Characteristic Curves for the
Numbers 593 Random~Effects Model Analysis of
19-4.3 Antithetic Rmldom VarianCe 623
Numbers 593 Table IX C:itical Values for the WUcoxon Two-
19-4.4 Selecting the Best Sample Test 627
Syst= 593 Table X Critical Values for the Sign
19-5 SlllIllDZIY 593 Test 629
19-6 Exercises 594 Table XI Critical Values for the Wdcoxon Signed
Rank Test 630
Appendlx 597 Table XTI Percentage Points of the Studentized
Range Statistic 631
Table I Cum\llative Poisson Table xm Factors for Quality-Control
Distribution 598 Cbarts 633
Table II Cumulative Standard Nonna:! Table XIV k Values for One-Sided and Two-Sided
Distribution 601 Tolerance Intervals 634
Table ill Percentage Points of the Table XV Rmldom Numbers 636
y} Distribution 603
Table IV Percentage Points of the References 637
t Distribution 604
Table V Percentage Pomts of the Answers to Selected Exercises 639
F DiJ;tribution 605
CbartVI Operating Cl:iaraJ;teristic Curves 610
Index 649
Cbart\lI Operating Characteristic Ccrves for the
F)X£d-Elfects ModelAnalysiJ; of
V:u:iance 619
Chapter 1
An Introduction to Probability
1-1 INTRODUCTION
Since professionals working in engineering and applied science are often engaged in both
the analysis and the design of systems where system component characteristics are nonde-
tetnlinistic. the understamling and utilization of probability is essential to the description.
design, and analy&is of such systems. Examples reflecting probabilistic behavior are abun-
dant, and in fact. true deterministic behavior is rare, To illusrrate. cons:der the description of
a variety of product quality or perfonnance measurements: the operational lifespan of
mechanical and/or electronic systems; the pattern of equipment failures; the occurrence of
natural phenomena such as sun spots ortomados; particiecouots from a radioactive source;
travel times in delivery operations; vehicle accident counts during a given day on a section of
freeway; or customer waiting times in line at a branch bank,
The termprobability has come to be widely used in everyday life to quantify thc degree
of belief in an event of interest. There are abundant examples, such as the statements that
"there is a 0.2 probability of rain showers" and "the probability that brand X personal com-
puter will survive 10,000 hours of operation without repair is 0.75." In tltis chaprer we intro-
duce the basic structure. elementary concepts, and methods to support precise and
unambiguous statements like those above.
The formal study of probability theory apparently originated in the seventeenth and
eighteenth centuries in France and was motivated by the study of games of chance. WIth lit-
tle formal mathematical understructure, people viewed the field with some $kepticism~
however, this view began to change in the nineteenth century, when a probabilistic model
(description) was developed for the behavior of molecules in a liquid. This became known
as Brownian motion r since it was Robert BrO\vn, an English botanist, who firs.t observed the
phenomenon in 1827. In 1905, Albert Einstein explained Brownian motion under the
hypothesis that particles are subject to the continual bombardment of molecules of the sur-
rounding medium. These results greatly stimulated interest in probability, as did the emer-
gence of the telephone system in the latter part of the nineteenth and early twentieth
centuries. Since a physical connecting system was necessary to allow for the interconnec-
tion of individual telephones, with call1engths and interdemand intervals displaying large
variation~ a strong motivation emerged for developing probabilistic models to describe this
system's behavior.
Although applications like these were rapidly expanding in the early twentieth cenrury,
it is generally thought that it was not until the 19305 that a rigorous mathematical structure
for probability emerged. This chapter presents basic concepts leading to and including a
definition of probability as well as some results and methods useful for problem solution.
The emphasis th:roughout Chapters 1-7 is to encourage an understanding and appreciation
of the subject, with applications to a variety of problems in engineering and science. The
reader should recognize that there is a large, rich field of mathematics related to probabil-
ity (hat is beyond the scope of this book.
1
2 Chapter 1 An Introduction to Probability
Indeed, our objectives in presenting the basic probability topics considered in the cur~
rent chapter are threefold. First, these concepts enhance and enrich our basic understanding
of the world in which we live. Second, many of the examples and exercises deal with the
use of probability concepts to model the behavior of real-world systems. Finally, the prob-
ability topics developed in Chapters 1-7 provide a foundation for the statistical methods
presented in Chapters 8-16 and beyond. These statistical methods deal with the analysis
and interpretation of data, drawing inference about populations based on a sample of units
selected from them, and with the design and analysis of experiments and experimental
data. A sound understanding of such methods will greatly enhance the professional capa-
bility of individuals working in the data-intensive areas commonly encountered in this
twenty-first century.
The set whose elements are the integers 5, 6, 7, 8 is a finite set with four elements. We could denote
this by
A=[5,6,7,8).
Note that 5 E A and 9 ~ A are both true.
If we say thatA is the set of all real numbers between 0 and 1 inclusive, we might also denote A by a
defining property as
A= [x:x E R, O$x'; 1),
where R is thc set of all real numbers.
In the real plane we can consider points (x, y) that lie on a. given lineA, Thus, the condition for inclu-
sion for A requires (x, y) to satisfy ax..;.. by = c, so that
A = (x, y): x E R, y E R. ax+by=c).
where R is the set of real numbers.
The universal set is the set of all objeets under consideration; and it is generally
denoted by u: Another special set is the null set or empty set, usually denoted by 0. To mus-
trate tbls concept> consider a set
A={X:XE R,x'=-l).
'The universal set here is R. the set of re-al numbers. Obviously, set A is empty, since there
are no real numbers having the defining property :? :;:: -1. We should point out that the set
(0)0'0.
If two sets are considered, say A and B, we call A a subset of B, denoted A c B, if each
element in A is also an element of B. The sets A and B are said to be equal (A = B) if and
only if A c B and B c A. As direct consequences of this we may show the following:
An interesting consequence of set equality is that the order of element listing is imma-
terial, To illustrate, let A (a, b, c) and B = (c, a, b). Obviously A ; B by our definition.
Furth=ore, when defining properties are used, the sets may be equal althougb the defin-
ing properties are outwardly different. As an example of the second consequence. we let.4. : : :
{x:.x E R, where x is an even~ prime number} andS """ {x: x+ 3 = 5}. Since the integer 2 is
the only even prime, A B. =
We now considet some operations on sets. Let A and B be any subsets of the universal
set U. Then the following hold:
1. The complemernof.4 (with tespectto U) is the set made up of the elements of Uthat
do not belong ro A. We denote this complementary set as A. That is,
A=(X:XE u,x>!c A}.
2. The intersection of A and B is the set of elements that belong to both A and R We
denote the intersection as A ('j R In other words,
AnB (x:xEAandxEB).
We should also note that A n B is a set, and we could give this set some designator,
such as C.
3. The union of A and B is the set of elements that belong to at least one of the ,etsA
and B. 1f D represents the union, then
'E~~ie'l;~
Let Ube the set of leiters ttl the alphabet, that is, U = (*; * is a letter of the English alphabet}; and let
A {"': ;;0 is :a yowell and B '=' {*: * is one ofthe1etrers a, b. c}. As a consequence of the definitions,
:E~l!leI1
If the unlversal.setis defined as U =- {1, 2, 3.4, 5, 6, 7), and threesubsets,A = {l. 2. 3},B= {2.4, 6}.
C::;: (1. 3, 5, -"}}. arc defined, then we see immedia:ely from the de:fi.citions that
A= (4.5.6. II. B= {U.5. 71=C.
A c, B= (1. 2. 3, 4. 6). AUC=!1.2.3.5.7}. BvC=U,
A"8~{2I, o4nC=(l,3), 80C=0.
The Venn diagram can be used to illustrate certain set operations, A rectangle is drawn
to represent the universal set U. A s~setA of U is represented by the region, within a cir~
de drawn inside the rectangle. Then A will be represented by the area of the rectangle out-
side of the circle, as illustrated in Fig. I-I. Using this notation, the intersection and union
are illustrated in Fig. 1-2.
The operations of intersection and union may be extended in a straightforward manner
to accommodate any finite number of sets. In the case of three sets, say A. B, and C, A u B
u C has the property that Au (B u C) = (A u B) u C, which obviously holds since both
sides have identical members. Similarly, we see that A n B n C = (A n B) 0 C = A n
(B " C). Some important laws obeyed by sets relative to the operations previously defined
are listed below.
Identity laws: Au0=A. AnU=A,
AuU=U. An0=0.
De Morgan's law:
-- - -
Au B =A n B, A 0 B =A u B.
Associative iaws: A u (B u C) = (A u B) u c,
A nCB n C)=(A nB) n C.
Distributive laws; A U (B n C) = (A u B) n (A u C),
A n (B u C) = (A n B) u (A n C).
The reader is asked in Exercise 1~2 to illustrate some of these statements with Venn dia~
grams. Fonnal proofs are usually more lengthy.
u u
(aJ (bJ
Figure 1~2 The interseetion and union of two sets in a Venn diagram.
(a) The intersection shaded. (b) The union shaded.
In the case of more than three sets, we usc a subscript to generalize. Thus, if n is a pos-
itive integer, andE j , E 2, .. , , En are given sets, thenE j nE 2 n ... n En is the set of elements
belonging to all of the sets, and E I U E2 U ... U En is the set of elements that belong to at
least one of the given sets.
If A and E are sets, then the set of all ordered pairs Ca, b) such that a E A and bEE is
called the Cartesian product set of A and B. The usual notation is A x B. We thus have
AxB~{(a, b):aE AandbE B).
Let r be a positive integer greater than 1, and let A j ' ••• ,Ar represent sets. Then the Carte-
sian product set is given by
Al xA 2 x .. ' xA r = {(ai' az, ... ,a r): ajE AJorj= 1, 2, ... ,r}.
Frequently, the number of elements in a set is of some importance, and we denote by
n(A) the number of elements in set A. If the number is finite, we say we have afinite set.
Should the set be infinite, such that the elements can be put into a one-to-one correspon-
dence with the natural numbers, then the set is called a denumerably infinite set. The non-
denumerable set contains an infinite number of elements that cannot be enumerated. For
example, if a < b, then the set A = {x E R, a':::; x':::; b} is a nondenumerable set.
A set of particular interest is called the power set. The elements of this set are the sub-
t,.
sets of a set A, and~ common notation is {O, 1 For example if A = {I, 2, 3 }, then
~ ~ (O, (I), (2), (3), p, 2), (1,3), (2,3), (1,2,3)).
necessity. The· set of possible outcomes is called the sample space, and these outcomes
define the particular idealized experiment. The symbols <=& and :t are used to represent the
random experiment and the associated saITlple'space·:-
Following the terminology employed in the review of sets and set operations, we will
classify sample spaces (and thus random experiments). A discrete sample space is one
in which there is a finite number of outcomes or a countably (aenumerably) infinite
number of outcomes. Likewise, a continuous sample space has nondenumerable (uncount-
able) outcomes. These might be real numbers on an interval or real pairs contained in
the product of intervals, where measurements are made on two variables following an
experiment.
To illustrate random experiments with an associated sample space, we consider several
examples.
:Ji~i!#~i~;i~,'
~: ·"T~~.s_~~e coin and observe the "up" face.
ff,: {H,n
'&2: Toss a true coin three times and observe the sequence of heads and tails.
ff,: {HHH, HHI; HIli, HIT, THH, THT, TTH, TIT},
~3: Toss a true coin three times and observe the total number of heads.
ff,: CO, 1,2, 3}.
'1g5: An automobile door is assembled with a large number of spot welds. After assembly, each
weld is inspected, and the total number of defectives is counted.
'!i5: (0, 1,2, ... ,X}, whereX= the total number of welds in the door.
1-3 Experiments and Sample Spaces 7
~G: A cathode:ay rube is manufactured. put on life test. ar.d aged to failure. The elapsed time
(in hoUts) at faih.t..-e is recorded.
Ef,: {UE R, ,;"0).
This set is uncountable.
'tj7: A monitor records the emission count from a 7adioactive source in one minute.
Ef,: {O, 1. 2, ... }.
"£8: Two key solder joints On a printed circuit board are inspected with a probe as well as-visu-
ally, and each joint is classified as gqod. G, or defective, D, requiring rework or scrap.
Ef,: {GG, GD,DG,DD).
e.g9: In a particular chemical plant the volume produced per day for a particular product ranges
between a minimum value, fl, and a maximum value, c, which corresponds to capacity. A
day is randomly selected and the amount produced is observed.
Ef,: (X:XE R, b:;;x';c).
i,O: An extrusion plant is engaged i::l making up an order for pieces 20 feet long. Inasmuch as
the trim ~eration crea~es scrap at both ends, the extruded bar rr.us~ exceed 20 feet.
Because of costs involved, the amount of scrap is critical. A bar is extruded. trimmed, and
iinished. and the totalleagth of Smtp is measured.
:flO: {r.xeR,x>O}.
~! 1: In a missile launch, the three components of velocity are mentioned from the ground as a
function of time. At 1 minute after launch these are printed for a control unit.
9' .l: «v..... V,Y' v): Vx' vJ1 v; arc real numbers}.
i 11: In the preceding example, the ve:ocity components are contin:.l.ously recorded for 5
minutes.
9'12: The space lS complicated here. as we have all possible realizations of the functions v.>:(t).
and vt(t} for 0 S tS 5 minutes to consider.
8 Chapter 1 An Introduction to Probability
All these examples have the characteristics required of random experiments. With the
exception of Example 1-19, the description of the sample space is straightforward, and
although repetition is not considered, ideally we could repeat the experiments. To illustrate
the phenomena of random occurrence, consider Example 1-8. Obviously, if ~l is repeated
indefinitely, we obtain a sequence of heads and tails. A pattern emerges as we continue the
experiment. Notice that since the coin is true, we should obtain heads approximately one-
half of the time. In recognizing the idealization in the model, we simply agree on a theo-
retical possible set of outcomes. In ~ I' we ruled out the possibility of having the coin land
on edge, and in ~6' where we recorded the elapsed time to failure, the idealized sample
space consisted of all nonnegative real numbers.
1-4 EVENTS
An event, say A, is associated with the sample space of the experiment. The sample space is
considered to be the universal set so that event A is simply a subset of ':t. Note that both 0
and ':t are subsets of ':t. As a general rule, a capital letter will be used to denote an event. For
a finite sample space, we note that the set of all subsets is the power set, and more gener-
ally, we require that if A c ':t, then A' c ':t, and if AI' A1 , ••• is a sequence of mutually exclu-
sive events in ':t, as defIned below, then U~"'l Ai c g. The following events relate to
experiments 'i; I' '"t;;1' ••• , 'i; 10' described in the preceding section. These are provided for illus-
tration only; many other events could have been described for each case.
Definition
If an experiment %has sample ~ce ff and an event A is defined on ff, then peA) is a real
number called the probability of event A or the probability of A, and the function P(·) has
the following properties:
1.. 0:0; P(A):O; 1 for each event Aof ff.
2" PC::I') = 1..
3. For any finite number k of mutually exclusive events defined on fJ,
(k 1 k
pi UA, 1= lp(AJ
,,1:::::1 J IC::d
P(QA}~P{AJ
Note that the properties of the definition given do not tell the experimenter how to
assign probabilities; however. they do restrict the way in which the assignment may be
accomplished. In practice, probability is assigned on the basis of (1) estimates obtained
from previous expeJ;ience or prior observations, (2) an analytical consideration of experi-
mental conditions, or (3) assumption.
To iliustrate the assigmnent of probability based on experience, we consider the repe-
tition of the experiment and the relative frequency of the occu.rrence of the event of interest.
This notion of relative frequency has intuitive appeal., and it involves the conceptual
repetition of an experiment and a counting of both the number of repetitions and the num-
ber of times the event in question occurs. More precisely. ~ is repeated m times and !:\.vo
events are denoted A and B. We let mA and mB be the number of times A and B occur in the
m repetitions.
Definition
The value fA = mA1m is called the relative frequency of event A. It has the following
properties:
1. 0 ~/A::; I.
2, IA = 0 if and only if A never oeeurs, and I" =I if and only if A occurs on every
repetition.
3, If A and B are mutually exclusive events, thenl" ~8 =1" +la.
10 Chapter 1 An Introduction to Probability
As rn becomes large, fA tends to stabilize. That is, as the number of repetitions of the
experiment increases, the relative frequency of event A will vary less and less (from repetition
to repetition). The concept of relative frequency and the tendency toward stability lead to one
method for assigning probability. If an experiment ~ has sample space g and an event A is
defined, and if the relative frequency fA approaches some number PA as the number of repeti-
tions increases, then the number PA is ascribed to A as its probability, that is, as rn """'""7 00,
(1-1)
i == 1, 2, .. , , n,
while
PI +P2 + '" +Pn == 1,
P(A)= L,Pi' (1-2)
i:Ci GA
This is a statement that the probability of event A is the sum of the probabilities asso-
ciated with the outcomes making up event A, and this result is simply a consequence of the
definition of probability. The practitioner is still faced with assigning probabilities to the
outcomes, ej , It is noted that the sample space will not be finite if, for example, the elements
ei of g are countably infinite in number. In this case, we note that
Pi~O, i= 1, 2, ... ,
and n(A) outcomes are contained in A. Counting methods useful in determining n and n(A)
will be presented in Section 1-6,
1-5 Probability DefInition and Assignment 11
Suppose the coin in Example 1-9 is biased so that the outcomes of the sample space g' = {HHH, Hm;
HTH, HIT, THH, THT, TTH, TIT} haveprobabilitiesPJ =-f.?,pz =
4 8
P3 = f;-, -
17,
P4 =~'P5=fr'P6=1:;,
.
P7=Ti'PS =21' where e1 =HHH. ez =HHI, etc.lfwe let event A be the event that all tosses YIeld the
I 8 I
same face, then P(A) = 27 + 27 = 3'
e-2 ·21-1
i=I,£, ...
Pi = (i-I)l
=0. otherwise.
where Pi is the probabili[)' that the monitor will record a count outcome of i-I during a I-minute inter-
val.lfwe consider event A as the event containing the outcomes 0 and 1, then A = {O, I}, and P(A) =
PI + P2 = e- 2 + 2£-2 = 3[2 =. 0.406.
Consider Example 1-9. where a true coin is tossed three times, and consider event A, where all coins
show the same faco. By equation 1-3,
since there are eight total outcomes and hvo are favorable to event A. The coin was assumed to be true,
so all eight possible outcomes are equally likely.
•_ _ .: ;_'- __ .~v~-"-,
Assume that the dice in Example 1-11 are true, and consider an event A where the sum of the up faces
is 7. Using the resulrs of equation 1-3 we note that there are 36 outoomes, of which six are favorable
to the event in questiOI!, so that P(A) = i.
Note that Examples 1-22 and 1-23 are extremely simple in two respects: the sample space
is of a highly restricted type, and the counting process is easy. Combinatorial methods fre-
quently become necessary as the counting becomes more involved. Basic counting methods
are reviewed in Section 1-'6, Some important theorems regarding probability follow.
Theorem 1-1
If 0 is the empty set, then P(0) = 0.
Proof Note thatg= g u 0 andg and 0 are mutually exclusive. ThenP(9') =P(9') + P(0)
from property 4; therefore P(0) = 0.
Theorem 1-2
P(A) = 1 - P(A).
12 Chapte' 1 An Introduction to Probability
Proof Note that g=A 'J A and A and A are mutually exclusive. ThenP(9') =P(A) .,.P(A)
=
from property 4, but from property 2, P(9') 1; therefore P(A) 1 - PCA). =
Theorem 1·3
P(A V B) = PCA) + P(B) - peA n B).
Proof Since A vB =A v (B nA), where A and (B nA) are mutually exclusive, andB=
(A n B) V (B n A), where CA n B) and (B n A) are mutually exclusive, then peA v B) =
P(A) + PCB n A\ and PCB) ~ PeA n B)+ P(B nA). Subtracting, peA v B) - PCB) = P(A)-
peA n B), and thus P(A v B) = P(A).,. P(B) - P(A n B).
The Venn diagntm shown in Fig. 1-4 is helpful in following the argument of the proof for
Theorem 1-3. We see that the ~double counting" of the hatched region in the expression
PCA) + PCB) is corrected by subtraction of P(A n B).
Theorem 14
PCA v B v C)= P(A).,.P(B) +P(C) -P(AnB) - p(A n C) -PCB n C) + PCA n B n C).
Theorem 1-5
k. k k
P(At vA2 v· .. vA,)= 2,p(A;)- 2,p(A; nAj )';' 2,P(Ai nAj nA, )+ ...
i=l i</=2 £<)<.r-.:.3
Theorem 1·6
!f A c B, then PCA):;; PCB).
Proof
!fA cB, thenB =A v (A n B) andP(B) =P(A) +P(A nB) ?;P(A), since peA () B);;' O.
CD
L-_ _ _~_ _ _ _~, Figure 1-4 VeM diagram for two events.
1-5 Probability Definition and Assignment 13
;$~~~pi~I:b1
If A and B are mutually exclusive events, and if it is known that peA) = 0.20 while PCB) = 0.30, we
can evaluate several pr~biliti~~·
1. peA) = I - peA) = 0.80.
2. p(S) =I - PCB) =0.70.
3. peA u B) =P(A) + PCB) = 0.2 + 0.3 =0.5.
4. P(AnB)=O.
5. peA nB)=~(A uB), by De Morgan's law_ V
=1-P(AuB)
= I - [peA) + PCB)] = 0.5.
Suppose events A and B are not mutually, exclusive and we know that P(A) = 0.20, PCB) = 0.30, and
peA n B) = 0.10. Then evaluating the s~;~ilities as before, we obtain
1. peA) = 1- peA) =0.80.
2. p(S) = I - PCB) =0.70.
3. peA u B) =P(A) + PCB) - peA n B) = 0.2+ 0.3 - 0.1 = 0.4.
4. peA n B) = 0.1.
5. peA n B) =peA u B) =I - [peA) + PCB) - peA n B)] =0.6.
!;®R~!~!t~fii:
Suppose that in a certain city 75% of the residents jog (1), 20% like ice cream (1), and 40% enjoy
music (M). Further, suppose that 15% jog and like icc cream, 30% jog and enjoy music, 10% like ice
cream and music, and 5% do all three types of activities. We can consolidate all of this information
in the simple Venn diagram in Fig. 1-5 by starting from the last piece of data, P(J n In M) = 0.05,
and "working our way out" of the center.
1. Find the probability that a random resident will engage in at least one of the three activities.
By Theorem 11, -----,.,,-
P(l u lu M) = pel) + pel) + P(M) - P(l n l) - P(l n M) -pel n M) + P(l n I n M)
=0.75 + 0.20+ 0.40 - 0.15 - 0.30 - 0.10 + 0.05 =0.85.
This answer is also immediate by adding up the components of the Venn diagram.
2. Find the probability that a resident engages in precisely one type of activity. By the Venn dia-
gram, we see that the desired probability is
P(l nl n M) +P(j n In M) + p(J n I n M) = 0.35 + 0 + 0.05 = 0.40.
/
/
Figure 1-6 A tree diagram for tossing a true coin three times.
Suppose we toss a true coin and cast a true die. Since the coin and the die are true, the two outcomes
to '"if: l' SOl = (H, T}. are equally likely and the six outcomes to "&2' S02 = {1,2,3,4,5.6 J, are equally likely.
Since n l = 2 and TI..z = 6, there are 12 outcomes to the total experiment and all the outcomes are equally
likely. Beeause of the simplicity of the experiment in this case, a tree diagram permits an easy and
complete enumeration. Refer to Fig. 1-7.
A manufacturing process is operated with very little "in-process inspection." When items are com-
pleted they are transported to an inspection area, and four characteristics are inspected, each by a dif-
ferent inspector. The first inspector rates a characteristic according to one of four ratings. The second
inspector uses three ratings, and the third and fourth inspectors use two ratings each. Each inspector
marks the rating on the item identification tag. There would be a total of 4 . 3 . 2 . 2 = 48 ways in
which the item may be marked.
1-6.3 Pennutations
A permutation is an ordered arrangement of distinct objects. One permutation differs from
another if the order of arrangement differs or if the content differs. To illustrate, suppose we
go = ({H, 1), (H, 2), (H, 3), (H, 4), (H, 5), (H, 6),
(T, t), (T, 2), (T, 3), (T, 4), (T, 5), (T, 6))
T~~
=--4
5
6
Figure 1-7 The tree diagram for Example 1-27.
16 Chapter I An Introduction to Probability
again consider four distinct chips labeled a, b, c, and d. Suppose we wish to consider all per-
mutations of these chips taken one at a time. These would be
a
b
c
d
If we are to consider all pemlUtations taken two at a time, these would be
ab be
ba cb
ac bd
ca db
ad cd
da de
Note that permutations ab and ba differ because of a difference in order of the objects,
while pennutations ac and ab differ because of content differences. In order to generalize,
we consider the case where there are n distinct objects from which we plan to select per-
mutations of r objects (r::; n). The number of such pennutations, p~, is given by
P," =n(n-l)(n-2)(n-3)·····(n-r+l)
n!
= (n-r)!'
This is a result of the fact that there are n ways to select the first object, (n - 1) ways to
select the second, "', [n - (r- 1)] ways to select the rth and the application of the multi-
plication principle. Note that P~ = n! and OJ = 1.
;~~~i1ipj".i?ir
A major league baseball team typically has 25 players. A line-up consists of nine of these players in
a particular order. Thus, there are P~ == 7.41 x lOll possible lineups.
1-6.4 Combinations
A combination is an arrangement of distinct objects where one combination differs from
another only if the content of the arrangement differs. Here order does not matter. In the
case of the four lettered chips a, b, c, d, the combinations of the chips, taken two at a time,
are
ab
ac
ad
be
bd
cd
We are interested in determining the number of combinations when there are n distinct
objects to be selected r at a time. Since the number of permutations was the number of ways
to select r objects from the n and then permute the r objects, we note that
1-6 Finite Sample Spaces and Enumeration 17
(1-4)
In the illustration witll the four chips where r~ 2, the reader may readily verify thatY, ~ 12
and (~) : : : 6, as we found by complete enumeration.
For present purposes, e) is defined where n and r are integers such that 0 ::::; r:S; n~ how-
ever. the terms (~ may be generally defined for real n and any nonnegative integer r. In this
case we write
n~ = "(n-l)(n - 2)·· · .. (n- r+ I)
(
,r), r.I
The reader will recall the binomial theorem:
n f n..
(a+by = L
rooO\' r
'ja'b'·" 0-6)
(1-7)
and
(1.8)
and to develop the re;"lt shown in equation 1-8, we expand the right-hand side and collect
terms.
To verify that a finite collection of n elements has 2' subsets, as indicated earlier, we
see that
from equation 1-6. The right side of this relationship gives the total number of subsets since
Gl is the number of subsets with 0 elements. (;) is the number v.;th one element, ..., ""d tl
is the number with n elements.
18 Chapter 1 An Introduction :0 Probability
A production lot of size 100 is known to be 5% defective. ArandoDl sample of 10 items is selected
without replacement. In order to determine the probability that there will be no defectives in the sam-
ple. we resort to counting both the number of possible samples a:od The number of samples favorable
to e'/entA. where eYent A is taken to mean that there are nO defectivc.,_ The number of possible sam~
I!l.
pIes is (:l~;;;;; l~::' The number "favorable to A" is ~1 SO that
To generalize the preceding example, we consider the case where the population has N
items of which D belong to some class of interest (such as defective). A random sample of
size n. is selected without replacemenL If il denotes the event of obtaining exactly r items
from the class of interest in the sample, then
fDr N - D\
~n)
Problems of this type are often referred to as hypergeometric sampling problems,
E:¥.Di~~~);~~
An NBA basketball team typically has 12 players. A starting team consists of five of these players in
no particular order, Thus, there are (~2) = 792 possible starting teams.
:F;{·inji~"~~:ji.
One obvious application of counting methods lies in calculating probabilities for poker bands. Befo:e
proeeeding, we remind the reader of some sta."ldard ~erminology, The rank. of a particu1~ c~d drawn
from a standard 52-card deck can be2, 3, "', Q. K,A, while the possLble suUs are +, 0, v,~. In poker,
we dravo' five cards at random from a deck. The number of poss~ble hands is G)
=2.598,960.
1. We fitst calculate the probability of obtainir.g (V.i'O pairs, for example Av, A~, 3v, 30, !.O~.
We proceed as follows.
(a) Select tv,,,
ranks (e.g.• A. 3). We can do this (;) ways.
(b) Select two suits for first pair (e.g" \? ~), There are W ways,
(e) Select two suits for second pair (e.g., <:', 0). There are Wways.
(d) Select remaining card to complete the hand, There are 44 ways.
Thus, the number of ways to select two pairs is
and So
0.0475.
1-6 Finite Sample Spaces and Enumeration 19
2. Here we calculate the probability of obtaining a full house (onc parr, one three-of-a-kind), for
exampleA'Y,Ao!o, 3'Y, 3<>, 3~.
(a) Select two ordered ranks (e.g., A, 3). There are PJJ ). ways. Indeed, the ranks must be
ordered sinee "three A's, two 3 's" differs from "two A's, three 3's."
(b) Select two suits for the parr (e.g., 'Y, +). There are@ways.
(e) Select three suits for the threc-of-a-kind (e.g., 'Y, <>, ~). (~) ways.
Thus, the number of ways to select a full house is
and so
3744
P(full house) ~ 0.00144.
2,598,960
3. Finally, we calculate the probability of a flush (all five cards from same suit). How many
ways can we obtain this event?
(a) Select a suit There are (~) ways.
(b) Select five cards from that suit. There are (~3) ways.
Then we have
n!
(1-10)
Consider the word "TENNESSEE." The number of ways ro arrange the letters in this word is
n! _9_!_=3780
nT!nE!nN!nS! 1!4!2!2! '
where we use the obvious notation. Problems of this type are often referred to as multinomial sam-
pling problems.
The counting methods presented in this section are primarily to support probability
assignment where there is a finite number of equally likely outcomes. It is important to
remember that this is a special case of the more general types of problems encountered in
probability applications.
20 Chapter I An Introduction to Probabilirj
PtA) =.2.
10
and
If the first item is replaced before the second item is selected, the conditional probability
P(BIA) ~ P(B) = ~, and the events A and B resulting from the two seleetion experiments
comprising % are said to be jr.dependent, A formal definltion of conditiooal probability
(a) (b)
Figure 1·8 Conditional probability, (a) Initial sample space. (b) Reduced sample space,
1-7 Condit:o'&al Probability 21
P(AIB) will be given later, and independence will be discussed in detail, The following
examples will help to develop some intuitive feeling for conditionallfObability.
Exampiel.:34
, ....... '.,
'. ~ /.
Recall E.x.ample 1-11 where two dice are tossed, and ass~e that each die isJ:roe. The 36 poss[b!e Out-
comes were enumerated. If we consider two events,
.' ~a,nk\e~:3~:
In World V;,rar II an early operations research effort in England was directed at establishing search pat-
terns for U-boats by patrol flights or sorties. Fo~ a time, there was a tendency to co:r.cer.tratc the flights
on h!-shore areas, as it had been believed Lim more sighti.."lgs took place in-shore" The research group
studied 1000 sortie :recor~ witll the following result (the data are fictitious):
Sighting 80 20 100
No sighting 820 80 900
.-~"---
Definition
We may define the conditional probability of event A given event B as
1. 0" P(AIB) ,a
2. NflB) = 1.
and
P(A n B) = P(A) . P(BIA), P(A) > O. (H2)
The second statement is an obvious consequence of equation 1-11 with the conditioning on
event A rather than event B.
It should be noted that if A and B are mutually exclusive as indicated in Fig. 1-9, then
A n B = 0 so that P(AIB) = 0 and P(BIA) = o.
In the other extreme, if B cA, as shown in Fig. 1-10, then P(AIB) = 1. In the first case,
A and B cannot occur simultaneously, so knowledge of the occurrence of B tells us that A
does not occur. In the second case, if B occurs, A must occur. On the other hand, there are
many cases where the events are totally unrelated. and knowledge of the occurrence of one
has no bearing on and yields no information about the other. Consider, for example. the
experiment where a true coin is tossed twice. EventA is the event that the first toss results
in a "heads," and event B is the event that the second toss results in a "heads." Note that
-t, t,
P(A) = since the coin is true, and P(B\A) = since the coin is true and it has no memory.
The occurrence of event A did not in any way affect the occurrence of B, and if we wanted
to find the probability of A and B occurring, that is, P(A n B), we find that
00
Figure 1-9 Mutually exclusive Figure 1-10 Event B as a subset of A.
events.
1·7 Conditional Probability 23
Definition
A and B are independent if and only if
PtA n B) = PtA) . P(B). (1-13)
An immediate consequence of this definition is that if A and B axe independent events, then
from equation 1-12,
P(AIB) =PtA) and P(BIA) =P(B). (1-14)
The following theorem is sometimes useful. The proof is given here only for the first part.
Theorem 1-7
If A and B are independent events, then the following holds:
1. A and B axe independent events. - I
Proof (part 1) (:
PtA n B) = PtA) . P(BIA)
J.y
= PtA)
= P(A)
. [1 - P(BIA)]
. [1 - P(B)]
~i
=PtA) . PCB).
In practice, there axe many situations where it may not be easy to determine whether
two events axe independent; however, there are numerous other cases where the require-
ments may be either justified or approximated from a physical consideration of the exper-
iment. A sampling experiment will serve to illustrate.
r, );:~"';;~'ei:~~
Suppose a random sample of size 2 is to be selected from a lot of size I ~O, and it is known that 98 of
the I 00 items are good. The sample is taken in such a manner that the first item is observed and
repl~ced before the second item is selected. If we let
A: First item observed is good,
B: Second item observed is good,
and if we want to determine the probability that both items are good, then
98 98
P(AnB) =P(A)· P(B) =_ . - =0.9604.
100 100
If the sample is taken "without replacement" so that the first item is not replaced before the seeond
item is selected, then
98 97
P(AnB) =P(A)' P(~A) = -·-=0.9602.
100 99
24 Chapter 1 An Introduction to Probability
The results are obviously very dose, and one coromon practice is to assume the events independent
when the sampling fraction (sample size/population size) is small, say less than 0.1.
EX3]l,Iiiej:~i .
The field of reliability engineering has developed rapidly since the early 1960s. One type of problem
encountered is that of estimating system reliability given subsystem reliabilitics. Reliability is defined
here as the probability of proper functioning for a sta~ed period of time. Consider the structure of a
simple serial system, shown in Fig. 1-11, The system functions if and only if both subsystems fullC~
tion. If the subsystems surVive independectly, then.
where Rl and R2 are the reliabilities for subsystems 1 and 2, respectively. For exampte, if R j ;;: 0,90
and R, = 0.80, then R, = 0.72.
Example 1-37 illustrates the need to generalize the concept of independence to more
than two events. Suppose the system consisted of three subsystems or perhaps 20 subsys~
terns. W'hat conditions would be required in order to allow the analyst to obtain an estimate
of system reliability by obtaining the product of the sub'J'stem reliabilities?
Definition
The k events AI' Az, ... , A, are mutually independent if and only if the probability of the (
intersection of any 2, 3, ... , k of these sets is the product of their respective probabilities.
Stated more precisely. we require that for r = 2, 3 ..... k,
= rrp(~I)'
j=t
In the case of serial system ",liability calculations where mutual independence may
reasonably be assumed, the system reliability is a product of subsystem reliabilities
(1-15)
In the foregoing definition there are 2" - k - 1 conditions to be satisfied. Consider three
events, A, B, and C, These are independent if and only if P(;!. n B) = peA) , PCB), PC.4 n C)
P~4) . P(C), PCB n C) = PCB) • P( C), and peA n B n C) = peA) . PCB) . P( C). The fol-
lowing example illustrates a case where events are pairwise independent but not mutually
independent,
System
0;,~!~P!:'i.~#~:
Suppose the sample space, with equally likely outcomes, for a particular experiment is as follows:
It follows that
P(B, n CJ ) ~± ~ P(B,).P( Cj ).
1
p(Ao nBo n Co) ~-" p(Ao)·P(Bo)·P( Co).
4
Definition
If peAl n A,) ~ peAl) . peA,). then '&1 and '&2 are said to be independent experiments.
Definition
When the experiment is performed, one and only one of the events, B,. occurs if we
have a partition of :1.
26 Chapter 1 All IJlIroduetion to Probability
A particular binary "word" consists of five "bits," b l • b2• b1> b4,' Os. where bl~O.l.l;;;;t 1, 2. 3, 4, 5. Au
experiment consists of tranSmitting a "word," and it follows that there are 32 possible words, If the
events are
Bi = (CO, 0, 0, 0, 0), (0,0,0,0, OJ,
{(O, 0, 0, 1,0), (0, 0, 0, 1, I), (0, 0, 1,0,0), (0, 0, 1,0, I), (0, 0, 1, 1,0), (0, 0, 1, 1, 1)],
E, = (0, 1. 0, 0, 0), (0,1,0,0,1), (0, 1, 0, 1, 0), (0, 1,0, 1, 11, (0,1, 1,0,01, (0,1, 1,0,11,
(0,1,1,1,0), (0,1,1, 1,1)},
E4 = {(I, 0, 0, 0, 0), (I, 0, 0, 0, 1), (1,0,0, I, 0), (I, 0, 0, I, I), (1, 0,1,0,0), (1, 0,1, 0, 1),
(1,0,1, 1,0), (1, 0, I, 1, 1»),
E, = [(I, 1, 0, 0, 0), (1,1,0,0, I), (I, 1,0, 1,0), (1, 1,0, 1, 1), (1,1,1,0,0), (1,1,1,0,1),
(1, 1, 1, 1, OJ),
B,= [(1, I, 1, 1, 1)],
so that
since the events (A n B;) are pairwise Illutually exclusive, (See Fig, 1-12 for k 4,) It does
not matter thetA n B; = 0 for some or all of the i, since P(0) = 0, .
Using the results of equation 1-12 we can state the follov.'ing theorem.
Theorem 1-8
If B" B" "', B, represents a partition of:'! and A is an arbitrary event on ff', then the total
probability of A is given by
,
PtA) = P(B,) peA I B,l + peA '.1Jz)~·" + P(B,) , PtA i B,l L,P(E,)P(A I Bil.
i"'1
The result of Theorem 1-8, also known as the law of total probability, is very useful, as there
are numerous practical .imations in wbicb P(A) cannot be computed directly, However,
1-8 Partitions, Total Probabiliry, and Bayes' Theorem 27
with the infonnation that Bi has occurred, it is possible to evaluate P(AIB) and thus deter-
mine P(A) when the values P(B;) are obtained.
Another important result of the total probability law is known as Bayes' theorem.
Theorem 1-9
If B 1• B2, • •• , Bic constitute a partition of the sample space ';j and A is an arbitrary event on
Ef. then for r= I, 2•... , k,
(1-16)
Proof
_p(B,nA)
P(B,A
I )- P(A)
p(B,)PI:AIB,)
k
2.: p(BJ p(AlB, )
i==l
The numerator is a result of equation 1-12 and the denominator is a result of Theorem 1.8.
Three facilities supply microprocessors to a manufacturer of telemetry equipment. All are supposedly
made to the same specifications. However, the manufacturer has for several years tested the micro-
processors, and records indieate the following information:
0.Q2 0.15
2 0.01 0.80
3 0.03 0.05
The manufacturcr has stopped testing because of the costs involved, and it may be reasonably
assumed that the fractions that are defective and the inventory mix are the same as during the period
of record keeping. The director of manufacturing randomly selects a microprocessor, takes it to the
test department, and finds that it is defective. If we letA be the event that an item is defective, and Bi
be the event that the item came from faciliry i (i = 1, 2, 3), then we can evalnate PCB)A). Suppose, for
instance, that we are interested in determining P(B31A). Then
p(s,).p(AIs,)
p( S,iA) p( ~ )p(AIB\) + p(s,). p(Ais,) + p( s,) p(AIs,)
(0.05)(0.03) 3
(0.15)(0.02)+ (0.80)(0.01) +(0.05)(0.03) 25
28 Chapter 1 An Introduction to Probability
1-9 SUMMARY
This chapter has introduced the concept of random experiments, the sample space and
events, and has presented a formal definition of probability. This was followed by methods
to assign probability to events. Theorems 1-1 to 1-6 provide results for dealing with the
probability of special events. Finite sample spaces with their special properties were dis-
cussed, and enumeration methods were reviewed for use in assigning probability to events
in the case of equally likely experimental outcomes. Conditional probability was defined
and illustrated, along with the concept of independent events. In addition, we considered
partitions of the sample space, total probability, and Bayes' theorem. The concepts pre-
sented in this ehapter form an important background for the rest of the book.
1-10 EXERCISES
1~1. Television sets are given a final inspection fol-
(a) A n B.
lowing assembly. Three types of defects are identi-
fied, criticaF.'majo~ and minb'r defects, and are coded
(b)AuB.
A, B, and C, respectively, by a mail-order house. Data (e)AnE.
are analyzed with the following results. (d) A n (B n C).
(e) An (B u C).
Sets having only critical defects 2%
1-4. A flexible circuit is selected at random from a
Sets having only major defects 5%
production run of 1000 circuits. Manufacturing
Sets having only minor defects 7% defects are classified into three different types, labeled
Sets having only critic:S' A, B, and C. Type A defects occur 2% of the time, type
and major defects\O 3% B defects occur 1% of the time. and type C occur
Sets having only critieal 1.5% of thc time. Furthermore, it is known that 0.5%
and minor dcfccts U Q., 4% have both type A andB defects, 0.6% have bothA and
Sets having only major C defects, and 0.4% have B and C defects, while 0.2 %
and minor defects V 3% have all three defects. What is the probability that the
Sets having all three types of defects 1% flcxible circuit selected has at least One of the three
types of defects?
(a) What fraction of the sets has no lefects? O.!1~
1~5. In a human-factors laboratory, the reaetion timcs
(b) Sets with either critical defects or major defects
of human subjects are measured as the elapsed time
(or both) get a complete rework. What fraction from the instant a position number is displayed on a
falls in this category? 0 _1% digital display until the subject presscs a button
1 ~2. illustrate the following properties by shadings or located at the position indicated. Two SUbjects are
colors on Venn diagrams. involved, and times are measured in seconds for each
(a) Associative laws: A u (B u C) = (A u B) u C, A subject (t 1, t2). What is the sample space for this exper-
nCB n C)=(AnB)n C. iment? Present the following events as subsets and
mark them on a diagram: (t l + tz)fl::; 0.15, max (t 1, t2J
(b) Distriburive laws: A u (B n C) = (A u B) n (A u
C), A n (B u C) = (A n B) u (A n C).
;;0.15, It,-r,1 ;;0.06.
(e) IfAcB,thenAnB=A.
1~6. During a 24-hour period, a computer is to be
(d) IfAcB,thenAuB=B. aecessed at time X, used for somc processing, and
(e) IfAnB=0,thenAcB. exited at time Y:2: X. Take X and Y to be measured in
(I) IfAcBandBcC,thenAcC. hours on the time line with the beginning of the 24-
hour period as the origin. The cxperiment is to observe
1-3. Consider a universal set consisting of the inte- XandY.
gers 1 through 10, 9r U= [1,2,3,4,5,6,7,8,9, 1O}.
LetA= {2, 3, 4}, B= {3, 4, 5}, and C= {5, 6, 7}. By (a) Describe the sample space ff.
enumeration, list the membership of the following (b) Sketch the following eVents in the X, Yplane.
sets. (i) The time of use is 1 hour or less.
1-10 Exercises 29
(li) The access is before tl and the exit is after t2, probability that the sample will contain no more than
a
where s: tl < t2 $" 24. .two cefective units?
(iii) The time of use is !ess than 20% of the
"!-!s/-k inspecting incoming lots of merchandise, the
period. ~Jollowing inspection rule is used where lhe lots con-
1-7. Diodes from a batch a.-.oe tested one at a time and tain 300 ulllts, A random sample of 10 items is
marked either defective or nondefective, This is con- selected. Ifthere.is no more than one defective item in
tinued until either tv.\) defective items are found or the samp1e, the lot is accepted. Otherwise it is
five items have been~ted. Describe the sample space returned to the vendor. If the fraction defective in the
for this experiment. original lot is p', determine the probability of accept-
ing the lot as a function of p'.
1MK A set has four elements. A :::: {a, b, c, d}.
Describe the power se' {O, J lA. 1~16. In a plastics plant 12 pipes empty different
chemicals into 2. mixing vaL Each pipe has a five-posi-
1-9. Describe the sample space for each of the fol-
tion gauge lhat measures the rate of flow into the 'lat,
low..ng experiments. One day, while expc--r1menting with various mixtures,
(a) A lot of 120 battery lids for pacemaker cells 1S a solution is obtained that emits a poisor.ous gas. The
knov;'U to contain a number of defectives because settings on the gauges were not recorded. W"hat is the
of a problem with 'JJ.e barrier materia: applied to prObability of obtaining this same solution when ran~
the glassed-in feed-through. Three lids are ran- domlyexperimenting again?
domly selected (without replacement) and are
carefully inspected following a cut do\VD. 1~17. Eight equally skilled men and WOmen are
app;ying for two jobs. Because the two new emp:oy-
(b) A pallet of 10 castings is known to contain one
ees must work closely together, their personaliti~·s
defective and nine good Ull.its. Four castfugs a..re
should be compatible. To achieve this. the personnel
randomly selected (without ~eplacement) and
manager has administered a test and must compare
inspected,
the scores for each possibility. How .wany compar~
1-10. The production manager of a ce:rtain company isons must the ~ager make?
is interested in testing a fInished product, which is
available in lots of size 50. She would like to rework a 1·18. By accident. a chemist combined two labora~
lot if she can be reasonably sure that 10% of the items tory substances that yielded a desirable product.
are defective. She decides to select a random sample Unfortunately, her assistant did not record the names
of 10 iter!'.s wilhom replacement and rework the lot if of the ingredients. There are forty substances avail2.ble
it contains one or more defecrivei:ems. Does this pro--
in the lab. If the twO in question must be located by
cedure seem reasonable? successi"'e triakmd~error experiments. what is the
m.aximum number of tests that might be made?
1-11. A trucking firm has a conb:lictto ship a load of
1~ 19. Suppose, in the previous problem. a known Cat~
goods from city W to city Z There are no direct routes
connecting W to Z, but there are six roads from W to X alyst \\'iiS used in 'JJ.e first accidenta:: reaction. Because
and five roads from X to Z, How many rotal routes a.re of this, the order in wxch the ingredients axe mixed is
there to he considered? important. "Wllat is 'JJ.e maximum. number of tests that
might be made?
1~12. A state has one million registered vehicles and
is considering using license plates with six symhols 1-20. A co:npany plans to build five additio.u.al ware~
where the fIst three are letters and the last three are houses at new locations. Ten locations are cnder con-
/digit~ Is this scheme feasihle? sideration. Hov.' many total possible choices are there
for the set of five locations?
( 1-13~)The maLagcI of a small plant wishes to deter~
mine the number of wa)'S he ca:a assign workers to the 1..21. Washing maclllnes can have five kinds of major
first shift. He has 15 workers who can scrve as opera- ru:d five kinds of minor defects. In how many ways
tors of the production equipment, eight who can serve can one oajor ane! one rrtinor defect occur? In how
as maintenance personnel, and fot;I' who can be super- many ways can two major and two minor defects
visors. If the shift requires six operators, two mainte~ occur?
nance personnel, and one supervisor, how many wa.ys 1~Z2. Consider the diagra1!1 at the top of the oext page
can the first shift be :na!l1led? of a::J. electronic system, which shows the probabilities
1..14. A production lot has 100 units of which 2Q of the system co:nponents operating properly. The
are kno>t'D to be defective, A n:,uldom sample of four entire system operates if assembly m and at least
units is selected 'hithout repla.cemenL 'What is the one of the coopOoents in each of assemblies I and IT
30 Chapter 1 An Introductiou to Probability
II III
operates. Assume that the components of each assem- units arc approximately 0.995, 0.993, ~d 0.994,
bly operate indepeudently and that the assemblies respectively. Estimate the system reliability.
operate independently. 'What is the probability that the
1-27. Two balls are drawn from an urn containing m
entire system operates?
balls numbered from I to m. The first ball is kept if it
1-23. How is the probability of system operation is numbered I and returned to the urn otherWise. 'What
affected if, in the foregoing problem, the probability is the probability that the second ball drawn is num-
of successful operation for the component in assembly bered 2?
ill changes from 0.99 to 0.9? 1-28. Two digits are selected at random from the dig-
1-24. Consider the series-parallel assembly shown its 1 through 9 and the selection is without replace-
below. The values Rj (i = I, 2, 3, 4, 5) are the reliabil- ment (the same digit cannot be picked On both
ities for the five components shown, that is, R j = prob- selections). If the sum of the two digits is even, find
ability that unit i will function properly. The the probability that both digits are odd.
components operate (and fail) in a mutually inde- 1-29. At a certain university, 20% of the men and 1%
pendent manner and the assembly fails only when the of the WOmen are over 6 feet tall. Furthermore, 40% of
path fromA to B is broken. Express the assembly reli- the students are Women. If a student is randomly
ability as a function of R 1• R2, R), R4, and R5 • picked and is observed to be over 6 feet tall, what is
the probability that the student is a woman?
1-25. A political prisoner is to be exiled to either
Siberia or the Urals. The prObabilities of being sent to 1-30. At a maehine center there are four automatic
these places are 0.6 and 0.4, respectively. It is also screw machincs. An analysis of past inspection
known that if a resigent .9f Siberia is sejected at ran- records yields the following data.
dom the probability is 05,that he.will be wearing a fur
coat, whereas the probability i~ 0.7 that a resident of
thc Urals will be wearing one. Up-on arriving in exile, Perccnt
the first person the prisoner sees is not wearing a fur Percent Defectives
coat. 'What is the probability he is in Siberia? Machine Production Produced
1-26. A braking device designed to prevent automo- 15 4
bile skids may be broken down into three series sub- 2 30 3
systems that operate independently: an electronics
3 20 5
system., a hydraulic system., and a mechanical activa-
4 35 2
[Or. On a particular braking. the reliabilitics of these
A 8
HO Exercises 31
Machines 2 and 4 are newer and more production has point (in which case he wins), or until he throws a 7
been assigned to them than to machines 1 and 3. (in which case he loses). ¥lhat is L'le probability
Assume that the current inventory mix reflects the that '.he player with the dice \\w
event'..:.ally win the
production perce:ltages indicated. game?
(a) If a screw is randomly picked from inventory. 1-37~ The industrial engineering depa."'1ment of
what is the probability L"lat it will '!::Ie defective? the XYZ Company is performing a work sampling
(b) If a screw is picked and found to be defective, study on eight technicians, The engineer wishes to
what is the probability that it was produced by randomize the order in which he visits the techni-
machine 3? cians' work areas. In how many ways may he arrange
1~31. A point is selected at random inside a cirele,
these visit~?
What is the probability that the point is closer to the 1 ~38. A hiker leaves poin: A shown in the :figure
center than to the circumference? below. choosing at random one path froi.'. A.B, AG,
1 ~32. Complete the details of the proof for Theorem AD, a..'l.dAE, At each subsequenljunction she chooses
1~4 in the text. another path at r.mdom. What is the probability tr,.at
she arrives at.: poir.t X1 ") I ;.?
1~33. Prove Theorem 1~5.
1~39. Three primers do work for the publications
1~34. Prove the second aud third parts of T!leorem office of Georgia Thch, The publications office does
1-7. not negotiate a contraCt penalty for late work. and the
1",35. Suppose there are n people in a room. If a list is
data below reflect a large amount of experience '.vith
these pm:ters.
made of all their birthdays (the specific month and
day of the month). what is the probability that tv.'o or
more persons have'the same bir.hday? Assume there
Fraction of
are 365 days in the year and that each day is equally
Fraction of Deliveries
likely ro occur for any person's birthday. Let B be the
Contracts Held More tha..'1
event that two or' more persons have the same binh~
day, FindP(B) andP('B) for n= 10, 20, 21, 22, 23, 24, Printer i Printer i One Month Late
/---
25, 30, 4<), 50, and 60.
'1~36.Jn a certain dice game, players continue to throw
1
2
0,2
0.3
0,2
0.5
tw:;;dice until they either win or lose. The player 'Nins 3 0.5 0.3
on the first throw if the sum of the two uptumed faces
is either 7 or 11 an9 loses if the sum is 2, 3, 0::- 12.
Otherwise, the sum of the faces becomes the A department obscrves that its recruiting bookIe: is
player's "pOint." The player continues to throw llt;til more th2n a month late. 'Wha: is the probability t'lat
the first succeeding throw on which he makes his the contract is heM by printer 3?
B c
x
32 Chapter I An. Inrroduction to Probability
1-40. Following aircraft accidents, a detailed invesLi.ga~ due to structural failure is 0.2. If 25% of all aircraft
tion is conducted. The probability that an accident accidents are due to structural failures, :find the proba-
due to structural failure is correctly identified is 0.9 bility that an aircraft accident is due to structural failure
and the probability that an accident that is not due given that it has been diagnosed as due to structural
to structural failure being identified incorrectly as failure.
Chapter 2
2·1 DlTRODUCTIO~
The objectives of this chapter are to introduce the concept of random variables. to define
and illustrate probabilitydisoributions and cumulative disoribution functions and to present
useful character:izatiolls~for va..-iables" ~ - _.- -- - -
When describ.ing the sample space of a random experiment, it is not necessary to spec-
ify that an individual outcome be a number. In several examples we observe this, such as in
Example 1·9, where a true coin is tossed tlrree times and the sample space is '1' = {HHB,
HIfT, HnI, HIT, THH, THT, 7TH, TlTj, or ExampJe 1·15, where probes of solder joints
yield a sample space '1'= {GG, GD, DG, DD}.
In most experimental situations. however, we are interested in numerical outcomes. For
example, in the illustration involving coin tossing we might assign some real number x to
every element of the sample space. In general, we want to assign a real number x to every out-
come e of the sample space if. A functional notation will be used initially, so that x = X(e),
whCreX is the function,. The domain of X is '!:f. and the numbers in the range are real numbers.
The func.ti0n Xis called arandom vari.ble. Figure 2-1 illustrates the nature of this fu'1cnon,
Definition
If't; is an experiment having sample space ~~ and X is a/unction that assigns a real num-
ber X(e) to every outcome e E if, tbenX(e) is called a random variable.
~,,~~pI~~";(
Consider the coin-tossing eXperiment discussed in the preceding paragraphs, If X is me fnrnbcr of
heads showtng, thee X(HHlI) ~ 3, X(HlflJ ~ 2, X(liTH) =2, X(HT/) 1, X(THH} =2, X(TH1) =1.
X(Tm}; I, and X(TTI) =0. The ta.1:ge spa.ceRx= {x:x= 0, t 2. 3} in t,1is example (see Fig. 2-2),
Rx
e.----~------------------+_----~ • X(el )
(e)
~ (b)
Figure2-1 The concept of a random variabk. (a):l': The sample space of'i!l. (b)R,: The range
space of X.
33
34 Chapter 2 One~Dimensional Random Variables
Of Rx
TIT·
TTH· ·0
THT-
-1
HIT·
HHT-
HTH·
THH-
-2
HHH- -3
X
Figure 2-2 The number of heads in three coin tosses.
The reader should recall that for all functions and for every element in the domain,
there is exactly one value in the range. In the case of the random variable, for every outcome
e E g there corresponds exactly one value X(e). It should be noted that different values of
e may lead to the same x, as was the case where X(ITH) = I, X(THTJ = I, and X(HTl) = I
in the preceding example.
Where the outcome in g is already the numerical characteristic desired, thenX(e)::= e,
the identity function. Example 1-13, in which a cathode ray tube was aged to failure, is a
good example. Recall that g = (t: ' " OJ. If X is the time to failure, then X(t) = t. Some
authors call this type of sample space a numerical-valued phenomenon.
The range space, Rx. is made up of all possible values of X, and in subsequent work it
will not be necessary to indicate the functional nature of X. Here we are concerned with
events that are associated with Rx. and the random variable X will induce probabilities onto
these events. If we return again to the coin-tossing experiment for illustration and assume
the coin to be true, there are eighl.:qually likely outcomes: HHH, HHT, liTH, HIT, THH,
THY, TTH, and ITT, each having probability t.
Now suppo~~ ~ i~ the event '~exacJly two
heads" and, as previously, we let X represent the number of h.~ads..(see Fig. 2-2), The event
- ••- - I . .,,"- - 3,
that (X = 2) r.ehltes ~_Rx,_no!3j_ho",ev_er,px(X = 2) = peA) = 8' since A = {HHT, liTH,
THHris the equivalent event in g, and probability-was 'defined--on events in the sample
f
space. The random variable X induced the probability of to the event (X = 2), Note that
parentheses will be used to denote an event in the range of the random variable, and in gen-
eral We will write Px(X =x).
In order to generalize this notion, consider the following definition.
Definition
If g is the sample space of an experiment ~ and a random variable X with range space Rx
is defined on g. and furthermore if event A is an event in g while event B is an event in-R~~
then ~ .~~ B ar~ equivalent events if
A={eE g:X(e)EBj ..
Figure 2-3 illustrates this concept.
More simply, if event A in g consists of all outcomes in g for whichX(e) E B, then A
and B are equivalent events. Whenever A occurs, B occurs; and whenever B occurs, A
occurs. Note that A and B are associated with different spaces.
2-1 Introduction 3S
RX
B
x
Definition
If A is an event in the sample space and B is an event in the range space Rx of the random
variable X, then we define the probability of B as
Px(B) = P(A), where A= leE 9':X(e)E B}.
With this definition, we may assign probabilities to even~ in Rx in terms of probabili-
ties defined on events in ::1, and we will suppress the function X, so that P x(X = 2) = in f
the familiar coin-tossing example means that there is an event A = (HHT, HTH, THH} = {e:
X(e) = 2} in the sample space with probabilityt. In subsequent work, we will not deal with
the nature of the function X, since we are interested in the values of the rangc space and
their associated probabilities. Vihile outcomes in the sample space may not be real num-
bers, it is noted again that all elements of the range of X are real numbers.
An alternative but similar approach makes use of the inverse of the function X. We
would simply define X-'(B) as
X-'(B) = (e E 9': X(e) E B}
so that
P x(B) =P(X-l(B)) =P(;!').
Y=6 {(l, 5), (2, 4), (3, 3), (4, 2), (5, I)}
,
36
Y=7
,
36
Y=8
{(I, 6), (2, 5), (3, 4), (4, 3), (5, 2), (6, I)}
{(2, 6), (3, 5), (4, 4), (5, 3), (6, 2))
j6
,
Y=9 {(3, 6), (4, 5), (5, 4), (6, 3))
j6
,
j6
J
Y= 10
",,
{(4, 6), (5, 5), (6, 4))
Y= 11 {(5, 6), (6, 5)} j6
Y=12 {(6,6)) j6
36 Chapter 2 One~Dimensional Random Variables
The following examples illustrate the sample space-range space relationship, and the
concern with the range space rather than the sample space is evident, since numerical
results are of interest.
.··.'E:l<.~I'~~2:2··
Consider the tossing of two true dice as described in Example 1-11. (The sample space was described
in Chapter 1.) Suppose we define a random variable Yas the sum of the "up" faces. Then Ry = {2, 3,
1
4,5,6,7, 8, 9, 10, 11, 12}. and the probabilities arc (36' -ft' 16, k,
;6' ~, ;6' ~, 3~' ;6' fc;), respee~
tively. Table 2-1 shows equivalent events. The reader will recall that there are 36 outcomes, which,
since the dice are true, are equally likely.
One hundred cardiac pacemakers were placed on life test in a saline solution held as close to body
temperature as possible. The test is functional, with pacemaker output monitored by a system pro-
viding for output signal conversion to digital form for comparison against a design standard The test
was initiated on July 1, 1997. Vlhen a pacer output varies from the standard by as much as 10%, this
is considered a failure and the computer records the date and the time of day (d, t). If X is the random
variable "time to failure," then ~ = {(d, t): d = date •.t = time} and Rx = {x: x;:: a}. The random vari-
able X is the total number of elapsed time units since the module went on test. We will deal directly
with X and its probability law. This concept will be discussed in the following sections.
In the case of the coin-tossing experiment, the random variable X assumed four values, 0, 1, 2, 3, with
probabilities t, t; t. t.
We can state FJ...x) as follows:
F,(x) ~ 0, x<O,
t
~8' 0::::x<1,
4.
t~x<2,
~8'
7
~8'
=1, x;::3.
F(x)
1
7/8 ----.0
4/8 ---.0
1/8
o 2 3 x
Figure 2-4 A distribution function for the number of heads in tossing three true coins.
fi~~p!~~~~~j
Again recall Example 1~ 13, when a cathode ray tube is aged to failure. Now :f = {t: t ~ O}, and if we
let X represent the elapsed time in hours to failure, then the event (X::;; x) is in the range space of X. A
mathematical model that assigns probability to (Jf_~x) is
FX<x) = 0, x:::; 0,
= l-e-,1.r, x>o,
where;" is a positive number called the failure.J!H~.....cfailureslhour). The use of this "exponential"
model in practice depends on certain assumptions -abo~t the failure process. These assumptions will
be presented in more derail later. A graphical representation of the cumulative distribution for the time
to failure for the CRT is shown in Fig. 2 ...5.
A eustomer enters a bank where there is a common waiting line for all tellers, with the individual at
the head of the line going to the first teller that becomes available. Thus, as the customer enters, the
waiting time prior to moving to the teller is assigned to the random variable X. If there is no one in
line at the time of arrival and a teller is free, the waiting time is zero, but if others are waiting or all
tellers are busy, then the waiting time will assume some positive value, Although the mathematical
form of FX depends on assumptions about this service system, a general graphical representation is as
illustrated in Fig. 2...6. ~
Cumulative distribution functions have the following properties, which follow directly
from the definition:
1. 0'; FX<x):;; 1, -oo<x<oo.
F,(x)
o x
Figure 2 ...5 Distribution function of time to failure for CRT.
38 Chapter 2 One~DimensioDal Random Variables
LProbability
, of no wa:t
oL---~o~------------------~x
Figure 2-6 Waiting-time distribution function.
2. limx~X<x) = I,
lim"..,J'x<;tl = O.
3. The function is nondecreasing. That is, if Xl S; x" then F;;.Xt ) ,; F;;'X;;.
4. The function is continuous from the rigbl That is, for all X and Ii> 0,
pm [Fx(x+ Ii) - Fx(x)] = 0.
0->0
In reviewing the last three examples, we note that in Example 2-4, the values of x for
which there is an increase in F "x) were integers, and where x is not an integer, then F .j.x)
bas the value that it had at the nearest integer X to the left. In this case, Fx(x) bas a saltus or
jump at the values 0, I, 2, and 3 and proceeds from 0 to I in a series of such jumps. Exam-
ple 2-5 illustrates a different situation, wbere F x<;t) proceeds smoothly from 0 to 1 and is
continuous everywhere but not differentiable at x = O. Finally Example 2~6 illustrates a sit-
uation wbere there is a saltus at X = 0, and for x > O. F;;'x) is continuous.
Using a simplified [ann of results from the !:z..besgue decomposition thect~ it is noted
that we can represent F X<x) as the sum 6!' two component functions, say GP) and HP), or
FP) = Gx(x) + Hx(x), (2-2)
where GX<x) is eontinuous and Hix) is a right-band continuous step function with jumps
coinciding with those of FP), andHX<- =) =O.lf Gx<;t) =0 for all x, then X is called adis-
crete random variable, and if Hx(x) " 0, then X is called a continuous random variable.
Where neither situation bolds, X is called a mixed type random variable, and although this
was illustrated in Example 2-6, this u.'Xt will concentrate on purely discrete and continuous
random variables, since most of the engineering and management applications of statistics
and probability in this book relare either to counting or to simple measurement,
Suppose !hat the number of worldng da:ys in a particular year is 250 and that the records of employ~
ees are marked for each day they are absent from work An experiment consists of randomly selecting
2-3 Discrete Random Variables 39
a record to observe the days marked absent. The random variable X is defined as the number of days
absent, so that Rx = {O, 1,2, ... , 250}. Ibis is an example of a discrete random variable with a finite
number of possible values.
A Geiger counter is connected to a gas tube in such a way that it will record the background radiation
count for a selected time interval [0, tJ. The random variable of interest is the count. If X denotes the
random variable, then Rx = {O, 1, 2, .. ,' k, ... }, and we have, at least conceptually, a countably infi-
nite range space (outcomes ean be placed in a one-to-one correspondence with the natural numbers)
so that the random variable is discrete.
Definition
If X is a discrete random variable, we associate a numbe~pX<x;) = Px(X =:xi)lwith each out-
come Xi in Rx for i = 1, 2, ... , n, ... , where the numbers PTxi) satisfy the following:
I. p,,(x,) ? 0 for all i.
2. 2::, Px(x,) ~ I.
We note immediately that
I
and
Fx(x,) ~ px(x,;; Xi) ~ 2:px(x). (2-4)
x:Sx,
The function Px is called the probability function or probability mass f!!:.nction or prob-
ability la'll. of the random variable, and fu,,'collection of pairs [(xi' p,,(x')0;;-'i;-2, ..-:ris
called the probability distribution of X. The functionpx is usually presented in either tabu-
Jf!r, gra.p!:!~91, or matk~"!.~tjc..Cl? f?Im, as illustrated in the following examples. ~-
~~jmf2]
For the coin-tossing experiment of Example 1-9, where X = the number of heads, the probability dis-
tribution is given in both tabular and graphical form in Fig. 2-7. It will be recalled tha~~:= _{O~;~~}} .
.'
Tabular Presentation Graphical Presentation
o 1/8
1 3/8
2 3/8 1/8 1/8
3 1/8
o 2 3 x
Figure 2-7 Probability distribution for coin-tossing expcriment.
40 Chapter 2 One-Dimensional Random Variables
Suppose we have a random variable X with a probability distribution given by the relationship
( ,
Px(x)=i "JpX(I-P)'-x, x=O, l,. .• • ,n,
"
=0, otherwise, (2-5)
where n is a positive integer and 0 < p < 1. This relatiouship is known as the binomial distribution,
and it -will be studied in more detail later. Although it would be possible to -di;piay this model~
graphical or tabular form forpartieular n andp by evaluating pf.;;.) for x =0. 1,2, ...• n. this is seldom
done in
Recall the earlier discussion of random sampling from a finite population ~i.thout replacement. Sup-
pose there are N objects of which D are defective. A random sample of size It is selected Vlithout
replacement, and if we let X represent the number of defectives in the sample, then
(DYN-DI
Px(x)= 'It:)') X =0, 1, 2.... , min (n, D),
=0, otherwise,
In the event that either tabular Or graphical presentation is desired, this would be as shown in Fig. 2-8;
hOVfeve.r, U!uess there is some special reason to use these forms, we will use the mathematical. relationship.
x p,(x)
0.B051
0 (~) (~5)/('~0) '" 0.805
i
m(~5) / C~O) """0.178
3 m(9,5);C~0) =Q,003
0.178
(~) (9;) / ('~) '" 0.000 0,Q14
4 0.003
------"'--------
0 1 2 3 4 x.
'j~~~ir~1,~}'
In Example 2-8. where the Geiger countel;' was prepared fol;' deteetir.g the background radiation count,
we might use the following relationship. which bas been experimentally sbown to be appropriate:
pix) = .->-e),.t'!!x!, x = 0, 1,2.... , i, > 0,
=0 oilie..-wise, (2-7)
This is called the f!zj§§QJLJ!i~1J:i.bU.tiO!1, and at a later point it will be derived analytically_ The param.-
eter ). is the mean rate~m' "hits"'per unit time. and x is rhe number of rhese "hits."
These examples have illustrated some discrete probability distrihutions and alternate
means of presenting the pab:s [(Xi,P,,(Xi))' i = 1.2, , .. ], In !.ater sections a number of proba-
bility distributions will be developed, each from a set of postulates motivated from consid-
erations of/redl-';""orzdphino~~\
A general graphical presentation of the discrete distribution from Example 2-12 is
given in Fig. 2-9. This geometric interpretation is often useful in developing an intuitive
feeling for discrete distributions. There is a close analogy to mechanics if ¥"e consider the
probability distribution as amass of one unit distributed over the real line in amounts Px(xi)
at points XI' i = 1,2, ...• n. Also, utilizing equation we note the following useful re.....ult
where b2 a:
(2-8)
(2-9)
Px(x)
~--~X~,----~X2----~X~3----X~4---\\--~XJn_-,----x~n--------~!
Figure 2~9 Geometric interpretation of a probability distribt.::tion.
42 Chllprer 2 One-Dimemional Random Variables
(2-11)
We also note the close correspondence in this form with equation 2-4 with an integral
replacing a summation symbol, and the following properties off,!,;;):
1. f,!.;;);:'O forallxE Rx.
J
2. fx(;r)dx = L
'.
3. f,Ax) is piecewise continuous.
4. f,Ax) = 0 if x is nOt in the range Rx.
These concepts are illustrated in Fig. 2-10. This definition ofa density function stipu-
lates a function fx defined on Rx such that
(2-12)
where e is an outcome in the sample space. Vv'e are concerned only with Rx and lx- It is
important to realize thatfx(x) does not represent the probability of anything, and that only
when the function is integrated between two points does it yield a probability.
Some comments about equation 2~9 may be useful, as this result may be counterintu-
itive. If we allow X to assume all values in some interval, then Px(X =xo);::::; 0 is not equiv-
alent to saying the event (X = xo) in Rx is impossible. Recall that if A = 0, then Pj.A) = 0;
however, although PiX =xol =0, the fact that the setA = (x: X =xo) is not empty clearly
indicates that the converse is not true.
All immediate result of this is that Px(a ,; X'; b) = Px(a < X'; b) = Px(a < X < b) =
Px(a'; X < b), where X is continuous, all given by Fx(b) - Fx(a).
~
!i¥N'!\l~~3!i~i
The time to failure of the cathode ray tube described in Example 1-13 has the follOwing probability
density function:
'''' O.
otherwise.
where A > 0 is a eonstant knov.-'U as ilie fa.i!ure rate. This probability density function is called the
exponentiol density, and experimental evidence has indicated that it is appropriate to describe the time
~(x)
x
Figure :2-10 HJTOthetical probability density function.
2-4 Contim.:.ous Random Va..'"iables 43
to failure (a real-world occurrence) for some types of components, In this example suppose we want
'" find P,(f?IOO hours). This is equivalent to stating P,(100" =), and T<
PT (T?100) r k-"dt
=h,
=e- too ,,".
We might again employ the cOncept of conditional probability and determine P"T21OOIT:> 99), the
probability the tube lives at least 100 hours given that it has lived beyond 99 hours. From our earlier
wo~k.
e-1(0). _.
-9?:t ;;;: e •
e
A random variable X has the triangular probability density fuuction given below and shown graJ?rl~
cally in Fig. 2-11:
1. p/-l<X<!\f'O""+I'!2
'\ 2) -I ()
xd>:=.1..
8
( 3\
2. pxX';;-.=
\.
f
2;...,.,
0",,+ 1,'xd>:+ 1'12(2-x)""
0 1
I
1 ( x'\'1' 7
=°:2+ 2X-Z"J =g ('
3. P,,(X"; 3) = 1.
")
4. P",X" 2.5) = O.
\.4
(1 3)
2 Jj4
r
5, Px -<X<- = xd.Hf,,~ (2-x)""
J! ' 'n
15 3 27
32 8
Returning to the coin·tossing experiment where X represents the number of heads and the probabil·
ity distribution is as sho'NIl in Fig. 2·7, the calculation of J1 yields
f
1'= ±X;PX(XI)=0·m+l-(%J+2HJ+3(~J=%,
,=1
as indicated in Fig. 2-13. In this particular example, because of symmetry, the value J1. eould have been
easily detenninedfrom inspection. Note that the mean value in this example cannot be obtained as the
output from a single trial,
;E#¥~r;,.1'!~:1
In Example 2-14, a density Ix was defined as
AN = X. O';;X';; I,
=2->; I,;;x ';;2,
=0, otherwise.
a b c x
Figure 2~12 A density function.
2-5 Some Characteristics QfDistribunons 45
316 3/8
118 1/8
o 2 3 x
~ 3/2
Figure 2·13 Calculation of the mean.
JJ.= f~x.Xdt>-fX.{2-X)dx
+ f·Odx+ft·Odx=4
another result that we could have dete:rmined via symmetry,
Another measure describes the spread or dispersion of the prebability associated with
elements in RJ(: 'This measure is called the variance, denoted by cr,
--.:..t and is defined as
~ ____ ~ ___ ~_~~ _ _ -i_
follows:
for discrete X,
This is the second moment about the mean, and it corresponds to the moment of .inertia.in
mechanics. Consider Fig. 2-14, where two hypothetical discrete distributions are shown in
graphical fonn. ~ote that the mean is one in both cases. The variance for the discrete ran-
dom variable shov.n in Fig. 2-14a is
WIDch is four times as great as the variance of the random variable shown in Fig. 2·14a.
If the units on the random variable are, say feet, fuen the units of the mean are the same,
but the units on the variance would be feet squared. Another measure of dispersion, called the
standard deviation, is defined as the positive square root of the variance and denoted o;where
(j={;;2. (2-15)
It is noted that the units of (jare the same as those of the random variable, and a small value
for (jindicates little 'dispersion whereas a large value indicates greater dispersion.
46 Cbllpter 2 One-Di:mensiorull Random Variables
PxIA
1;2
114 114
o 2 x -1 o 2 3 x
(al,u~' and 0'0 0,5
Figure 2-14 Some hypothetical distributions.
for discrete X,
1-
= _x 2 f x (x)dx-/1 , for continuous X. (2-16)
This simply indicates. that the second moment about the mean is equal to the second
moment about the origin less the square of the mean. The reader familiar with engineering
mechanics will recognize that the development leading to equation 2-16 is of the same
nature as that leading to the theorem of moments in mechanics.
(f
2 :- 2 1 2 3 2
=lO '8+ 1 '8+ 2 8+ 3 '8 - \2 ='4'
3 2 1J (3)' 3
which is only slightly easier.
2. Binomial distribution-Example 'J.alO. From equation 2-13 we may show that jJ.= "P, and
or
,,' Jtx"(
I
n11'"(1_ p)n-z1-(np)2,
,:x
r=<l I
o •x
and
9x(X)
o x
rnen
and
Note that the mean is the same foe the densities in parts 3 and 4. with part 3 having a variance
twice that of part 4,
In the development of the mean and variance, we used the terminology "mean of the
random variable" and ~'variance of the random variable." Some authors use the terminology
"mean of the distribution" and "variance of the distribution:' Either terminology is accept-
able, Also, where several random variables are being considered, it is often convenient to
use • subscript on If. and (Y, for example, If.x and (Yx'
In addition to the mean and variance. other mOments are also frequently used to
describe distributions, That is, the moments of a distribution describe that distribution,
measure its properties. and, in certain circumstances? specify it. ~1oments about the origin
are called on'gin moments and are denoted J.4. for the kth origin moment, where .
48 Chapter 2 One-Dimensional Random Va.'iables
for discrete X,
for continuous X.
(2-18)
2
Note that the mean j.J. == j.J.~ and the \'ariance is 0 -:= i12. Central moments may be expressed
ill terms of origin moments by fue relationship /'"
, ik\
ILk=I,H)'l,jpJp;-j, k=O,I,2",,, (2·19)
j"'O J.1
Thwrem 2·1
Let X be a random variable (discrete or continuous), and let k be some positive number,
Then
(1,..20)
Since
it follows that
(f
1 ~ JP'./K (x- IL)1 'fx(x)dx+ JW ~(x- IL)1 . fx(x)dx.
-- 'J.t+..JK
or
1
Px(ll-kcr<X<Il+kcr) ~ I-I?'
is often useful.
The usefulness of Chebyshev's inequality stems from the fact that so little knowledge
about the distribution of X is required. Only J1 and 0- 2 must be knO\Vll. However, Cheby-
shev's inequality is a weak. statement, and this detracts from its usefulness. For example,
px(ix - 111 ~ cr)" 1, which we knew before we started! lithe precise form oflx(x) or Px(x)
is known, then a more powerful statement can be made.
From an analysis of company records, a materials control manager estimates that the mean and stan-
dard deviation of the "lead time" required in ordering a small valve are 8 days and 1.5 days, respee-
tively. He does not know the distribution of lead time, but he is willing to assume the estimates of the
mean and standard deviation to be absolutely correct. The manager would like to determine a time
interval such that the probability is at least ~ that the order will be received during that time. That is,
1 8
1--=-
k' 9
so that k = 3 and j1 ± ka gives 8 ± 3(1.5) or [3.5 days to 12.5 days]. It is noted that this interval may
very well be too large to be of any value to the manager, in which case he may elect to learn more
about the distribution of lead times.
2·7 SUMMARY
TItis chapter has introduced the idea of random variables. In most engineering and man-
agement applications, these are either discrete or continuous; however, Section 2-2 illus-
trates a more general case. A vast majority of the discrete variables to be considered in this
book result from counting processes, whereas the continuous variables are employed to
model a variety of measurements. The mean and variance as measures of central tendency
and dispersion, and as characterizations of random variables, were presented along with
more general moments of higher order. The Chebyshev inequality is presented as a bound-
ing probability that a random variable lies between Ji- ka and Ji + ka.
50 Ch2.pter 2 One-Dimensional Rando:::a Variables
2-8 EXERCISES·
2-1. A five-a.rd poker hand may contain from zero to
f~~ aces, If X is the random variable denoting the x=0.1.2 .....
number of aces, enumerate the nmge space of X Vlhat
are the probabilities associated with each possible
=0, otherwise.
vitue of Xl Find the probability that he will have suits left over at
2~2. A car rental agency has either 0. 1, 2. 3, 4, or 5 the season's end. £: '3 ,
'2 J
", ·th
ears returned each U4y, Wi pro
bahil"lUes 6'
I I , J
6' J' 12' z"lO. A ra.:1dom variable X has a CDF of the form
t. -h.
and respectively. Fmd the mean and the vari~
(1 V+1
ance of the number of cars returned..
Fx(x) = 1-(2:) • x=0,1,2, ....
" .~~ random variable X has the probability density
function ce-x, Find the proper value of c, assuming 0 $ =0. x<{}.
X < Z<:I. Find the mean and the variance of X (a) Find the probability function for X.
2.4t The cumulative distribution function that a tele- (h) Find PodO < X ~ 8).
vision tube will fail in! hours is 1 - e-.:t, where c is a
2Ml1. Consider the following probability density
parameter dependent on the manufacturer and t 2::: 0. f';![Lction:
Find the probability density function of T. the life of
fxf.x)=/a; O~x<2,
the tuhe.
=k{4-x). 2';x';4.
~Consider the three functions given below. Deter- := 0, o1heJ1lf1.se.
mine which functions are distribution functions
(a) Fmd the value of k for which f is a prob.bility
(CDFs). 0-7J...
demity function.
>" (a) Fodx) 1 O<X<<><>.
(b) Find the mean and variance of X
_~. (h) GJ.x) ~ e~. O.s:;x<oo,
(c) Find the C1J.l):J,ulative distributionfuncdon,
=0, x<O.
2--12. Rework the above problem, except let the prob-
" (e) HJ.x)=e'. -_<xS;O.
ability density function be defined as
= 1. x>O.
f/..x)=lo:. O::;'x<a"
2,6. Refer to Exercise 2-5. FInd the probability den- = k{2a -x). a";x'; 2a.
sity function corresponding to the functions given, if .' ;,:. ; ~:) '/ 0, otherwise
they are distritaltion functions. ~:
.a.:l~~The manager of a job shop does not "know the
(l!J)Wbich of the f'ollowillg functions are ~screte pr05ability distribution of the time required to com~
probability distributioDs? " plete an order. Howe'ler. from past pcdo:rmance she
( 1 has been able to estimate the mean and variance as :!.4
va) Px () x =-3' x=O. "
deys and 2 (deys) • respectively. Fmd an interval such
, 2 that the probability is at least 0.15 that an. order is fin-
x=l,
, ··0
-'1l
=0, othenvise.
ished during that time.
2~14. The continuous random variable Thas the prob-
(
(5\1 2" ability density funcnonj(') = ki'- for -I ,; ,:5 O. Find
J,3 j l::d .
/1,5-,1:'
,/(h) Px(x) = x x = 0-1.2,3,4,5. 1hc following:
'i (a) The appropriate value of k.
=0. otherwise.
(h) The mean and variancc of T.
2..8. The demandfof a product is-l. 0, +1) +2 per day (c) The cumulative distribution function.
W1'th prob a biliti' . 1
'es SJ 'IO'5'lD,respectlvey.
J 2 3 Ad emand
of -1 implies a Ul1it is returned. Find the expected 2--15. A discrete random variable X has probability
demand and the variax:.ce, Sketch the distribution functionpX<x), where
~n(CDF). pl.x) = k(1/2t. x=I.2.3,
•~.
:r f"
"~ manager 0 a men s clothing store 1$ con~
. =0, otherwise.
\~'"Iled over the inventory of suits, which is currently (a) Fmd k;
30 (all sizes). The number of suits sold from nOw to (h) Fmd the mean and variance of X.
t.~e end of the seasor. is distributed as (c) Fmd the cumulative distribt.>tion ~ctionFxf.x).
2-8 Exercises 51
2-16. The discrete random variable N (N =. 0. 1, ... ) 2-.21. A continuous random 'fanabIe X has a density
has probabilities of OCC1i.'TCnce of kf1 (0 < r < 1). Find function
the appropriate value of k.
/-~ 2x
. ,,2--11:I'he postf. service requires, on the average, 2 f.(x)~9' o<x<3,
days to deliver' a letter across town. The variance is
~O otherwise,
estimated to be 0.4 (dayyz. If a business executive
wants at least 99% of his letters deL'vered on time, (a) Deve!op the CDF for X. V
how early should he mail them?
I,l1) Find the ~ of X and the VlIIWnce of X . /
2-18. Two different real estate developers, A and B, :!lc)~ Find 14. ':.- jr,...
O""'n pa..--cels of land being offered for sale, The proba-
bility distributions of selling prices per parcel are
'rd) F1nda value m such that PI-X l!:m) =Px(X 5m),
This is called the median of X 5f,f-2.,
sho'W"TI in the following ~ab1e,
2-22. Suppose X takes on the values 5 and -5 with
Price t.
probabilities Plot the quantity P[JX - 1'1'" k<r] as a
ftmction of k (for k > 0). On the same set of axes, plot
Sl000 $1050 Sl100 $1150 $1200 $1350 the same probability determined by Chebyshev's
A 0.2 0.3 0.1 0.3 0.05 0.05 inequality.
B 0.1 0.1 0.3 0.3 0.1 0.1 ~:;l;:mLnd the cumulative distribc.tion function asso-
dated with
Assuming that A and B are operating independently,
X If 2
x \
co:rrqJute the fo:Cowi."lg: fx(x)='::r exp - - 2 :" t>0,;;:2:0,
• \ 2t I
(a) The expected selling price of A and of B.
O. otherwise.
(b) The expeeted selling price of A given that the B
selling price is $1150.
2~24. Find the cumulative distribution function asso-
(c) The probability marA and B both have the same ciated with
selling price.
2-19. Show that the probability function for the sum of
values obtained L"l tossing two dice may be written as
x-I
Px(X) =36' x=2.3, ... ,6,
2-25. Coasider t.i.e probability density :unctioa
13-x f,{y) ~ k sin y. 0 ;; y S 7If2. What is the appropriate
x~7.8 .... ,12.
~36' ruue of k? Find the mean of the distribution,
2~2(}. Fbd the mean and variance of
the rando.!1! vari· 2-26. Show tba~ central momen!S can be expressed in
able whose probability fur.ction is defined in the pre- of origin moments by eq:;:aUon 2-19, Hint See
ter!'!1s
vious problem. Chapter 3 of Kendall and Stuart (1963).
Chapter 3
3-1 INTRODUCTION
Engineers and management scientists are frequently interested in the behavior of some
function. say H, of a random variable X. For example, suppose the circular cross-sectional
area of a copper wire is of interest. The relationship Y = 1tX2/4, where X is the diameter,
gives the cross-sectional area. Since X is a random variable, Y also is a random variable, and
we would expect to be able to determine the probability distribution of Y =H(X) if the dis-
tribution of X is known. The first portion of this chapter will be concerned 'With problems
of this type. This is followed by the concept of expectation, a notion employed extensively
throughout the remaining chapters of tlris book. Approximations are developed for the
mean and variance of functions of random variables, and the moment-generating function,
a mathematical device for producing moments and describing diytributions, is presented
with some example illustrations. (
s RX Ry
x +----"'----f-+ • y = H(x)
52
3-2 Equivalent Events 53
Definition
If X is a random variable (defined on ff) having range space Rx , and if H is a real-valued
function, so that Y = H(X) is a random \lariable with range space R y. then for any event
C c R", we define
(3-1)
It is noted that these probabilities relate to probabilities in the sample space. We could
write
P,.(C)=P«(ee 9': H[X(e)) E C)).
However, equation 3-1 indicates the method to be used in problem solutiou. We find the
event Bin Rx that is equivalent to event C in Ry; then we find the probability of event B.
:!:~~~~fl@~
In the case of the cross-sectional area Yof a ..vire. suppose we know t.'mt the diamete: of a Vvire has
density function
, ~)'
PxliXi';~1.01
" '
r;-;-) ~
=PxILO:;;X';vLOI
,
W·200dx
1.000
OI
0,9975,
rj;:1¥#Yl~_'~~J
In the case of the Geigercouater experiment of Example 2-12. we used the distribution given in equa-
tion 2-7: ~
= O. or.."lerwise.
Recall that .A.. where A > 0, represents the Clean "hit" rate and l is the tir!J.e interval for which the
counteI: is operated. Now suppose we wish to find P~Y:5 5), where
Y~2X+2,
The event {XE Rx: xst} ism the rnngespace ofX. and we have thefuoctlonpx to work with in that
-=---..- . .- . .- - - - - - - - - - - -.. . .
space,
~
54 Chapter 3 Functions of One Random Variable and Expectation
For example, in Fig, 3·A where s, = 4, the probability of y,ls phD = Px(x\) + px<xJ + Pl.x,,)
+ p,!.xJ. '
In the special case where H is such that for each y there is exactly one x, then
ph) = Px(x). where y, = H(x;). To illustrate these concepts, consider the following examples.
In the coin-tossing experiment where X represented the number of heads, recall that X assmned four
values, 0, 1,2, 3. ""'ith probabilities t. t. t, t. If Y:::: 2X ~ I, then the possible values of Y a.::e - 1, 1.
3, 5. and p,(- 1) =t. py(l) ~ 1. py(3) =1, py(S) =}. In this case. H is such that for each y there is
exactly one x.
i¥}~21~,~i1 V
Xis as in the previollS example; however. suppose now that Y:::: ~ - 21, so that the possible values of
Ya:e 0, 1, 2, as indiCllted in Fig. 3·3. In this case
py(O) = Px(2)=t,
4
py(l) = Px(l)+ px(3)=S'
1
pA2)= Px(O)=g'
Rx
H Ry
Xi; •
I
) \!
I Xi::!_
i
I Xl::;_
I
Xil\_
Ax y=H{x)=:x-21 Ay
(~+=============rrl!~"'101
8 ~
\111'
In the event that X is continuous but Y is discrete, the formulation for p.f.y) is
(3-3)
where the event B is the event that is equivalent to tlre eveot (Y = y) in Ryo
··iX;;~~I~~-.~l \//
Suppose X has the exponential probability density function given by
N?:) = kr", X" 0,
=. 0 otherwise.
Furthermore, if
y=o forXS l/)"
for X > 1/}...
and
lfX is a continuous'random variable with probability density functionf", and His also con-
tinuous, tben Y =H(X) is a continuous random variable, The probability density function for
the random variable Ywill be denotedfyand it may be found by performing three steps.
1. Obtain the CDFof Y. F,(y) = P,(Y ';'y). by finding the eventBinRx• which isequiv-
alent to the ovem (Y;; y) in Ry.
2. Differentiate F,(y) with respect to y to obtain the probability density functionf,ty).
3. Find the range space of the new random variable.
56 Chapter 3 Functions of One Random Variable and Expectation
~x~~ie.§
Suppose that "fue random variable X has the following density function:
If y= H(X) is therendom variable for wbich the density ilis desired, and H(>:) =0:+ 8, as shov;min
Fig. 3-4, then we proceed according to the steps given above.
•,... f,h(y'
}
= R'(Y')
r = 32 41
3. If .x-=O.),= 8, and ifx=4,)':=. 16, so then we have
y 1
tr(Y) = 32 -4' g<'y<,16,
• = 0, otb.:TNise.
Y= H(x) 1
I
I Hex) 2x+8
16~
i
y= H(x}
.:§~~l~;~:i;
Consider the random variable X defined in Example 3-6, and suppose Y=H(X) = (X - 2?, as sbown
in Fig. 3-5. Proceeding as in Example 3-6, we find the following:
1. Fh) ~ P!.Y~y) ~ Px«X- 21' ~y) ~px(-.JY ~ X -2 ~.JY)
~Px(2-.JY ~H2+.JY)
2+-5 x x 2 [+-!Y
~t-!Y8dx~16 c
2-'0'
~!...JY.
2
=0 othenvise.
In Example 3-6, the event in Rx equivalent to (y" y) in Ry was [X'; (y - 8)12]; and in
Example 3-7, the event in Rx equivalent to (Y'; y) in Ry was (2 X ,; 2 -.JY ,;
In the +.JY).
first example, the funetion H is a strictly increasing function of x, while in the second
example this is not the case.
Theorem 3-1
If X is a continuous random variable with probability density function/x that satisfies/X<:<)
> 0 for a < x < b, and if y = H(x) is a continuous strictly increasing or strictly decreasing
function of x, then the random variable Y ~ H(X) has density function
. l~fX~~·r£I,I (3-4)
with X~H-l(y) expressed in terms ofy. IfHitiiicreasmg, then/,,(y) > 0 if H(a) <y <H(b);
and if H is decreasing, then/,,(y) > 0 if H(b) < y <H(a).
Proof (Given only for H increasing. A similar argument holds for H decreasing.)
F,,(y) ~ p,(Y'; y) ~ P,(H(X)'; y)
~pX[X';Wl(y)]
=Fx[W1(y)].
'() dFx{x) dx
/y ()
y = Fy y =~. dy' by the chain rule,
3-5 EXPECTATION
If X is a random variable and Y = H(X) is a function of X, !bell the expected value of H(X)
is defined as follows:
\
(3-5)
. E[H(X)] =
----~-,.~---
rH(x)· fx(x)dx for X conti~':..
.. _.--- -~ ~-" - ---~....--~
(3-6)
In the case where X is continuous, we restrict H so that Y;;:;:. H(X) is a continuous random
variable, In reality. we ought to regard equations 3-5 and 3·6 as theorems (not definitions),
and these results have come to be known as the law of the unconscious statistician.
The mean and variance, presented earlier, are special applications of equations 3-5 and
3·6. If H(X) = X, we see that
E[H(X») =E(X) =/1. (3·7)
Therefore, the expected value of !be random variable X is just lhe mean, p..
If H(X) = (X - J1.?, then
E[H(X») = E«X - /11') = d'. (3-8)
Thus, lhe variance of the random variable X may be defmed in terms of expectation. Since
the variance is utilized extensively. it is customary to introduce ~a variance operator V that
is defined in terms of the expected value operator E;
V[H(X)] = E([H(X) - E(H(X»)]~. (3-9)
Again, in the case where H(X) = X,
and
ilk = E[(X - E(X))''j.
There is a special linear function H that should be considered at this point. Suppose that
H(X) = aX + b, where a and b are constants. Then for discrete X, we have
-._____c
'I E(aX + bl = I'(ax, + b )Px(x,)
- ,
= a 2,xi Px(x')+b 2,Px(x,)
i, .. ~
faE(X)+b'J (3-12)
~= [(ax+b)fx(x)dx
=a[ xfx(x)dx+b[fx(x)dx
=aE(X)+b, (3-13)
and
V(b) = O. (3-16)
These results show that a linear shift of size b only affects the expected value, not the variance.
Suppose X is a random variable such that E(X) = 3 and VeX) = 5. In addition, let H(X) = 2X -7. Then
EIH(X)] =E(2X -7) = 2E(X) -7 =-1
and
V[H(X)] = V(2X -7) = 4V(X) = 20.
Suppose a contractor is about to bid on ajob requiring X days to complete, where X is a random vari-
able denoting the number of days for job completion. Her profit, P, depends on X; that is, P = H(X).
The probability distribution of X, (x, p/..x)), is as follows:
60 Chapter 3 Functions of One Random Variable and Expectation
x Px(xl
[
3
8
5
4
ii
5 ~
8
otherwise, O.
Using the notion of expected value, we calculate the mean and variance of X as
1 5 2 33
E(X) = 3·-+4·-+5·-=-
8 8 8 8
and
V(X)=[3,.2.+4,.~+52.~J_(33)2 = 23
8 8 8 8 64
If the function HUG is given as
x H(x)
3 $10,000
4 2,500
5 -7,000
and the contractor would view this as the average profit that she would obtain if she bid this job many,
many times (actually an infinite number of times), where H remained the same and the random vari~
able X behaved according to the probability function Px. The variance of P = H(X) can readily be
calculated as
A well-kn~ simple inventory problem is "the newsboy problem," described as follows. A newsboy
buys papers for 15 cents each and sells them for 25 cents each, and he cannot return unsold papers.
Daily demand has the following distribution and each day's demand is independent of the previous
day's demand.
Number of customers x 23 24 25 26 27 28 29 30
Probability, pIx) 0.01 0.04 0.10 0.10 0.25 0.25 0.15 0.10
If the newsboy stocks too many papers, he suffers a loss attributable to the excess supply. If he stocks
too few papers, he loses profit because of the exeess demand. It seems reasonable for the newsboy to
3·5 Ex.pec:ation 61
stock some number'bf ;tapers so as to ;ninirnize the expected loss. If we let s represent the number of
papers stocKed, Xtbe daLy demand, andL(X, s) the newsboy's loss for a pa..>1:icular stock level $, then
the loss is simply
For 3=26.
Fors=27.
E[L(X, 27)] = 0.15i(27 - 23)(0.01) + (27 - 24)(0.04) + 07 - 25)(0.10)
+ (27 - 26)(0.10) ... (27 - 27)(0.25): ... 0.10[(28 - 27)(0.25)
+ (29 - 27)(0.15) + (30 - 27)(0.10),
= $0. 154().
For $ 28,
BIL(X, 28)J =0,15[(28 - 23)(0.01) + (28 - 24)(0.04) + (28 - 25)(0.10)
(28 - 26)(0.10) + (28 - 27)(0.25) + (28 - 28)(0.25)1
+ 0.10[(29 - 28)(0.15) + (30 - 28)(0.10))
$0.1790
Thus, the newsboy's policy should be to stock 27 papers if he desires to minimize his expected loss.
At least one of the units must function, the redundancy is standby (meaning that the second unit does
not operate until the first falls), sMtching is perfect, and the system is nonmaintained. It can be
shown that ~det' certain conditions, when the time to failure for each of the units of this system has
an exponential distribution, then the time ttl failure for the system has the following probability den·
sity function.
62 Chapter 3 Functions of One Random Variable and Expectation
Hence, the redundancy doubles the expected life. The terms "mean time to failure" and "expected
life" are synonymous.
Using only the first two terms and grouping the third into the remainder so that
Y =HC!l) + (X - Il) . H'C!l) + R I,
where
The surface tension of a liquid is represented by T (dyne/centimeter), and under certain eonditions,
Too 2(1- 0.005X)1.2, where X is the liquid temperature in degrees centigrade. If X has probability den~
sity functionjx, where
3-6 A;>proximations to E,H(X)] and V(H(X)] 63
and
V(T)
.r~, 4(I-D.005x)'·' .3000x-4dx-(E(T»)' .
In order to determine these values, it is necessary to evaluate
r(1-0-OO5x)" ax.
J:o X4
Since the evaluation is difficult., we use the approximations given by equations 3-17 and 3-13. Note
that
Since
H'(lS) =- 0.012,
li"(15) = O.
Using equations 3-17 and 3-18
and
V(1) = [H'(I5)]" d'= [- 0.012]'·75 = 0.Q108.
Theorem 3·2
Suppuse that X is a random variable with CDF F ,{x). Then the random variable F,{X) has
a uoiforrn distribution on [0, 1]; that is, it has probability density function
Jul.u) = 1, OS uS I,
=0 otherwise. (3·19)
Proof Suppose that X is a continuous random variable. (The discrete case is similar.)
Since X is continuous, its CDF Fx(x) has a unique inverse. Thus, the CDF of the random
variable F x(X) is
P(FiX) S u) = P(X S F!i(u) = F,{F!i(u» = u.
Taking ilie derivative wiili respect to u, we obtain ilie probability density function of F ,{X),
d
duP(Fx(X),;;u)~I,
Suppose that U is unifO,ITn on [0,1]. We will show how to use the inverse transfonn method
to generate aD exponential random variate with parameter ).,. that is, thc continuous distribution
3~ 7 l.c:.e Moment~Genera.ting Funetion 65
having probability density functionfx.Cx) :;; A,e-.tr. for x 2: 0. lne CDF of the exponential random
va..-iable is
where we obtain the inverse after a bit of algebra. In other words, -(1/A.) In(1 - [1) yields au expo-
nential random va..-iable with parameter A.
We remark that it is not always possible to find Pi in closed form, so the usefulr.ess
of equation 3-20 is sometimes limited. Luckily, many other random variate generation
schemes are available, and taken together, they span all of the co=only used probability
distributions.
Definition
Given a random variable X, the moment-generating funetionMJt) of its probability distri-
bution is the expected value of ex. Expressed mathematically,
M,ft) = E(ex) (3-21)
For certain probability distributions, the moment-generating funetion may not exist for
all real values of t. However, for the probability distributions treated in this book, the
moment-generating function always exists.
Expanding etX as a power series in t we obtain
(3·24)
66 Chapter 3 Functim:.s of One Random Variable and Expectation
Thus, we see that when ~41yj..t) is written as a power series in r~ the coefficient of fir! in the
expansion is the 7th moment about the origin. One procedure, then, for using the moment~
generating function would be the following:
1. Find Mxit) analytically for the particular distribution.
2. Expand M,(t) as a power series in t and obtain the coefficient of fir! as the 7th origin
moment.
'The main difficulty in using this procedure is the expansion of Mit) as a power series in t.
If we are only interested in the first few momenTS of the dislribution, then the process
of derenninlng these moments is usually made easier by noting that the rth derivative of
Mxil), with respect to t, evaluated at t= 0, is just
d'
_.Mx(t)1
dt T t=
0 = EflX'e"'] t=O = J.l;, (3-25)
assuming we can interchange the operations of differentiation and expectation. So, a sec-
ond procedure for using the moment-generating function is the following:
1. Determine M/..,) analytically for fue particular distribution.
2. Find J.l; = !', Mx(tll,.o·
.Moment-generating functions have many interesting and useful properties. Perhaps fue
most important of these properties is that the moment·generating function is unique when
it exists, so that if we know the moment-generating function. We may be able to determine
the form of the distribution.
In cases where the moment-generating funetion does not exist, we may utilize tbe
characteristic function, C/"t), which is defined to be the expectation of ,fiX, where i = H.
There are several advantages to using the characteristic function rather than the moment-
generating function. but the principal One is that Cx{t) always exists for all t. However. for
simplicitYj we will use only the moment·generating function.
!i[(~mRf~~tI~~:j
Suppose that X has a binomial distrfbutif:m., that is.
(It~ .. _ ..
Px(x) =l X'
/
p (I-p) , x:= 0,1,2,. '".!t,.
=0, otherwise.
where 0 < p ~·l and n is a pOSitive integer, The moment-generating fimction MJ,t) is
This last summation is re<:ognize<! as the binomial expansion of[pe' + (I-p))". so that
and
M~(t) = npe'(l- p +npe')(: - p(e' -l)l~'.
Thus
and
14 =M~(t)I",,= np(l- p'" npl.
The second cenrral moment may be oblained using cr;;;;;; J.4 - [il;;;;;; np(1- p).
;i!~p!~~~~E'
Assume X to have the following gamma distribution:
b-l-~
b
a
fxt")=·····-x e O:5x<-,a>O,b>O,
reb) ,
== O. otherwise.
,
Mx(t)=_a_;;;;;;J( 1-!.. )-' fort < a.
(a-t)' \ a
(1--'- r.
aJ
we find
u' b ,_0(0+1)
. 1 and !1i ---,-.
a a
3-8 SUMl\HRY
This chapter first introduced methods for determining the probability distribution of a ran-
dom variable that arises as a function of another random variable with known distribution.
That is, where Y=H(X), and either Xis discrete with knowndistributionp.,{x) or X is COIl-
68 Chapter 3 Functions of One Random Variable and Expectation
tinuous with known density !x(x), methods were presented for obtaiIDng the probability dis-
tribution of Y.
The expected value operator was introduced in general terms for E[H(Xl], and it was
shown that E(Xl = J1, the mean, and E[(X - J1l'] = ri', the variance. The variance operator V
was given as V(Xl = E(X') - [E(Xl]'. Approximations were developed for E[H(Xl] and
VlH(K)] that are useful when exacr methods prove difficult. We also showed how to use the
inverse transform method to generate real.izations of random variables and to estimate their
means.
The moment~generating function was presented and illustrated for the moments f.l~ of
a probability distribution. It was noted that E(X') = 1';.
3-9 EXERCISES
3~1. A robot positions 10 units .t.... a chuck for machinv
(b) If the profit per sale is $200 and thc replacement
iog as the chuck is indexed. If the robot positions the of a picture tube eosts $200, tind the expected
unit improperly, the unit falls away and the chuek profit of the business.
position rCl'.l:lains open, thus resulting in a cyele that 34~ A contractor is going to bid a project, and the
produces fewer than 10 units. A study of the robot's number of days, X, required for completion fonows
past performance indicates that if X =number of open the probability distribution given as
positions,
pix) =0.1. x= 10.
px(x)=O.6, x=o, =0.3, x= 11,
=0.3, x= 1, =0.4, x=12,
0.1, x 2, =0.1, x=13,
=0.0, otherwise. =0.1, x=14.
'= 0, otherwise.
If the loss due to empty positions is given by Y= 20;(-!,
The contractor's profit is Y"", 2000(12 - X).
find the following:
(a) Find the probability distribution of Y.
(a) p,.iy).
(b) Find E(X). V(X). E(y). and V(Y).
(b) E(Y) and VCy).
3-5. Assume thaI: a continuous random variable X has
3-2. The content of magnesium in an alloy i.s a random probability density function
variable given by the following probability density
function: fx(x) =2xer r , X~O,
=0, otherwise.
bighwtest and sells for 98 cents per gallon, Find the Evaluate E(A) :and V(A) by using the approximations
expected re'>'enue per gallon wberefA(a) == 1. O!::a S 1; derived in this chap~.
otherwise.k(a) O. 3~13~ Suppose that X bas the uniform probability
3..8. The probability function of the random variable X. density function
- I -(ve)(x-p) 1 :5'x $; 2,
f x (x) -Be ,
otherwise.
=0, othenvise
(aJ Flnd the probability density funetion of Y = H(X)
is knO'tVtl as the two~parameter exponential distrfbu, wbere H(x) = 4 - x'.
non. Find the moment-generating function of X Eval~ (b) Find the probability density function of Y= H(X)
uate E(X) and V(X) using the moment-generating where H(;<) = e,
function.
3-14. Suppose that X has the exponer.'tial probability
3--9~ A random variable X has the following probabil~ density function
ity density function:
x:::: 0,
fl.x) = a-', X:>O. otherwise.
O. othel;'Wise,
Find the probability density function of r = FI(X; where
(al Develop the density function for Y= 2 X'.
(b) Develop the density for V = X·n. 3
H(X)=~),.
(cJ Develop the density for U = In X tl+x
I •
3--10. A two-sided rotating antenna receives signals. ~ used-car salesman finds that he sells either 1.
The rotanonal position (angle) of the antenna is denoted 2,3,4,5, or 6 ca.."'S per week with equal probabilIty.
X. and It may be assumed that this position at the tlme ~ind the moment-generallng function of X.
a IUgnal is received is a ra.'1dom variable with the den~ (b/usin~ the moment~ger.e:ating function fincC.E('X':
sity below. Actually, the randomness lies in the signal, ~ and VgJ/
'.,..>' ' I
r--'
1,4·1 ) INTRODUCTIO:--;
In many situations we must deal with two or more random variables simultaneously. For
example, we might select fabricated
~--'--- .....
sheet steel specimens
-.----~- - - and
-_ measure
.•.. shear
.._ strength __
- 'and
..
2!]'ld.diameter 01 sP9.!.lilelds. Thus, both weld !'lieru:.str!'llW..?,ll..Q weld di.'!!lleter:."!.e the ran-
~jlbl"'u;>.f mt",re5t. Or we may select people from a certain population and
their height and weigbL '. . . . .. -- - - - - . ' , . . ...'
mcasure .
-The objective "filii, chapter is to formulate joint probability distributions for two or
more random variables and to present methods for obta1.Ung both marginal and coruItiwnal
distributions. Conditional ~~"~~(tas w.tl.ll~!pe regre~~91tpfm[~~!1' We
also present a definition of irul.ependence for random variables, and covariance and corre-
lation are defined. Functions of two or more random variables are p~:;;Sentecf~~d a specIal
~f linear combinations is presented with its corresponding moment~generating func-
tion. Finally the law of large numbers is discussed.
Definition
If if is the sample space associated with an experiment ~, and Xl' X1;> •• '. Xk are functions.
each assigning a rea! number X,(e), .'I:,(e) .... Xie) to every outcome e, we call [Xl' X" ....
XJ a k-dimensional rtJJ]dom vector (see Fig. 4-1).
The range space of the random vector [Xl' X" "', X,l is the set of all possible values of
the random vector. This may be represented as Rx, XX1Y.". xXe where
Rx,xx~x ... xx.. = ([Xl'Xz., ... , x,J:XI E Rx,>.t2 E Rx~. ""Xk E Rx)
This is the Cartesian product of the range space sets for the componentS, In the case where
k =2, that is, where we have a two-dimensional random vector, as in the earlier illustrations,
RXI xX. is a subset of the Euclidean plane.
e.~~__~__________~~__~~R~
r-~./~_.,Rx,
~
71
72 Chapter 4 Joint Probability Distributions
Consider the case where weld shear s::rength a.'1d weld diameter are meas:lred. If we let Xl represent
diameter in inches andX2 represen.t strength In pounds, and if we know 0 ~x! <.0.25 inch while 0 ~
x, S 2000 poUlJds, then me ra:nge space for [X" X,] is the ser {[x" x,]: 0 ~ XI < 0.25, 0 "-"z ;; 2000].
This space is shoWD graphically in Fig. 4~2.
1'E,~~~~~1:'
A small pump is inspe..-red for four quality-co;ltrol characteristics. Each charaCteristic is classified as
good, mino::: defect (not affecting operation) or IIJ.ajor defect (affecting operation). A pump is to be
selected and the defects counted. IfXt :::::: the number of minor defects andX2 = the number of major
defects, we know that Xl =. 0, 1, 2. 3. 4 and Xz =. O. I, ,.,' 4 - X I because only four characteristics are
inspected. The range space for [X" X,] is thus ([O, 0], [0, I), (0, 2], [0, 3), (0,4). (1, 0], [I, IJ, (1,2],
[1,3], [2, 0], [2, 1], [2, 2], [3,0], [3, 1], [4, 0]). These possible outcomes are shown in Fig. 4-3.
X2 (pounds)
2,000
2~'--~--~--+---+---
In presenting the joint distributions in the definition that follows and througbout the
remaining sections of this c!tapter, where no ambiguity is introduced. we shall simplify the
notation by omitting the subscript on the symbols used to specify these joint distributions.
Thus. if X = [X,. X,J.Px(x,. x,) = p(x" x.J and/x(x,. x.J =/(x;, x,).
Definition
Bivariate probability functions are as follows.
o
~
!?iscrete c"'!i!; To each outcome [x;, Xzl of [X,. X,l. we a'sociate a number.
p(x,.x.;J =P(X, =x, and X, =x,),
where
and
(4-1)
The values ([x,. x,l. p(X,. Xz») for all i, j make up the probability distribution of
[X,. X,J.
(3) Euclidean
Contim{OHS ~I! [Xl' Xi] is a continuous random vector with range space R in the
plane, tlienf, the joint density fwtction, has the follO\Ving properties:
for all (x" x.,) E R
and
'"
Figore4-4 A bivariate den..~ty·funct:ion. w!:.ere peal ::;;X:::; oJ' az S:Xz ~07) is given by the shaded
volume.
74 Chapter 4 Joint Probability Distributions
It should again be noted thatj(x" x,) does not represent the probability of anything. and
the convention thatj(x\, x,) ~ 0 for (Xl'>;') '" R will be employed so that the second prop"
erty may be ",'ritten
In the case where [X,. X,J is discrete, we migbt present the probability distribution of
[Xl' X 2] in tabular, graphical, or mathematical fonn. in the case where [Xl' X,J is continu-
ous, we usually employ a mathematical relationship to present the probability distribution;
however, a graphical presentation may occasionally be helpful.
;ll~~~f~'
A hypothetical probability distribution is show::t in both tabular and graphical form in Fig. 4--5 for the
random variables defined In Example 4-2.
~i>l;;;~:;(
In the casc of the weld diameters represented by Xl and tensile strength represcnted by Xl' we might
have a uniform distribution as by
!~ o I 1 2 3
L=-J
I 0 1/30 ! 1/30 2/;!{) I 3/30 I 1/30
, 1 1/30 i 1/30 3130 i 4/30 I \
:
2 1/30 , 2/30 .'lI3Oi I !
: 3 1/30 I 3/30 i
,
i
i 4 ,3/30 ,
I I I
(a)
pix;, xz}J.
3/30 X,
----- -
1/30
.'lI3O
----- -----
3/30
-----
4/30
-
~30-- ~.
0
2 3 4 x,
(b)
Figure 4~5 Tabular and graphical presentation of a bivariate probability distribution. (a) Tabulated
values are p(x1• xJ. (b) Graphical presentatiQIl of discrete bivariate distribution.
Uf
4-3 Ma..--ginal Distributions 75
"
O$X, <0,25,0$'2 ~2000,
;:; 0, other"<Y1Se,
The range spare was shO'Vi'D. in Fig. +2, and if we add another dimension to display graphically
y;. fix l ,;0, then the distribution would appear as in Fig, 4-6.ln the univariate case, area corresponded
to probability~ in the bivariate case. volume under the surface represents the probability.
For example, suppose we wish to find P(O, 1 ,,0,2, 100 ~X2:;; 200), Tnis probability would
befound by integratingj(x" x,) over Ihe region OJ "X,
,,0,2, 100:5 X, :> 200, That is,
rlOO r"~ 1 1
J100 JO•1 500
/-
,G MARGDlAL DISTRIBUTIOKS
,~
Having defined the bivariate probability distribution, sometimes called the joint probabil,
ity distribution (or in the continuous case the joint density), a natural question arises as to
the distribution of Xl Of X 2 alone. These distributions are called marginal distributions. In
the discrete case, the marginal distribution of Xl is
r
I
p,h)= LP(X',x,) 1for all
,r, -J
x, (4-2)
(4-3)
Jf~~R!~1f~;
In Example 4-2 we considered the joint discrete distribution shov.~ in Fig, 4-5. The marginal distri-
butions are shown in Fig. 4-7. We see that [Xl' Pl(X;)] is a univariate distribution a11d it is the distribu~
tiQn of XI (the number of mL.1.OC defects) alone. Likewise [.l'.1. p:.(~)] is a unl\<1.irlate distribution and it
is the distrib>J.lion (the number of major defects) alone.
(4-4)
= 2.000
Ni 0 1 2 3 4 p,(x.) I
i
1
o I 1130 liS(] 2130 8130 1/30 : 8130 I
P,(X,»)
La
P,('I) +
ai30 9130
! /7130
o 1 2
(b)
3
,1/30
4 X,
lCL
o 1
1
.
L
6130
2 --3:--- 4
(0)
4130
1 j3/30,
'2
Figure 4~7 Marginal distributions for discrete [XI' X:.J. (a) Marginal distributions-tabular fann, (b) Mar~
ginal dislribution (x,.p,(x,», (el Marginal distribution (x,. p,(x,),
(4-5)
The function/, is the probability density function for Xl alone, and the fUllctionh is the den-
sity function for X2 alone.
'~§l?!~*,6)
In Example 4-4, the joint donalty 01: LX" X,] was given by
otherwise.
and
otherwise.
These are shown graphically in Fig, 4-8,
4-3 Marginal Disu;butions 77
fl(X:~ .
LL_ o
(a)
0.25 Xl
1/2000
o
(b)
2,000 "z
Figure 4~8 Marginal distributions for bivariate uniform vector [Xl' Xzl. (a) M..arginal>:iistribution of
X" (b) Marginal distribution of X"
Sometimes the maro-.Jnals do not come out so nicely, This is the case, for example,
when the range space of [Xl' X,) is not rectangular.
E~~l~L~f;s
Suppose that the jOint density of [X;, Xii is given by
~
t,(XI)~ [. f(x"x,)dx,
:= f6x
x, t dx 2
forO<x\<':"
and the m~.-ginal of Xz is
= 1: 6x z
1 d:t1
TIle expected values and variances of X, and X, are determined from the marginal dis-
ttibutions exactly as in the univariate case, 'Where [X:. X2 ] is discrete, we have
= LLx~p(XI'X,)-.u~, (4-7)
" x,
78 Chapter 4 Joint Probability Distributions
and, similarly,
(4-8)
(~9)
In Example 4~5 and Fig. 4--7 marginal distributions for XI and ~ were given. Working with the ntar-
gjnal distribution of Xl shown in Fig. 4-lb, we may calculate:
7 7 S 7 1 g/,/
E(X)=
1
"1 =0·-+1·-+2·-+3·-+4·-=-
~ 30 30 30 30 30 5
V
and
103
~75
The mean and variance of X2 could also be detennined using the m.aIginal distribution of ~~:
Equations ~6 tlJrough ~9 show that the mean and variance of X, and X" respectively,
may be determined from the marginal distribetions or directly from the joint distributioIL
In practice, if the marginal distribution has already been determined, it is usually easier to
make use of it.
In the case where [Xl' X21 is continuous, then
(~11)
and
= r xifz(xlJdx, - J.Li
(4-13)
4-4 Conditional Distributions 79
Again, in equations 4~10 through 4~I3, observe that we may use either the marginal densi-
ties or the joint density in the calculations.
;~l?:4,!l,;
1L E.xample 4-4. th.e joint density ofwe1d diameters, XJ' and shear strength. X'z. was given as
1
f(x"x,> SIlO' 0';", < 0.25, 0,; x, s2000,
;;;;; 0, otherolf..se.
and the marginal densities for Xl and X2 \\-'ere given in Examp!e 4--6 as
O~xl<O.25, V
and
1
A(",) = 2000' 0';" < 2000,
;;;; O. otherwise.
Wo.dd:J.g with L.1.e [!')arginal densities, the mean and V3..."iance of XI are- thus
and
~
rg
\...-
CONDITIONAL DISTRIBUTIONS
When dealing with two'jointly distributed random variables it may be of interest tcfind..the.
~stribution-Of one of these variables, given a particular value of th~oili~r. That IS-I ·we may
~.~istribution OL~l g:ive.J)~2 s: For example, what is the disttltru&in
of a person's weightg(venfhat he is a particularneight? This prObability distribution would
be ealled the conditional distribution of X, given that X, = "'.
Suppose that the random vector [X" X;J is discrete. From the definition of conditional
probability, it is easily seen that the conditional probability distributions are
P(Xl,X2) (4-14)
p,(x,)
and
(4·15)
i~~I~~~~,t1j
Consider the COU:lting of minor and major defects of the small pumps in Example 4-2 and Fig. 4-7.
There will be five conditional distributions ofXZ, one for each value of Xl' They are shown in Fig. 4-9.
The distribution PXJ'rJ(x,J, fot XI == O,is shown in Fig. 4-9a, Figure 4-9b shows the disaibution Px~: (.xz),
Other conditional distn'butions could likewise be determined for XI 2,3. and 4, respectively. The
distribution of Xl given that Xl == 3 is
otherwise.
(4-16)
and
(4-17)
,~jam~I~*j~:
Suppose the joint density of [Xl' X2] is the functionfpresented here and shown in fig. 4-10:
x, 2
°
p{O,O) p(O,l) p(0,2) p(O, 3)
4
p(O,4)
p,(O) . p,(O) . --p;(O) p,(O) P1(O}
(a)
x, o 2 3 4
- -...- - - - - - - - - - - - - - - - - - -
p{1,0) p(1,1) p(1,2) p(1,3) p(1,4)
P',i ,(x,) = p(l, x 2 )fp,(1) p;(1) p,Cl) . p;(1)
p,(1) p,(1)
(b)
Figure 4--9 Some examples of conditional distributions,
4-4 Conditional DistributioIlS 81
f(x,. x"l
otherwise,
O<xzS;2,
othervrise.
and
",(3x, +x,)
1' /" 1+(x,/2) •
otherwise.
Note that for fx,;,x/.J:;;. there are an i.nfinite number of these conditional densities, one for each value
o<Xl < L TWo of these.fx#!J2)(Xz} andfx:jl(x.z), are shown in Fig, 4-12.A1so, forfx;!xz(,xl)' there are an
o
Figure 4-11 The matginal deusities for Example 4-11.
82 Chapter 4 Joint Probability Distributions
Figure 4-1.2 Two conditional densities iX1jltl and fX~l> from Example 4-11.
o x,
infinite nU!lloer of these conditional densities. one for each va1ue 0 S; Xz ~ 2, Three of these are shov.n
in Fig. 4· 13.
-------------------------------------
]!~pI{£t2
Consider the probability distribution of the discrete random vector [XI' Xz), where X: represer.ts the
number of orders for a large turbine in July and X2 represents the number of orders in August The
joint distributon as wee as the marginal distributions are given in FIg. 4-14. We cor-sider the tJree
conditional distributions PX:IO' PX2:j. and PXzf/. and the conditional e:x.:pected values of each:
I \ 1
'PX~jl(X2J:;:;: 10'
5 1
Xl"'" 1, Xz =1,
10 '4'
3 1
=-
10' ='4'
1
.%2 = 3, =- =0, Xl :.:::3,
JO
:;:;: O. Otherw1Se., =0, =0, otherwise,
, 1
E(x,12)o,O -'-+1'-
2 4
1-:-2'±+3"C~O.75
(4-20)
and
(4-21)
and in each case there will be an infinite number of values that the expected value may take.
In equation 4-20, there will be one value of E(X,ix,l for each value X, and in equation 4-21,
there will be one value of E(X,Ix,) for each value x"
'1!~ii'Ri~,~~l,3.;'
In Example 4-11; we considered a. joint density I, where
\ - 2 +-3-'
J(Xj,X:u-Xj X,x, O<x;S:1,O::';:xz S:2,
=0, otherwise,
X){1 2
0
°
0,05 0,05 O.~O
P2(XiJ
0.2
Figure 4-14 Joint and margiaal distrib..1ion, of [X" X,]. Values in body of table are p(x" X;;,
84 Chapter 4. Joint Probability Distributions
and
(/ )="I'
E\X,!x,
, I)
I 3x, +x,
x 2 ,-·---dxz
2 3x, + 1 ='
+4 '"
=
9x!+3
It should be noted that this t<; a function of Xl' For the two conditional densities shov.-n in Fig. 4~ 12,
t i;
where Xl ;:;: and Xl ::;< 1. the corresponding expected values are ECX'Ji) = andE{X211) ~ #.
Since E(X,ixll is a function of x,, and x, is a realization of the random variable XI'
E(X2iX,l is a random variable, and we may consider the expected valne of E(X2iX,), that is,
E[E(X,IX,l]. The inneroperatoris the expectation of X2 given X, = xl' and the outer expec,
tation is with respect to the marginal density of X,. This suggests the following "double
expectation~' result.
Theorem 4·1
E[E(X,iX,l] = E(X,) f12 (4,22)
and
(4-23)
Proof Suppose that Xl and Xz are continuous random vMiables. (£he discrete case is sim-
ilar.) Since the random _iable E(X,iX,) is a function of X" the law of the unconscious stat,
istician (see Section 3,5) says that
= [, x,[,f(x"x,)dx:dx2
= [ , x'/2(x,)dx2
= E(X2~
which is equation 4-22. Equation 4,23 is derived similarlY, and the proofis complete,
Consider yet again the join.t probability density function from Example 4-: I.
f(.J:l,.:t2);;:.:t~+ -"1;2.
=0, otherv.'ise.
In that example, we derived expressions for the ma..rginal densities f&~l) a."1df:?;(X:!). We also derived
E(X,Jx:)~ (9x, + 4)/(9x, .,. 3) in ExampJc 4-13. Thus.
=
9
Note that this. is also E(XiJ. since
The variance operator may be applied to conditional distributions exactly as in the uni-
variate C3Se.
j[~!3~:f:!5;l
In Example 4-13 we found E~p:) for the bivariate density of Example 4-11, that is,
X.X2
"1"3' O<X1 ~1,O$x2 $2,
=0 otherwise.
(a) (b)
Figure 4--15 Some regression curves. (a) Regression of X2 on X;. (b) Regr-:!Ssion of Xj on X 2 •
86 Chapter 4 Joint Probability Distributions
/
Regression will be discussed further in Chapters 14 and 15.
V\
'4-7/ INDEPENDENCE OF RANDOMVARRBLES
//
The notions of independence and independent random variables are very useful and impor-
tant statistical concepts. In Chapter 1 the idea of independent events was introduced and a
fonnal definition oftbis concept was presented. We are now concerned with defining iru1e~
pendent random variables. Vlhen the ontcome of one variable. say Xt' does not influence
the outcome of Xl, and vice versal we say the random variables Xi and XL are independent.
Definition
1. If [Xl' K,J is a discrete random vector, then we say that Xl and X, are independent
if and only if
(4-24)
Theorem 4-2
1. Let [Xl. Xzl be a discretc random vector. Then
Px,~,(x,) = p,(x,)
and
px~,(X) = PI(X)
and
Ix,,,,(x,) = !; (x,)
for all x, and"" if and only if X, and X, are independent.
.t; (x,)
if and only if
ft.x ,• x;J = I, (x,)f;,(x,) for all Xl and x.,
if and only if Xl and X, are independent, and we are done.
Note that the requirement for the joint distribution to be factorable into the respective
marginal distributions is somewhat similar to the requirement that, for indepeudent events,
the probability of the intersection of events equals the product of the event probabilities.
A city trrulSit service receives calls from broken-down buses and a \VTeeker crew must haul the buses
in for service. The joint distribution of the number of calls received on Mondays and Tuese:ays is
given in Fig. 4-16, along 'n-irh the marginal distributions. The variable Xl represents the number of
calls on Mondays and X2 represents the number of calls on Tuesdays. A quick inspection u.ill show
thatX! andXl are independent. since the joint probabilities are the product of the apprOpriate ~'Ull
probabilities.
X){l 0 2 3 4 p,(x,)
Definition
If (Xl' X2] is a two-dimensiona1 random variable; the c<JVariance denoted Ci,,2' is
j
\ j
The cQvariance is measured :in the units of Xl times the units of X2. The correlation
coefficient is a dimensionless quantity that measures the linear association between two
random variables. By performing the multiplication operations in equation 4-26 before dis-
tributing the outside ex lue operator across the resulting quantities, we obtain an
alternate form for tb covariance follows~
(4-28)
Theorem 4-3
ik If X, andX2 are independent, then p = o.
Proof We again prove the result for the continuous case. If XI and X1 are independent,
= [ [ -'lA(x,).x,tz(Xz)dxJdxz
Thus Cov(X" X,) =0 from equation 4-28, aod p =0 from equation 4-27. A similar argument
would be used for [X" X,] discrete.
The converse of the theorem is not necessarily true, and we may have p = 0 withom the
variables being independent. If p = 0, the random variables are said to be uncorrelated.
Theorem 4-4
The value of p will be on the interval [-1, +1], that is,
-!$p$+l.
Proof Consider the function Q defined below aod illustrated in Fig. 4-17.
Q(t) = Ei(X, - E(X,) + t(X, - E(X,))]'
= E[X, - E(X,)]' + 2tE[(X, - E(X,»(X, - E(X,))]
+ "E[X, - E(X,)]'.
Since Q(J) ~ 0, the discriminant of Q(t) must be :;; 0, so that
{ZE[(X! - E(X,»(X, - E(X,))]}' 4E[X, - E(X,)]'E[X, E(X,»)' $ O.
4-8 Covariance and correlation 89
It follows that
4[Cov(X" X,)]' -4V(X,)· V(X,)" 0,
so
[COV(X1,X2)]' ,,1
V(X;)V(X,)
and
-I~p~+l. (4-29)
A correlation of p <::: 1 indicates a ~'hlgh positive correlation." such as one might find
between the stock prices of Ford and General Motors (two related cOtopanies). On the
other baud, a ''high negative correlation," p ~ -I, might exist between snowfall and tem·
perature.
~~"li'4?i~
"_";..."'-:;:,.t~L=;W"",,,,,\.
Rocall Example 4-7, which examined a continuous random voctor rXj' X2] with the joint probability
density function
O<X1 <X1 < 1.
otherwise.
The marginal of Xl is
f,(x,l =6x,([ -x;), forO <""; < [,
-:::;: 0, otherwise,
£
£"(xl) = 3xIdx, = 3/5,
V(X,) = E(xi)-(E(X,)]' =0.39.
Further. we have
A continuous random. vector [XI' Xz] has density functionjas given below;
f(x" x..) = 1, -~<XI<+~.O<x;:<l.
=0, Qtb.envisc.
This function is shown in Fig. 4-18.
The .marginal densities are
.!;IX,) = I-x, f0r.0<x!<l,
=l+xl fOI:-·1 <Xl < O.
=0. otherwise.
and
\
\
\
\
X,
SinceftxJ> xz}:P !;(x1) :fieX;J, the variables are not independent. If we calculate the covariance, We
obtain
and thus p = 0, so the variables are uncorTelated although they are not independent.
(4-32)
~~ii1~~:~1; !
Suppose XI and Xl have the following density:
=6xf.xJ.
S.O<xj<landX:!>l.
F(Xt,X2) = I:,t-r'24tlt:zdtzdt1
= 6x:~-8x~+3xi,
;,~~~~~3Q~
If Xl represents the number of defective units produce~ by machine No.1 in 1 hour andXzrepresents
the number of defective units produced by machine No.2 in the same hour, then the joint distn1mtion
might be presented as in Fig. 4-21. Furthermore. suppose the random variable Y "" H(X!! X,;). where
H(x x,) ='lx: + ",. I, follows that R y =(0, 1, 2, 3, 4, 5, 6. 7,8, 9j, In order to determine, say, P(Y
=0)"= p,(O), we note that ¥= 0 lfand only if X: = 0 andX,= 0: therefore,p,(O) =0.D2,
We Dote that Y"" 1 if and only if Xl = 0 and X1 -= 1; therefOl;e py(l) = 0,06, We also note that Y= 2
lfand onlyifeitberX, =0, X, =2orX, = I,X, = 0; sop,.(2)=OJO+ 0.03 0,13, Using similar logic,
we obtain the r~1. of the distribution, as follows:
YI Py(Y,)
0 0.D2
0,06
2 0,13
3 0,11
4 0.19
5 0,15
6 0.21
7 0.D7
8 0,05
9 om
otherwise
°
In the case where the random vector is continuous with joint density functionftx:, xJ
and H(x" x,) is continuous, then Y = H(X p X,) is • continuous, one-dimensional random
variable, The general procedure for the determination of the density function of Y is out-
lined below.
1. We are given Y ~ H,iX;, XJ, .(
2, Introduce a second random variable Z = H,(Xl , X,). The function H, is selected for
convenience, but we want to be able to solve y:;;; Hl(XI'~) and z = flz(x 1• Xz) for XI
and x., in ternis of y and z. "
"1
3. Find = G, (Y, zj, and x., = G,(y, z).
4. Find the follov.ing partial derivatives (we assume they exist and are continuous):
X~i p,(x,)
° 0,03
2 3
0,01 O.~
° 0.02
0,06 0.09
0.04
5. 'The joint density of [Y, 2J. denoted t(y, z), is found as follows:
t(y, z)=f(G1(y, z). G,(y, zll· iJ(Y, z)l. (4-33)
where J(y, z). called the Jacobian of the tnuJsformation. is given by the determinant
IOxtfa y oxt/az!
J(y,,) = axzldy ax,/o,I' (4-34)
!1~iap:.~€~z~I
Consider the contin,uous rax.dom \-'eCtor [Xl' Xi! \O.'ith the foll(,)\-\'ing density:
j(Xl,~)"""'4e-U.Y,+.l'>j. xj>O,x,>O,
= O. otherwis.e.
Suppose we are interested in the distributio:J. of Y =XjlJ[J!~ We will let y:: xlx2 and ehoose z;;; Xl +-'1:
sothatx,=yzl(l+y)~x,=zI(l+y).Itfo];owsthat . -
and aXlr(JZ=~Y~.
l+y
1
and dX,{aZ=-.
l+y
Therefore,
-y-
l+y z zy Z
---+-----
1 -(1+y)' (l+y)'-(I+y)'
l+y
and
I[G/,y, z), G'iy, ill =4e!-~1+yjt-d(l+Y)!1
::;; 4e:-:<Z.
Thus,
. )_- 4e ." '-.-'-I
ily,' z . (
[l+y J
and
y>O,
otherwise.
that the variables are continuous; however, the results may readily be extended to the dis~
crete case by substituting the appropriate summation operations for integrals. Vle assume
the existence of a joint density I such that
(4-36)
and
Thus,
(4-37)
Definition
The variables [Xl' x,..... XJ are independent random variables ifand only if for all [xpx" •. .• x,J
1(x,. x,.... ,x,) =II(X,!' !,(x.,)' .... I,(xn).
The expected value of. say, XI is
---- (4-38)
111,= :(X,) = [J:r XI' l(x,.x, ... ,xn )dxl dx2 • .. dx, (4-39)
(4-40)
We recognize these as the mean and variance, respectively, of the marginal distribution of Xl'
1:1 the two-dimensional case considered earlier, geometric interpretations were instruc~
tive; however, in dealing with rz--dimensional random vectors, the range space is the Euc1id~
ean n-space, and graphical presentations are thus not possible. The marginal distributions
are, however in one dimension and the conditional distribution for one variable given val-
j
ues for the other variables is in one dimension. The conditional distribution of Xi given va1~
ues (x." x,. "., x"J is denoted
(4-41)
E(X;'X2. X' ..... x,) = J': Xl' Ix,I", .....', (x')dxl· (442)
96 Chapter 4 Joint Probability Distributions
The hypothetical graph of E(XJx,. x,. _.. , x,) as a function of the vector tXt, x,. ___ • x,] is
x, x
called the regression of on (X::!, 3• •• ,' X,,).
5li2~ifu!ItJ,£~~6
Four resistors are connected in series as shown in Fig. 4-22. Each resistor has a resistance that is a ran.-
dom variable. The resistance of the assembly may be denoted Y, where Y=X 1 + Xz + X3 +Xlj'
;]~g!?'!;23]
Two parts are to be assembled as shov.'ll. in Fig. 4-23. The clearance can be expressed as Y=X\ -Xl
or Y::::::. (l)X\ + (-1)X2; Of eourse, a negative clearance would mean interference. This is a1i.1.ear com-
bination ""i:11 0c ;:;::: 0, a 1 = t and a l = -1.
A sample of 1() items is randomly selected from the output of a process that manufactures a small
shaft used in electric fan motors, and the diameters are to. be measured with a ·...alue called :he sample
mean, calculated as
- 1, )
X=io\X,+X,+"'+X" -
The valueX=~! +~ + ... +~lois a linear combination withac=O and at =a:z = •.. =ow=w.
Let us next consider how to determine the mean and variance of linear combinations.
Consider the sum of two random variables,
Y=X, +x,. (4-44)
or
, 2 2
(7y =0': +0'2 -:-2a:2_ (4-46)
as follows:
n
E(Y) Co + L ",E(X,)
. ~--11 ..... - .. - ..
or
98 Chapter 4 Jomt Probability Distributions
In Exa,"l1ple 4-22. four resistors were connected in series so that Y = Xj + Xl + X, + X4 was the .Ie:>'ist~
ance of the assembly, where Xl was the resistance of the first resistor, and so on. 'The mean and van-
a.'1ce of Y m terms of the me3llS and variances of the components may be easily calc'J.lated. If the
resistors are selected randomly for the assembly, it is reasonable to assume that Xlt Xl' X3• and ~ are
independent,
We have said nothlng yet about the distribution of Y; however, given the mean and variance of XI'
Xz> X3> and X4• we may readily calculate the mean and the variance of Y smce the variab!es are
independent.
~~p!e.·~~2ijf
ill Example 4-23, where two components were to be assexnbled, suppose the jobt distribution of ,
[X"X,J is
N" x.,) = 8e-<'"' ""', XI ~O,Xz~O,
=0. otherwise.
Sinceflxl. x...J can be easily factored as
fl.x" x.J = [2['''] , [4e-''']
=it(x,) 'f,(x.,),
XI andXz are independent. Furthermore, E(X j ) = PI =~. and E(X:0:::;: J12 =~ We may calculate the vari-
ances
r ' 2-'''d>:
x,e
o :
(1 \' _1
__ 1 _
'2) - 4
and V(}')=
In Example 4-24, we might expect me random variables Xi> Xi' , .. , X1C to he ind~ndent because of
the random. SalIlpling process. Furthermore, the distribution for each variable X! is identical. This is
sho,,"n in Fig. 4~24.1n me earlier exar::.ple. the linear combination of interest \\"3.8 the sample mean
It follO'NS that
-' 1 (X)
1 I ) 1 ' ) ~,,·+-·E
E(XI=;t-=-·E,X
, x 10 . ' +-·E\X,
10 • 10
"
1 1 1
=w;t+ 1O;t+ "+101'
=JL.
r 4-l4 The Law of Large Numben; 99
~~ t!1'" j.! X1
Figure 4-24 Some identical distributions.
Furthermore,
I
4-13 MOMENT·GENERATING FlNCTIONS
and
Y=X,+X,+,,·+X"<
Thus Y represents the number of successes in n trials, and YIn is an approximation (or esti~
mamr) for the unkn<Yl\'Il probability p. For convenience, we will let p YIn. Note that this
value corresponds to the lenni, used in the relative frequency definition of Chapter 1.
The law of large numbers states that
I
(4-53)
or equivalently
(4-54)
(4-56)
so if
We mayuow fix both E and ain equation 4-57 and determine the value of n required to sat-
isfy the probability statement as
> p(l- p)
n_--
z-· (4-58)
€ a
4-16 Exercises 101
(~pIe~¥~:
A manufacturing process operates so that there is a probability p that each item produeed is defective,
and p is uu.mo'Wll, A random sample of n. items is to be selected to estimate p. The estimator to be used
isp=YIn., where
; 0, if the jth item is good,
and
Xj
t
= ,... ifthejth item is defective,
Y:::XI+Xz+""+X~.
It is desired that the probability be at least 0.95 that the error, 1ft - pi,
not exceed O.OL In order to
determine the required value of n, we Dote that € = 0.01, and a"", 0.05; however, p is unknown. Equa~
tion 4-58 indicates that
n ~ ~p",(lr--,,-p)i
(0.01)' ·(0.05)
Since p is unknQV;n, the worst possible case must be assumed [note thatp(1 - p) is maximum when
p=~. This yields
(0.5)(0.5)
11 >- '2 50, ODD,
(0.01) (O.OS)
a very large number indeed.
Example 4-28 demonstrates why the law of large numbers sometimes requires large
sample sizes. The requirements of € = 0.01 and a= 0.05 to give a probability of 0.95 of the
departure 1ft - pI being less than 0,01 seem reasonable; however, the resulting sample size
is very large. In order to resolve problems of this nature we must know the distribution of
the random variables involved (ft in this case). The next three chapters will consider in
detail a number of the more frequently encountered distributions,
4-15 SlJMMARY
This chapter has presented a number of topics related to jointly distributed random variables
and functions of jointly distributed variables. The examples presented illustrated these top-
ics, and the exercises that follow will allow the student to reinforce these concepts.
A great many situations encountered in engineering. science, and management involve
situations where several related random variables si{nultanoously bear on the response
being observed. The approach presented in this chapter provides the structure for dealing
with several aspects of such problems,
4-16 EXERCISES
+1. A refrigerator manufacturer subjects his fir.ished table, where X represents die occu..~ence of finish
prodUcts to a final inspection, Of interest are two cat~ defects and Yrepresents the occummce of mechanical
egories of defects: scratches or flaws iD the porcelain defects.
futish, and mechanical defectS, 'Ib.e number of each <a) Find the margmal distributions of X and Y.
type of defect is a random variable, The results of (b) Find the probability distribution of mechanical
1ru:.--pecting 50 refrigerators are shO'Wl11n the following defects given that there are no finish defects.
102 Chapter 4 Joint Probability Distributions
;~ i j 2 _3_1-_4-,'_5_+-_6_ i _7_1-
' _8-1'_9-1
1
t
i 1 1101100'51100 31100 21100 111100 1/10011110011/100 '1/100
+---;~-r~+-~~~--4
i 2 181100 51100; 31100 21100 1'/100 1/100 i 1/100 I I
,
i 3 181100 5/100:21100 1I1001~1100 I I ,
4-16 Exercises 103
~ S1 52 53 54 55 [ g(uz)h(u)lu,du.
S1 0.06 0.05 0.05 0.Q1 , O.Q1
Hint: LetZ=.\'lY and U = Y, and find the Jacobian for
52 0.01 0.05 0.01 • 0.Q1 I om . tlle transforrr:..ation to the joint probability density
! 53 0.05 ! 0.10 ! 0,10 0.05 ! O.OS : function ofZ:md U, say r(z. u). Then ir.tegrate r(z:. u)
0,0'1 0.01 with respect to u.
54 0.05 • 0.02 0.03
....._-
:-ss 0.05 : 0,06 0.05 0.01 i 0,03
4~14. Suppose we have a simple electrical circuit in
which Obm's law V"" IR holds. We v;ish to find the
probability distribution of resis!.ance given that the
value of the score on test No.1 given the score or. test probability clistribuciollS of voltage (11) and current (l)
No.2. are ktiOWll to be
4-10. Let
g,,(y) e~. V;?O,
, - ::::0, otherv.rise;
j{xl.Xz} =4x1Xze-\r1 H t, X;, x.z > 0
:=. O. otherwise. Vi) i~O,
pay, Yis also a random variable. Assume that the joint 4-2&. For the bivariate distribution,
dis:ribution of [X,:(I isir", y). k(l+x+ y)
(a) Under what citcumstances will a sale take place? f(x,y) O';x<-,O$y<~,
(1+x)'(lTy)' ,
(b) Write an expression for the probability of a sale
;;;:. 0, otherwise.
tlkiJ:g place.
(c) Write an expression for the expected prk'e of the (a) Evaluate the con.'itant k..
transaction. (0) Fl.'1d the marginal distribution of X
4~23. Let [X y) be UIllformly distributed over the 4~29. For the bivariate distribution.
semicircle in the following diagram. Thusf(>; y) = 21"
if ~x, y) is in the semicircle, fIx,,) k ,,;;'0,y2:0,n>2,
(l+x+yr
(a) Find the marginal distributions of X and Y.
=0. otherwise.
(b) Find the conditional probability distributions,
(a) Evaluate the constant k.
(c) Find the conditional expectations.
(b) Fi::1d F(x; y).
y 4-30. The manager of a small bank vvisbes to deter~
mine the proportion of the ti..me a particular teller is
busy. He decides to observe the teller at n randomly
spaced mwrvals. The estimator of the degree of gain~
ful employment is to be YIn, where
_ (0, if on the ith observation. the teller is idle,
Xi - ~1, if on the ith observation, the teller is busy,
Sand D, whose probability densities are (b) .1\>; y) = 3i'y-\ 0'; x'; y'; L
(e) 1(>;y) 6(l+x+y)~, x;;:0,12:0.
fs(s) = 3~' lO~s';40,
4·32, Ler}\>; y, t)= h(x)hfylh(z). x 2: 0, Y,,0, z" 0,
otherwise; Detennine the probability th2;t a point drawn at ran-
.
gn (d)= 20'
° l
WSd';30,
dom will bve a coordinate (x, 'I. z) thaI docs not sat~
isfy either x>y > z or x <y < z.
5-1 INTRODUCTION
In this chapter we present several discrete probability distributions, developing their ana-
lytical form from certain basic assumptions about rea1~world phenomena. We also present
some examples of their application. The distributions presented have found extensive appli-
cation in engineering, operations research. and management science. Four of the distribu~
tions. the binomial, the geomerric, the Pascal, and the n.egative binomial, stem from a ,
random process made up of sequential Bernoulli trials. The hypergeometric distribution,
the multlnomial distribution., and the Poisson. distribution will also be presented :in this
chapter.
When we are dealing with one random variable and no ambiguity is introduced, the
symbol for the random variable will once again be omitted in the specification of the prob~
ability distributions and cumulative distribution function; thus, px<x) = p(x) and FIx) ~
F(x). This practice wili be continued throughout the rext.
and
Xj'=l, j 1.2.... ,~
Xj =0. j=I.2, .... n, (5-1)
.otherwise,
106
5-2 Bernoulli Trials and thc Bernoulli Distribution 107
~ ~
0=\..;;,i rl----X..:.)----t{.
..... hl~X_'<j,((SS;)"~
) FF"j I \ 0 X 'F) .
LV ~
Figure 5-1 A Bernoulli trial
For one trial, the distribution giveD in equation 5-1 and Fig. 5~2 is ealled the Bernoulli
distri.bution.
The mean and variance are
and
VeX) = [(0'. q) + (I'. p)]- p' = p(l-p) = pq. (5-2)
:~p.~Y:2
Suppose we consider a to.allufactUl'ing process in wbich a stI.ta1.1 steel part is produced by an auto-
matic machine. Furthermore, each part in a production nm of 1000 parts may be classified as defec~
live or good when inspected. We can think of the production of a part as a single trial that results
in success (say a defective) or failure (a good item). 1£ we have reasOD to believe that ffie machine
is just as likely to produce a defecth:e on one run as on another, and if the production of a defec~
tive on one run is neither more nor less likely because of the results on the previous runs, then it
would be quite reasonable to assume that the production run is a Bernoulli process \vith 1000 trials,
The probability, p, of a defective being produced on one trial is called ~ process CNerage fraction
defective.
Note that.in the preceding example the assumption of a Bernoulli process is a matheff
matical idealization of the actual real-world situation. Effects of tool wear, machine adjust-
ment,. and instrumentation difficulties were ignored. The real world was approximated by a
model that did not consider all factors, but nevertheless, the approximation is good enough
for useful results to be obtained.
We are going to be primarily concerned with a series ofBemoulli trials. In this case the
experiment % is denoted [('1:1' 'S" .",11:,,): 'S; are independent Bernoulli trials,j = I, 2, ... ,
n}. The sample space is .
~ = (Xl' ""X/I): Xi= S or F. i= I •...• n}.
Suppose all experiment consists of three Bernoulli trials and the probability of success is p on each
trial (see Fig. 5-3). The random variable X is given by X = L~"'l Xj' The distribution of X can be
determined as follows:
"
Q P{FFF) =q' q' q=cl
1 P{FFS}+P{FSFj +P{SFFj 3p'!'
2 P{FSS)+P(SFS}+P{SSF}=3p'q
3 P{SSS) =p'
()
P." 'l'nl
x)pX( I-p )'-X • x=O,I,2,,,.,.,
= 0, ofuerwise. (5-4)
Example 5-2 illustrates a binomial distribution wifu n = 3. The parameters of the binomial
distribution are n andp, where n is a positive integer and 0 ~p S 1, A simple derivation is
outllned below. Let
p(X) = P["x successes in n trials").
S Rx
IP
X
'0
FFS- I
I\
FSP· -j I
I
)
I
"-I
SFF-
FSS-
SFS-
\
)
(
'2
)
i
SSF- I
i SSS-
e) Figure S~3 Three Bernoulli trials.
5~3 The Binomial Distribution 109
The probability of the particular outcome in ~ with Ss for the first x trials and Fs for the
last n - x trials is
= 0, otherwise.
Since q =-1 - p, this last expression is the binomial distributiun.
so that
E(X) = np. (5-5)
E(X(X-l)) = ±.x(x-l)n!p"r"
""0 x! (n - x)!
so that
VeX) = E(X2) - (E(X)),
= E(X(X - 1)) + E(X) - (E(X))'
= n(n - l)p2 + np - (np)'
=npq. (5-6)
116 Chapter 5 Some Important Discrete Distributions
An easier approach to find the mean and variance is to consider X as a sum of n inde~
pendent Bernoulli random variables, each with meanp and variance pq, so that X Xl ~ X, =
+ ... -rXfj' Then
E(X) = p + p + ... .,. p = np
and
VeX) =pq +pq + --. + pq=npq.
The moment-generating function for the binomial distribution is
;~~~i¥~~~;
A production pro<;ess represented schematically by Fig. 5-4 produces thousands of parts per day. On
the average. 1% of the parts are defeetive and this avernge does not vary with time. Every hour, arm-
dom sample of 100 parts is selected from a oonveyor and severnl characteristics are observed and
measured on each part; however, the .inspector classifies the part as either good or defective. If we
consider the sampling as n 100 Berooulli trials with P =0,01, the total number of defectives in the
sample, X, would have a binomial distribution
100'
p(X) =( x jeO.Ol)'(O.99)'OO-', x=O,1,2, ... ,100,
=0 otherwise.
Suppose the inspector has instructions to stop the process if the sample has more than two defectives.
=
Then,P(X> 2) 1- P(X'; 2), and we maycalcuJate
, (100\
P(X52) = 1, ~O.OI)'(O.99)'OO-'
r<>O" X)
Thus, the probability of the inspector stopping the process is approximatdy 1- 0.92::: 0.08. The mean
number of defectives that would be found is E(X) =np = 100(0.01) ~ 1, and :he variance is VeX) = npq
=0.99.
(5-8)
I
Conveyor
):I Warehouse
I
I
I ,
{. Hourly, n= 100 Figure 5-4 A sampling situation with attrib-
samples J ute measurement
5-3 The Binomial Distribution 111
The function is readily calculated by such packages as Excel and :Minitab. For example,
suppose that n ~ 10,p ~ 0.6, and we are interested in calculating F(6) ~ P(X" 6). Then the
Excel function call BINOMDIST(6,10,0.5,TRUE) gives the result F(6) ~ 0.6177. In
Minitab, all we need do is go to CalcIProbability DistributionslBinomial, and click on
"Cumulative probability" to obtain the same result.
In order to evaluate, say, P(p :5" Po), where Po is some number bernreen 0 and 1, we note that
(5-13)
From a flow of product on a conveyor belt between production operations J and J + 1, a random sam-
ple of 200 units is taken every 2 hours (see Fig. 5-5). Past experience has indicated that if the unit is
not properly degreased, the painting. operation will not be successful, and, furthermore, on the aver-
age 5% of the units are not properly degreased. The manufacturing manager has grown accustomed
to accepting the 5%, but he strongly feels that 6% is bad. performance and 7% is totally unacceptable.
He decides to plot the fraction defective in the samples, that is, p. If the process average stays at 5%,
he would know that E(p) :::: 0.05. Knowing enough about probability to understand that p will vary,
he asks the quality-control department to determine the P(j > 0.07 j p = 0.05). This is done as follows:
:~!,!§pi~~1~
An indusmal engineer is concerned about the excessive "avoidable delay" time that one machine
operator seems to have. The engineer considers two acti.-ities as '"avoidable delay time" and "not
avoidable delay time." She identifies: a t.i.n1e-dependent variable as follows:
X(t) = 1, avoidable delay,
=0, otherwise.
A pMticular realization of X{t) for 2 days (%0 minutes) is shown in Fig, 5~6,
Rather thanbave a time study technician contln'.lously analyze this operation, the engineer electS
to USe '"work sa.'llpling," randomly selects rt points ou the 960~minute span., and estimates the fraction
of time the '"avoidable delay" category rusts, She lets XI::: 1 if XCr) 1 at the time of the ith obser-
vation and X" =:; 0 if X(:) =:; 0 at the time of the ith observation. The statistic
Ix;
P=~
"
is to be evaluated, Of course, Pis anmdom variable having a mean equal to p. variance equal to pqlrt.
and a standard de.iation equal to ~. The procedure oudined is not necessarily the best way to
go about such a study, but it does illusrrate one utilization of the rmdom variable P.
In summary~ analysts must be sure that the phenomenon they are studying may be rea-
sonably considered to be a series ofBemoulli trials.in order to use the binomial distribution
to describe X. the number of successes in n trials, It is often useful to 'visualize the graphi-
cal presentation of the binomial distribution. as shown in Fig. 5-7. The values Pix) increase
to a point and then decrease. More precisely, Pix) > Pix - I) for x < (n + I)p, andp(x) <
p(x -1) for x > (n+ I)p.lf (n + I)p is an integer, say m. thenp(m) = p(m- I).
XC')
o
llLDDlL~ 480 min, 960 min. t
Figure 5-6 A realization of X(I). Example 5-5.
5-4 T,::e Geometric Distribucon 113
'---OO---''--2:'--c'3-~'\-'n-_L2=--n-'_''C1--nL--~x
Figure 5-7 The binomial distribution.
and range space for X are illustrated in Fig. 5·8. The range space for X is Rx= (1,2,3, .. ),
and the distribution of X is given by
.x I, 2, . .. ,
othenvise. (5·14)
It is easy to veciiy that this is a probability distribution since
Lpq'-l
x",t
p fq' =p.[_1_J=1
k=O l-q
and
p(X) <: 0 for ali x.
or
d: q ] 1 (5-15)
!l P dql1-q =p'
~~______x______~~
£;~) )::.~6
)' FFFFFS.-t(==================~: )
'\ FFFF::!,,/,F>iS °l-
° i ( ) ·7 I
~ 0
Figure 5,.8 Sample space and range space for X
114 Chapter 5 Some Important Discrete Distributions
(5-16)
pc'
Mx(r)=--, . (5-17)
I-qe
A certain experi:Lent is to be performed until a successful result is obtained. The ~a!s are ir.depend~
ent and the cost of pe:rfonni.1.g the experiment is $25,000; however. if a failure results, it costs $5000
to "set up" for the next trial. The experimenter would like to detennine the expected COS! of the proj~
ect. If X is the number of trials required to obtain a successful experiment, then the cost function
would be
Then
I)
r 000· - -·5000.
; 130,
L P
If the probability of success on a single trial is, say, 0.25, then the E[C(X)]; $30,00010.25 - 55000 =
SI15,000, 'I'his mayor may not be acceptable to the ex.per'..mcnter, It should also be recognized that
it is possible to conti.."me indefinitely without having a successful experiment. Suppose that the exper-
imenter has a maximum of $500.000. He may wish to find the probability that the experimental work
would cost more tha:a this amount., that is.
=pl'X>~1
30,000 )
=P(X> 16.833)
I-P(X'; 16)
\6
=1- 2:;O.25(0.75)~1
= O.O!.
The experimenter may Dot be at aU willing to run the risk (probability 0.01) of spending the available
$500,000 without getting a successful run.
The geometric distribution decreases, that is, p(x) < p(x - 1) for x = 2, 3. " .. This is
shown graphically in Fig. 5-9.
An interesting and useful property of the geometric distribution is that it has no mem-
ory, that is,
The geometric distribution is the only discrete distribution having this memoryless properry.
5-5 The Pascal Dostribution 115
P(X)I
LetX denote 'the number of tosses efa fair die until we ob.~erve a 6, Suppose we have already tossed
the rlie five times without seeing a 6. The probal:ility that more than tvlO adCiticnal tosses will G'e
requlrOO is
I
P(X> 7 X> 5)=P(X>2)
I-P(X$2)
,
=1- Lp(x)
x_,
1--11 ( 1+-5\
(j \6)
I
25
=-
36
I
Lj
! .11. _,.~
34567
~-------:::-\
x r---
Figure 5-10 An example of the Pascal distribution.
cr:rq/p', (5-21)
and
pi Y
Mx{t)= ( - - , I· (5-22)
1- qe )
.tlciIhpie ~.'8·.
The president of a large corporation makes decisions by throwing darts at a board. The center section
is !:larked "yes" and represents a success. The probability of his hitting a l'yes" is 0.6, and this pro!)..
ability remains constant from throw to throw. The president con.tinues to throw until he has three
"tits." We cenoteX as the number of the trial .on which he experiences:he third hit. The mean is 3/0.6
:5, mea:.ri.rig that on the average it 'Will ta.l(e five throws. The president"s decision rule is simple. Ifhe
gets three hits on or before the fifth throw be decides in favor of the. question. The probability that be
will decide in favor is :herefore
P(X" 5) = p(3) - p(.) + p(5)
(3\,( )3. )' l/4\ )" ,
=l2(21J,0.6 )'r,0.4)' +l2) 0.6. (0.". + 2fo.6 \0.4)
=0.6826.
P ( XI'X'''''');.
ll
=
r n!
Xl!X2!'''Xk~J
ll:!x" Xl
PI P2' ,,,Pk (5-23)
,
for Xl:::::: 0> 1,2, "., n, i:::::: 1,2. ... , k, and where LX:;:: n.
i=.l
k It should be noted that X,. x" ... , X, are not independent random variables, since
L x It for any n repetitions,
i"": It turns out that the mean and variance of Xi' a particular component, are
5*7 The Hypergeometric Distribution 117
~1~lli~§~9'!
Mechanical peucils are manu:actured by a process involving a large amount of labor in the assembly
operations. This is bighly repetitive work and L'1.Ceutive pay is invo!ved. Final inspection has revealed
that 85% of the product is good, 10% is defective but may be reworked. and 5% is defective and must
be scra.pped. These percentages remain constant oyer time. A random samp~e of 20 items is selec:ed,
and if we ler
X; ;;;;; number of good iteI11S.
A; : : : number of defective but reworkable items,
X) : : : number of items to be scrapped.
then
Suppose we want to evaluate t.;!::; probability function for Xl := 18, ~ = 2. andx~ =0 (we must have Xl
+-'-l +xJ =20); !ben
'ZOi!
Pilg 2 O,=_~i-'·-(O.8S'J"(0.10)'(O.05)'
\ ' • I (18 11 ')10'
r"'" ,
0.102.
D~(N -DJ
() (
_xJn-x x = 0, I, 2, ... , rnin(n, D),
p x - (~J
=0 otherwise. (5-26)
The hypergeometric's probability mass function is available in many popular software pack~
ages, For instance. suppose that N : : : 20, D = 8~ n :::.- 4, and.;t = 1, Then the Excel function
E(x)=n·r~l (5·27)
and
, [DJ [ D) [N-nJ
.(Xl=n· N ' 1- N . N-I' (5.28)
In a receiving inspeetion department. lots of a pump shaft are periodically received. The lots contain
100 units and the following acceptance sampling plan is used. A random sample of 10 units is
selected wi:hout replacement. The lot is accepted if the sample b..a..1'; no more thai: one defective. Sup-
pose a lot is received that is p"(lOO) percent defective. What is the probability that it will be accepted?
~ (lOOp'UIOO[l- p'JI
£../, x 1\
lO-x
P(accept lot) =P(X';l) = =01. (l~O\ )
\ 10)
(100P'X100[1- P'J}(lOOP"(I00[l- p'~
" 0 l 10 , 1) 9 ,
eOO) ,,'
,!O)
Ob. . iously :he probability of accepting :he lot is a function of the lot quality, p~
If p' = 0,05, then
The first assumption is that the number of arrivals during nonoverlapping time inter-
vals are independent random variables. Second, we make the assumption that there exists a
positive quantity A such that for any small time interval, I1t, the following postulates are
satisfied.
1. The probability that exactly one arrival will occur in an interval of width I1t is
approximately A' ill. The approximation is in the sense that the probability is (A'
ru) + o,(ru) where the function [o,(ru)/ru]--+ 0 as ru --+ O.
2. The probability that exactly zero arrivals will occur in the interval is approximately
1 - (,1, . ru). Again this is in the sense that it is equal to 1 - (,1, . ru) + o,(ru) and
[02(ru)/ru]--+ 0 as ru --+ O.
3. The probability that two or more arrivals occur in the interval is equal to a quan-
tity o,(M), where [o,(ru)/ru] --+ 0 as ru --+ O.
The parameter Ais sometimes called the mean arrival rate or mean occurrence rate. In
the development to follow, we let
so that
and
(5-30)
For x> 0,
so that
and
(5-32a)
and
x= 1, 2, .... (5-32b)
120 Chapter 5 Some ImpOrtant Disc:ete Distributions
= 0, otherwise, (5-34)
Note that this distribution was developed as a consequence of certain assumptions; thus,
when the assumptions hold or approximately hold, the Poisson distribution is an appropriate
modeL There ,are many real,world phenomena for which the Poisson model is appropriate.
If we let np = c, so that P = c/n and I - p = 1- c/n = (n - e)/n, and if we then replace terms
involving p with the corresponding terms involving c, we obtain
p(x) = Xl
l- l -c1"-'
n{,,-I)(n- 2) ... (n - x+ I) r eJ'~ n-
n n ...
C'[ ( IV 2) f x-I'l( c\"f c',-'
=-~ (1)11--,,1-- ",,1--) 1--),1--' . (5·35)
x! \ n/\ n \ n J n \ n/
In lettingn --> = andp --> Oin such a way that op =cremainsfixed, the terms (I _1), (J --"),
1> 1 c C tl
... , (I - -=-) all approach 1, as does (1 - -T". Now we know that (1 -j" --> tf as n -->
!'! II "
=.
tl
Thus, the limiting fonn ofequatlon 5,35 is pix) = (e'lx!) . e-<, which!s the Poisson distribution.
::::::C. (5·36)
Similarly,
so that
V(X) = E(X') (E(X»)'
c. (5,37)
5-8 The Poisson Distribution 121
Theorem 5-1
If Xl' X2• "'jXk are independently distributed random variables, each having a Poisson dis~
tribution with parameter c," i;::; I, 2, .. _, k, and Y = Xl + X2 + ,,- + Xkj then Yhas a Poisson
distribution with parameter
Suppose a retailer determines tha: the number of orders for a certain home appliance in a particular
period has a Poisson distribution with parameter c. She would like to determine the stock level K for
the beginning of the period so that there will be a probability of at least 0.95 of supplY.ng all cus-
tomers who order the appliance dwing the ~od, She does ;:lot wish to back-order merchandise or
me
re..<;upply warehouse du.."':ing the period. If X represents the number of orders.,. the dealer wishes to
determine K such that
P(X'; Kl ,,0.95
or
P(X>Kl ";0.05.
so that
x:X+!
The solution may be determined directly from tables of the Poisson distribution and is obviously a
fuuction of c,
The attainable sea.'litivjty for e1ectromc amplifiers and app~ is limited by noise or spontaLeous
current fluctuations. In vacuum t',:ibes, one noise source is shot noise due to the random emission of
122 Chapter :; Some Important Discrete Distributions
electrons from the heated ca.thode. Assume that the potential difference between the anode and cath·
ode is large enough to ensure that all electrons I'.:mitted by the cathode hav\'.: high velocity-high
enQugh to preclude spare charge (accumulation of electrons between the cathode and anode). Under
these conditions, and cefining an arrival to be an emission of an electrode from the cathode. Daven-
port and Root (1958) showed that the number of elect:cons, X, emitted from the cathode in time t has
a Poisson distribution given by
p (x) = (AtYe-:tyx!. X= 0,1. 2, "',
0, otherwise.
The parameter 1 is the mean rate of emission of electrons from the cathode.
'~~;Ppl~5jl~'~
A prod..:ction lot of 200 units has eight defectives. A random sample of 10 units is selected, and we
want to find the probability that the sample will contain exacdy one defective, The true probability is
p(X=l) Gf: J 2
0.288.
(~~)
SiDcer.!N=~=O,05 is small. we letp =~= O,(}4 3!l.d use the binomial approximation
p(I)= ( I
10'
r· 04 )'(O.96)' =0.277.
In the case of the Poisson approximation to the binomial, we indicated earlier that for large
n and small p, the approximation is satisfactory, In utilizing this approximation we let c:::;
np. In general, p should be less than 0.1 in order to apply the approximation. The smallerp
and the larger n, the better the approximation,
'~~pIۤi~]
The probability that a particular rivet in the wing surface of a new aircraft is defective is 0.001. There
are 4000 rivets in the wing. \Vhat is the probability that not more than six defective rivets ",ill be
inst2lJed?
and
,
P(X~6)~ 2:,-4
,m,
4 '/x! = 0.889.
5·11 SUMMARY
The distributions presented in this chapter have wide use in engineering. scientific, and
management applications. The selection of a specific discrete distribution \\ill depend on
how well the assumptions underlying the distribution are met by the phenomenon to be
modeled. The distributions presented here were selected because of their wide applicability.
A sllIl1Jl1JlI)' of these ilistributions is presented in Table 5·1.
S.U EXERCISES
5~ 1. An experiment consists of four independent 5~2. Six independent space missions to the moon are
Bernoulli trials with probability of success p on each planned. The estimated probability of success on each
trial. The random ...ariable X is the number of suc- mission is 0.95. What is the probability that at least
cesses. Enume:rate the probability distribution of X. five of the planned missions will be successful?
,..
~
Q
.g
T"ble 5-1 Summary or Discrete Distributions !\'
,"
Momcnt~
Distribution
..
Parameters
Probability
Function p(x) Mean Variance
Generating
Futiction i
,-, ,
-~- -------~-----
otherwise
P
r----------
pq pe'+ q
-----~
f
ninonuro 11=1,2 •...
O<p<l
p(x) ~(:}Xq"-x.
=0.
x:::::O,I,2 •. , •• 1I
oilienvi!>c
"P
-_. f--. . -
llpq
-~-
(pel -1- q)~
-------
f
if
,-,
Geometric O<p< I p(x) pq. x-0,1,1, .. , lip qlP' pe'l (I qe') ~
.. otherwise
-,--- - ---- ----- ~'
Pascal
(Neg. binomial)
O<p< 1
r~ I. 2.... (r> 0)
r~lI) prqx-r,
p(x)-= (x.
~O,
x=r. r+1,r+2, ..
otherwjse
rIp rqlp<
[ r pe'
1 qe l
·0 otherwise
Poisson c>O (). -, Xl-I
pX-f: C .'C. x=O,~2 ....
,
C c e'(r-I)
O. otherwi~e
I
5~12 Exercises 125
S~3. The Xi'Z Company has pla!lD.ed sales presenta- 5-12. The X1'Z Compa.'1Y plans to ...isit potential cus-
tions to a dozen important customers. Tne probability tomers until a substantial sale is made. Each sales
of receiving an order as a result of such a pre~ntation presentation costs $1000. It costs $4000 to travel to
is esti.rnated to be 0,5, What is the probability of the nex.t customer and set up a new preseatation,
receiving fOUI Or more orders as the result of the meet- (a) What is the expected cost of making a sale if the
ings? probability of lMking a sale after any presentation
54. A stockbroker calls her 20 most important cus~ is O.1O?
tomerS every morning, If the probability is one in (b) If \he expeeled profit from each sale is $15,000,
three of making a transaction as the result of such a should the trips be undertaken?
call, what arc the chances of her h.a:tdli:lg 10 or more (c) If the budget for advertising is only $100,000,
transactions? what is the probability that this sum will be spent
5~5. A production process that manufactures transis~ Vlithout getting an order']
tors operates, on the average, at 2% fraction defective. 5~11. l'1nd the mean and variance of the geoJ.lletric
Ever; 2 hours a random sample of size 50 is taken distribution usiog the momeut-ge!1erating function.
frOIll the proeess. If the sample contai:::ts more than 5~14. A snbmarine's probability of sinking an enemy
t\VO defectives the process must be stopped. Deter-
ship with anyone of its torpedoes is 0.8. If the
Ii'lir.e the probability that the process will be stopped firings a:::e ir.depe:ode:1t.. determine the probability of
by the sampling scheme. a sinking within the first two firings. \Vithin the Erst
5~6. Find the mean and variance of the binocrial dis- three.
tribution using the moment~generating function (see 5·15. In Atlar.ta the probability that a thunderstorm
equation 5~7). will occur on any day during the spring is 0.05.
5-7. A production process manufacto.ri.Dg tum~indica~ Assuming independence, what is the probability that
tor dash lights is known to produce lights that are 1% the first thunderstorm occu..rs on April 25? Assume
defective. Assume this value remains unChanged and spring begins on March 21.
assunle a sample of 100 such lights is randomly 5-16. A potential customer enters :an automobile deal-
selected, FindP(p $ 0,03), where Pis the sample frac- ership every hour. The probability of a salesperson
tion defective. concluding a transaction is OJO. She is detennined to
5-8. Suppose a random sample of size 200 is ta.'cen keep working until she has sold th.••..ee cars. 'Nhat is the
frorr.. a process that is 0.07 fraction defective. \\'llat is probability that she will have to work exactly 8 hours?
the p:obabilit}' that p will ex.ceed the L"'Ue fraction More than 8 hours?
defective by one standard deviation? By two standard
5~17. A persor.ncl manager is interviewing potential
deviations? By three standard deviations? employees in ordct to fill twO jobs. The probability of
5-9. Five cruise missiles have beea bc.i1t by an aero~ an interviewee having the necessary qualifications a::.d
space company. The probability of a successful firing accepting an offer is 0.8. VIhat is the probability :hat
is, on anyone test, 0.95. Assuming independent fu- exactly four people must be inteniewed? "'hat is the
ings. what is the probability that thefust failure occurs probability that fewer than four people must be inter-
on the fifth firing? viewed?
5-10. A real estate agent estimates his probability of s..18. Show that the moment-generating function of
sellin.g a house to be 0,10. He has to see fOUI clients the Pascal tandom. variable is as given by equation
today. If he is successf.d on the :first three calls, what 5-22. Use it to deten:.n.i:ne the mean and variance of the
is the probability that his fourth call is unsuccessful? Pascal distnoution.
5-11. Suppose five independent identical laboratory 5 ..19. The probability that an experiment has a suc-
expe..1mcnts are to be undertaken, Each experiment is cessful outcome is 0.80. The experiment is to be
extremely sensitive to en...ironmental conCitions. and repeated until five successful outcomes have occurre,t
there is only a probability p that it will be completed What is the expe..."1ed numbe:- of repetitions required?
successfully, Plot, as a functio:l of p, the probability W:Jat is the variance?
:hat the:fifth experiment 15 the first failure. Fir.d math- 5~20. A military co=mander wishes to destroy an
ematically the value of p that maximizes the probabil- enem.y bridge. Each fligflt of planes he sends out has
ity of the fifth trial being the first unsuccessful a probability of 0.8 of scoring a direct hit on the
e"'periment. bridge. Ir takes four direct hits to completely destroy
the bridge, If he can mount seven assaults before the
126 Chapter 5 Some Important Discrete Distributions
bridge becomes tactically unimporta:J.t. what is the 5~29~ The number of antomobiles passing through a
probability that the bridge will be dcstrOyed? pa.'1icular intersection per hour is estimated to be 25.
,5..21. Three companies, X, Y. and Z, have probabilities Find the probability t:.hat fewer than ; 0 vehicles pass
of obtaining an order for a particular type of mer- through during any 1~hour interval. A",sume that the
chandise of 0.4, 0,3, and 0,3, respectively, Three number of vehicles follows a Poisson distribution.
orders are to be awarded independently, What is the 5-30. Calls arrive ata telephone switchboard such that
probability that one company receives all the orders? the number of calls per hour follows a Poisson distri-
5-22. Four companies are interviewing five college bution with a mean of 10. The current equipment can
students for positions after graduation. Assuming all ha::tdle up to 20 calls \Vithout becoming overloaded.
five receive offers from each company and assuming What is the probability of such an overload occurring?
the probabilities of the companies hiring a new 5-31. The number of red blood cells per square unit
employee ate equal, what is the probability that one v-lsible under a microscope follows a Poisson distribu-
company gets all of the new employees? ~one of tion with a mean of 4, Find the probability that more
diem? than five such blood cells are visible to the observer.
5-23. We are interested in the weight of bags of feed.
5,.32. LetX/ be the number of vehicles passing through
Specifically, we need to know if any of me four events
an intersection during a length of time t. The random
below has occurred:
variable X/is Poisson distributed with a parameter ?t.
TJ ;(X~ 10), p(T1) =0:2, Suppose an auto!'l'latic counter has been installed to
T,;(lO<X'; 11), p(T,) 0.2, count the number of passing vehicles, However, this
counter is not functioning properly, and each passing
T,;(ll <X,; 11.5), p(T,) = 0.2,
vehicle has a probability p ofnot bei:tg counted. Let Yt
T4 ;;; (11.5 < X). p(T,); 0.4.
be the number of vehicles counted during t. Fmd the
probability distribution of ~.
If 10 bags are selected at random. what is the proba-
bility of four being less than or equal to 10 pounds. 5 ..33. A large insurance company has discovered that
one being greater than 10 but less than or equal to 11 0.2% of the U.S. population is injured as a result of a
pound.<;. a!ld two bcing greater than 11.5 pounds? pa.~cuJ:ar t;>pe of accident.:nlis company has l~.OOO
5·24. In Problem 5·23 what is the probability that :ill policyholders carry"ing coverage aga..inst such an acci-
10 bags weigh more than 11.5 pounds? 'What is the denL What is the probability that three or fewer claims
probability that five bags weigh more than 11.5 will be filed against those policies next year? Five o~
pounds and the remaining five weigh less than 10 Qore claims?
pounds? 5,,34. Maintenance crews arrive at a tool crib request-
5-25. A lot of 25 color television :;ubes is subjected to ing a particular spare part according to a Poisson dis-
an acceptance testing procedure, The procedure con- ti.bution with parameter 1;;; 2. T.hree of these spare
sists of drawing five tubes at random, withourreplace w parts are nOI'I!lally kept on hand. If more than three
ment. and testing them. If two or fewer tubes fail. the orders occur. the crews must journey a conside..""abte
remaining ones are accepted. Otherwise the lot is distance to central stores.
rejected, Assume the lot contains four defective tubes. (a) On a given day. what is the probability that such a
(a.) What is the exact probability of lot aeceptance? journey must be made?
(b) \Y'bat is the probability of lot acceptance com- (b) Wbat is the expected demand per day for spare
puted from the binomial distribution with p :=: ?:s partS?
5-26. Suppose that in Exercise 5-25 the lot size had (c) How many spare parts must be carried if the tool
been 100. Would the binomial approxilnation be satis- crib is to service all incoming crews 90% of the
factory in this case? time?
5-27. A purchaser receives small lots (N:= 25) of a Cd) "''hat is the expected number of crews serviced
daily at the tool crib?
high~precision device, She wishes to reject the lot
95% of the time iiit contains as many as seven defec- (e) Wbatis the expected number of crews makiLg the
tives, Suppose she decides that the presence of one journey to central stores?
defective in the sample is sufficient to cause rejection. 5--35. A loom experiences one yam breakage approx-
How large should bet sample size be? .i.mately every 10 hours, A panicular style of cloth is
5 ..28. Show that the moment-generating function of the being produced that will take 25 hours on this loom.lf
Poisson random variable is as give:.:. by equation 5.-38, ti:u:ee or more breaks are required to tender the prod-
5-12 Exercises 127
uct unsatisfactory. find the probability that this style of (a) Find the probabilities of 1, 2, 3, 4. or 5 accidents.
cIo'll is finished wi:h acceptable quality. (b) The Poisson distribution has been free:y aPPlied
5-36. The ::mmber of p,eople boarCing a bus at each in the area of industrial accider:ts. However. ir fre-
stOP follows a Poisson disnibution with parameter A.. que~tly provides a poor ''fit'' to actual hismrical
TIle bus company is surveying its usages for scbedul~ data. Why rnigl".t this be t:1le? Hint: See Kendall
ing purposes and bas installed an automatic counter and Stuart (1963), pp. 128-130.
on each bus. However, if more than 10 people board at 541. ese either yo',;! favorite co;nputer language or
any one s~op, the counter cannot record the excess and the random integers in Table XV in ':he Appendix (and
merely registers 10. If X is the number of riders scale them by mUltiplying by 1(}$ to ge::. uniform :0,1]
recorded, find the probability distribution of X. random nu.::nberrea.liz.ations) to do the following:
5-37. Amathemadcs textbook has 200 pages on which (a) Produce five realizations of a binomial random
typographica: errors in the equations could occur. If variable with n "'" S.p = 0.5.
there are in fact five euors randomly dispersed among
(b) Produce ten realizations of a geometric distribu.-
these 200 pages. what is the probability that a :andom
tion 'With p "'" OA.
sample of 50 pages will contain at least one error?
How large must the .random sample be to assure that at (c) Produce five realizations of a Poisson random
least three ertO('S will be found with 90% probability? variable with c "'" 0.15.
5-38. The probability of a vehicle baving an accident at 5-4~ If Y"", XI .3 a:1(:1 X follows a geometric discribution
a particular intersection is 0.0001. Suppose that 10,000 with a wean of 6, use uniform lO,l} random number
vehicles per day travel th.vugh this intersection. \\"tat realizations, and produce five realizations of Y.
is the probability of no accidents occurring? What is
5-43. \Vith Exercise 5-42 above, use your computer to
the probability of two or more accidents? do the following:
5..39. If the probability of being involved in an auto (a) Produce 500 realiz.ations of Y.
accident is 0.01 during any year, wbat is d:!e probabil-
(b) Calculate)i =~, +)'1 + ........ Ysoo), the mean
ity of baving two Or :Dare accidents during any 10-
from this sz.mple.
yem: driving period?
540* Suppose that the number of aecidents to 5~44. Prove the meworyless prope:ty of the geometric
,
G
1
/i.
r
\/
Chapter. 6
Some Important
Continuous Distributions
6-1 INTRODUCTION
We will now study several important continuous probability distributions. They are the uni~
form. exponential, gamma, and Weibull distributions. In Chapter 7 the normal distribution,
and several other probability distributions closely related to it, will be presented. The nor-
mal distribution is perhaps the most important of all continuous distributions. The reason
for postponing its study is that the normal distribution is important enough to warrant a sep-
arate chapter, ,
It has been noted that the range space for a continuous random variable X consists of
an interval or a set of intervals. This was illustrated in an earlier chapter, and it was observed
that an idealization i') involved. For example, if we are measuring the time to failure for an
electronic component or the time to process an order through an information system, the
measurement devices used are such that there: are only a finite number of possible out-
comes; however, we v.-ill idealize and assume that time may take any value on some inter-
val. Once again we will simplify the notation where no ambiguity is introduced, and we let
f(x) = fx(x) and F(x) = F,!.x).
f(x) =,f-.
p-a
(6-1)
where a and {3 are real COnStants with a < {3 The density function is shown in Fig. 6- L
Since a uniformly distributed random variable has a probability density function that is
f(x)t
!
1 ,
p-a!
,_ _ .... ~L-_ _ _ _-L._ _ _.~
128
6-2 Toe Uuiforrr. Distribution 129
constant over some inte:val of definition, the constant must be the reciprocal of the length
of the interval in order to satisfy the requirement that
[fix) dx~1.
A uniformly distributed random variable represents the continuous analog to equally
likely outcomes in the sense that for any subinterval [a, b], where a" a < b ,; f3, the
P(a,,;X"; b) depends only on the length b - a,
(6·2)
(64)
For a uniformly distributed random variable, the distribution function F(x) P(X'; x)
is given by equation 6-.5~ and its graph :is shown in Fig. 6-.2.
F(x) 0, x < a,
}: - - - - -
~ x-{):
J.f3-a- f3-0:'
a';x<f3,
1, (5·5)
F(X)t
o a
/ f3 X
Figure 6-2 Distnoution function for the uniform random variable.
130 Chapter 6 Some Important Continuous Distributions
:E§.¥~!;;t6!J
A point is chosen at random on the interval [0. 10). Suppose we wish to find the probability that the
t t·
point lies between and The density of the random yariableX isf(x) =7:0,
O:S;x::5: 10, andf(x):::O
otherwise. Hence, pet s X s t) =fa.
:~~i~;6;~.:
Numbers of the form NN.N are "rounded off" to the nea..'<!St integer. The round-off procedure is such
that if the decimal. part is less than 0.5, the round is "down" by simply dropping the decim.:ll part~
however, if the decimal part is greater than 0.5. the round is UP. that is. the new number is livlV.NJ 1 "'I'"
whcre LJ is the "greatest integer contaix.ed in" function. Ii the decimal part is exactly 0.5. a coin is
tossed to determine which way to r~:lUnd. The round.-off error,x, is defined as the difference between
the number before roundiD-g and the number after rounding, These errors are comroonly distributed
according to the uniform distribution on the interval [-1),5, + 0.5;. That is,
j(xl I, -<l.5 "X'; ~ 0.5,
== O. otherwise,
~.K<"'aiIl"'1;'(i:3'
"",, ...,,-~," R-"·,~,~"",,w
One of the special features of many simulation languages is asiruple automatic procedure for using the
unifonn distr..bution. The user declares a mean and modifier (e.g.• 500, 100). The compiler immediately
creates l! routine to produce realizations of a random variable X uniformly distributed on (400, 600],
f(xlf
o
~~ x Figure 6-3 The expouetttial density function.
6-3 T.'1e Exponential Distribution 131
If we now let time vary and consider the random variable T as the time to occurrence, then
,,,,0. C\;/ )(6·10)
""lI,{,.
And since jtt) = F(I), we see that the density is "" ",
e"J, j}
f(t)=)..,:"", no, :V'· X)' ~ J/
=0, otherwise. yi'~ (6-11)
This is the exponential density of equation 6·6. Thus, the relationship between the
exponential and Poisson distributions may be stated as follows: i: the number of occur~
rences has .a Poiss'on distribution as shown in equation 6-7. then the time between succes-
sive occurrences has an exponential. distribution as shown in equation 6-11. For example, if
the number of orders for a certain item received pel," week has a Poisson distribution, then
the time betv.reen orders would have an exponential distribution. One variable is discrete
(the count) and the othe, (tiroe) is continuous.
In oreer to verify thatfis a density function, we note that/(x)" for all x and °
l ~ k -'" dx
o
._"'15
-e Q = 1:)l-'<
/~
,
V
/
. (:) / ",
... /~ / )< 'V
(6.12)
and
V(X) = J; x2k-"'dx-(l//,)2
= [-x e-"'I; +zJ; xe-"'dx]-(l/:<)' =1/:<'
2 (6.13)
132 Chapter 6 Some lIUportant Continuous Distnoutions
F(x)t
+--
~ _______, Fig= 64 Distribution fulletion for the
x expoaential.
The standard deviation is 1/A., and thus the mean and the standard deviation axe equal.
The moment-generating function is
( . yl
\ '.j
Mx(t)=i 1-":"1 (6-14)
provided t < ?.
The cumulative distribution function F can be obtained by integrating equation 6-6 as
follows:
X<O,
An electronic component is known to have a useful life represented by an exponential density 'X-ith
failure rate of 10-' failures per hour (i.e., .1= 10-5). The mean t:i;me to fwtL""e, E{X,), is thus 10> hours.
Suppose we want to determine the fraction of such components that would fail before the mean life
or expected-life:
This result holds for arty value of J. greater than zero. In our example. 63.212% of the iteI:1S
would fail before 10' hours (see Fig. 6-5).
fix)
f(x) = Ae-Ax, X2 0:
:::: 0, ot'1erwfse
A
Figure 6-5 The mean of an exponential dlstribution.
6~3 The Exponential Distribucon 133
Suppose a designer IS to make a d.ecision between two manufacturing processes for the rna!:ufacture
of a certain component Process.4 cos~ C dollars per unit to manufactcre a component P:::ocess B
costs k· C dollars per unit to manufacrure a component, where k> 1, Cor.::ponents J:-..ave an exponen-
tial time to failure density with a failure rate of 200-: failures per hour for process A. while compo-
nents from process B have a failure rate of 300- 1 failures per hour. The mean lives are rhus 200 hours
and 300 hou~, respectively, for the tWo processes, Because of a warrar.ty clause, 1: a componen: lasts
for fewer than 400 hours, the IDOllufacC'.:.ter rou.st pay a penalty of K dollars. Let X be be time to fail-
ure of each component. Thus, the component costs 3..4;;
ifX~400,
UX <400,
and
UX;24OO,
ifX<4oo.
The expected costs are
= (C - K){_e-,noc
40
'0..J
'l+ C~_e-Xl)!fJl- 1
L 400..,;
=(C+K)[l-e-']+C[e-'j
=C+K(I-e-2)
, ,
and
=(kC + K)[1_e-if3]+kC[e-4l3]
. =kC-Kll-e"''''j,
Therefore, if k <: 1- KlC{[2 - e~), then E( CA ) > E(Cfj), and itis likely that !"1.e designer would se:ect
process B.
so that
P(X",,'" six> x) = P(X>s), (6-16)
For example, if a cathode ray tube has an exponential time to failure distribution and at time
x it is observed to be still functioning, then the remain.in.g life bas the same exponential fail-
ure distribution as the tube had at time zero_
134 Chapter 6 Some Important ConriDuous Distributions
forn>O. (6-17)
~J\n important recursive relationship that may easily be shown on integrating equation 6~ 17
by parts is
r(n) = (n - I)f(n - I). (6-18)
since reI}
= J;
e -~dx::;:;: L Thus, the gamma function is a generalization of the factorial.
The reader is asked in Exercise 6-17 to verify that
r~
r,/1'
-I~ J, x-lf2e-xdx~,.fii. (6-20)
\2/ 0
The parameters are r > 0 and A. > O. The parameter r is usually called the shape parameter"
and "lis called the scale parameter. Figure 6-6 shows several gamma distributions, for ;1. == 1
and various r. It should be noted thatj(x);, 0 for alb, and
f(x)
ple, Gk\11<LWIST(15, 5.5, 4.2, TRUE) returns a value of F(15) = 0.2126 fur the case r =
5.5, .<. = 1/4.2 = 0.238. The Excel function GAMMAIl\'V gives the inverse of the CDP.
=0, othervrise,
and where the Xj are mutually independent, then X has 11e density given in equation 6-21.
b many appli::;ations of the gamma distribution that we will consider, r will be a positive
integer, and we may use this knowledge to good advantage in developing the distribution
function, Some authors refer to the special case in which r is a positive integer as the
Erlang distribution.
and
V(X) rl.<.". (6-24)
Equations 6~23 and 6-24 represent the mean and variance regardless of whether or not Tis
an integer; bow ever, wben T is an integer and the interpretation given in equation 6-22 is
made, it is obvious that
,
E(X) = 2:E(Xj) rJj.<.=r/.<.
j=!
and
,
V(X) = 2: V{X))=r<1/.<.2 =r/.<.2
j=l
from a direct application of the expected value and variance operators to the sum of inde-
pendent random variables.
The moment-generating function for the gamma distribution is
(6-25)
136 Chapter 6 Some Important Continuous Distributions
,
RecalUng that the moment-generating function for the exponential distribution was
[1 - (IIi.)r', thls result is expected, since
~~~~!~~i~2
A redundant system operates as shownm Fig, 6-7. Inirially unit 1 is On line, whlleuuit 2 and unit 3 arc
on standby. 'When uniT 1 fails r the decision switch (OS) switches unit 2 on until it fails and then unit 3 is
switehed on. The decision switch i" assumed to be perfect, so that the system life X may be represented
as the sum of the subsystem lives X::= Xl . :. . -l2 + XJ , If the subsystem lives are independent of one
another, and if the subsystems each have a life.x:,,,j== I, 2. 3. ha..,r.ng density g(x) = (1!l00)e-!I1l00, X ~ 0,
then X will have a garr.:.ma density with r~3 and ,l=O,Ol. That is,
The probability that the system 'Hill. operate at least.~ hours is denoted R(x) and is;;;alled the reliabil~
ity ftmction. Here,
R(x)=l-F(xl= ±e"'01'(o.OlX)'/k!
"'"
= e-'()'Ol'[1 + (O.Olx) + (O.01x)' /2]'
~~~pX;;¥:Z;
t
For a garn.."I:a distribution with ;t= and r = v!2. where v is a positive integer. the chi-square distri~
lJution with v degrees offreedom results:
f (,x) 1
2 vtZ r\(y/2) x
(,{2)-: -'I' 0
x> ,
,. It •
: : : 0, otherwise.
f3 > 0, the shape parameter. By appropriate selection of these parameters, this density func-
tion will closely approximate many observational phenomena.
Figure 6-8 shows some Weibull densities for r= 0, 8 = 1, and f3 1, 2. 3, 4. Note that =
when y= 0 and f3 = 1, the Weibull distribution reduces to an exponential density with }. =
1/8. Although the exponential distribution is 3 special case of both the gamma and Weibull
distributioust the gamma and Vleibull in general are noninterchangeable.
(6·30)
0.8 2.0 x
Figure 6..s Weibull densities for r=O, 0= 1. and j3 = 1, 2, 3, 4.
138 Chapter 6 Some Impor:ant Continuous Distributions
and
(6-31)
F(x)=l-exp -: - - r (x-r)Pl
'- \ (j
J' (6-32)
The Weibull CDF F(x) is conveniently provided by software packages sueh as Excel and
Minitab, For the case .< = 0, the Excel function VlEIBlTLL (x, (3, y, TRIJE) retnms F(z).
~~i~~:~)
The tirue~to-failure distribution for eleCtronic subassemblies is known to have a Weibu!l density with
y=O, p=t. and 0= 100. The fraction expected to survive to, say, 400 hours is fuus
~X;;:~~C6::~~
Berrettoni (1%4) presented a number of applications of the Weibull distribution.. The following a...-e
examples of natural processes having a probability law closely approximated by the Weibull distrl~
bution. The randol'!) variable is denoted X in the examp:es.
1. Corrosion resistance of magnesium alloy plates,
X: Corrosion weight loss of 10: mg/(cml)(day) when magnesium alloy plates are immersed
in an inhibited aqueous 20% solution ofMgBr2'
2. Return goods classified according to number of we<"-ks after shipment.
X; Length, of period (10-\ weeks) until a customer returns the defective product after
shipment.
3~ Number of downtimes per shift.
X: Number of downtimes per sblft: (times 10-1) occurring in a CoIltinuoUS automatic and C()m~
plicated .assembly line.
4. Leakage failure in dry-cell batteries.
X: Age (years) when leakage starts.
5. Reliability of capacitors.
X: Life (hours) of 3,3-,uF, 50-V. solid tantalum capacitors operating at an ambient tempera-
ture of 125°C, where the rated catalogue voltage is 33 V
---
If we desire to prod"ce realizations ofa uniform random variable on [a, ill. this is sim-
ply accomplished using
X, a+ u,({3- a), i~ I, 2, .... (6-33)
If we seek realizations of an exponential random variable with parameter A. the inverse
transform merlwd yields
-I
xi =TIn(Ui}' i= 1.2 .... (6-34)
Similarly, using the same method, realizations of a Weibull random variable with P2.....""ame-
ters 11, {3. <5 are obtained usIng
(6-35)
The generation of gamma variable realizations usually employs a technique known as
the acceptance-rejection mefr..od~ and a variety of these methods bave been used. If we wish
to produce realiz,ations from a gamma variable with parameters r > 1 and }.. > 0, one
approach, suggested by Cheng (1977), is as follows:
For more details on these and other random-variate generation techniques, see
Cbapter 19.
6-7 SUMMARY
This chapter bas presented four widely used density functions for continuous random vari~
abIes. The uniform, exponential, gamma, and Weibull distributions were presented along
with underlying assumptions and example applications. Table 6-1 presents a summary of
these distributions.
6-8 EXERCISES
6*1. A point is chosen at random on the line segrr:ent this profit is distributed between SO and
[O,d]. Whatjs the probability 1Jatitlies between and t $2000, find the probability distribution of the broker's
1"47
3 Between 2'4a::.d
" 3t1 total fees.
6--2. The opening price of a particular stock is uni~ 6-5. lise the moment~generating function for the uni-
fonn1y distributed on the interval [35{, 44{-J. What is form density (as given by equation 64) 10 generate
the probability that. on any given day. the opening the mean and VJrian.ce.
price is less than 401 Between 40 and 427 6-6. Let X be uniformly distributed and symmetric
6-3. The random variable X is uniformly distributed about zero with 'lariance 1. Find the appropriate v:li-
on the interval [0, 21. find the distribution of the ran- ues for a and fl.
dom variable Y;;;;; 5 + 2X. 6~7. Show how the uniform density function can be
.-4. A real estate broker charges a fixed fee of $50 used to generate variates from the empirical probabil~
plus a. 6% cqmmission on the lando\iiller's profit. If it}' distribution desctibed below:
....
:!3
Q
.§
!l
'"
Table 6-1 Summary of Continuous Di\'isiolls '"~
, ~ ---------r------------------------,---~---------_r-----------------------I
Moment~
Generating
i
Density I ~----~-.
Pnrl1meters
+-- --1
Density Punctionf(x) Mean Varianee Funetion i
Unifonn a,fJ fix) ~ fJ.a' a';x~fJ (a, fl)12 (fJ- a)' 112
e" - etC' ~
I(f! a) g
f3>a g
:::;; O. otherwise,
~
-----.- ----
n'xpouential ,bO f(x)~M-)·" 00 1/ ,t II ;I.' (1-llAl' t1
);;.
a.~
Gamma 00 fix) ~r(r)(A'r·,eu"', X>O rl;l. rl ).2 (l-ll;tr ~
,bO
~o, otherwise.
\ - - - - - - -- - - - - - - - - - - - - - + --~-
6-11. For the data in Exercise 6~ 10, what percentage 6-22. The life of an electronic system is Y == Xl + Xz . ;.
of C3I'S will experience failure ir. the engine and drive X3 +~, the su:rn of the subsystem component lives,
train du..~g the fi..""St 6 months of use? The subsystems are independent, each having expo-
nential failure densities with a mean time between
6·12. Let the length of time a machine will operate be
failures of 4 hours. "Wnat is the probability '.:hat the
an e,~ponentially distributed random variable with
system wffi opet<1te for at least 24 hom:s?
probability density functionj(t) = ee-!)', t;' O. Suppose
an operator for this machine must be hired for a pre-- 6~23. The repl~hment time for a certaic product is
determined and fixed length of time, say Y. She is paid known to be gamma distributed \\'ith a mean of40 and
d dollars per time period during this interval. The net a variance of 400. FInd the probability that an order is
profit from operating this machine. exclusive of labor received wirhin the first 20 days after it is ordered.
costs, is r dollars per time period that it is operating. Within the first 60 days.
Fmd the value of Y that maximizes the expected total 6-24# Suppose a gamma distributed random variable is
profit obtained. . defined over the interval u :$ x < .;;<) with density
:tCi~ time to failure of a television tube is estt- function
.,
IDa~ to be exponentially distributed \vith a mean of
3 years. A company offers insurance on these tubes
f(x) == :(r) (x_uy-i e-J.{;t-,,) , x 2:u,.;~~O. r> O.
for the f.."'St year of usage. On what percentage of poli-
= 0, otherwise.
cies will they have to pay a claim?
Find the mean of this three~parameter gamma
(fl4Jrs there an exponential density that satisfies the
disoibution.
fc;m5wing condition?
6--25. Tne beta probability distribution is defined by
.P(X ~ 2) = tp{H 3}
If so, fuld the value of ':L f ') r(~+r) "-I (I-x )~I ,O~x~I,)'>O,r>O,
\X - r(),)!'(r) X
6-15. Two manufacturing processes are under consid-
eration. The pet'-umt cost for process I is C. while for = 0, otherwise.
process IT it is 3C Products from both processes have (al Graph the distributiou fOLt> 1, r> L
exponential time~to-failure densities with mean rates (b) Grapll the distribution for ;I, < I, r < L
of 2S-! failures per nour and 35-: fai:ures per =:'our (e) Graph the distribution for ,:( < 1, r;' L
from I and II, respectively. If a product fails before 15
(d) Gmpll the distribution for ,:( ~ 1, r < L
hours it must be replaee'd at a cost of Z dollars. Vlhich
process would you recommend? (e) Gmpll the distribution for,:(= r.
142 Chapter 6 SOUle Important Continuous Distributions
6~26. Show that when i..:= r.:::: 1 the beta distribution 6~36. A manufacturer of a commercial television mou-
reduces [0 the unifoffi1 distribucion, itor guarantees the pic:ure tube for 1 year (8760
hours). The monitors are used in airport terminals for
6~27.Show that when A=2. r= 1 Of,t= 1. r=2:he
flight schedules. and they a.!1~ in continuous use with
beta distribution reduces to a triar.gular probability
power on. The mean life of11e tuhes is 20,000 hours,
distribution. Graph the density fu.1..ction.
and they follow an exponential time-to~failure density,
6~28. Show that if,t ::;:: r :;; 2 the beta distribution It costs the manufacturer $300 to make. sell. and
reduces to a parabolic probability distribution. Graph deliver a ClOnitof that will be sold for $400, It costs
the density functioD. $150 to replace a failed tube, including materials and
6~29. Find the mca."1 and variance of the beta labor. The manufacnuer has no replacement obliga.-
tion beyond the first replacement V1hat is the ma;:lU~
distn'bution.
faeturer's ex.pected profit?
6 ..36. Flnd the mean and variance of the Weibull
6..37. The lead time for ordcts of diodes from a certain
distribution.
manufacturer is known to have a gamma distribution
6 ..31. The d:!ameter of steel shafts is Wcihull distrib.- with a mean of 20 days and a standard deviation of 10
uted with parameters r = 1.0 inches, j3 = 2, and S:::: days. Determine the probability of receiving an order
0.5. Find the probability that :a randomly selected within 15 days of lhe placement date,
shaft 'Nill not exceed 1.5 inches in diameter. 6~38~ Use random numbers generated frOUl your
6~32. The time to failure of a ce::tair. transistor is fav6ritc co:nputer 1an.gu.age or from scaling the ran-
known to be Wei.bull distributed with parameters r:::::: dom integers in Table XV of the Appendix by multi-
0, f3 = ~, and {j = 400. Find the fraction expected to plying by 1O-~. and do the following:
survive 600 hours. (a) Produce 10 realizations of a 'lariable that is uni-
for:n 0= [10, 20}.
6~33. The time to leakage failure in :a certain type of
dry-cell battery is expected to have a Wcibull distribu~ (b) Produce five realizations of an exponential ran-
tion wit.~ panuneters r= t,
0, 13= and 0=400. What dom va..riable with a parameter of ),,= 2 x la-so
is the probability that a battery ""ill survive beyoJid (c) Produce five realizations of a ga.mma variable
800 hours of use? withr=2andi.::::4.
(d) Produce 10 real.izations of a Weibull variable \'lith
6-34, Graph the Weibull distribution with r= 0,8 =1,
and f3 = 1, 2, 3, and 4, r=O,f3 lfl,0=100.
6~39.Use the random number generation schemes
6-35. The time to failure density for a small e01l?-puter
suggested in Exercise 6-38. and do the following:
o. i,
system has a Wetoci1 density v.'ith r:;; fj :::: and
(a) Produce 10 realizations of Y =2X"'! where X fo1-
0= 200,
lows an exponcutial distribution with:a mean of 10.
(a) Vlhat fraction of these units will survive to 1000
hours? (b) Produce 10 realizations of Y:::: fX: I'[:Z;, where
(b) What is the mean time to failure? X; is gamma with r 2, i.::; 4 and ~ is uniform
on [0, 1].
Chapter 7
1: I': f(xleb: I
; required of all density functions.
2. f(x):2:0 for all xj (7·2)
3. lim,..,.,f(x); 0 and lim"-+-f(x) ; O.
4. f(Jl. + x) ~ f(Jl. - x). The density is symmetric about J1.
5, The maximum value off occurs at x; }l.
6. The points of inflection off are at x ~ }l ± cr.
143
144 Chapter 7 The Normal Distribution
(Xii
L. ~_ . J-!
I'
_-=._
x
Figure 7·1 The normal disrribution,
1= 2- 1- e-(l;~)y' dy.
"2,, -
Our proof that [,i( x) dx = 1 will consist of showing thatI' = 1 and then inferring that
1= 1, since/must be everywhere positive. Defining a second normally distributed variable,
Z. we have
[' = !
1 J~ e -(1f2!,' dy ;..- I~ e -(lI')" dz
'12" - ...,2" -
=-
1 Joo 1- e-(1j2)(y'+," 'dydz.
21& -= ----«>
On changing to polar coordinates with the transformation of variables y:::::; r sin e and z == r
cos 91 the integral becomes
Since the integrand of the first integral is that ofa normal density withJl = 0 and <I' = 1, the
value of the first integral is one. The second integral has value zero, that is,
7~2 The Normal Distribution 145
and thus
E(X) =)1[1] + 0":0],
=)1 (7-3)
In retrospect, this result makes sense via a symmetry argument.
To find the variance we must evaluate
so that
v(X) = 0-'. (7-4)
In summary the mean and variance of the normal density given in equation 7-1 are)J and
0-', respectively.
The moment-generating jimcrion for the normal distribution can be sho\l.'I1 to be
,po
r
Mit) = exp ,t)1 + -t-J. (7-5)
(7-7)
-
146 Qmptel' 7 The Normal Distribution
is that of a normal distribution with mean 0 and variance I; that is, Z - N(O, 1), and we say
that Z has a standard nonnal distribution, A graph of the probability density function is
shown in Fig, 7-2, The corresponding distribution function is 4>, where I
and this function has been well tabulated, A table of the integral in equation HI has been
provided in Table II of the Appendix, In fact, many software packages such as Excel and
Minitab provide functions to evaluate <l)(z), For instance, the Excel call NOR.\1SDIST(l)
does just this task. As an example, we find that NOR.\1SDIST(l ,96) = 0,9750, The Excel
function NOR.\1SlNV returns the inverse CDE For example, NOR.\1S!NV(O,975) = 1.960,
The functions NORMDIST(x, J1, (J, TRL'E) and NORMIl'N give the CDF and inverse
CDF'of the N(p., <i') distribution.
or
F(l04) = <r>(2).
<#(z)
u=1
Appendix Table II contains cumulative standard nonna! probabilities for various values of
z. From this table, we can read
4>(2) 0,9772.
Note that in the relationship z = (x - /1)1cr, the variable z measures the departure of x
from the mean 11 in standard deviation (cr) units. For inst:ance. in the case just considered,
F(104) = 4>(2), which indicates that 104 is Mo standard deviations (<1= 2) above the mean,
In general. x = J1 + O';Z. In solving probleIP.s. we sometimes need to use the symmetry prop-
erty of 'P in addition to the tables, It is helpful to make a sketch if there is any confusion in
determ.ining exactly which probabilities are required, since the area under the curve and
over the interval of interest is the probability that the random variable will lie on the
interval
The breaking strength (in newtons) of a synthetic fabric is denoted X, and i: is distributed as N(800,
:44), The pwxhaser of the fabric requires the fab:ic to have a strength of at least 772 nt.A fabric sam-
ple is randomly selected and tested. To find p(X;;:: 772). we first calculate
=P(z<
~ ? :;', ~-
=4>(-2,33)=0,01. ~/:..
Hence the desired probability. P(X ~ 772), equals 0.99. Figure 7~3 shows the calculated probability
re~a:iveto both X and Z. We have chosen to' work with the randO'm variable Z because its distribution
function is tabulated.
~~i\'/['t:i
T!le time required to repair an automatic loading machine in a complex food-packaging operation of
a production process is X minutes. Studies have sho~"n that t.l:\e approx.imationX - N(120, 16) is quite
goO'd. A sketch is shown ~ Fig. 74. If the process is dO\\-u for more than 125 minates. ail equipment
cr=1
~
. ." y"':/ : - - - _ .
772 I' = BOO x z
~=4
must be cleaned, with the loss of all product in process. The total cost of product loss ,and c1eaning
associated with the long downtime is $10.000, In order to deter.:J.I'li.ne the probability of this occurring,
we proceed as follows:
= 1- <1>(1.25)
=1-0.8944
=0.1056.
Thus. given a breakdov.lI. of the packaging machine, the expected cost is E(C)::: 0,1056(10.000+ CRI)
+ 0.8944(CRj ), where C is the total cost,and CRI is the repair cost. Simplified, E(e)::: CR + 1056. SuP'"
!
pose the management can reduce the mean of the service time distribution to 115 minutes by adding
more mainte::lancc personneL The new Cost for repair will be CRl > CRI; however,
=1-<!>{2.5)
=1-0.9938
=0.0062,
so that the new expected cost would be Cftz + 62, and ODe would logically make the decision to add
to the maintenance crew if
or
CR, C" <$994.
It is asstl.OC.cd that the frequency of breakdo~:ns remains unchanged.
The pitch diameter of the thread on a:fitting is nonnally distributed with a mean of 0.4008 cmanda stan-
dJtrd deviation of 0.0004 CllL The designspecitiClltiOns are O.4000±O.OOlO em. This is illustruedillFig.
7"5. Notice that the procc.<>s is openrting with the mC3J1 not equal to t:he nominal specifications, We desirc
to detmrrine what fr:action of product is \\itbin tolerance. Using the approach employed previously.
0.3990 0.4010
Lower spec, Upper spec.
limit limit Figure 7-5 Distribution of thread pitch diameters.
7-2 The Normal Distribution 149
As process engineers study tbe result') of such calculations, they decide to replace a worn cutting tool
and adjust the machine producing the fittings so that the new mean falls: dirwJy at the nom.i:nal value
of 0.4000. Then.
P(0.3990 ~ X S 04{)lO) ~ Pi' 0.3990- OA ,; Z:5 0.4010 -0.41
\ 0.0004 0.0004 ~
~ P(-2.5'; Z S+2.5)
~ <I;>{2.5)-<t>(-25)
~ 0.9938-0.0062
=0.9876.
We see that with the adjust1:nt'::nts, 98,76% of the fittings will be ~ithin tolerance. The distribution of
adjusted maclrine pitch diameters is shown in Fig. 7~6,
Another type of problem involving the use of ta.bles of the nonna! distribution sometimes arises. Sup-
pose, for example, that X ~ N(50, 4). Furthermore, suppose we want to determine a value of X. say x.
such that P(X > x) "'" 0,025. Then, ,
,J X-50) =0.025
P(X>x)='\.Z>-2-
or
.~.:::50 =L96=<I;>~1(0.975)
2
There are several symmetric mrervals that arise frequently. Their probabilities are
P(jJ.- LOOo-;fX".u+ 1.000) =0.6826,
P(jJ.- L6450-"X" Ii + 1.6450) =0.90,
P(jJ. - 1.960-" X "J.I + 1.960-) = 0.95,
P(jJ. - 2.570-;f X;f J.I + 2.57 d) = 0,99,
P(jJ.- 3,()I)o-"X:;; Ii + 3.00d) =0.9978. (7-9)
and
n
v(f)=ai=Lor
1''"'1
Using moment-generating functions, we see that
(7-12)
Therefore,
(7-13)
~1~~~r~~i;~z
An assembly consists of three linkage components. as sho\Vn in Fig. 7-7. "The properties X:. ;s,;, and
X) are given below, with means in centimeters and variance in square centimeters.
XI - N(12, 0,02)
X, - N(24, 0,03)
X, - N(lS, 0.(4)
Links are ~dU~bY different machine.'> andopemtors, so we have reason to assume thatX\.X2> and
XJ axe ~ent Suppose we want to deter"'~e P(53.& S Y S; 54.2). Since Y;; XI + X2 + XJ' Y is
distributednomuilly withme.m,uy= 12+ 24+ 18 ~54lmd variance rr= a:
~ cr; + aj~O,02 + 0,03
+ OJ14 =0.09. Thus,
( (0) <0) )
·)(.-..f--_·xv'---+--)(3---i
rt!~ilI~I~I~j
A shaft is to be assembled into a bearing, as shown in Fig. 7-8. The clearance is Y;;;;; X! -;X,. Suppose
x, - N(1.500, 0.0016)
and
x, - N(I.480. 0.0009).
Then,
Jly= oJ', + a,J1,
(1)(1.500)+ (-1)(1.480;
=0.02
~:,,;+a;:0
and
= (1)'(0.0016) + (-1)'(0.0009)
=0.0025,
so ::hat
r--
( Vlhen the parts arc assen:bled. there 'Will be interference if Y< 0, so
x,[~,,-.___
Z, = .
Y- k
,
""'/1.,
,'<I _-
I"" (12
(7-15)
il£'" I
l i=l
bas an approximate N(O,!) distribution as n approaches infinity. If F, is the distribution
function of Zit' then
(7-16)
The Hgeneral conditions!) mentioned in the theorem are informally summarized as fol~
lows. The terms Xi' taken individually, contribute anegIigible amount to the variance of the
SUlli, and it is not likely that a single term makes a large contribution to the sum.
The proof of this theorem, as well as a rigorous discussion of the necessary ass"uinp~
!ions, is beyond the scope of this presentation. There are, however. sevenU observations that
should be made. The fact that Y is appro>imately normally distributed when the Xi terms
may have essentially any distribution is the basic underlying reason for the importance of
the nonnal distribution. In numerous applications, the random variable being considered
may be represented as the sum of n independent random variables t some of whicb may be
. measurement error. some due to pbysical considerations. and so on, and thus the normal
distribution provides a good approximation.
A special case of the central limit theorem arises when each of the components has the
same distribvtion.
Theorem 7-2
X:z, .... , Xn is a sequence of n independent. identically distributed random variables
If Xl'
withE(X,) ~ J1 and VeX,) ~ <1", and Y=X, +X2 + ... + X" then
Z =Y-"!!: (7-17)
fI (1'~
practical standpoint, some very crude rules of thumb can be given where the dis.tribution of
the X, terms falls mto one of three arbitrarily selected groups, as follaws:
1. Well bebaved-The distribution of Xi does not radically depart from the nonnal dis-
tribution. There is a bell~shaped density that is nearly symmetric. For this case,
practitioners in quality control and other areas of application have found that n
should be at least 4, That is, n " 4,
2. Reasonably bebaved-The distribution of X, has no prominent mode, and it appears
much as a uniform density. In this case, n ~ 12 is a commonly used rule,
3. TIl behaved-The distribution has most of its measure in the tails, as in Fig. 7-9. In
this case it is most difficult to say~ however, in many practical applications, n ¢! 100
should be satisfactory.
,~;W;p'!e1\?':
Small parts are packaged. 250 to the crate. Part weights are independent random variables \';Iith a mean
of 0.5 pound and a standard dC\-i.atiou of O.lQ pound, Twent'j crates are loaded to a pallet. Suppose
we wish to fiDd the probability that the parts Q!: a pallet 1,\-ill exceed 2510 pounds b weight. (Neglect
both pallet and crate weight,) Let
Y=X j .... X2 + ,., +XSI)X}
represent the total weight of the parts. so that
and
P(Y>25lO)=P(Z> 2510-2500\
7.071 )
=1-<1>(1.41) =0.08.
Note that we did not know the distribution of the individual part weights.
t(x)
(al (b)
Figure 7-9 ill-behaved distributions.
154 Chapter 7 1M Normal Distribution
Table 7~1 Activity Mean Times and Variances (in Weeks and Weeks~
'[~hlIz~J
In a construction project, a network of m.ajor activities has been eonstucted to serve as the basis for
pla."'lUing and scheduling. On a critical path there are 16 activities, The means and variances are given
in Table 7-1.
The activity times may be considered independent and the project time is the sum of the activ-
ity times on the critical path, that is, Y==-X 1 + Xz -:- .,. + XIS' where fiO) the project time and XI is the
ti:m.e for the ith activity. Although the distributions of the XI arc unknown. the distributions are fairly
well behaved. The contractor would like to know (a) the expected completion time, and (b) a project
time corresponding to a probability of 0.90 of having the project completed. Calculating Jly and ~
we obtain
/.J.1';;;;;: 49 weeks,
er!.; 16 weeks2"
The expected completion time for the project is thus 49 weeks. In detennining the tfrne Yo such that
the probability is 0.9 of having the project completed by thaI time, Fig. 7-10 may be helpful
We may calculate
so that
and
y;=49+ 1.282(4)
~ 54.128 weeks,
t
O'y= 4 weeks
Yo Y
Figure 7~10 Distribution of project times,
7~5 The Norma1 A.pproximation to the Binomial Distribution 155
as n -7 IXI. Using Stirling's formula to approximate the terms involving n! in the binomial
model, we eventually find that,{~ Ilirgen,
(7·20)
sO that
(7-21)
This result makes sense in ligbt of the central1intit theorem and the fact that X is the sum
of iudependent Bemotill:r1:rial~ ES<1·tb~t E(X) = np and V(X) = npq). Thus, the quantity
(X -npl/,jnpq approximately llJlj; aN(O,I) distnDutiolL IT p is close to and n > 10, the t
approximation is fairly good; however, for other values of p, the value of n must be larger.
In general, experience indicates that the approximation is fairly good as long as np :> 5 for
1 . 1
P S '2 or when nq > 5 whenp > 2' , _,.'
"'>.~ .,r: ..' ,
I'r )0>5 - .,
~i>~~El
L
156 Chapter 7 The Normal Distributio:l
pex=x)
P(X" x)
P(x:;'x)
p(a"Xsb)
rlHffi§~:t~!!j
Using the data from Example 7-9, where we had n = 100 and p = 0,2. We evaluate P(X 15).
P(X:; IS), peX < 18), P(X ~ 22). aIld P(18 < X < 21),
0. P(j~=P(14.5"
~-=":
X 515.5)= <1>(155-20)_<1>(14.5-20\
\4 4)
= <1>(- Ll25) -<1>(-1375)= 0,046,
Vz. "'"
P{X S 15) = <I>(!2:5-20\j = 0,)30,
\ 4
I7
v.P(X <IS) =P(X" 17)=i .5-2'?! = 0.266,
~ " 4 ;'
~ "",
@''1'(Xi'22)=I-<I>(21.5 -20)=0,354, ---~-'-
4
JI P(lS<X<21)=P(1%X,,20)
=<I>( 20.5 - 20 l_dI85- 20)
" 4 } \ 4
_ _ ,, _ _ =0550-0.354 =0,196. _ _ _ _ _ _ _ _ _ _ _ _ _ ,
~,--c''-'-'---'-~-'--'-
7·6 11:e Lognormal Distribution 157
As discussed in Chapter 5,' the random variable p =XI" where X has a binomial distribution
with parameters p and n, is often of mte:Tst Interest.in this quantity stems primarily from
sampling applications, where a randQID sample of n observations is made, with each obser-
vation classified success or failure. and where X Is the number of successes in the sample,
The quantity p is simply the sample fraction of successes. Recall that we showed that
E(j;) =p
and
v(p)
has an approximate N(O, I) distribution. This result has proved useful in many applications,
including those in the areas' of quality control, work measu:ement, reliability cmgineering.
and economics. The results are ouch more useful than those from the law of large numbers.
Instead of timing the activity of a maintenance mechanic over the period of a week to determine the
fraction of his time spent in an activity classification called "secondary but necessary," a technicia.n
elects to use a work sampling st'.1dy, randomly picking 400 time point'> over the week. taking a flash
observation at each, and classifying the ;:activity of the maintenance mecb.ru.tic. The value X will rep~
resent the number of times tbe mechanic was involved in a "secondary but necessary" activity ar~dp
;:;; Xl400. If tbe true fraction of time that he is involved iT. this activity is 0.2. we determine !he prob~
ahility tharp. the estim.atcd fraction, falls bet\\'een 0,15 and 0.25. That is,
/(x) --;~e
1 {-1!2)[{InJ:-/ly)!CTy t, x>O,
xO"y.JZ1!
=0 oilieI'W'ise. (7-24)
The lognormal distribution is shown in Fig. 7 -II. Notice that, in general, the distnbution is
skewed, with a long tail to the right. The Excel funetions LOGNORMDIST and LOGINV
provide the CDF and inverse CDF, respectively, of the lognoIlDlll distribution.
(7-26)
In some applications of the lognormal distribution, it is important to know the values of the
median and the mode. The median, which is the value x such that P(X :> Xl ; 0.5, is
x=eUl', (7-27)
The mode is the value of x for which/(x) is maximum, and for the 10gnoIlDlll distribution,
the mode is
(7-28)
Figure 7~11 shows the relative location of the mean, median, and mode for the
lognormal distribution. Since the distribution has a right skew, generally we will find thai
mode < median < mean.
x
Median Flgnre 7·11 The lognormal disuibution.
7~6 The Lognormal Distribution 159
()i.YI' q),j = 1, 2, '''' n., respectively, and {a) I is a sequence of constants while b
e" is a single constant, ilien ilie product
,
W~bTIX?
j=l
and
4~ If Xl' X-z, .. "' X,t are independent lognorm.al variates; each with the same parameters
(fly. ~), ilien ilie geometric mean
(IT Xi)"'
\j=l
has a lognormal distribution wiili parameters)i., and a;ln.
i~~~iiicz;r~
The: random variable Y = In X has a NO O. 4) distribution, sO X has a lognormal distribution with a
mean and varia.:nce of
E(X) =: e IC +:l'l14 = ell",", 162,754
.
and
VeX) = e[1(10)+4)(e4 - 1).
e'4(e4 _ 1) = 53.59Se",
,\J!!~~!t1~!~'"
Suppose
Y,=InX,-N(4,1),
Y, =InX,- N(3, 0.5),
Y, = InX, - N(2, 0.4),
Y. =InX, - N(l, 0.01),
and fu."therrnore suppose XI' Xz, X,:, and X4 are independent rando:n variables. The random va\"iable
W defined as follows represents a critical perfor::nance .'3riable 0:1 a telemetry system:
W =: eL5rx~x;.1x;·7X!1}.
(7-29)
x,
X,
.'f--""",=
relation coefficient between XI and X2• such thar - oc < fJ.; < OCt oc < f.l2 < OX'. 0'1 :> 0, 0': >
0, and-l <p< 1.
The marginal densities/, and/, are given, respectively, as
£(Y ) -
.. 1 ....: -
r-
""""
'I x x
.1 \ :' 2
lax2 -- _1_e-{~2)[(', -#,)!c.]'
r:;---
0'1 ",;2%
(7-31)
(7-32)
and
X, - N(Ji1, a~),
so that
E(X,) -)J,l'
E(x;) = iLz,
VeX,) ~ a;,
veX2 ):=: O'~. (7-34)
The correlation coefficient p is the ratio of the covariance to [at' 0';1. The covariance is
Thus
(7-35)
162 Chap,er 7 The Normal Distribution
The conditional distributions Ix,,,,(x;) and/x,",(x,) are also important. These condi-
tional densities are nonnal, as shown here:
(1-37) ,
for- 00 < Xl < 00. Figure 7~13 illustrates some of these conditional densities.
We first consider the distribution!XZU:l' The mean and variance are
E(X,!>:,) ~ i11. + p(a.ja)(x, -1',) (7-38)
x,
E(Xzlx,)
x,
x,
(b)
Figure j·13 Some typical conditional distributions. (a) Some example conditional distributions of
X2 for a few values of x~. (b) SO:rtle e:<.ample conditional distributions of Xl for a few values ofX:!.
I
, and
a;O - p')
163
(7-39)
The locus of expected values of X:r for xl' as shown in equation 7-38, is called the
regression ojX1 on Xl' and it is linear. Also, the variance in the conditional distributions is
constant for all Xl'
In the case of the distribution/x!;,;:. the results are similar, That is.
E(X,1x,) = Ji., + p(O',/IY,)(x, - .11.,), (7-"1)
and
(7-43)
b the bivariate nonnal distribution we observe that if p:= 0. the joint density may be fac-
tored into the product of the marginal densities and so X j and Xz. are independent. Thus, for
a bivariate normal density, zero correlation and independence are equivalent. If planes par-
allel to the X l ,.l2 plane are passed through the surface shown in Fig, 7~12, the contours cut
from the bivariate normal surface are ellipses. The stt:dent may wish ~o show m:s property.
In an atterr.pt to substitute a nondestructive testing procedure for a destructve test, an extensive st<.ldy
was made of shear strength. Xz., and we~d diameter, Xl' of spot welds, with the foUo'N.ing findings,
1. [XI' Xl) ~ bivariate normaL
2. J.ll = 0.20 inch. Jlz = 1100 pounds, a; "" 0.02 inch:" pounds'. and p=O,9,
The regression of X2 on Xl is ,thfis-.....\
=
= 1-1>(-1,71) 0.9564,
a-:d he recommends a policy such that if the weld diameter is not less tea..-: 0.18, the weld will he clas-
sified as satisfactory.
164 Chapter 7 The Norm.al Distribution
F;x;,n,~i~7;':J.S:
In developing an admissions policy for a large university, the office of student testing and evaluadon
has noted rl:at X l the combined score on the college board examinadons, and X:. the student grade
<
point average at the end of the freshman year, have a bivariate normal distribution, A grade point of
4.0 correspond'! to A. A study indicates that
1', = 1300,
",,=2,3,
0;=6400,
~=O.25.
p=O,6,
AJ:J.y student with a grade point average less than 1.5 is au:omatically dropped at the end of the fresh~
man year; however, an average of2.0 is considered to be satisfactory.
Ax.. applicant takes the college board exams, receives a combined score of 900. and is not
accepted. A.n irate parent argues that the student will do satisfactory work and, specifically, will have
better than a 2.0 grade point average at the end of the freshman year. Considering only the proba~
bilistic aspects of t.'1e problem. the director of admissions wants to determine P(~ t:: 2.0rc: 900).
Noting that
1_ <1,(2,0-0,8) 0,0013=
;, 0,4 '
which predicts only a very sli.m chance of the parent's claim being valid.
WIth this procedure, we would begin by generating 12 uniform [0, 1] random number real-
izations. adding them and subtracting 6. This entire process is repeated until the, desired
number of realizations is obtained.
Although the dire<ot method is exact and usually preferable, approximate values are
sometimes acceptable.
7·9 SUMMARY
This chapter has presented the normlll distribution with a number of example applications.
The normal distribution, the related standard nannal, and the lognormal distributions are
univariate, while the bivariate nonnal gives the joint density of tvlo related normal random
variables.
The normal distribution forms the basis on which a great deal of the work in statistical
inference rests. The wide application of the normal distribution makes it particularly
important.
7·10 EXERCISES
/X/
(7:"i. Let Z be a standard no::n:..al random variable and 7~e personnel manager of a large company
"calculate the follo .....ing probabilities. llSing sketches , requires job applicants to take a certain test and
where appropriate: achieve a score of 500. If the test scores are aorn:tall.y
<a) p(0" Z" 2). distributed \Vith a mean of 485 and a standard devia-
(b) p(-I:$Z;;+I). tion of 30, what percentage of the applicants pass the
test?
(e) P(Z ~ 1.65).
7~8. Experience indicates that thf': rhwelonment time
(d) P(Z~-1.96).
for a photographic printing pal
(e) p(IZI> 1.5). N (30 seconds, 1.21 seeonds2).
(I) P(-1.9';; Z;; 2). The probability that X is at 1eas-. _.
(g) P(Z;; 1.37). probability that X is at most 31 seconds. (c) The prob-
(b) pclZl,;; 2.57). ability that X differs from its expected value by more
than 2 seconds.
7·2, Let X - N(lD, 9). Find P(X:!: S), P(X~ I2).PCz,; C':\
X S; 10). t/ ,: 1.9jA ccrtain type of light bulb has an output known
~ roOe norroaily distributed with mea.'1. of 2500 end foot-
, 'Ji? In ea~'part. b'elow, find the value of c that makes candles and ~ standard devi~~n of 75. e~d footean-
/ tile probability statement true. dles. Determ:ne a lower speCt::lcation limit such that
(a} ~(c);;;;; 0.94062. only 5% of the :na:mfactured bulbs will be defe.....-tive.
(b) p(jZl
:!:e) = 0.95. 7·10. Show that the moment·generati!:g function for
(c) P(1Zi5C)=0,99. the nO!."""~ distribution is as given by equation 7-5.
(d) p(Z ~ c) = 0.05, Use it to generate the mean and variance.
~lfX-N(p. 0-'), show that Y=aX+b. where a
14~ 1£ P(Z ~ Zg) = a., detemrlne Zc for a 0 0225
. ~;:r:.:;.¥
and b is
are real comta:nts, also normally distributed.
. ~"=Q.l)'l5, «=0.05, and «=0.0014. . Use the methods outlined in Chapter 3 ..
(,1.0:: ; X' N(80, 10'), compu"' the following.a; ?n 1",12. The inside diameter o:(~_piston ring is normally
Ce) P(X;; 100). distributed with ame;m of 12 em and a standard <levi-
(b) p(X'; SO). arion of 0.02 Glli.
(e) p(75';X;; 100). (a) Wbat freedon of the piston rings will buve dimn-
(d) P(l5';; X). etern exceeding 12.05 em?
(e) p(1X - 801,; 19.6). (b) Wb.at inside diameter value e has a probability of
Q,90 being exceeded?
7";;. The life of a particular type of dry-<ell battery is
normally distributed with a mean of 600 days and a (e) t is the probability that tile iLside diameter
standard deviation of 60 day'S, What fraction of these ? ."".._:fur between 11.95 and 12.051
batteries would be expected to survive beyond 680, \ 7 3. plant manager orders a process shutdown and
days? W"'....at fraction would be expected to fail before 'etting readjustment whenever the pH of the fb.a1
560 days? product falls above 1.20 or below 6.80, T.::e sample
166 Chapter 7 The :Sonnal Distribution
).~" r
"
pH is normally distributed with unknown j.1 and a :(7)\:wh~c~the acceptable range is as in (a) and the
'/
standard deviation of 0"= 0.10, Detennine the follow~ ':;/' hardness of each of cine randomly selected spec~
ing probabilities: imens is independently deten:rined, what is the
~ a) Of readjusting \Vhen the process is operating as , expected llumber of acceptable specimens among
intendedwithj.1::::7.0. '1 ;:~- ()- L ,:..::.::-~ the n.ine specimens?
.......,.-- \. ? ' ,:>,' ,
(b) Of readjusting when the process is slightly off 7..20. Prove that E(Z~) "'" 0 and V(Zt\):::: 1, where z., is
ta..rget with the mean pH 7.05. as cL<>filled in Theorem 7·2,
(c) ?ffailing to readjust ",:hen~e process is too aIka~ l~tXi(i """ 1, 2, ... , n) be independent and iden~
line and the mean pH IS J1. - 725. i'ruCiily distributed random variables with mean j.1 and
(d) Of falling to readjust when the process is too variance cr.
Consider the sample mean,
acidic and the mean pH is f.1 "'" 6.75.
7~14. The price being asked for II certain security is
distributed normally \I,'1tt a mean of $50.00 and a sta.n~
dan! deviation of $5.00. Buy"" are willing <0 pay an Show thatE(X) = Ii and V(X) = d'in.
amount iliac is also nonnaTIy :n~tributed W:th. a m~ 7-22. A sbaft with an outside diameter ~O.D.) -N(L20,
of $45.00 ~~ a standard de-:atlo~ of $2.50. v.~at 1S 0.0016) is inserted into a sleeve bearing having an
the probability that a ~Ctlon ",ill take place, ~in de diameter (I.D.) that is N(L25. 0.0009). Detet~
W a c specifications for a capacitor are that its lif,~' ,,' e the probability of interference,
must be bet-.veen 1000. ~d 5000 ~ours. The life i _ ,,..2 An assembly consists of three components
0:
known to be normally d~,tnbutcd WIth a mean 300.0 (,placed side by side. The length of each component is
hours.. The revenue ~alize~ frmn e3;ch capacItor IS non::nally distributed v.':ith a mean of 2 inches' and a
$9.00; however, a failed urut must be replaced ~t a standard deviation of 0.2 inch. Speeifications re<;iuire
cost of $3.00 to the compa~y. Two .manW:actunng that allrsemblies be between 5.7 and 6,3 inches long.
processes can produce capacltors hav";ng sallsfactory How JPany assemblies v.ill pass these requirements?
mean lives. The standard deviation for process A is \. !
1000 hom and for process B it is 500 hours. How- ,\241 Find the mean and variance of the linear
ever, process A manufacturing costs arc only half co'nlbinatior. '
those for B. What value of process manufacturing cost Y "" Xl + 2Xz + X3 + X4 ,
is critical, so far as dictating the use of process A or B? where Xl ... N(4. 3), X2 ~ N(4. 4), X, ~ Nfl, 4). and
7~16. The diameter ofa ball bea.-ing is • normally dis- ~X.
--1,;(3,2). \Vbat is the probability tbat 15 S; YS; 207
tnbuted random vanable With mean j.1 and a standard ~ound~off error has a uniform distribution on
deviation of 1. s~ecifiC,atio~ f?I" the ffi:a~ete: are, .~ +O.5J and round-off erron; are independent A
6."!: X."!: 5, and a ball beanng \\'1thin these limits )'lelds ';~um of 50 numbers is calculated where each is
<, a proSt of C.doUars. However. if X.: 6. then the profit rounded before adding. \-Vhat is the probability' that
/ f is -R j dollars, or if X > S. L.1e profit is -Rz dollars. the total rO~'1d-off error exceeds 57
Fied the value of jl that maximizes the expected
profit 7-26. One hundred small bolts are packed in a box.
Each bolt weights 1 ounce, with a standard deviation
'Z..dVIn tt:e preceding exercise, find the Optimum of 0.01 O'.L.1.ce, Find the probability that a box. weighs
vatue of j1. if R~ :R2 == R. more than 102 ounces.
7~18. Usc thc results of Exercise 7~ 16 with C= 58.00,
·S~27)An automatic machine is used to fill boxes 'With
Rl "" $2.00, and R1. = $4,00. What is the value of Ii that s;;[;'p powder. Specificati(Iit, require that the boxes
~
;nizes the expected profit? weigh bet"Neen ~l.&and ;I2.2, ounces. The only data
7',~e Rockwell hard::ess of a particuJ.ar alloy is available about ~inachine peiformance concern the
rmally distributed with a mean of 70 ar.d a standard average content of grot:p~(nine boxes. It is l$noy,,"D.
. tion of 4. that the average eontent i~ 11.9l?llUces with a standard
(a) If a specimen is acceptable only jf its hardness is deviation ofl O.OS .ounce.- ~ fraction of the boxes
between 62 and 72, wha~ is me probability that a produced is ddcctive? Where shOUld the me~ be
tandomly chosen specimen has an acceptable located in order to minimize this fractioa defective?
hardness? Assume the weight is normally distributed.
(b) If the acceptable range of hardness was (70 - c. 7~28. A bus-J;r:eve1s between !:VIa cities, but visits six
70 + c), for what value of c would 95% of all Intermediate cities on thero-at.e. The means and st3L-
specimens have acceptable hardness?" dard deviations of t!:e traveL times are as follows:
7~10 Exercises 167
-----------------
Mean StandMd
'=D:Drne ""dom variable Y = In X has a.'l(SO, 2S)
dist..--i.bution. Find the mean, va."':ance, mode, and
City Tune Deviation
median of X
PaL., (hours) (hours)
7-3&. Suppose independent :random variables YJ , Y:;,
1-2 3 0.4 Y~ are such that
2-3 4 0,6 Y, - N(4, I).
3-4 3 0,3
In X, - N(3, I).
4-5 5 1.2
In X3 - N(2, 0.5).
,('1,5-6 7 0,9
0,4 Find the mean and variance of W = eZX~X~SX~,:!ll.
16-7 S
Determine a set of specifications L and R such that
0- 7-8 3 0.4
,r, /.----.., P(L;; W ~ R); 0,90.
~~yts the probability that the bus cOC1p:etcs it" jour- '2::Yhow:ba: the densi"! function for a lognormally
\!"
"
" n·b.~,\~!itbin 32 hottts? distributed ra.'1dom variable X is given by equation
~
.r' : ~se randoLl numbers generated from your - 0'2
avorite computer package or from scali."lg the random Show that the two new random variables are
J integers in Table 'XV of the Appendix by multiplying 0 'er:.dent.
by 10-5 to generate six realizations of a N(100, 4)
~'The life of a tube (X,.) and the filament cliameter
variable, using the follovdng:
(X1) are distributed as a bivwte normal random vari-
(a) The di..rect method.
ahle with the parameters fi.l 2000 hours, fJz::::; 0,10
(b) The approximate methO<t inch, 0-; =2500 houn;', <f, ~ 0.01 inch', and p ~ 0.87.
7~32. Consider a linear SOmbiDation Y::::; 3Xt - 2X;;, The quality-control manager wishes to deteani..'ie the
where XI is N(lO. 3) ~d X1. is unifo::::::::lly distributed life of eaeh tube by illcasuring the filament dim.eter.
on .[01°1. Generat{six rea~tions of the random If a filar:r::.ent diameter is 0.098, what is the probability
va;;al1'fe Y, whetc XI andX:r. are mdependent. that the tube will last 1950 hours?
~ Z - N(O, 1), generate five :ealizations of Zi. 7-42~
A college professor has noticed that grades on
,;;:;.;. ~y = In X a;J.d Y - N(jJ.y, v;), develop a proce-
/mr ger:.era~:lr:wz~tions of X. 2
each of t'\IO quizzes have a bivariate norma: clistribu-
tion with the parameters fi.; 75. fJz ;;;; 83, ~~ 25,
t'~1~j~If y =:: X/~,X;, wnere XI - N(J1.p v t ) and 0; = 16, and p = 0.8. If:a student receivC$ a
/
~' .2 grade of 80 on the first quiz, what is the probability
x,.i~ N(J1,. a,) and where X, and
_
Xo-are
.
independent.
_
thatsh e will d0 be tterOn th csecond ' HOWlS
one: . th e
deVelop a generator for producmg realizations of Y; aft ed -,"- " 8'
~wer ect bv mQ.li.ll1g p..:.:.. --v. .
7~36. The brightness bulbs is normally dis- ( . . . ':\ •
tributed. with. a mean of 2500 foor.caodles and a stan- '-~ Consider the surface y X), whercfis the
dard deviation of 50 footcandles. The bulbs are tested bivariate normal density function.
and all those brighter than 2600 footcandles are (a) Prove that y = constant cuts the surface in an
placed in a special high-qUality lot. What is the prob- ellipse.
ability distribution of the remaining bulbs? \Vbat is (h) Prove that y = constant with p
their expcc:ed hrightness? cuts the surface as a circlc.
and ° 0: ;: : a;
168 Chapter 7 The Normal Distribution
L
WL.ere Xi - lVI " and"IS Ind
"(0,cr) dCll.
epen t bution with n degrees of freedom.
7~46~ Let the independent random variables XI ~ G'LetX- NCO,I), Define anew random variable Y
N(O, 0'-) for i ~ 1, 2. F'l!ld the probabiJity distribution ~!XI. Tneu, firul the probability distribution of Y. This
of is often called the hali'~It()rmLll distribution. .
[ ,
Chapter 8
Introduction to Statistics
and Data Description
8·1 THE FIELD OF STATISTICS
Statistics deals with the collection. presentation, analysis, and use of data to solve prob!ems.
make decisions, develop estimates, and design and develop both prOduCIS and p:ocedures.
An unde:Slllnd:ing of basic statistics and statistical methods would be useful to anyone in
this information age; however, since engineers, scientists. and those working in m:mage~
:ment science are routinely engaged with data, a knowledge of statistics and basic statistical
methods is particularly vilal. In this intensely competitive, high-tech world economy of the
first decade of the tv.:enty-first century. the expectations of consumers regarding product
quality, perrorIllllllce, and reliability have increased significantly over expectations of the
recent past. Furthermore, we have come to expect high levels of performance from logisti-
cal systems at a111eve1s, and the operation as well as the refinement of these systems is
largely dependent on the collection and use of data. While !.hose who are employed in var-
ious "service industries" deal with some\Vbat different problems, at a basic level they, tOO,
are collecting data to be used :in solving problems and improving the "serv.ice" so as to.
become more competitive in attracting market share,
Statistical methods are used to present, describe, and understand variability. In observ-
ing a Y,ariable value or several variable values repeatedly, where these values are assigned
to units by a process, ~:e note that these repeated observations tend to yield different results.
\Vhile this chapter wHl deal with data presentation and description issues; the fonowing
chaplers will utilize the probability concepts developed in prior chapters to model and
develop an understanding of variability and to utilize this understanding in presenting infer-
ential statistics topics and methods.
Virtually all real~world processes exhibit variability. For example, consider situatio:lS
where we select several castings from a manufacturing process and measure a critical .
dimension (such as a vane opening) on each part. If the measuring instrument has sufficient
resolution, the vane openings will be different (there will be variability in the dimension).
Alternatively, if we count the number of defects on printed circuit boards, we will find vari-
ability in the counts, as some boards will have few defects and others will have many
defects. This variability extends to all environments. There is variability in the thickness of
oxide coatings On silicon wafers, the hourly yield of a chemical process, the number of
errors on purchase orders, the flow time required to assemble an aircraft engine, and the
therms of natural gas billed to the residential custon:ters of a distributing utility in a given
month.
169
110 Chapter 8 Introduction to Statistics and Data Description
Almost all experimental activity reflects similar variability in the data observed, and in
Chapter 12 we will employ statistical methods not only to analyze experimental data but
also to construct effective experimental designs for the study of processes.
Why does variability occur? Generally, variability is the result uf changes ill the con-
ditions under which observations are made. In a manufacturing context,. these changes may
be differences in specimens of material, differences in the way people do the work. differ-
enceS in process variables~ such as temperature, pressure, or holding tice, and differences
in environmentalJactors, such as relative humidity. Variability also occurs because of the
measurement system. For example. the measurement obtained from a scale may depend on
where the rest item is placed on the pan. The process of selecting units for observation may
also cause variability. For example, suppose that a lot uf 1000 integrated circuit chips has
exactly 100 defective chips. If we inspected all 1000 chips, and if our inspection process
was perfect (no inspection or measurement error)t we would find aU 100 defective chips.
However, suppose that we select 50 chips. Now some of the chips will likely be defective;
and we would expect the sample to be about 10% defective, but it could be 0% or 2% or
12% defective, depending on the specific chips selected.
The field of statistics consists of methods for describing and modeling variability, and
for ma.ki.ng decisions when variability is present. In ir¢erential statistics, we usually want
to make a decision about some population. The term population refers to the colleCtion of
measurements on aU elements of a universe about Which we wish to draw conclusions or
make decisions. In this text, we make a distinction between the universe and the popula~
tions, in that the universe is composed of the set of elementary units or simply units, while
a population is the set of numerical or categorical variable values for one variable associ~
ated with each of the universe units. Obviously, there may be several populations associated
with a given universe. An example is the universe consisting of the residential class cus~
tomers of an electric power company whose accounts are active during part of the month of
August 2003. Example populations might be the set of energy consumption (kilowatt-hour)
values billed to these customers in the August 2003 bill, the set of customer demands' (kilo-
watts) at the instant the company experiences the August peak demand. and the set made up
of dwelling cat~gory, such as single~family unattached, apartments, mobile home, etc.
Another example is a universe which consists of all the power supplies for a personal
computer manufactured by an electronics company during a givenpctiod. Suppose that the
manufacturer is interested specifically in the output voltage of each power supply. We may
think of the output voltage levels .in the power supplies as such a population, In this case,
each population value is a numerical measurement, such as 5.10 or 5.24. The data in this
case would be referred to as measuremem data. On the other hand, the manufacturer may
be interested in whether or not each power supply produced an output voltage that conforms
to the requirements. V.le may then visualize the population as consisting of attribute data,
in which each power supply is assigned a value of one if the unit is nonconforming and a
value of zero if it conforms to requirements. Both measurement data and attribute data are
called numerical data. Furthermore, it is convenient to consider measurement data as being
either continuous data or discrete data depending on the nature of the process assigning
values to unit variables. Yet another type of data is called categorical data. Examples are
gender, day of the week when the observation is taken. make of automobile. etc. Finally, we
have unit identifying data. which are alphanumeric and used to identify the universe and
sample unitS. These data might neither exist nor have statistical interpretation; however, in
some situations they are essential identifiers for universe and sample units. Examples would
be social security numbers, account numbers in a bank., YIN numbers for automobiles, and
serial numbers of cardiac pacemakers. In this book we will present teclmiques far dealing
with both measurement and attribute data; however, categorical data are also considered.
8-1 The Field of Statistics 171
In most applications of statistics, the available data result from a sample of units
se!ected from a universe of interest, and these data reflect measurement or classification of
one or more variables associated with the sampled units. The sample is thus a subset of the
units, and the measurement or classification values of these units are subsets of the respec-
tive universe populations.
Figure 8-1 presents an overview of the data acquisition activity. It is convenient to think
of a process that produces units and assigns values to the variables associated with the units.
An example would be the manufacturi.I.;g process for the power supplies. The power sup-
plies in this case are the units, and the output voltage values (perbaps along with other vari-
able values) may be thought of as being assigned by the process. Oftentimes, a probabilistic
model or a model with SOme probabilistic components is assumed to represent the value
assignment process. As indicated earlier, the set of units is referred to as the universe. a.'1d
this set may have a firrite or an infinite membership of units. Furthermore, in some cases this
set exists only in concept~ however. we can describe the elements of the set without enu-
merating them, as was the case with the sample space, 'if, associated v,.ith a random exper~
iment presented in Chapter I. The set of values assigned or that may be assigned to a
specific variable becomes the population. for that variable.
Now, we illustrate with a few additional examples that reflect several universe struc-
tures and different aspects of observing processes and population data. First. continuing
with the power supplies, consider the production from one specific shift that consists of 300
units, each having some voltage output. To begin, consider a sample of 10 units selected
from these 300 units and tested with voltage measurements, as previously described for
sample units. The universe here is finite. and the set of voltage values for these universe
UIJits (the population) is thus finite. If our interest is simply to describe the sample results,
we have done that by enumeration of the resnlts, and both graphical and quantitative meth-
ods for further description are given in the following sections of this chapter. On the other
Process
Observation
Df~
Data
Unit
generation '--</
Or
/
/
/
/
/" o
'--/
/
/
/
/
/
/
/
/
/
Sample units I
Q)®®"'@
Figure 8--1 From process to data..
172 Chapter 8 in'JOduction to Statistics and Data D::-:scription
hand, if we wish (0 employ the sample results to draw statistical inference about the popu-
lation consisting of 300 voltage values, this is called an enumerative studyl and careful
attention must be given to the method of sample selection if valid inference is to be made
about the population. The key to much of what may be accomplished in such a study
involves either simple or more complex forms of rar.dom sampling or at least probability
saT1".pling. \VhUe these concepts will be developed further in the next chapter, it is noted at
this point that the application of simple random sampling from a finite universe results in
an equal inclusion probability for each unit of the universe. In the case of a finite universe,
where the sampling is done without replacemem and the sample size, usually denoted n. is
the same as the universe Si7.e, N, then this sample is called a census,
Suppose Our interest lies not in the voltage "talues for units produced in this specific
shift but rather in the process or process variable assignment model. Not only must great
care be given to sampling methods, we must also make assumptions about the stability of
the process and the structure of the process model during the period of sampling. In
essence, the universe unit variable values may be thought of as a realization of me process"
and a random sample from such a universe, measuring voltage values. is equivalent to a rau~
dom sample on the process or process model voltage variable or population. With our
example, we might assume that unit voltage. E, is distributed as N(u. aZ) during sample
selection. This is often called an analytic sruciy, and our objective might be, ior example, to
estimate J1. and/or cr. Once again, random sampling is to be rigorously defined in the fol~
lowing chapter.
Even though a universe sometimes exists only in concept, defined by a verbal descrip-
tion of the membership, it is generally useful to think carefully about the entire process of
observation activity and the conceptual universe. Consider an example where the ingredients
of concrete speci.r:lens have been specified to achieve high strength with early cure time. A
batch is formulated, five specimens are poured into cylindrical molds, and these are cured
according to test specifications. Following the cure, these test cylinders are subjected to lon~
gitudinalload until rupture, at which point the rupture strength is recorded. These test units,
with the resulting data, are considered sample data from the strength measurements associ~
ated with the universe units (each with an assigned population value) that might bave been,
but were never actually, produced. Again, as in the prior example, inference often relates to
the process or process variable assignment model. and model assumptions may become cru-
cial to attaining meaningful inference in this analytic study.
Finally, consider measurements taken on a fluid in order to measure some variable or
characteristic. Specimens are draw:> from well-mixed (we bope) fluid, with eacb specimen
placed in a specimen container. Some difficulty arises in universe identification. A conven-
tion here is to again view the universe as made up of all possible specimens that might have
been selected, Similarly. an alternate view may be taken that me sample values represent a
realization of the process value assignment model. Getting meaningful results from such an
analytic study once again requires close anention to sampling methods and to process or
process model assumptions.
Descriptive statistics is the branch of statistics that deals with organization, summa~
rizatioo. and presentation of data. Many of the techniques of deseriptive statistics bave been
in use for over 200 years, y;ith origins in surveys and census activity, Modem computer
technology, particularly computer graphics, has greatly expanded the field of descriptive
statistics in recent years. The techniques of descriptive statistics can be applied either to
entire finite populations or to samples and these methods and techniques are illustrated in
j
the following s""tions of this chapter. A wide selection of software is available, rdllging in
focus, sophisticatio~ and generality from simple spreadsbeet functions &'Uch as those found
in Microsoft Excel"", to the more-comprehensh'e but "user friendly" Minitab®, to large)
comprehensive, flexible systems such as SAS. Many other options are also available.
8-3 Graphical Preser.tation of Dam 173
8-2 DATA
Data are collected and stored electronically as well as by hUJrulD observation using tradi-
tional records or files. Formats differ depending on the observer or observational process
and reflect individual preference and ease of recording, In large~scale studies, where unit
identification exists~ data must often be obtained by stripping the required information
from several files and mergillg it into a file format suitable for statistical analysis.
A table or spreadsbeet-type format is often convenient, and it is also compatible with
most software analysis systems. Rows are typically assigned to units observed, and columns
present numerical or categorical data on one or mare variables. Furthermore, a column may
be assigned for a sequence or order index as well as for other unit identifying data. In the
case of t.~e power supply voltage measurements. no order or sequence was intended so that
any permutation of the 10 data elements is the same. In ather contexts. for example where
the index relates to the time of observation, the position of the elementin the sequence may
be quite important. A common practice in both situations described is to employ an index.
say i = 1,2; , .. , n, to serve as a unit identifier where n units are observed. Also, an alpha~
bedcal character is usually employed to represent a given variable value; thus if €i equals t.~e
value of the ith voltage measurement in the example, e~ = 5.10, e2 5.24, e3 = 5.14.. ."
e lC = 5.11. In a table format for these data there would be 10 rows (II if a headings row is
employed) and one or!\Vo columns, depending on whether the index is included and pre-
sented in a column.
'Where several variables and/or categories are to be associated with each unit. each of
these may be assigned a colUIDIl. as saoVffi in Table 8-1 (from Montgomer:y and Runger,
2003) which presents measurements of puJl strength, wire length, and die height made on
each of25 sampled units in a semiconductor manufacturing facility, In situations where cat~
egorical classification is involved, a columo is provided for each categorical variable and a
categor:y code or identifier is recorded for the unit. An example is shift of manufact1:..""e with
identifiers D, N, G for day, night, and graveyard shifts, respectively.
9.95 2 50
2 24.45 8 110
3 31.75 II 120
4 35.00 10 550
5 25.02 8 295
6 16.86 4 200
7 14.38 2 375
8 9.60 2 52
9 24.35 9 100
10 27.50 8 300
11 17.08 4 412
12 37.00 11 400
13 41.95 12 500
14 11.66 2 360
15 21.65 4 205
16 17.89 4 400
17 69.00 20 600
18 10.30 ·585
19 34.93 10 .540
20 4<5.59 15 250
21 44.&8 15 290
22 54.12 16 510
23 56.63 17 590
24 22.13 6 100
25 21.15 5 400
in the data set, dot plots are effective displays. Figure 8-2 shows the univariate dot plots or
marginal dot plo", for each of the three variables with dall! presented in Table 8-L These
plots were produced using Minitab".
Where we wish to jointly display results for two variables, the bivariate equivalent of
the dot plot is called a scatter plDt. We construct a simple rectangular coordinate graph,
assigning the horizontal axis to one of the variables and the vertical ""is to the other. Each
observation is then plotted as a point in this plane. Figure 8~3 presents scatter plots for pull
strength vs. wire length and for pull strength vs. die height for the datl! in Table 8-L In order
to accommodate data pairs that are identical, and thus that fall at the same point on the
plane, one convention is to employ alphabet characters as plot symbols; so A is the dis-
played plot point where one data point falls at a specific point on the plane. B is the dis-
played plot point if two fall at a specific point of the plane. etc. Another approach for this.
which is useful where values are close but not identicaL is to assign a randomly generated,
small, positive or negative quantity sometimes called ji..l1er to one or both variables in order
to make the plo", better displays. Wbile scatter plots show the region of the plane wbere the
data points fall, as well as the data density associated with this region, they also suggest
possible association between the variables. Finally, we n.ote that the usefulness of these
plots is not limited to small data sets.
8-3 Graphical Presentation of Data 175
.. ~ !~ .. •• , . .,
10 20 30 40 50 60 70
Pull strength
5 10 15 20
Wire length
T •• t· .., ! • t • •• T
Figure 8-2 Dot plots for pull strength, wire length, and die height.
70 70
60 60
" 40
50
" ~
50 ..
. . .. .'
~
c c
~ 30 ~ 40
"5 "5 30
~
20
10 ..'. '
~
20 .. ' :
0 10
0 10 20 o 100 200 300 400 500 600
Wire length Die height
Figure 8-3 Seatter plots for pull strength vs. wire length and for pull strength vs. die height (from
Minitab'"l.
To extend the dimensionality and graphically display the joint data pattern for three
variables, a three-dimensional scatter plot Dlay be employed, as illustrated in Figure 8-4 for
the data in Table 8-1. Another option for a display, not illustrated here, is the bubble plot.
whicb is presented in two dimensions, with the third variable reflected in the dot (now
called bubble) diameter that is assigned to be proportional to the magnitude of the third
variable. As was the case with scatter plots, these plots also suggest possible associations
between the variables involved.
~ttn-~~~-l-1~o
406'°0
O'---'-4'"-8.L.-':'12~---;;;,-
8
__?;; ~.e.\~
200300 «,\
100
12 16 20 0 <:)\0
Wire length
FigureS-4 Three~ilimensiona1 plots for pull strength, wire length, and die height.
Table,g..2 Bursting Strength in Pounds per Square Inch for 100 Glass, I-Liter, Nonreturnable Soft
Drink Bottles
265 197 346 280 265 200 221 265 261 278
205 286 317 242 254 235 176 262 248 25G
263 274 242 260 281 246 248 271 260 265
3rJ7 243 258 321 294 328 263 245 274 270
220 231 276 228 223 296 231 301 337 298
268 267 300 250 260 276 334 280 250 257
260 281 208 299 300 264 230 274 278 210
234 265 187 258 235 269 265 253 254 280
299 214 264 267 283 235 272 287 274 269
215 318 271 293 277 290 283 258 275 251
Or "what percentage of the bottles burst below 230 psi?" are not easy to answer when the
data are presented in this form,
A frequeney distribution is a more useful summary of data than the simple enumera-
tion given in 1able 8-2. To construct a frequency distribution, we must divide the range of
the data into intervals, which are usually called class intervals. If possible, the class inter-
vals should be of equal width, to eohance the visual infonnation in the frequency distribu-
tion, Some judgment must be used in selecting the number of class intervals in order ro give
a reasonable display. The number of class intervals used depends on the number of obser-
vations and the amount of scatter or dispersion in the data. A frequency distribution that
uses either too few or too many class intervals will not be very informative. \Ve generally
find that between 5 and 20' intervals is satisfactory in most cases, llIld that the number of
class intervals should increase with n. Choosing a number of class intervals approximately
equal to the square root of the nwnber of observations often works well in practice.
A frequency distribution for the bursting s!l'ength data in Table 8-2 is abown in Table
8-3. Since the data set contains 100 observations, we suspect that about ,fWo = 10 class
intervals will give a satisfactory frequency distribution. The largest and smallest data val~
ues are 346 and 176, respectively, so the class intervals must cover at least 346 - 176 = 170
psi units on the scale. If we want the lower limit for the first interval to begin slightly below
tile smallest data value and the upper limit for the last cell to be slightly abeve the largest
data value, then we might start the frequency distribution at 170 and end it at 350. This is
an interval of 180 psi units. Xine class intervals, each of width 20 psi, gives a reasonable
frequency distribution, ani! the frequency distribution in Table 8·3 is thus based on nine
class intervals.
8-3 Graplrical Presentation of Data 177
Table 8-3 Frequency Distribution for the Bursting Strength Data in Table 8-2
The fourth column in Table 8-3 contains the relative frequency distribution. The rela-
tive frequencies arc found by dividing the observed frequency in each class interval by the
total number of observations. The last column in Table 8-3 expresses the relativ~ frequen-
cies on a cumulative basis. Frequency distributions are often easier to interpret than tables
of data. For example: from Table 8-3 it is very easy to see that most of the bottles burst
between 230 and 290 psi, and that 13% of the bottles burst below 230 psi.
It is also helpful to present the frequency distribution in graphical form, as shown in
Fig. 8-5. Such a display is called a histogram. To draw a histogram, use the horizontal axis
to represent the measurement scale and draw the boundaries of the class intervals. The ver-
tical axis represents the frequency (or relative frequency) scale. If the class intervals are of
equal width, then the heights of the rectangles drawn on the histogram are proportional to
the frequencies. If the class intervals are of unequal width, then it is customary to draw rec-
tangles whose areas are proportional to the frequencies. In this case, the result is called a
.
40
,-
30
i;'
c
•o r-
~ 20
u.
-
10 r-
-
,--I
170190210230250270290310330350
n
Bursting strength (psi)
Figure 8-5 Histogram of bursting strength for 100 glass, I-liter, nonreturnable soft drink bottles.
178 Chapter 8 Introductioc to Statistics and Data Description
density histogram, For a histogram displaying relative frequency nn the vertical axis. the
rectangle height' are calculated as
'gh
rectangIe hel& t:::::
class relative frequency.,_
class width
When we find there are multiple empty class intervals after grouping data into equal-width
intervals, one option is to merge the empty intervals with contiguous interVals. thus creat~
ing some wider intervals. The density histogram resulting from this may produce a more
attractive display. However, histograms are easier to interpret when the class intervals are
of equal width. The histogram provides a visual impression of the shape of the distribution
of the measurements, as wen as information about centering and the scatter or dispersion
of tho data.
In passing from the original dam to either a frequency distribution or a histogram. a cer~
tain amount of information has been lost in that we no longer have the individual observa-
tions, On the other hand, this information loss is small compared to the ease of
interpretation gamed in using the frequency distnoution and histogram, In cases where the
data assume only a few distinct values, a dot plot is perhaps a better graphical display,
Vlhere observed data are of a discrete nature, sucb as is found in counting processes~
then two choices are available for constrUcting a histogram, One option is to center the rec~
tangles on the integers reflected in the count data, and the other is to collapse the rectangle
into a vertical line flaced directly over these integers. In both cases, the height of the rec-
tangle or the length of the line is either the frequency or relative frequency of the occurrence
of the value in question.
In summary, the histogram is a very useful gr1ljOhic display, A histogram can give the
decision maker a good understanding of the data and is very useful in disflaying the shape,
location, and variability of the data. However, the histOgram does not allow individual data
points to be identified, because all observations falling in a cell are indistinguishable.
The resulting stem-ana-leaf plot is presented in Fig. 8-6. Inspection of this display immediately
reveals that most of ti:.e bursting strengths lie bet'Neen 220 and 330 psi. and that the central value is
soroc'Nhere bet\llccn 260 and 270 psi. Furtheun.ore, the but'Sting strengths are distributed approxi~
mately symmetrically about the central value. Therefore, the stem-and~leaf plot., like the histogram.
ci1OV1S US to determine quickly some important fuatures of the data that were not immediately obvi~
o";!s in the original display. Table 8-2, Note that here the original cumbers. arc not lost, as occurs in a
3-3 Graphical Presentation of Data 179
Stem Leaf
17 6
18 7
19 7
20 0,5,8 3
21 0,4,5 3
22 1,0,8,3 4
23 5,1,1,4.5,5 6
2A 2,8,2,6,8,3,5 7
25 4,0,8,0,0,7,8,3,4,8-1 11
26 5,5,5,1,2,3,0,O,5,3,8,7,O,O,~,5,9,5,4,7,9 21
27 8,4, 1,4,0,6,6,4,8,2,4, 1,7,5 14
28 0,6,1 ,0,1,0,0,3,7,3 10
29 4,6,8.9~9,3.0 7
30 7,1,0,8 4
31 7,8 2
32 1,8 2
33 7,4 2
34 6 1
100
Figure 8--6 Stem~and-leaf plot for the borJe-bursting-strength data in Table 8~2.
histogram. Sometimes, in order to assist in finding percentiles. we order the leaves by magnitude, pro-
ducing a:: ordered stem~and-leaf plot, as in Fig. 8-7. For instance, since n =. 100 is an even number,
the median, or '''midd1e'' observation (see Section 8-4.1), is the average of the tviO observations with
ranl;s 50 and 51, or
Stem Leaf
17 6 1
18 7 1
19 1
20 0,5.8 3
21 0,4,5 3
0,1,3,8
22
23 1,1.4,5,5,5 "6
24 2.2,3,5,6,8,8 7
25 0,0,0,1,3.4,4,7,8,8,8 11
26 0,0,0.0,1,2,3,3,4,4,5,5,5,5,5.5,7,7,8.9,9 21
27 0,1,1,2,4,4,4,4,5,6,6.1,8,8 14
28 0,0,0,0,1,1,3,3,6,7 10
29 0,3,4,6,8,9,9 7
30 Q,l,7,& 4
31 7.8 2
32 1,8 2
33 4,7 2
34 6 1
100
Figure g~1 Ordered stcm~and~leaf plot for the bott1e~bUX5ting-streugth data.
percentile or the median) Q2 ~ i. A Une at either end extends to the extreme values. These
lines, sometimes called wbiskers, may extend only to the 10th and 90th percentiles or the
5th and 95th percentiles in large data sets. Some authors refer to the box plot as the box"
and-whisker plot. Figure 8-8 presents the box plot forthe bottle-bursting-strength data. This
box plot indicates that the distribution of bursting strengths is fairly symmetric around the
central value, because the left and right whiskers and the lengths of the left and right boxes
around the median are about the sarne.
The box plot is useful in comparing two or more samples. To illustrate, consider the
data in Table 8-4. The data, taken from Messina (1987), represent viscosity readings on
three different mixtures of raw material used on a manufacturing line. One of the objectives
of the study that Messina discusses is to compare the three mixtures. Figure 8-9 presents the
box plots for the viscosity data. This display permits easy interpretation of the data. Mix-
ture I has higher viscosity than mixture 2, and mixture 2 has higher viscosity than mixture
3, The distribution of viSCOsity is not symmetric, and the maximum -viscosity reading from
mixture 3 seems unusually large in comparison to the other readings. This observation may
be an outlier. and it possibly warrants further exantination and analysis.
I
I : I
I I
175 200 225 250 275 300 325 350
Figure 8~~ Box plot for the bottic-bursting-streng'J> data.
8-3 Graphical Presentation of Data 181
27.0
26.67
25.8 25.70
25.19
m 24.7
".g-"
24.62 24.67
24.32
0
rn 23.68
~ 23.5
23.29
'"8
-;;;
22.62
22.3 22.37
5 22.02
21.88
21.49
21.2
21.08
20.33
20.0
2 3
Mixture
Figure 8-9 Box plots for the mixture-viscosity data in Table 8-4.
500 100
400 80
sao 60 1:
1: e
8 200 80
-40 a.
10e 20
0 0
737 m 757 767 717 747 MD·11
oC\..\.1! on metal parts used in a structural component of an automobile doorfrarne. Notice bO\V
the Pareto chart highlights the relatively few types of defects thaI are responsible for most of
the observed defects in the part. The Pareto chart is an imponanl parr of a quality-improve-
ment program because it allows management and engineering to focus attention on the most
critical defects in a product or process. Once these critical defects are identified, corrective
actions to reduce or eliminate these defects must be deVeloped and implemented. This is eas-
ier to do, however, wben we are sure that we are attacking iii legitimate problem: it is much
easier to reduce or eliminate frequently occ~ng defect'> than rare ones.
81 I
-----_. __ . iOQ
.,
'if
75 .3" e
ll'
~ ~
Q
.!!l
m
'0
50 ~
"
cr
'0 Q
~
Q
.0
E so j
~
g
z
3°1 21 L25 0
Q
2°l
l ;;;
1lla:
10
~6 5 5 4 4
O-~.-- J
Out of Missing !
Parts net
contour hoi(JS!slcts igreased
Pans Outet Parts r,ot Oth.,
ur.der·!rlmmcd sequonce dro.;rred
Defect typo
J !
50
c 2 4 6 8 10 12 14 16 18 20 22
Hour
Figure 8-12 Summer peak doy hourly wow.tt (KW) demand data for an office building.
l
184 Chapter 8 Introduction to Statistics and Data Description
we will refer to the arithmetic mean as the sample mean. If the observations:in a sample of
size n are x;. x 2• •• '. X'I' then the sample mean is
f::= X1 +XZ +,,,+xn
n
(8-1)
}.>
- i~i _ 26, 406 "64 06
X
lOa -J:Oil - ..
From examination of Fig. 8-5, it seems that the sample mean 264,06 psi is a "typicall1 value
of bursting strength. since it occurs near the middle of the data, wbere the observations are
concentrated. However, this impression can be lIlisleading. Snppose that the histogram
looked like Fig, 8~13. The mean of these data is still a measure of central tendencYI but it
does not necessarily imply that most of the observations are concentrated around it. In gen-
eral, if we think ofllie observations as having unit mass, the sample mean is just the center
of mass of the data. This implies that the histogram will just exactly balance if it is sup-
ported at the sample mean.
x
The sample mean represents the average value of all the observations in the sample.
\Ve can also think of calculating the average value of all the observations in a finite popu~
larion. This average is called the population mean, and as we here seen in previous chap-
ters, it is denoted by the Greek letter J1.. \,nen there are a finite number of possible
observations (say Iv) in the population, then the population mean is
_ 1', (8-2)
/1- N I
'\;'N
where l'x = .L<i""l Xi is the finite population total for the population.
In the followmg chapters dealing with statistical inference. we will present methods for
making inferences about the population mean that are based on the sample mean. For exam-
ple, we will nse the sample mean as a point estimate of fl.,
Another measure of central tendency is the median, or;he point at which the sample is
divided into two equal halves, Let X~l)' XC'l)l ••• , ~It) denote a sample arranged in increasing
order of magnitude; that iSt x(li denotes the smallest observation, X(2) denotes the second
smallest observation, ... , and x~l!) denotes the largest observation. Then the median is
defined mat'Iematically as
" n odd,
t«(r,+I)/2)
x~ (''') +2(,;2)+1) 11 even. (8-3)
The median bas the advantage that it is not influenced very much by extreme values. For
example, suppose that the sample observations are
1,3,4,2,7,6. and 8.
The sample mean is 4.43, and the sample median is 4. Both quantities give a reasonable
measure of the central tendency of the data. Now suppose that the next~to-Iast observation
is changed, so that the data are
1,3,4,2,7,2519, and 8.
For these dara. the sample mean is 363.43. Clearly, in this case the sample mean does not
tell us very much about the central tendency of most of the data. The median, however, is
still 4, and this is probably a much more meaningful measure of central tendency for the
majority of the observations.
Just as;; is the middle value in a sample, there is a middle value in the population. We
define fl as the median of the population; that is, fl is a value of the associated random vari-
able sucb that half the population lies below i1 and half lies above.
The mode is the observation that occurs most frequently in the sample. For example,
the mode of the sample data
2,4,6, 2) 5~ 6. 2, 9, 4, 5, 2, and 1
is 2, since it occurs four times, and no other value occurs as often. There may be more than
one mode.
If the data are symmetric~ then the mean and median coincide. If. in addition~ the data
bave only one mode (we say the data are unimoda1)~ then the mean, median; and mode may
all coincide. If the data are skewed (asymmetric, with along tail to one side), then the mean,
median, and mode will not coincide. Usually we find t.'13t mode < median < mean if the dis-
tribution is skewed to t'Ie righ~ while mode> median> mean if the distribution is skewed
to the left (see to Fig. -8-14).
The pistribution of the sample mean is well-known and relatively easy to work with.
Furthermore. the sample mean is usually more stable than the sample median, in the sense
that it does not vary as much from samplp to sample. Consequently. many analytical ,"tatis-
tical techniques use the sample mean. However. the median and mode may also be helpful
descriptve measures.
(8-5)
The formula for S;r:x presented in equation 8-5 requires only One computational pass; but
care must again be taken to keep enough decimals to prevent round-off error.
To see how the sample variance measures dispersion or variability, refer to l'1g. 8-16
which shows the deviations Xi -:x for the second sample of six bottle-bursting strengths. The
greater the amount of variability in the bursting~strength data., the larger in absolute mag-
nitude some of the deviations'x; -x, Since the deviations Xi - iwill always sum to zero, we
must use a measure of variability that changes the negative devi.ations to nonnegative quan-
tities. Squaring the deviations is the approach used in the sample variance. Consequently,
if ,? is smaR then there is relatively little variability in the data, but if SZ is large, the vari~
ability is relatively large.
The units of measurements for the sample variance are the square of the original units
of the variable. Thus. if,X is measured in pounds per square inch (psi). the units for the sam~
pIe variance are (pSi)2.
~~;.pi~;~1'
We wi11 calculate tte sample v'ariance of the bottle~bursting stre::J.gths for the second sample in F1g.
8-15. The deviations Xl -.i for this sample are shown in Fig. 8-10.
o o o o 0 o
, , , "1" ~.~~:"-:~_---o'_
180 200 220 240 r~-2W--2BO 300 320
x,
o
x,o x,
o
Xo Xs
o 0
x,
o
180 320
Figure 8-16 How the sample variance measures variability thro'.lgh ':he deviations Xi -;t:.
Observations
We may also calculate SkY: from the forn::w.lation given in equation 8-5, so that
II " 1r 01 \2
s2. = S;c __
lx, --llx, . ~ 367.534- "J488) ;6
1",1 n. i,.,1 j
1
:502{psi)2.
'
n-I n-I 5
If we calculate the sample variance of the buISting strength for the Sample I values, we
find that;' 158 (psi)'. This is considerably smaller than the sample variance nf Sample 2,
confirming OUI initial. impression that Sample I has less variability than Sample 2.
Because;' is expressed in the square of the original units~ it is not easy to interpret
Fmthermore, variability is a more difficult and unfamiliar concept than location or central
tendency, However, we can solve the "curse of diroensioo.ality'~ by working with the (posi-
tive) square root of the vari.ance~ $, called tlle sample standard deviation.. This gives a meas-
ure of dispersion e.xpressed in the same units as the original variable.
The sample standard deviation of the bott1e~bursting strengths for the Sample 2 bottles in Rumple 8~2
and Fig, 8-15 is
Compute the sz.mple: variance and sample standard deviation of the bottle-bursting-strength data in
Table 8-2. Note that
'00 lC.
2:>:;. 7,074,258.00 LX,
,., :26,406.
Consequently,
S=; 7,074,258.00 - (26,406)'1;00; 101,489.85 and ,2; 101.489,85199 = 1025.15 psi',
so that the sample suu:dard deviation is
\\'hen the population is finite and consists of N values we may define the population Yari~
ance as
.v
L(x, -,ul
0'2 = 1=:1___ _ (8-6)
N '
which is simply the mean of the average squared departures of the data values from the pop-
ulation mean. A closely related quantity, if', is also sometimes called the population vari-
ance and is defined as
N
---0'-,
,
(8-7)
N-I
Obviously. as N getS large, a: -1 (f1., and oftentimes the use of iT simplifies some of the
algebraic formulation presented in Chapters 9 and 10, \\'bere severa! populations are to be
observed, a SUbscript may be employed to identify the population characteristics and
descriptive measures, e,g., f.Lx, O'~, s!. x
etc., if the variable is being described,
We noted that the sample mean may be used to make inferences about the population
mean. Similarly. the sample variance may be used to rn.ake inferences about the population
variance, We observe that the divisor for the sample variance,?t is the sample size minus
I, (n - I). If we actually knew the true value of the population mean!l, then we could define
the sample variance a,.;; the average squared deviation of the sample observations about J.l.
In practice, the value of J.l is almost never kno'W'n, and so the sum of the squared deviations
x
about the sample average must be used instead_ However! the observations Xi tend to be
closer to their average i than to the population mean j..L, so to compensate for this we use as
a divisor n 1 rather than n.
Another way to think about this is to consider the sample variance s2 a,.., being based On
II I degrees affreedom. The term degrees offreedom results from the fact thac the n devi-
ations XI - X. X:!: - x, .. ,' XII - x always sum to zero, so specifying the values of any 11 1 of
-<
these quantities automatically determines the remaining one. Thus. only n - 1 of the n de\1.-
ations Xi - X are independent.
Another useful measure of dispersion is the sample range
(8-8)
The sample range is very simple to compute, but it ignores all the information in the sam-
ple between the smallest and largest observations, For small sample sizes, say II :;; 10, this
;nformation loss is not too serious in some situations. The range traditionally has had wide-
8-4 ~wr.erical Des..-ription of Data 189
spread application in statistical quality control, where sample sizes of 4 or 5 are common
and computational simplicity is a major consideration; however, that advantage has been
largely diminished by the widespread use of electronic measurement, and data storage and
analysis systems, and as we will later see, the sample variance (or standard deviation) pro~
vides a "better" measure of variability. We will briefly discuss the use of the range in sta-
tistical quality-control problems in Chapter 17.
Cak'::ate the ranges of the tv,.'o samples 0: bottle-bur:sting~strength data from Section 8-4. shown in
Fig. 8-15. For the first sample, we find that
}(, 265-230=35.
whereas for the second sample
}(, =305 -190= 115.
Note that the range of the second sample is much larger than the range of me first, implying that the
second sample has greater variability than the first.
(8-9)
The coefficient of variation is useful when comparing the variability of two or more data
sets that differ considerably in the magnitude of the observations. For example, the coeffi-
cient of variation might be useful i1. compa.."ing the variability of daiiy electricity usage
within samples of single-family residences in Atlanta, Georgia, and Butte, Montana, dnring
July.
where. in the case of finite populations. the kth central moment is defined as
N
2:(xi-4
(8-11)
N
As discussed earlier, skewness reflect" the degree of symmetry about the mean; and nega*
tive skew results from an asymmetric tail toward smaller values of the variable, while posi-
tive skew results from an asymmetric tail extending toward the larger values of the variable,
190 Chapter 8 Introduction to Statistics and Data Description
Symmetric variables, such as those described by the normal and uniform distributions. hav~
skewness equal zero. The exponential distribution, for example, has skewness /33 = 2.
Kurtosis describes the relative peakedness of a distribution as compared to a normal
distribution, where a negative value is associated with a relatively flat distribution and a
positive value is associated with relatively peaked distributions. For example, the kurtosis
measure for a unifonn distribution is -1.2. while for a nonnal variable. the kurtosis is zero.
If the data being analyzed represent measurement of a variable made On sample units.
the sample estimates of f3, and /3, are as shown in equations 8-12 and 8-13. These values
may he calculated from Excel~ worksheet functions using the SKEW and KURT functions.
n
skewness /3, = . =~-.,...---; n>2.
(n-l}(n-Z)
where Sa is as shown in equation 8-5 and Syy is similarly defined for the y variable, replac-
ing x with yin equation 8-5, while
/11 \J.ln\/Jt\
1--n Ix,).: Iy, i.
11-
(8-17)
respectively.
A similar situation arises where original data have been either lost or destroyed but the
grouped data have been preserv'ed. In such cases, we can approximate the important data
moments by using a convention, which assigns each data value the value representing the
midpoint of the class interval into which the observations were classified, so that all sam-
ple values falling in a particular .interval are as.qigned the same value. TIris may be done eas~
ily, and the resulting approximate mean and variance values are as follows in equations 8-18
and 8.19.lfmj denotes the midpoint ofthejth class interval and there are c class intervals,
then the sample mean and sample variance are approximately .
if m if m
j j
)=1
j j
(8-18)
n
and
C
~ im'
L..iJJ
(~ fm, J'
I'
--I'n~J.
C
~~~iiii1!~.f~;.
To illustrate the use of equatio!ls 8-18 and 8~19 we compute tb.e mean and variance of bursting
strength for the data in the :frequency distribution of Table 8~3. Note that there are c 9 class inter-
vals, and ilia, m, ISO,f, =2, m,= 200,!i =4, m, =220,1, =7, m, =240.1. = 13, m, =2f1J,f, =32,
m, = 2lI0,f, = 24, m, = 300'/1 = 11. m, = 320,fs = 4, m, = 340, andj, = 3, Thus,
9
L,JjmJ
- j-' 26,460 2'"
x=---=---:m: \)"+.
60 pst.
n 100
and
'When the data are grouped in class intervals, it is also possible to approximate the
median and mode. The median is approximately
(8-20)
where LM is the lower limit of the class interval containing the median (called the median
c1ass),fM is the frequency in the median class, T is the total of all frequencies in the class
intervals preceding the median class, and A is the width of the median class. The model say
MO, is approximately
MO = L,w +
, a+b
(_a_)c., (8-21)
where L MO is the lower limit of the modal class (the class interVal with. the greatest fre-
quency), a is the absolute value of the difference in frequency between the modal class and
the preceding class, b is the absolute value of the difference in frequency between the
modal class and the following class, and C. is the width of the modal class,
8-5 SUMMARY
This chapter has provided an introduction to the field of statistics, including the notions of
process, universe, population, sampling, and sample results called data. Furthermore, a
variety of commonly used displays have been described and illustrated. These were the dot
plot, frequeru:y distribution, histogram, stem-and-Ieaf plot, Pareto cbart, box plot, scatter
plot, and time plot.
We have also introduced quantitative measures for summarizing data. The mean,
median and mode describe cen.tral tendency. or location, while the variance, standard devi~
ation, range, and interquartile range describe dispersion Or spread in the data. Furtbermor~
measures of skew and kurtosis were presented to describe asymmetry and peakedness,
respecti¥ely. We also presented the correlation coefficient to describe the strength of linear
assoctztion between two variables. Subsequent chapters will focus on utilizing sample
results to draw inferences about the process or process model or about the universe.
8-6 Exc::"Cises 193
8.6 EXERCISES
8~1. Theshelf life of a high~speed photographic film is 86.6 94.1 93.1 89.4 973 83.7
being inves~ated by the manufacturer. The fo!1owing 91.2 97.8 94.6 88.6 96.8 82.9
datl are available, 86.1 93.1 96.3 84.1 94.4 87.3
90.4 86.4 94.7 82.6 96.1 86.4
Life Life Life Life 89.1 87.6 9U 83.1 98.0 84.5
(days) (days) (days) (days)
3 6 4 7 6 7
Construct a histogrnm and comment on the properties 4 7 8 2 4
of tile data. 2 9 4 6 4 8
8-2. The percentage of cotto:!. in a material used to 5 10 10 9 13 7
manufacture men's shirts is givc., below. Construct a 6 14 10 12 3
histogram for t.'lC data Comment on the properties of 10 13 8 7 10 6
the data. 5 10 12 9 2 7
4 9 4 16 5 8
34.2 33.6 33.8 34.7 37.8 32.6 35.8 34.6
3 5 11 .7 4
33.1 34.7 34.2 33.6 36.6 33.1 37.6 33.6
II 10 14 13 10 12
34.5 35.0 33.4 32.5 35.4 34.6 37.3 34.1
9 3 2 3 4 6
35.6 35.4 34.7 34.1 34.6 35.9 34.6 34.7
2 2 8 13 2 17
34.3 36.2 34.6 35.1 33.8 34.7 35.5 35.7
7 4 6 3 2 5
35.1 36.8 35.2 36.8 37.1 33.6 32.8 36.8
8 6 10 7 6 10
34.7 35.1 35.0 37.9 34.0 32.9 32.1 34.3
4 4 8 3 ~
8
33.6 35.3 34.9 36.4 34.1 33.5 34.5 32.7 '10
2 10 6 2 9
6 8 4 9 S J1
8--3. The :following data represent the yield On 90 con~ 5 7 6 4 14 7
secutive batches of ceramic substrate to which l! metal 4 14 15 :3 6 2
coating has been applied by a vapor·deposition 3 13 4 3 4 8
proeess. Const.ract a histogra::n for ::hese data ane 2 12 7 6 4 10
comment on the properties of the data,
8 5 5 5 8 7
10 4 3 10 7 4
94.1 87.3 94.1 92.4 84.6 85.4
9 6 2 6 9 3
93.2 84.1 92.1 90.6 &3,6 86.6
II 5 6 7 2 6
90.6 90.1 96.4 89.1 85.4 91.7
91.4 95.2 88.2 88.8 89.7 87.5
88.2 8ciJ 86.4 86.4 87.6 84.2 8 ..5. Consider the shelf life data in Exercise 8~ 1, Com-
86.1 94.3 85.0 85.1 85.1 85.1 pute the sample mean. sample va.."'iance. sad sample
95.1 93.2 84.9 84.0 89.6 905 standard deviation,
90.Q 86.7 87.3 93.7 90.0 95.6 8~6. Consider ::he cotton percentage da':a in Exercise
92.4 83.0 89.6 87.7 90.1 88.3 3-2. Find the sample =ean. sa=ple variar.c~. sa.."'!1ple
87.3 95.3 90.3 90.6 94.3 84.1 sr.:andard deviation, samp!e median, ar:.d sample mode,
l
194 Chapter 8 Introduction to Statistics and Data Description
8~7. Consider tl:e yiClddata in Exercise 8-3. Calcula!e (a) CoIilltruct a ste:n-and-leaf pIal
the sample mean. samp~e variance, and sample stan- (b) Construct a frequency distribution and bi<;togram.
dard deviation. (c) Calculate the sample mean, sample va.....iance. and
S-S. An article in Computers r:ma' Industrial Engineer- sample standard deviation.
ing (2001; p. .51) describes the time-to-failure data (in (d) Find the sample median and sample mode.
hours; for jet engines. Son::.e of the data are repro- (e) Dete:J:.::.ine:he ske\1lIless and kurtosis measures,
duced below.
8~11. Consider the shelf-life data in Exercise 8-1.
Cor.struct a stem~and-leaf plot for these data. Con.,
Failure Failure
struct an ordered stem-and-leaf piOL Use this plot to
Engine # Time Engine # Tlille
find the 65th and 95th percentiles,
I 150 14 171
&.12. Consider the cotton percentage data in Exercise
2 291 15 197 8-2,
3 93 16 200
(a) Cons:ruC! a stem-and-leaf plot.
4 53 17 262
5 2 [3 255 (b) Calculate the sample mean, sample va:..ance, and
286 sample standard deviatiou.
6 65 19
7 183 20 206 (c) Cons-r..::ct an ordered stem~and-leaf plot.
8 144 21 179 (d) Fmd the median and the fin;t and third quartiles.
9 223 22 232 (e) Fmd the interquartile range.
10 !97 23 165 (f) Detcmirle the ske\\'ness a;)d kurtosis measures.
:1 187 24 155 8~13. Consider ':he yield data in Exercise 8-3,
12 197 25 203
(a) Construct an ordered stcm-and-leaf plot.
13 213
(b) Find the median and the first a."ld third quartiles.
(c) Calculate the inlerquartile range.
(a) Construct a frequcr.cy distn'bution and his.togra.'1l
for these data. 8~14. Construct a box plot for the shelf-life data in
(b) Calculate the sample mean. sample median, sam~ Exercise 8-L Interpret the data using this plot.
pIe variance, and sa.r.:lple standard deviation. 8~15. Construe!. a box plot for the cotton perce:ttage
8-9. For the time·to.-failure data in Exercise 8-8, sup- data in Exercise 8-2. Interpret the data using this plot,
pose the ffth o'::>scrvation (2 hours) is discarded. Con- 8~16. Construct a box plot for the yield data in Exer-
struct a frequency distributioa and a histogram for the cise 8-3. Compare it to the histogram (Exercise 8~3)
remaining data, and calculate the sample mean, sa.'"n- and the stem~and-le3f plot (Exercise 8-13). btc:rpret
pIe median, sample variance, and sample standard the data.
deviation. Compare ':he results with ::hose o:.tained LTl
Exercise 8-8. Wbat impact has removal of this obser- 8~17. An article in the Electrical Manufacturing &
vation had on ':he s.um.ma.ry statistics? coa Winding Conference Proceeding:; (1995, p. 829)
presents the results for the number of returned ship-
8-10. An article in Tecnnol7U!rrics (Vol. 19, 1977, p.
ments for arecord-of~:he-month club. The company is
425) presents the followi:lg data on motor fuel octane interested in the reason for a returned shipment The
ratings of 5eY'eral blends of gasoline:
result') are shown below. Construct a Pareto chart and
88.5,87.7,83.4,86,7,87.5,91.5,88.6, 100.3, interpret the data.
95.6,93.3,94,7,91,1,91.0,94.2, 87.8, 89.9,
88.3,87,6,84.3,86.7,88,2, 90,8,88.3,98,8,
Reason Number of CustOt:l.efS
94.2,92.7,93,2,91.0,90.3,93,4,88.5,90.1,
89.2,83.3,85.3,37.9,88.6,90.9,89.0,96.1, Refused 195,000
93.3,91.8,92,3,90.4,90.1, 93,0, 88,7, 89,9, Wrong selection 50,000
89,8,89.6,87.4,88.4,83,9,91.2,89.3,94.4, Wrong answer 68,000
92,7,91,8,91.6,90,4,91.1,92.6,89,8,90,6,
Canceled 5,000
91.1,90.4,89.3,89.7,90.3,91.6,90.5,93.7,
Other 15,000
')2.7,92,2,92,2,91.2,91.0,92.2,90.0,90,7.
r 8~18. The following table contains the frequency of
&-.6 Exercise'S
occurrence of final letters in an adcle in the Atlcmta II, where a and b are nonzero constants. Find the rela-
Journal. Construct a histogram from these data. Do tionship between x and y, and between Sx and Sy'
any of the numerical descripto;:s in this :.:hapter have
any meatiliig for these dam? 8~2S. Consider the quantity ",~ L,-..l (Xi _a)2. For what
value of a is '.his quantity m.i.nimized?
a 12 n 19
8~26. The Trimmed l\>lean. Suppose :hat the data are
b II 0 13
arranged in ir.creasing order, L~% of the observations
c Jl p 1
removed from each end, and the sample mean of L.1.e
d 20 q 0
remaining numbers calculated. The resulting q::.antit,Y
e 25 r 15 is c.alled a trimmed mean. The trimmed mean gener-
f 13 s 18 ally lies between the sample mean i and t,i:e sample
g 12 20 median i: (why?).
h 12 u 0 Ca) Calculate the 10% trimmed mean for the yield
8 v 0 da.ta in Exercise S-3.
j 0 w 41 (b) Calculate the 20% trimmed mean for the yield
k 2 x 0 data in Exercise 8-3 and compare it with the quan~
11 Y 15 tit)' found in part (2.).
12 z
m
(e) Determine the sample mean, median, and mode (h) Determine the skewness and kurtosis measures for
for kwh, kw, and sq. ft. kwh, kw, and sq. ft., and compare to the measures
(f) Determine the sample standard deviation for kwh, for a normal distribution.
kw, and sq. ft. (i) Determine the simple (pearson) correlation
(g) Determine the first and third quartilcs for kwh, kw, between kw and sq.ft. and between kw and kwh.
and sq. ft. Interpret.
l
Chapter 9
Definition
X" X" ... , X, is a random sample of size n if (a) theXs are independent random variables,
and (b) eveq observation Xi bas the same probability distribution.
To illustrate this defmition. suppose that we are investigating the bursting strength of
glass, I-liter soft drink bottles, and that bursting strength in the population of bottles is nor-
mally distributed. Then we would expect each of the observations on b1l!Sting strength X"
~, .•. , XII in a random. sample ofn bottles to be independent random variables v:.1th exactly
the same normal distribution.
It is not always easy to obtain a random sample. Sometimes we may use tables of uni~
form random numbers, At otber times, the engineer or scientist cannot easily use formal
procedures to help ensure randomness and must rely on other selection methods. Ajudg-
ment sample is one chosen from the population by the objective judgment of an indivi~ual.
198
9-1 Random Samples 199
Since the accuracy and statistical behavior of judgment samples cannot be described, they
should be avoided.
Suppose we ",ish to take a random sample of five batches of raw material out of 25 available hatches.
We may n1l!llber the batches 'With the integers 1 to 25. Now, using Table XV of the Appendix, arbi~
tra."ily choose a row and colum."l as a st.arting point Read do';O.'D the chosen column, obtaining rNO dig-
its at a time, until five acceptable numbers are found (an acceptable number lies be!ween 1 and 25),
To illustrate, suppose the above process gives t:s a sequence ofnumben5 that rea.ds 37, ~8, 55,02.17,
61.70.43,21.82.13.13,60.25. The bold numbe;s spec:Iy which batches of raw material arc to be
chosen as the random s<L'llp:e.
We first present some specifics on sampling from finite universes. Subsequent sections will
describe various sampling distributions, Sections marked by an asterisk may be omitted
without loss of continuity.
Consider a univcn;e consisting offive units numbered 1,2,3,4"s.ln sampling without replacement. we
will employ a sample.of si;r.e two and enumerate the possible samples as
(1,2), (1,3), (1,4), (1.5), (2,3), (2,4), (2,5), (3,4). (3,5), (4,5),
Note that there are (~) =- to possible samples. we select one from these. where each has an equi!.l
IT
selection probability of 0.1 assigned, that is si:uplcrandom sampling. Consider these possible s~ples
to be numbered 1,2, ... ,0, where 0 represents num;", 10, Now, go to Table XV (iJ: the Appendix),
sho~.l1g n1;ldom integers, and vAth eyes closed place a finger down, Read the first digit from the five-
digit integer presented. Suppose we pick the integer which is in row 7 of columl14. The fust dig:t is
6. so the sample cor.sists of units 2 and 4. An alternate to us'..::.g the table is to cas~ a. sing;e icosohe-
dron die and pick the sample corresponding to the outcome, Kotice also that each unit appears in
exactly foUI of the possible samples, and thus the inclusion probability for each unit is 0":';', which is
simply rJN =-215,
It is usually not feasible to enumerate the set of possible samples. For example. if '
N = 100 andn = 25, there would be more than 2.43 X 1013 possible samples, so other selec-
tion procedures which maintain the properties described in the definition must be
employed. The most commonly used procedure is to first number the universe units from 1
to N, thcn use realizations from a random number process to sequentially pick numbers
from 1 to N, discarcful,g duplicates until n units have been picked if we are sampling with-
out replacement, kceping duplicates if sampling is with replacement. Recall from Chapter
6 that the term random numbers is used to describe a sequence of mutually independent
variables U 1, Uz, ... , which are identically distributed as uniform on [0,1). We employ a
realization u 1• ~, ... , which in sequentially selecting units as members of the sample is
roughly outlined as
(Unit Nurober),=LN· u)+ I, 1,2, ... ,1. i= I, 2, .... n, i<j,
where J is the trial number on which the nth, or finaJ.. unit is selected. In sampling without
replacement, J ~ n, and LJ is the greatest integer contained function, 'When sampling with
replacement, J = n.
nh =n
r.LN>·a·a
LNh
h
k
1 h = 1,2,.",L. (9-3)
Lh=l
The values Gh are standard deviation values within strata, and when designing the sampling
study, they are usually unknown for the variables to be observediu the study; however. they
may often be calculated for an explanatory or auxiliary va..<iable where at least one of these
is numeric, and if there is a "reasonable" correlation, Ipi > 0.6, between the variable to be
measured and such an auxiliary variable, then these surrogate standard deviation values,
denoted a~, will produce an allocation wbich will he reasonably close to optimal,
9-2 Statistics and Sampling Distributions 201
In seeking to address growing COnSUmer concerns regarding the quality of claim processing, a
national tnaruiged i:.ealth care:bealtb insurance company has identified several characteristics to be
monitored on a monthly basis. The most important of these to the company are, fi.rst, t..'!e size of the
m~ ''financial error," which is the absolute value of the error (overpay or underpay) :or a claim and.
second. the fraction of eorrectly rued claims paid to providers within 14. days. Three processing cen~
tern are eastern, mid~America, and western. Together, these eenters process about 450,000 claims per
month, and it has been observed that they differ in accuracy and tb::e11ne.ss. Fu.rt.hermore, the corre-
lation between the total dollar amount of the claim end the financial error overall is historically abou:
0,65-0.80. If a sample of size 1000 claims is to be drawn monthly, strata might be formed using cen-
ters E, MA, and Wand total claim sizes as $0-$200, $20h$900, $9(11-$= claim, Therefore there
would be nine strata. And ufor a given year there were 41.357 claims, these may be easily assigned
to strata. sinee they are natu...-ally grouped by center. and each center uses the same processing system
to record data by claim amount. thus allowing ease of classification by claim size, At this point, the
WOO-unit pla.1Jled sa.~1e is allocated to the ;tine strata as shown in Table 9-1, The allocation in ital-
ics, sho'W'tl. first, employs proportional allocation while the second one uses "optimal" allocation. The
values represent the standard deviation in the claim-amount met.:ic w.i:hin the stratum, since the stan-
dard deviation in "financial error" is unknown. After deciding on the allocation to use. nine inde-
pendent~ sirLple random samples, without replacement, would be selected, yielding 1000 claim. forms
to be inspected. Stratum identifying subscripts are not shown in Table 9·1,
It is nored that proportional allocation results in an equal inclusion probability for all
units, not just witb.in strata, while an optimal allocation draws larger samples from strata in
which the product of the internal variability as measured by standard deviation (often in an
auxiliary variable) and the number of units in the stratum is latge, Thus, inclusion proba-
bilities are the same for all units assigned to a stratum but may differ greatly across strata..
sample data makes considerable use of statistics. The procedures require that we understand>
the probabilistic behavior of certain statistics. In general, we call the probability distribu-
tion of a statistic a sampling distribution. There are several important sampling distributions
that wili be used extensively in subsequent chapters. In this section, we define and briefly
iliustrate these sampling distributions. Firs~ we give some relevant definitions and addi-
tional motivation.
A statistic is now defined as a value detennioed by a function of the values observed
in a sample. For example, if XI' X2" .. ') Xn represent values to be observed in a probability
sample of size n on a single t'dndom variable X, thenK and S', as described in Chapter g in
equations 8~1 and 8~, are statistics. Furthermore, the same is true for the median. the
mode, the sample range, the sample ske\\'Iless measure, and the sample kurtosis. ~ote that
capital letters are used here as this reference is to random variables, not specific numerical
results, as was the case in Chapter 8.
Zn = X-I'
I r>
th
en
lim F.(z)
,
l, (9-4)
cr!"n ..... ~ \IllZ)
where Fiz) is lhe cumulative distribution function of Z" and <I>(z) is the CDF for the stan-
dard normal variable, Z. Simply stated, as n -t ~, Z, -t N(O, I) random variable, and this
result has enormous utility in applied statistics. However1 in applied work, a question arises
regarding how large the sample size n must be to employ the N{O,I) model as the sampling
distribution of Z" or equivalently stated, to describe the sampling distribution of X as
NC/.!, cr'ln). This is an important question, as the exact form of the process variable assign-
mentmodel is usually unknO\\'Il. Furthermore~ any response must be conditioned even when
iris based on simulation evidence. experience. or "'accepted practice." Assuming process sta~
bility, a general suggestion is that if the skewess measure is close to zero (impJying that the
variable assignment model is symmetric or very nearly so), then X approaches normality
9-2 Statistics and Sampling Distributions 203
quickly, say for n ~ 10, but this depends also on the standaulized knrtosis of X, For exam-
ple, if fJ, ~ 0 and Ill,! < 0.75, then a sample size of n = 5 may be quite adequate for many
applications. However, It should be noted that the tail behavior of Xmay deviate somewhat
from that predicred by a nolllW model.
Wnere there is considerable skew present, application of the Central Limit Theorem
describing the sampling distribution must be interpreted with care. A Hrul e of thumb" that
has been successfully used in survey sampling, where such variable behavior is common
(fJ, > 0), is that
n > 25(fJ,)'. (9-5)
For instance, returning to an exponential process variable assignment model. this rule sug-
gests that a sample size of n > 100 is required for employing a nolllW distribution to
describe the behavior of X, since fJ, = 2.
Definition
The standard error of a statistic is the standard deviation of its sampling distribution. If the
standard error involves unknown parnrneters whose values caD be estimated, substitution of
these estimates intO the standard error results in an estimated standard error.
To illustrate this definition, suppose we are sampling from a normal distribution with
mean j.J. and variance fi1. Now the distribution of Xis normal with mean j.J. and variance
d'ln, and so the standard error of X is
If we did not know (J but substituted the sample standard deviation s into the above, then
the estimated standard error of Xi.
s
t~~~pg'[~j
Suppose we colicet data on the tension bond strength of a modified portland cement mortar, The ten
observations are
16.85,16.40, :7.21,16.35,16.52,1'7.04,16.96,17.15,1659,16.57,
whe~e tension bond strength is measured in units of kgf/cml , We assume that the tension bond
strength is well-described by a notnl31 distribution. The sample average is
x= 16.76 kgflcm'.
First, suppose we know (or are willing to assume) that the standard de'\'iation of tension baed strength
is 0'= 0.25 kgfIcm 2• Then the standard error of the sample average is
(J,/.,fn =0.25/~ =0.079 kgf/cm 2 •
If we are unwilling to assume that 0'= 0.25 kgfJcm\ we could use the sample standard deviation. s ==
0.316 kgfJcm2 to obtain the estimared standard error as follows:
G)
EO(8) = L1!,6k , (9·6)
k=l
where fI, is the value of the Statistic if possible sample k is selected. In simple random sam-
pling, recall that nk ::: 1J('~),
Now in the case of the sample mean statistic, X. we have EC(X) :::.; fl.;::. where I1r is the
finite population mean of the random variable X. Also, under simple random sampling
without replacement.
(9·7)
where a~ is as defined in ~uation 8-7. The ratio nhV is called the sampling fraction, and it
represents the fraction of the population measures to be included in the sample. Concise
proofs of the results shov,'ll in equations 9·6 and 9-7 are given by Cochran (1977, p. 22).
Oftentimes) in studies of this sort, the objective is to estimate the population total (see equa~
tion 8-2) as well as the mean. The "mean per unit,l1 Or mpu. estimate of the total is simply
Tx=N· X, and the variance of chis statistic is obviously W)=lv"'· V(X). Wlrile necessary and
sufficient conditions for the distribution of X to approach normality have been developed,
these are of little practical utility, and the rule given by equation 9-5 has been widely
employed.
In sampling with replacement, the quanitities E'(X) = p" and v"(X) = a~/n.
'tr'here stratification has been employed in a finite population enumeration study, there
are two statistics of common interest. These are the aggregate sample mean and the estimate
of population total. The mean is given by
(9-8)
(9·9)
In these formulations, Xh is the sample mean for stratum h, and WII• given by (Nh/b\ is
called the stratum weight for stratum h, Note that both of these statistics are expressed as
simple linear combinations of the independent" stratum statistics Xhl and both are mpu esti-
mators. The variance of these statistics is given by
(9-10)
I
I The withln~stratum variance terms,
9-3 The Chi-Square Distribution
S;~, for the respective strata, The sampling distributions of the 3&:,oregate mean and total
estimate statistics for such stratified, enumeration studies are stated only for situations
where sample sizes are large enough to employ the limiting normality indicated by the Cen-
tral Limit Theorem, Note that these statistics are linear combinations across observations
within strata and across strata. The result is that in such cases we take
(9-11)
and
~~~~~~j
Suppose the sampltng pla."l. in Example 9-3 utilizes::he "optimai" aliocatio;;. as shov.';l:n Table 9-L
The stratum ~ple means and samp!e staadard deviations on the fi.'1a!1cial error variable are cakJ~
lated with ::he results shown in Table 9-2, where the units of measurement are error $/claim.,
Then utilizing the results presented in equations 9-8 and 9-9, the aggregate statistics whe::l eval-
uated are.>'= $18.36, and f, = $8,103,3,5. Utilizing equations 9-10 and eoploying the within-stratum
sample variance ,,'Blues s!, o;h
to estimate the w-lt.1in~stratum variances the estimates for the variance
in the sampling distribl.rtior. of the sample mean and the esti.mate of total are V(fl) "'" 0.574 and
Hi.>:) "'" 1.118 x 101!, and the estimates of the respective standard errors are thus $0.785 and $334,408
for the mean and tOtal esrimator distributions.
Claim Amount
$0-$200 $201-$900 $901 - $Max Claim All
E N= 132,365 N=41,321 N~ 10,635 184.321
n=29 n::.-::54 n=371 454
.>'=6.25 x=34.10 x=91.65
----
206 Chapter 9 Random Samples illld Sampling Distributioru;
Theorem 9·1
Let Z" z" ... , Z, be normally and independently distributed random variables, with mean
= =
I' Q and variance d' I. Then the random variable
X~=.. i+ Z'·
, Z' Z,
21' .. · , . .... k
p{X~" x~,d = J~
z., f(u)du a.
Ims probability is shown as the shaded area in Fig. 9-2. To illustrate the use of Table m,
note that
p {X:, ;"Xi.".IQ} =P (X;," l8.3lJ =0.05.
That is, the 5% point of the chi-square distribution with 10 degrees of freedom is X;.",.lO =
18.31.
Like the nonnal distribution, the chi-square distribution has an important reproductive
property.
k=IA
p"l
PrOfl! Note that each chi~square random variable X~- can be written as the sum of the
squares of k standard nonn.al random variables, say
j
,
i=l. 2, ... , p,
Therefore,
.u P ki
Y=LX;
i=t
LLZ~
1",1 j=:
and since all me random variables ZfJ are independent because the X; are independent, Y is
just the S1..--n1 of the squares of k ~ kJ independent standard normal random variables.
='"
~'=l
From Theorem 9-1, it follows that Y is a chi-square random variable with k degrees of
freedom.
As an example of a statisti:: t:!:iat follows the chi-square distribution, suppose that Xl> x., .... , XI! is a
random sample from a normal population. with mean J1. and variance 01-. The functiQU of t.1C sample
variance
f(u)
o u
Figure 9~2 Percentage point X!..t of the chi-square distribution.
208 Chapter 9 Random S""'ples and Sampling Distributions
is distributed as X!.:' We will use this random variable extensively in. Chapters 10 and 11. We will see
in those chapters that becauSe the distribution of t:l.:ris random variable is chi square. we can construct
confidence interval estimates and test statistical hypotheses about the variance of a nor:wa1 population.
To illustrate heuristically why the distribution of the random variable in Example 9~6,
(n. _l)SlJo-i, is clJ.i~square, note that
(9-15)
~::lXi - 1')'
;"'!
,,'
isi, because each term (x,.- ]1)1<115 anindependeut standard normal random variable. Now consider
the following:
"
1')' = Y[(X,-X)+(X-Jl)]
,
,\..1
i. II f'
= Irx,-X)' +n(X-I')'.
10.. ;
Therefore,
or
"
I(X,-I')' 1- ,Z
,., (n-I)S' lX-I')
2 2 +-.,-,-. (9-16)
(r in
(j
"
Since X is normally distributed v.ith mean.}l and variance erIn, the quanti:y (5: fJif( fTln) is dis-
tributed as X!. Furt.iJ.errnore, it can be shown thaJ: the random Variables X anaS*·ire~ae'pendent.
Therefore. s.ince 2::1 (Xi - j..t)1 is distributed as X~, it seems logical to use the additivity proP"'
1(12
erty of the chi-~u.are distrihucion (Theorem 9~2) and conclude that t..1.e distribution of (n - 1)S2/til is
X~_I' '
Theorem 9-3
Let Z - N(O, I) and V be a chi-square random variable v>'ith k degrees of freedom. If Z and
V are independent, then the random variable
Z
T= r,-
'V V/k
9-4 The r DistrilY,;:tion 209
t= Z
and
u v
are
Z=t~f
and
v::::u,
The Jacobian is
..
IU
1= i'lk
I0 1
Thus,
ill= 1.<:
. ~k
and so the joint probability density fwletion of T and lJ is
Now. since v> 0 we must require that u > 0, and since - 00 < z < 00, then - 00 < t < "">, On
rearranging equ1ition 9-18 we have
-....
L
210 Chapter 9 Random Samples and Sampling Distribu:iODS
J(t)
-oo<t<=-,
Primarily because of historical usage, many authors make no dis'"...inction between the
random variable T and the symbol t, The mean and variance of the t distribution are p. = 0
and d' = k/(k- 2) for k > 2, respectively, Several t distributions are shown in Fig, 9-3, The
general appearance of the rdistribution is sllnilar to the standard normal distribution, in that
both distributions are symmetric and unimodal, and the maximum ordIDate value is reached
at the mean p. = 0, However, the t distribution has beavier tails than the normal; that is, it
has more probability further out, As the number of degrees of freedom k ..... ~, the limiting
form of the I distribution is the standard nonnal distribution, In visualizing the t distribution
it is sometimes useful to know that the ordinate of the density at the mean fJ.. := 0 is approx ~
imately four to five times larger than the ordinate at the 5th and 95th percentiles, r'Or exam-
ple, ""th I 0 degrees of freedom fort this ratio is 4.8, with 20 degrees of freedom this factor
is 4.3, and with 30 degrees of freedom this factor is 4.1. By comparison. for the normal dis-
tribution, this factor is 3.9,
The p,,!,centage points of the t distribution are given in Table IV of the Appendix. Let 1«,
be the percentage point or value of the t random variable with k degrees of freedom such that
p{TZ'a,.}=f J(t)dt=a,
'a>
This percentage point is illustrated in Fig. 94. Note that since the t distribution is symmetric
about zero, we find t1-<1.,.1: =:; -ta.,k' Tb.is relationship is useful, since Table N gives only upper-rail
t""
percentage points, that is, values of for a$ 0.50. To illustrate the use of the table, note that
P{T~ tQl)S,IO} =PIT:;' L812} =0,05.
Thus, the upper 5% point of the ,distribution with 10 degrees of freedom is ''os,,. = 1.812,
Similarly, the lower-tail point to..".:o ~ -t'"",10 = -I ,812,
;~~~!~~:J\i
As an ex:runple of a ra."1dom variable that follows the t distribution. suppose thatX:.~, ... , Xil is a xan.-
com sampic from a normal distribution withmcanfJ. and variance rr.and letX and 51. denote thesam~
pIe mean and variance. Corsica the statistic
o
Figure 9~3 Several: distributions,
9·8 The F DIstribution 211
(9-\9)
Divlding both the nurneretor and denominator of equ.ation 9-19 by c;. We obtain
X-/1
...3[y=_; ~~~ .
S'l/n -,;S-/(72
Sbee (X -11)/(c;I';;;) - N(O.1) and S2!dl_ X:~;/(n - 1), and since X and S2 are independent, we see
from Theorem 9-3 mat ~
(9-20)
follows a t distribution with v:::: n ~ 1 degrees of freedom.. In Chapter;, 10 and 11 we \\111 use the rdD-
ciom variable in equa:i.on 9-20 to construct confidence intervals and test hypotheses about the c;.ear.
of a normal distribution.
Theorem 9·4
Let Wand Ybe indepe_ndent chi-square random variables with u and v degrees of freedom,
respectively. Then the ratio
F W/u
Ylv
has the probability density function
- ·~(u)ltl2
r(u-+v - p'u(2)-1
h(t) " 2 ; v 0</<=,- (9·21)
and is said to follow the F distribution with u degrees of freedom in the numerator and v
degrees of freedom in the denominator. It is usually abbreviated FIJ,'i"
Proof Since Wand Yare independent, their joint probability density distribution is
212 Chapter 9 RJmdom Samples and Sampling Distributions
(,12)-1 (,/1)-1
W' Y -(w+y)/2
I(w,y) O<w,Y<QQ·
2"2r(%)z'f2r(~J e ,
Proceeding as in Section 4-10 t define the new random variable M ::;:;: 1': The inverse solutions
of/= (wlu)I(ylu) and m = y are
w= umf
v
and
y m.
Therefore. the Jacobian
J=~ ~L~m.
I~ ~ I v
Thus, the joint probability density function is given by
. (
UUfm
)(','2)-1 'ij2
g(j,m) -;:;l-;:; m -(m/2)(('i'lf+l)
2'f2 r( ~
\.2.:
12' r(x.)
J1
\2
e ,
0< j,m < «')
and since h(fJ = r; g(f, m) dIn, we obtain equation 9-21, compleling the proof.
The mean and variance of the F distribution are Il = v/(v - 2) for v > 2, and
2v'(u+v-2)
v>4.
u(v-2)'(v-4)'
Several F distributions are shown in Fig. 9-5. The Frandom variable is nonnegative and the
distribution is skewed to the right. The F distribution looks very similar to the chi-square
distribution ill Fig. 9-1; bowever, the parameters u and v pro"ide extra flexibility regarding
Shape.
u=5,v=5
The percentage points of the F disrribution are given in Table V of the Appendix. Let
F "-"-~ be the percentage point of the F distnlJution with u and v degrees of freedom. such that
the probability that the random variable F exceeds this value is
Th.is is illustrated in Fig. 9-6. For example, if u = 5 and v = 10, we find from Table V of the
Appendix that
PiFe. F,m.',lo) = PiF;' 3.33) = 0.05.
That is, the upper 5% point of F,.lC is FO,05~.JO = 3.33. Table V contains only upper-tail per-
centage points (values of F a,U,lo' for a:s; 0.50). The lower~tail percentage points F I~,v can
be found as follows:
1
fl-a,lt,v =~. (9-22)
a,v,1I
For example, to find the lower-tail percentage point F O•95,5,lOl note that
0.211.
4.74
A.s an example of a statistic that follows the F dis~bution. suppose we have two notmal popula1ons
with variances ~ and 0;. Let independent nm'dom samples of sizes n l and r"1; be taken. from popu1a
S;
!ions 1 and 2. respectively, and let S~ and be the sample variances. Then the rano
(9-23)
has an F distributitin -v.'ith n, - 1 numeratOr degrees of freedom and li.z 1 denominator degrees of
freedom. This follows dire.."tly from the facts that (nt -1)~10'~ - X:, ~ 1 and (n.z-l),f/O'~ "" X!l-! and
from Theorem 9...!, The nmdom variable in equation 9-23 plays a key role in Chapters 10 and 11,
where we address the preblems of confidence interval estimation and hypothesis testing about the
variances of two independent normal populatiOI15.
h(0
f
Figure 9-6 Upper and lower percentage points of the F distribution,
214 Chapter 9 Random Samples and Sampling Distributions
9-6 SL-:rvL\1ARY
This chapter has presented the eoncept of random sampling and introduced sampling' dis-
tributions.In repeated saml'.ling from a population, sample statisties of the sort discussed in
Chaptcr 2 vary from sample to sample, and the probability distribution of suell statisties (or
funetions of the Statistics) is called the sampling distribution, The normal, chi-square, Stu-
dent t, and F distributions have been presented in this chapter and will be employed exten-
sively .in later chapters to describe sampling variation.
9-7 EXERCISES
/~ Suppose that a random variabl: is normally dis- the vendor 2 observed resistances asscmed normaUy
\:.Juted with mean JJ. and variance if. Draw a random and independently distributed with a mean of 105 Q
sample of five observations. What is the joiut density an£! a standard deviation of 2,0 0.. "/hat is the saro~
~tioo of the sample? pling distribution oCr! -X2?
( 9~2_'TransistorS have a life that is exponentially dis~ 9~9. Consider the resistor problem in Exercise 9-8.
"tnbuted ",1.th pararr-.erer ;_ A random sampl::: of It :ran- Fmd the standard elTO! of X! ~ Xl'
slstors is taken. "/hat is the joint density function of 9~ to. Consider the resistor problem in Exerci$e 9~8. If
the sample? we could not assume that resistance is normally dis~
9~3. Suppose thar X is uniformly distributed 00 the tribut~ what could be said about the sampli:cg distri-
,interval from 0 to 1. Consider a ranqom sample of size bution ofX\
4 from X. ~at is the joint density function of the 9~11. SuppoSe that independent random samples of
sample? sizes n l and '1:1 are taken from wo normal populations
94. A lot consists of N transistors, and of these M with mean.., JJ.I and f.J..;. ar.d variances (j~ anda;,
(M $" N) are d:::fective. We randomly select twa tran- respectively" If X, and X2 are the sample means, find
sistors without replacement from this 10t and deter M
the sampling distribution of the statistic
rrine whether they are defective or ooodefectivc. The X, -:1'2 - (Ill -1'2)
ra.'1dom variable
if the iOO transistor is nondef~tive, .
Mlnl)+ ((j~ I~) ::
X-=
,
IlOI if the ith transistor is defective.,
.=L2, 9~ 12. A manufacturer of semico:r.ductor devices takes
a random sample of 100 chips and tests them. classi~
Detennine the joint probability function for Xl and Xl' fying each chip as defective or nondefe<!tive. Let
'Woat arc the marginal probability functions for XI and X, 0(1) if the lth chip is Ilondefective (defective).
~i? Ate Xl and X2independent random variables? The sample fraction defective is
,: 9~~A population of power supplies for a personal ~ Xl -tXZ+'''+XIOO
~puter has an outpUt voltage that is normally dis~ p 100
tributed with a mean of 5.00 V and a standard devia~
"What is the sampling distribution ofp?
tion of 0.10 V. A random sample of eight power
supplies is selected. Specify the sampling distribution 9 ..13. For the semiconductor problem in Exercise
of X 9-12.:find the standard error of p. Also find the esti-
mated standard error of p.
9.-6, Consider j}e power supply problem described in
Exercise 9~5. What is the standard error off? 9,.14. Develop the moment-generating function of the
'/\?'" chi-square distribution.
9-7. Consider the power supply problem described in
9~ 15. Derive the mean and variance of the chi·square
Exercise 9-5. Suppose that the papulation standard
random variable wij} u. degrees of f..-eedom.
deviation is nnknown. How would you obtain the esti-
mated standard error? 9~16. Derive the mean and variance of the I
distribution.
QtA procurement specialist has purchased 25 rcsis-
~from vendor 1 and 30 resistors from vendor 2. 9·17. Derive the mean and variance of the '1
Let X IP X!:2' .. " X j ,2j represent the vendor 1 observed distribution,
resistmce.s assumed normally and independently dis- 9--18. OrderStatistic.s. LetXj , X2, ""X>j be arando:rr
tributed with a mean of 100 .Q and a standard devia- sample of size n from X. a random v.d.riable baving
tion of 1.5 n, Similarly, letA;l'X:u, ... , XzJi)represent distribution function F(:;;). Rank the elements iu arde;
-g~ 7 Exercises 215
of increasing DIlffie:t:CaI magnitude, resciting in X(I)' variable with parameter A. Derive the distribution
XO) •. "., X{,,). where X(ll is the smallest sample element functions and probability distributions for X{l) and
(Xm=min [X1,Xl • .... X~}) andX(~) is the largest sam- X:,,), Use the results of E-'Ccrcise'""9-18.
ple element (X,,,) = max: {XI' Xl' "" X~}). A;,) is called 9-22. Let Xl' Xl' ... , X:, be a tandom sample of a con~
11e ith order statistic, Often. d:e distributio:;:; of some
tinuous random variable. Find
of the order Statistics is of interest. particularly the
;ninimum and maximurI.'.. sample valnes. XiI} and ~"l' E[F(X,,)l]
respectively. Prove that t.1je distribution functions of and
X(n and XC!!}' denoted respectively by FX(,j(t) and
Fx",,(t). are
FX{,,(t) = 1-:1 F(tW. 9-23. Using Table ill of the Appendix, lind the fol-
F",.,<r) ~ [F(r)]". lowmg values:
Prove that if X;;is concinuous y,'ith probability distribu- .~X:.;".
tionf(x), then the probability distributions of Xli) and (~X~,~.ll·
:1;" are (c) X~.cr.5,m'
f"",ttl ~ n[l- F(t)1-' fit), (d) X~:J such that P[X~J ~X~l:'l} =; 0.975.
ix",,{r) ~ n[F{tl]-' fi')·
9-24. Using Table IV of the Appendix. find the fol-
9 19. Continuation of Exercise 9-18. Let Xl' Xl' ....
R lowing values:_
X be a random sample of a Bernoulli random vaii:able
q (a) iO.Z!,lO'
with parameter p. Shov.' that .
P(X(,)~ 1)~ 1- (l-p)'.
-- (b) t;,."w
(c) ta,10 such that Pi tiC::; fa,lo} = 0.95.
P(X(l) 0) = 1 - p'.
9-25. Using Table V of the Appendix, find the follOV{~
tlse the results ofE.1tercise 9-18. ing \--alues;
9-20. Continuation of Exercise 9-18. Let X!, X'). • .... (a) FO.2S.4.9·
Xn be a random sample of a non:nal random variable (b) F01)'j,)s .•c"
\\ith mean J1. and va.>iance a2. Using the resclts of
(e) F O•95 .6.S'
Exercise 9~18, derive the density functions of X(l) and
. X:~)' (d) F O.90.24,24'
9--21. Continuation of Exercise 9-18. Let Xl' X'). • •.• , 9-26~ Let FI-4.!i,~ denote a lowerwtail point «(I. s:; 0.50)
Xi!!) be a tandom sample of a.'1 exponential r:mdom of the F..,v distribution. Prove that F;-tt,H,y =; lIFa,v,JJ'
Chapter 10
Parameter Estimation
Statistical inference is the proeess by which information from samp1e data is used to draw
oonelusions about the population from whieh the sample was selected. The teehniques of
statistical inferenee ean be divided into two major areas: parameter estimation and hypoth-
esis testing. This ehapter treats parameter estimation, and hypothesis testing is presented in
Chapter 11.
As an example of a parameter estimation problem. suppose that civil engineers are ana~
lyzing the compressive strength of eoncrete. There is a natural variability in th~' strength of
each individual eonerete speeimen.£onsequently, the engineers are interested in estimating
the average strength for .the population consisting of this type of concrete. They may also
be interested.in estimating the variability of eompressive strength in this population. We
present methods for obtaining point estimates of parameters such as the population mean
and variance and we alSQ discuss methods for obtaining certain kinds of interval eSP-males
of parameters called confidence intervals.
216
to-I Point Estimation 217
~~~!€~if:l
Suppose that X is a random variable with mean fland variance cr, I.etX1,XZ> " ' . X" be a randomsamw
pIc of sizen: from X Show that the sample llleanX and samplevanance S1. are unbiased estimators of
jJ. and ,"" respectively. Consi&.'T
L
218 Chapter 10 Parameter Estimation
1 '
E(X)=- ~>=Ii.
n ;",1
Therefore, the sample mean X is an unbiased .estimator of the population mean p. Now consider
~I'-'-)
1 £..
=-1:. X'+X'-2XX.
,,-1 ;..! \ I <
~_I Ji x1 -,&')
n-1 \)"'{
~~[iE(Xl)-nE(X')].
1 /..
It 1
MSE(e,)
MSE(e,) ,
If this relative effici~cy is less than one, we would conclude that a{ is a more efficient esti...
mater of () than is Oz. in the sense that it has smaller mean square error. For example,
suppose that we wish to estimate the mean ,il of a population. We have a random sample of
10-1 Point Estimation 219
l'! observations XI' Xz, ... , X.'I' and we wish to compare hvo possible estimators for J1~ the
sample mean X and a single observation from the sample, say Xi - Note that both X and j{,
..are
------ -.- .. ----- .. -.--~--- ~ ,
unbiased estimators of ).1; consequently, the mean square enor of both estimators: is sim-
ply the variance. For the sample mean, we have MSE(X) =' VeX) = <5'fn, whare <5' is the
population variance; for an individual observation, we have MSE()(,) = VeX,) <5'. There-
fore, the relative efficiency of Xi to X is
MSE(X)
MSE(X,)
Since (lfn) < I for sample sizes n "2, we would conclude that the sample mean is a better
estimator of !1 than a single observation Xr
Within the class of unbiased estimaton;, we would like to find t.~e estimator that has the
smallest variance. Such an estimator is called arninimum variance unbiased estimator. Fig-
ure 10-1 shows the probability distribution of two unbiased estimaton; 8, and with 81 hav- 9"
t\
ing smaller variance than~. The estimator 81 is more likely than to produce an estimate
that is close to the true value of the unknown parameter 8,
It is possible to obtain a lower bound on the variance of all unbiased estimators of ().
Let 8be an unbiased estimaror of the parameter 9, based on a random sample of n observa-
tions, and let fix, IJ) denote the probability distribution of the random variable X. Then a
lower bound on the variance of fJ is'
(10-4)
This inequality is called the Cramer-Rao lower bound. If an unbiased estimator satisfies e
equation lO~4 as an equality, it is the minimum variance unbiased estimator of 8.
;;!!~pI~!!8~1
We 'Will show that the sample mean X is the minimum variance unbiased estimator of the mean of a
nonna! dist;r!lmtion v.-ith known variance. ~.
-~
Figure 10~1 The probability distribution oct'.Vo unbiased estimators, 81 and (J2,
ICertain conditions on the functionf{X, B) a..--e required for obtaining the Cmn6'-Rao inequality (for example,
see Tucker ::962). These conditiOn:> are satisfied by most oCme standard probability distribution&.
220 Chapter 10 Parameter Estimation
From Example lO~l we observe that X is an U!lbiased est.irnator of p.. Note tha~
1
~--~~~
nE(X-l'r
0"4
=-.
n
Since we know ::hat, in general, the variance of the saaple mean is V(X):;:::: erln. we see that VeX} sat~
isfies the C:amer-Rao lower bound as an I!qmility. Therefore Xis the minimum variance unbiased
estimator of J.l. for the normal distribution where 0=' is knO\Vll,
(10-5)
Consistency is • lMge-sample property, since it describes the limiting beh.TInr of the esti-
mator (j as the sample size tends to infinity. It is usillilly difficult to prove that an estimator
r 10-1 Point Estimation 221
Figure 10~2 A biased estimator, 81, that has smaller variance than the unbiased estimamr. e~:
is consistent using the definition of equation 10-5. However, estimators whose mean square
error (or variance, if the estimator is unbiased) tends to zero as the sample size approaches
infinity are consistent. For example, Xis a consistent estimator of the mean of a normal dis-
tribution, since X is unbiased and limn~ VeX) = limll~ (a2/n) = O.
Note that the likelihood function is now a function of only the unknown parameter e. The
e e
maximum likelihood estimator (?vll..E) of is the value of that maximizes the likelihood
function L(fJ). Essentially, the maximum likelihood estimator is the value of that maxi- e
,mizes the probability of occurrence of the sample results.
where p is the parameter to be estimated. The likelihood function of a sample of size n would be
[ .. \
We observe that if p maximizes L(p) then p also maximizes lnL(p) sinee the logarithm is a monoto-
nieally increasing function. Therefore,
1
Now
an intuitively pleasing answer. Of COurse, one should also perfOCl a second derivative test, but we
have foregone '<hat here.
"~~I~~!,ii.~
Let X be normally distributed wi;n unknown mean f.1 and knO'Wtl variance cr, The likelihood function
of a sample of size n is
Now
dlna;ll) (a'r'L(x,-Il).
J"':
Equating tbls last reswt to zero and solving for f.1 yields
.. - -..- - - - - - - - - - - - - - - - -
It may not always be possible to use calculus methods to determine the maximum of
48). This is illustrated !n the follawing example.
:!~P!~!ji~~:
LetXbe uniformly distributed on the interval 0 to a, 'The likelihood function of a random sample X!,
x,.. "., X" of size n is
Notc that the slope of this function is not zero an)'\Vhcre, so we cannot use calculus methods to find
the maximum likelihood estimator a. Hov{ever, notice that the likelihood function increases as a
decreases. Therefore, we would maximize L(a) by settingd to the sIIULllest value that it could reason~
ably assume. OearIy. a can be no smaller than the largest sample value, so we wo;;Jd use the largest
observation as fi. Thus, d =maxi Xi is the !>.-1LE for a.
------
10-1 Point Estbation 223
The method of maximum likelihood can be used in situations where there are several
unknmvn parameters. say 81> 82, •. '. 8k' to estimate, In such cases, the likelihood function
is a function of the k unknown parametern 8" a" "., 13k and the IDa,mnum likelihood esti-
mators [Ili ) would be found by equaring the k first partial derivatives iJL(B" 8" . ", B,)liJ8;,
1,2. ,..• k, to zero and sol\ug the resulting system of equations.
:~~iF'~~2
LetXbenormally dist:ibuted v;'ithmean f.I. and variance cr. rr
where both ,uand are unknown, Find the
maxi.mw:n likelihood estimators of ,U and 02. Tue likelihood function for a nmdom sa.::cple of size n is
L(Ji.,()2) = tIO'~e-{;t!-f1)~llC12
,.\
1 e -IV2"')k~,(,-4
(z=,t
and
Now
dlnL(ll,a') I "
, :Lh-,u)=o,
all (J i=l
dlnL(!'.,,')
a(,,-)
"\
and
-2
(J
I ~( X; -X
=-""-' -)' .
n 1..,1
which is closely related to the unbiased sample variance Sl, Namely, 6"1 «n_l)/n)S2.
Maximum likelihood estimators are not necessarily unbiased (see the maximum like-
lihood estimator of d' in Example 10-6), but they usually may be easily modified to make
them unbiased. Further, the bias approaches zero for large samp1es. In general, maximum
likelihood estimators have good large-sample or asymptotic properties. Specifically, they
are asymptotically normally distributed, unbiased, and have a variance that approachcs the
Cramer-Ran lower bound for large "- More precisely, if 11 is thc maximum likelihood esti-
mator for 8. then .-yJ;(8- 8) is normally distributed with mean zero and variance
~,In(iI-e)l=v(~niJ}= I -2
E[:e 1nf(X,1I) J
for large n, Maximum likelihood estimators are also consistent, 1n addition, they possess the
invariance property; that is, if (J is the maximum likelihood estimator of () and u( If) is a func-
)
224 Chaprer 10 Parameter Estimation
tion of II that has a ,ingle-valued inverse, then the maximum likelihood estimator of u( IJ) is
u(fJ).
It can be shown graphically that the maximum of the likelihood will occur at the value
of the maximum likelihood estimator. Consider a sample of size n ;::::: 10 from a normal i
distribution:
14.15,32.07,32.30,25.01,21.86,23.70,25.92,25.19,22.59,26.47.
Assume that the population variance is known to be 4. The 11LE for the mean. J1., of a nor-
mal distribution has been shown to beX. For this set of data, x; 25. Figure 10-3 displays
the log~likelihood for various values of the mean. Notice that the maximum value of the
log~likelihood function OCC1Jl'S at approximately x= 25. Sometimes, the likelihood function
is relatively flat in the region around. the maximum, This may be due to the size of the sam..
pie taken from the population. A small sample size can lead to a fairly fl't log likelihood,
implying less precision in the estimate of the parameter of interest.
27
Sample mean
Figur< 10-3 Log lil:'elihood for various means.
lO~1 Point Estimation 225
The population moments {J1;} will, ill general. be functions of the k unknown parameters
( 8i ). Equating sample moments and population moments will yield k simultaneous equa-
tions in k unknowns (the (Ii); that is,
t= I, 2, ... , k. (10-9)
The solution to equation 10-9, denoted 8"e" ... ,tI" yields the moment estimators of 8" 8"
"., 9;:
~~~w~!~Ii917~
LetX - N(p., cr)
where 11- and crare unknol1."!l. To derive estimators for ,u and cr by the method of
moments. recaE that for the normal. distributio::l
11-; =. tt.
14 = cfl + ;t.
The sample moments are m{ :;;{l/n)'~ ,Xl and rn2::;C: (l/n) ~~ IX?, . From equation 10-9 we obtain
L...t.,.,. L...t"",
't!~R!~10}~
Let X be uniformly distributed on the interval (0, a). To find an estimator of a by the method of
moments, we note that the first population moment about zero is
The method of moments often yields est.imarors that are reasonably good. In Example
10-7, for instance, the moment estimators are identical to the maximum likelihood estima~
tors. In general. moment estimators are asymptotically normally distributed (approxi-
mately) and consistent. However, their variance may be larger than the variance of
estimators derived by other methods, such as the mcthod of maximum likelihood. Occa-
sionally! the method of moments yields estimators that are very poor, as in Example 10-8.
The estimator in that example does not always generate an estimate that is compatible with
our knowledge of the situation.. For example, if our sample observa.tions were Xl == 60, Zz :::::
10, and x) :::::5. then a= 50, whlch is unreasonable, since we know that a ~ 60.
226 Chapter 10 Parameter Estimation
(10-10)
The joint distribution of the sample and 6 in the numerator of equation 10-10 is the prod-
uct of the prior distribution of 6 and the likelihood, or
f(;t" x.,., ... , x,; 8) = f(8)· f(x" x.,.•••. , xn!8)·
The denominator of equation 10-10, which is the maIginal distribution of the sample, is just
a no.rm.alizing constant obtained by
(IO-Il)
, 0
Consequently, we may write the posterior distribution of eas
I )_ f(6)f(x x" ... ,x,18)
f (Srl'X,,,,,,x, - ( " , (10-12)
f xllx2""'Xn)
We nOle that Bayes' theorem has been used to tr.msforrn or update the prior distribution
to the posterior distribution, The posterior distribution reflectS our degree of belief about fJ
given the sample information. Furthermore, the posterior clistribution is proportional to the
product of the prior distribution and the likelihood, the constant of proportionality being the
normalizing constantf(x" x.,., .•. , x,).
HH Point Estimation 227
Thus, the posterior density for e expresses our degree of belief about the value of e
given the result of the sample.
The time to failure of a transisto:; is known to be exponentially distributed with pttameter A" For a ran-
dom sample of n transistors, the joint dens:ty of the sample elements, giver. A. is
-ltx 1
f (x!,XZ,·",xn .•.:t)_'
-kA, e->(2'>.+<) .
and the marginal density of the sample is
ar.d we see that the posterior density :ot ;. is a gzu:u.'Ul'i distribution 'With pa."'amcters n + 1 and IX; + k
,-
228 Chapler 10 Para..'!leter Estin:;.:1tion
The function B will be minimized if we can find a function d that minimizes the quantity
within the large braces in equation 10-14 for every set of the x values. That is, the Bayes
x,
estimator of &is a function d of the that minimizes
If the loss function e(1}, (1) is the squared-error loss (8_11)', then we may show that the Bayes
e,
estimator of e, say is the mean of the posterior density for &(refer to Exercise 10-78).
~};~~ii!~l;~f~'
Consider the situation in Example 10-9. where it wa.<; shown. that if the random variable X is expo~
nent:ially distributed with parameter;;" and if the prior distribution for A. is exponential with parame-
ter k, then the posterior distribution for).. is a gamma distribution. with parameters n ..k 1 and
~'~
,L..,=l
Xi + k. Therefore. if a squared-error loss function is assumed. the Bayes estimator for A. is the
mean of this gamma distribution.
i rJ,+1
iXi+k
'"I
Suppose that in the ti:m'e~to-failure problem in EXac1ple 10-9. a reasonable exponen'ial prior distri-
bution fox A. has parameter k::;;; 140. This is equiValent to saying that the prior estimate for)" is
,,10
0.07142. A r:mdom sample of size n== 10 yields A...J.4x,< =1500. The Bayes estimate of).. is
We may compare this with the results that would have been obtained by classical methods. The max"
imurn lilce1ihood es.timator of me parameter A. in an exponential distribution is
1().1 Point Estimation 229
Cor.sequently, the maximum likelihood estimate of t., based or. the foregoL'1g sa.-nple data, is
~ = 10
;I.' =
tx;
i=!
1500
0,06667,
Note that the results produced by the two methods differ somewhat. The Bayes estimate is slightly
closer to the prior estimate than is the maximum likelihood estimate.
Let X[, J;, ""XII be a random samplc from the normal density MID mea."l J.L and va..riance 1, where ,I.l
is unknown. Assume that the prior density for J.1. is nonnal with mean 0 and variance 1; that is,
-
f (xl·x2'··'Xr. If.l )- (2tr),,/2· e
1 -ll!,)r(X,-p)'
1 -(!/2Jfr";-2,J.tr ..,+nd"1
=---e' ,
(2,,:),"''Z
Thus, the joint dC!lsity of the sample and p:s
f'~xl,.x2···'x" )- 1
- (2n')~"-i)f2 L.!z x '} [ cXPu,
exp~ 2 ), '_2
t 2lln+1 J.L
f __ -l~'
jP'..xj J.l.
1
(n+l)Ii2(2")"f.l exp
[-.!(ZX'
2
_n',' 'I'
' n-'-l)j
using:he fact :hat the i:ategral is (2n:):t.l./(n ,.;..l)lfl (si:ace a normal density has to integrate to 1). Now
the pos;;erior density for lJ. is
230 Chapte: 10 Parameter Estimation
There is a relationship betvleen the Bayes estimator for a parameter and the maxi-
mum likelihood estimator of the same parameter. For large sample sizes the two are near,!}'
equivalent, In general, the difference between the two estimators is small compared to ;In, 1/
In practical problems, a moderate sample size will produce approximately the same estimate
by either the Bayes Or the maximum likelihood method, if the sample results are consistent
with the assumed prior information. If the sample results are inconsistent with the prior
assumptions, then the Bayes estimate may differ considerably from the maximum likelihood
estimate. In these clrcumstances, if the sample results are accepted as being correct, the prior
information must be incorrect. The maximum likelihood estimate would then be the better
estimate to use.
lf the sample results do not agree with the prior information, the Bayes estimator will
tend to produce an estimate that is between the maxbnUnl likelihood estimate and the prior
assumptions. If there is more inconsistency between the prior ioformation and the sample,
there will be a greater difference between the two estimates. For an illustration of this. refer
to Example 10-10.
(1()..l7)
If Gt! involves any unknown parameters, thee. if we substitute estimates of these param-
eters into equation 1()"17, we obtain the estimated standard error ofe, say6s. A small stan-
dard error implies that a relatively precise estimate has been reported.
·.Ex:#l'l~l~# •
An article in the Journal a/Heat Transfer (Trans. AS:ME. Scs. C, 96, 1974, p. 59) describes a method
of measuring the then::lal conductivity of Anneo iron. Using a temperature of 100"'F and a power input
of 550 W, the followi:.1g 10 measurement<: of thermal conductivity (in Btulnr-ft-l:IF) were obtained:
5. Let t. e;Is
= 100 Or 200 is often used).
e' = represent the sample mean of the bootstrap estimates.
6. The bootstrap standard error of i1 is found with the usual standard deviation
formula:
-8
_.)2
S; =
In the literature, B - 1 is often replaced by B; for large values of B, bowever, there is little
practical difference in the estimate obtained.
::~R~~!~::m:i~;
The failure times X of an electronic component arc known to follow an exponential distribution with
unknown paramete: Il A random sample of ten cot.lponents resulted ill the following fail~e times (in
hours):
195.2, 201.4, 183.0, 175.1. 205.1, 191.7, 188.6, 173.5,200,8,210.0.
The mea.."I. of the e.xpone;ritial distribution is given by E(X) =. 11)., It is also known that E(X; =. 1I1l A
reasonable est:imate for A &n is i = VX'. From the sa.'nple data, we find X 192.44, resulting in
l=: 1/192,44 =0.00520. B= 100 bootstrap samples of size 1'/.::::.10 were generated using Mi..'litab$ll with
fir, 0.00520) =: 0.00520e-o·lrn2fu, Some of the bootstrap estirr.ates are sho'Jlll in Table 10-1,
The average ofllie bootstrap estimares is found to be
ercoc of the estimate is
X' = Ii;/.1
1",1
'\00=0.00551. The standard
0.00169.
·1
~Ir-._ - - - l
a frequency interpretation; that is, we do not know if the statement is true for this specific
sample, but the method used to obtain ille interval [L, U] yields correct statements
100(1- a)% of the time.
The confidence interval in equation 10-19 might be more properly called a two-sided
conJidence interval, as it specifies both a lower and an upper limit on (J, Occasionally, a one-
sided confidence interval might be more appropriate. A one-sided 100(1- a)% lower-con-
fidence interval on "is given by the interval
L'; e, (10-20)
or
Comparing equations 10-24 and 10-18, we see that the 100(1- a)% two-sided confidence
interval on J1 is ./
(10-25)
t~p!~iij.;H:
Consider the thermal conductivity data in Example 10-12. Suppose that We want to find a 95% con-
fidence interval cn The mean thec:nal conductivity of Ar::nco iron. Suppose we know that the star:.dru:d
deviation of The.rmal conductivity at 100"F ar.d 550 W is 0'= 0.10 BtuIhr-ft-¢R If we assume that
thermal conductiv:ty is normally distributed (or that thc conditions of the Central Limit Theorem are
meo), then we can use equation 10-25 to construct the confidence interva1. A 95% interval implies tliat
1 - a = 0.95, so ex:::;; 0.05, I!n.d from Table IT in the Appendix Zan. """ Z:J,0512 """ ZO.02S """ 1,96, The lower
confidence limit is
L:::::i-Za,12 (j/Jn
= 4 '.924-1.96( 0.1 0)/-50'
=41.924-0.062
41.862
and the upper confidence limit is
U::::.x+Zat'1.vj..J;i
=~1.924+ 1.96(O.10)/-JiO
=41.92~+O.062
= 41.986,
Thus the 95% two-sided confidence intcrval is
41.862 ., 1''; 41.986.
This is our interval of reasonable wlues f-or mean thermal conductivity at 95% coniidence.
2(2.5&O";..Jn) = 5. 15a/..Jn.
10-2 Singlc-Sample CDnfidence Interval Estimation 235
The 99% confidence interval is longer tllan the 95% confidence interval. Tbis is why we
have a higher level of confidence in the 99% confidence interval. Generally. for a fixed sam-
ple size n and standard deviation a, the higher the confidence level, the longer the resulting
confidence interval
Since the length of the confidence interval measures the precision of estimation, we see
that precision is inversely related to the confidence level. As noted earlier, it is highly desir~
able to obtain a confidence interval that is short enough for decision-making purposes and
that also has adequate confidence. One way to achieve this is by choosing the sample size
n to be large enough to give a confidence interval of specified length with prescribed
confidence.
If the right-hand side of equation 10-26 is not an integer, it must be rounded up. Notice that
2 E is the length of the resulting confidence interval.
To illustrate the use of this procedure, suppose that we wanted the error in estimating
the mean thermal conductivity of Armco iron in Example 10-14 to be less than 0.05 BtuIhr
-ft-"F, with 95% confidence. Since r:r= 0.10 and ZU.D25 = 1.96. we may find the required
sample size from equation 10-26 to be
It is also possible to obtain ooe-sided confidence intervals for Il by setting either L = - ~'br
U = - and replacing Z"", by Z". The 100(1 - (X)% upper-confidence inteITal for Il is
- r
fJ5.X+Za G' "in, (10-27)
and the 100(1 - (X)% lower-confidence interval for Il is
X-Za(J .Jn~Il. (10-28)
t=~
X-u
Sf";n
is the t distribution with n - 1 degrees of freedom. We nOw show how the confidence inter~
val on f.l is obtained.
The distribution of r = (X - 1l)/(sj.Jn) is shown in Fig. 10-7. L.."11ing r""", __ , be the
upper al2 percentage point of the t distribution with n - 1 degrees of freedom, we observe
from Fig. 10-7 that
or
t~~\~~~fgl
An .article in. the Journal <if Testing and Evaluation (Vol 10, No.4, 1982, p. 133) presents the
following 20 measurements on residua! flame time (in seconds) of treated specimens of cbildren's
nightwcar:
9.85. 9.93, 9.75, 9,77, 9.67,
9.87, 9.67, 9,94, 9.85, 9.75,
9.83, 9.92., 9,74, 9,99, 9.88,
9.95, 9,95, 9,93, 9.92, 9.89,
We ms.i, to find a 95% confidence interval on the mean :residual flame time, The sample mean
and standard deviation are
x~9,8475,
s~O,0954.
From Table IV of the Appendix we find to.025 ,19 "'" 2,093. The lower and upper 95% confidence limits
are
L~ X-la/2"H s/~
~ 9,8475 - 2.093(0,0954)/-J20
=9,8029 seconds.
,
/'
238 Chapter 10 Parameter Estimation
and
- (r
U = x +tlX/'L.J1-1S/'Vn
= 9.8475+ 2.093(0.0954)/fjjj
9.&921 seconds.
Tnerefore the 95% confidence interval is
9.8029 sec ~)J. ~ 9.8921 sec
We are 95% confident that the mean residual fla:ne time is between 9.80"..5 and 9.8921 seconds.
Comparing equations 10·33 and 10·18, we see that a 100(1 - a)% two-sided confidence
interval for a' is
(n-l)S2 < 2 < (n-l)S'
2 (1-2 . (10·34)
%a/2,1I-1 %1-«/2,n-1
012
o X~-(l12,n-1
Figure 10-8 'The X" distribution.
10-2 Single-Sa."np]e Confidence Ir.terval EstL.'71ation 239
To fmd a 100(1 - a)% lower-confidence interval on a', set U = = and replace X~'_l with
" ..
X'U,I1-I> gl'nng
{n-I)S' ~ 2 (10-35)
2 ;::;,.(j ,
Xa,n-l
The 100(1 - a}% upper-confidenee interval is found by setting L = 0 and replacing
X.~-ttr2Jl-l with X~----I"L,n:-:' resulting in
2/(n-l)S2
0' ~ 2 (10-36)
Xt-a.n-t
A manufacturer of soft drink beverages is jntcrested in the uniformity of the m..'1.chine used to 'fill cans.
Specifically. it is desirable that the standard deviation 0' of the filling process be less than 0.2 fluid
ounces; otherwise there will be a bigher than allowable perce:ltage of cans that are under:filled. We
will assume that fill volume is approx:i.rr.ately normally distributed. A random sample of 20 cans result
in a s.arople variance of S1 = 0.0225 (fluid ouncesJ2. A 95% upper.cOI:iidence interval is found from
equation 10~36 as fo1:ows:
or
or
(10-37)
,
P { P--Z~i2V
Iil(l-ill
n
'
';p:5P+Za/~~
Ip(l=iij}_
n -1--a, (10-38)
, !il(I--P) , lil(l--il)
p--ZaI2~ n ,;p:5P+Zat2~ n . (10-39)
P
- -- Z
a',
~(l=iij.:5 p, (10-40)
, n
P
,;'p+ za~IIP'(111-- P') (10-41)
~~i?j~:!.t;.
In a random sample of 75 axle shafts, 12 have a surlace finish ~t is rougher than the specifications
will allow. 'Thetefore, a point estimate of :he proportion p of shafts in the population that exceed :he
roughness specifications is p:::::x/n:::;: 12nS =0.16. A 95% two-sided Confidence interval for p is CQIlJ.-
puted from equation 10-39 as
or
which simplifies to
0.08';;p:; 0.24.
10-2 Single-Sar:::ple Confidence Interval Estimation 241
This function is relatively flat from p:::: 0.3 to p = 0.7. An estimate of p is required to use
equation 10-42. If an estimate p from a pre-vious sample is available, it could be substituted
for p in equation 10-42, or perhaps a subjective estimate could be made. If these alternatives
are unsatisfactory, a preliminary sample could be taken, p computed, and then equation
10-42 used to determine how many additional observations are required to estimate p with
the desired accuracy. The sample size from equation 10-42 will always be a maximum for
p = 0.5 [that is, p(1 0.25], and this can be used to obtain an upper bound on Il. In other
words, we are at least I OO{ 1 - a)% confident that the error b estimating p using p is less
than E if the sample size is
n=(Z;'2 rrO.25).
1n order to maintain at least a 100(1 - a)% level of confidence the value for n is always
rounded up to the next integer.
!~3#p!~i()j8
Consider ihe data in Example 10-17. How large a sample is required if we want to be 95% con5dent
that the error in usingp to estimate p is less than 0.05? Usingp = 0.16 as an initial estimate of p, we
fiLd from equation 1042 that the required sample size is
We note that the procedures developed in th1s section depend on the nonnal approxi~
mation to the binomiaL In situations where this approximation is inappropriate, particularly
cases where n is small, other methods mUSt be used. Tables of the binomial distribution
could be used to obtain a confidence interval for p. If n is large but p is smaIl, then the Pois~
SOn approximation to the binomial could be used to construct confidence intervals. These
procedures are·illustrated hy Duncan (1986).
Agresti and Coull (1998) present 1I!1 alternative form of a confidence interval On the
population proportion, p, based on a large-sample hypothesis test on p (see Chapter 11
of this text). Agresti and Coull show that the upper and lower limits of an approxnnate
100(1 - a)% confidence interval on p are
'
". Z;;/2.... I' A( ')
.pl-p ,
Z;j2
p---_z
2n
/'l
0, .. ,
1---"'-,-
tz. 4n~
Z;!2
1+-'-
n
The authors refer to this as the score confidence interval. One-sided confidence intervals
can be constructed simply by replacing Zafl with Z".
,
242 Chapter 10 Parameter Estimation
To illustrate tins confidence interval, reconsider Example 10-17, whlch discusses the
surface finish ofa shaft with n =75 and p = 0.16. The lower and upper limits of a 95% con-
tidence interval using the approach of Agresti and Coull are
1+ (1.96)"
75
0.186 ± 0.087
=-
1.051
0.177 ±O.083,
The resulting lower- and upper-confidence limits are 0.094 and 0.260, respectively.
Agresti and Coull argue that the more complicated confidence interval has several
advantages over the standard large~sample interval (given in equation 10-39). One advan-
tage is that their confidence interval tends to maintain the stated level of confidence better
than the standard large-sample interval Another advantage is that the lower-confidence
limit: will always be non-negative. The J.a.."ge~sample confidence interval can result in neg-
ative lower-confidence limits, which the practitioner will generally then set to 0, A method
which can report a neg~tive lower limit on a parameter that is inherently non~negative (such
as a proportion, p) is: often considered an inferior method. Lastly, the requirements that p
not be close to 0 or 1 and !l be relatively large are not requirements for the approach sug-
gested by Agresti and CoulL In other words, their approach results in an appropriate confi-
dence interval for any combination of nand p.
is standard nonnal if X; andX, are normal or approximately standard normal if the condi-
tions of the Central Lintit Theorem apply, respectively. From Fig, 10-5, tins implies that
P{-Z"",;Z :>Z"nJ: 1- a
or
10-3 Two-Sample Confidence Interval Estimation 243
- - 100t
I 2 (y.,2 '}
~Xl -Xl +Za/l'\I-+~ =l-a. (10-43)
I nJ "'z
Comparing equations 10-43 and 10-18, we note that the 100(1 - a)% confidence interval
for 1'1 - 1"1 is
I 2 2 C:r--2
XJ-Xo-Z"
- al""'I1l!l+'" ""J-"-$X
,- ,-'1. 1 -X,+ZI
.. a 2'11£::..+"2, (10-44)
l~ ~ . \~ ~
One-sided confidence intervals on 1'1 - 1"1 may also be obtaiued, A 100(1- a)% upper-
confidence interval on 1'1 - 1"1 is
(10-45)
Tensile stre:lgth tests were perfoIDlcd on two different grades of alumint:.Il1 spars used in manufac-
turing the v,.ing of a commerciai t:rn:osport aircri.L~ From past experience with the spar manufacturing
process and the testing procedure. the standard deviations of tensile strengths are assumed to be
knOWll. The data obtained are shov,'U in Table 11)...2.
If fl.l and p-:. denote tile true mean tensile strengths for the two grades of spars, then we may 5nd
a 90% confidence ioterval on the difference in mean strength #1 - pz .as follows:
"-,--,-
- -;t.z
L ""';T;. - - Z r>
10'1 0',
1-+-
• a, .. ~ nl n.z
1,(1-0-',;;---(,-5,7,
=87,6-74.5-1.645)-'1_+..:::...L
i 10 ;2
=13,1-0,88
=12,22 kglmm',
If the standard deviations 0', and 0'2 are known (at least approximately), and if the sam·
pIe sizes n l and ft:t. are equal (n! ;;;; nz : : n, say), then we can determine the sample size
required SO that the error in estimating Iii - ,Uz using XI - X2 will be less than E at 100(1 -
a)% confidence, The required sample size from each population is
S;
and are estimates of the commOn variance 0", we may obtain a combined (or "pooled")
estimator of 0":
To develop the confidence interval for Jil - fLJ.t note that the distribution of the
statistic
ri-~1 '
+ t"/2"
I. f'"1
+,,_2S, 1-+- = I-a,
711 ~ J
~ (10-49)
Therefore) a 100(1- a)% two~sided confidence interva: for the difference in means Pi - Jl'J.
is
- - 11 1
$.J11 ~f.l2 ~XI-X2 +ta/21'1 +1I~~2Sp.,.i-+-. (10-50)
'I~'vnl~
(10-51)
(10-52)
In a batch chemical process used for etching printed circuit boards, two different catalysts are being
compared to de~r.mine whether they require differer:.t emersio!l times fo: re...l1oval of identical quan-
tities of photoresist material, Twelve batches were rn."l with catalyst 1, resti:ung in a sample mean
emersion time of x, ~ 24.6 minutes and a sample standard de..iation of s! = 0,85 minutes. Fifteen
batches were ron -with catalyst 2, resulting in a rce.a... emersioD time of ~ ;;: 22.1 minutes and a stan w
da.."; deviation of S2 "" 0.98 minutes. We will find a 95% confidence interval on the difference in means
246 Olapter 10 Parameter R~tim.ation !
PI - fI2, assuming that the standard deviations (or variances) of the two populations are equa::. The
pooled est.i.mate of the COIllllJ.on va..'iance is found using equation 10-48 as follows:
, (n,-I),; .;.('" -1)$;
s = -'-'--'--'--'-'-:-'-'-
p "1+"2- 2
= 11(0.85)' H4{O.98)'
12.,.15-2
=0.8557.
The pooled standard deviation:is sf! =JO.8557 =0.925, Since lQ/2.n,+".,....2 til1lZ5.2:i "" 2,060, we may
calculate the 95% lowet:~ and upper-confidence limits as ' -
/1 I
=24.6-22.1-2.060(0.925).,-+-
112 15
= 1.76 minuteS
and
t _ X, X, -(11, - Jl,)
'21 2 '
..js!!", +S,/nz
is distributed approximately as t with degrees of freedom given by
2 I ,2
Si in, +S21nz I
.., !
(
(10-54)
10-3 Two-Sample Confidence mre!>,al Esti.-nation 247
Upper (lower) oDe-sided confidence limits may be found by replacing the lower (upper)-
confi.dence limit with -00(00) and changing cd2 to a.
JO-2.2 to find the confidence interval on ,uo' By analogy with equation 10-30, the 100(1 -
a)% confidence interval on I1D = 11; - i11 is \,
(10-55)
where 15 and SD are the sample mean and sample standard deviaoon of the differences D"
respecovely. This confidence interval is valid for the case where (j~"* because So esti- 0';,
mates a; n
~ veX, X,). Also, for large samples (say 2: 30 pairs), the assumption of nOr-
mality is unnecessary.
We now return to the data in Table 10-3 concerning the time for n =: 14 subjects to parallel park two
cars, From the column of observed differences d, we calculate d:= :.21 and sa=- 12.68. The 90% Con-
fidence interval for f.tn:=: J.1, - P2 is found from equation 10~55 as follows:
Note that when pairing data, degrees of freedom are lost in comparison to the two-sample
confidence intervals. but typically a gain in precision of estireation is achieved because 8tl
is smaller than Sp'
Hence
1O~3 Two-Sar:\.ple Confidence Interval Estimation 249
Comparing equations 10-56 and 10-18, we see that a 100(1 - a)% two-sided confi-
dence interval for ~/a; is
sf c;f ~2
si Jli-CZ/2,/l.l-l./I.:-l:5; O'~ s:: s:f Fa:tZ,i'll-!:./l.l-l' (1O-57)
where the lower 1 - aJ2 tail point of the F"r'tJ:!!-; distribution is given by (see equation
9-22).
1
(10-58)
FI)./2,fI,l -l,n:-t
(10-60)
;:[~E~!,'i~2,!
Consider the batch chemical etching process described in Example 10-20. Recall that two catalysts
are being compared to measure their effectiveness in reducing emersioD times for printed circuit
boards. n; 12 batches were run with catalyst 1 and ll:z = 15 batches were rJ.!l with catalyst 2, yield-
ing $! = 0,85 mL.'lutes and $1 :;;;;; 0,9S minutes. We will:find a 90% cor.iidence interval on the ratio of
va.ri3x.\ccs ~!o;. Froro equation 1!)-'57. we find that
$1
2
ar $~
-rro,9)J4JI :S:-2 $21'0.05,14,11>
of2 0'2 $2
usmgthe factiliatFQ,95.14,H = lIFo.6~,II.A "'" If2.58 =0.39. Since this confdeace interval includes unity,
we could not claim that the staIl.d.ard deviations of the emersion times for the two catalysts are differ-
ent at the 90% level of confidence.
•• Z ,1,0,(1-,0,)+,02(1-,02)
P2 - a!2~i
PI -
I n, ""
r;::---c-.~-;
• • + Zo/2. ,p,(I- PI) . -,,0.2",-(I~p2,,-)
,,; p, - P2 ,,; PI - p, (10-61)
~ nl ll:2
An approximate 100(1 - a)% lower-confidence interval for p, - P2 is
••
PI-PZ -
Za Vlp'(I-p;)+p2(I-fi~L
-·~--~Pl P2' (10-62)
n1 n:z
and an approximate 100(1 - a)% upper-confidence interval for p, - p, is
.-
p~ P2
~n.-'
~ P2
Yl
+Za, ,lp,(I-p,Y~I-':'~
' . (10-63)
1", n.,
;:~~~~~19;~:;
Consider the data in Example 10-17. Suppose that a modification is made in the surface finishing
process and subsequently a second random sample of 85 axle shafts is obtained, The number of defec-
tive shafts.in this second sample is 10. Therefore. since!t: ;; 75, Pt 0,16, ~ = 85. and /)2:::: 10/85:;:;:;
0,12, we can obtain an approximate 95% co:nfidence interval on the difference in the proportions of
defectives produced under the !\Va processes from equation 10-61 as
10-4 Approx.ima~e Confidence Intervals in Max.imum Likelihood Esti.:nation 251
or
0.16_0.12_1.96.1°.16(0.84) + 0.12(0.88)
,75 S5
r, '~~=C::: ...
:> Pl _ p, <016-012
_.
1961°,16(0.84) 0.12(0,88)
, + , 'i 75 + ll5 .
This simplifies to
- 0.07 s: P, - Pz'; 0.15,
This U:.rervaJ includes zero, so, based on the sample datl, it seems unlikely that the changes made in
the st:.rface finish process have reduced the proportion of defective axle shafts being produced..
~,~iij!J~~
Recall Example 10-3, where it was shoW!;; that the maximum likelihood estimato= of the parameter p
of a Beruoulli distribution is p= (lIn.)""
Lr",l
XI = X. Using the Cr:un6:'-Rao lower bound, we may
verify that the lower bound for the variance offt is
"),,
V\P - 1 ,
1~ 10[pX(I- Pj'- l] X
1
- J~_ (I-X)]'
'~lp \l-p)
1
nE[X2 + (I-X)' _2;((I-Xl]'
p' (I-p)' ;(I-p)
For the Bernoulli distribution, we observe thatE(X) =P and E(X,) =p. Therefore, this last express:on
simplifies to
V(p\"
, ) [1 1 1
n-+--
P (l-p)
This result should not be surprising, since we know directly that for the Bernoulli distribution, V(X)
= VeX,)/. =p(l- p)/n. Jn any case, replacingp in VCP) oy p, the approxi=te 100(1- a)% contitlence
interval for p is found from equation 10-04 to be
252 Chap~er 10 Parameter Estimation
where 1 (Xi is the confidence level used in the ith confidence interval. In practice, we select
a yalue for the simultaneous confidence level 1 - a, and then choose the individual ai such
that ~ ~ at:::: a. 1.Jsually. we set a". = aim.
~I=l
As an illustration, suppose we wished to construct two confidence intervals on the
means of!VIo normal distributions such that we are at least 90% confident that both state-
ments are simultaneously correct. Therefore, since 1 - a 0.90, we have a:::; 0.10, and
since two confidence intervals are required, each of these should be constructed with
= ,=
a,= a12 0.1012 =0,05, 1,2. Thatls, two individual 95% confidence intervals on 1', and
III will simultaneously lead to correct statements with probability at least 0,90.
the Bayes kterval provides a subjective degree of belief about IJ radler tila:! a frequency
e
interpretation. The Bayes interval estimate of is affected by the sample results but is not
completely determined by them.
Suppose that the random variableXis normally distributed \:.-1m meanjJ. and variance 4. T..'1e value of
jJ. is :.mknOVl!l. but a reasona.ble prior density wocld be normal \Villi mean 2 and variance L Tha: is,
- 8 1'}
-+1)' ,J!2 r I(
eXP1-- '(
'::+1 j.i.- nx+ I
L 2 4 j n+4 j
the methods of Section 10-1.4. Thus, the posterior distribution for jJ. is normal with mean
S)/(r. + 4) and variance 4!(n "7 4). A 95% Bayes interval for jl, which is symmetric about the
posterior mear.;, would be
nX-8 2 nX+8 2
---z".'2S···~"J!~--+z".c" (10-66)
n+4 "Vn+4 n-4
If a ra:odom sample of size 16 is taken and we:5.nd that X:::;; 2.5, equation 1O~66 reduces TO
1.52~J!"3.28.
If we ignore the prior information, the :::lassical confidence interval for jJ. is
We sec tb.at the Bayes i.:.'"lterval is slightly shorter than the classical co:r6dence i..'"lterva1, because the prior
:.riormation is equiva1en~ to a slight increase in lle sample size if ne prior knOWledge was 3.';sumed.
L
254 O1apter 10 Parameter Estimation
A"l electronic device consists of four components. The time ~o failure for each component follows ar.
exponential distribution and the co:nponents are identical to and bdependen! of one another. The
elcctronic device will fail only after all four components have failed. ine times to failure for the elec-
tronic components have been collected for lS such devices. The total times to failure are
distribution and is defined as gamma(r, It). Therefore, r=4, bllt A. needs to be estir:lated. A bootstrap
estimate for Acan be found using the technique given in Section 10-1.6, Using the time-to·failure data
above. we find the average time to faibrc to be x::; 42.545. 1he mean of a gamma distribution is E(X)
""" riA 2.I1d A is calculated for each bootstrap sample. Ruuui.ng M:nilablS for B """ 100 bootstraps, we
found that the bootstrap estimate X"", 0.0949. Using the bootstrap estimates for each sample, the dif-
ferences can be calculated and some of the calculations are showr. in Table 10-4.
\\Then the 100 differences are arranged in increasing order, the 5th percentile and the 95th per~
centiles turn out to be -0.0205 and 0,0232, respectively. Therefore, the resulting confidence limits are
We are approximately 90% confident that the :rue value of Alles bet\1.'cen O.fYl 17 a:.d 0.1154. Figure
10-10 displays the histogram of the bootstrap estimates 1~ while Fig. lQ..ll depicts the Cift"'erences
A: - T. The bootstrap esf..JIlates are reasonable when the estimator is unbiased and the stmdm"d eITO!
is approximately constant.
r
20
- r-
-- I- ~
-
o r il--h D
0-
20
r- r--
-
-- _r-r-
o r-
,
Il-h
-0,03 -0,02 -0,01 0,00 0.01 0,02 0,03 0,04
differences
Figure 10~11 Histogram of the differences):.; - J'
aDd
Since Xltt1 and Xare independent, normally distributed random variables, the prediction
error is also Donnally distributed and
Z= \
iX,-X)-O
1\'1" _______ •
I t' 1'\
~
o-'l1+-j
n;
and is standard nonnal, If cT is unknown, it can be estimated by the sample variance, S',
and then
By rearranging the inequality we obtain the final fonn for the two-sided 100(1 - a)% pre-
diction interval:
-
X-ta./'l,ll-l;JS
(
I' 2 l'' -
i ,.
ll+-;;)::;XIl+~S;X+tai2.II-I\S 2I I)
(1+;. (10-67)
- f,(!'1
X-t~'"'l';S ll+~)5Xr.+l' (10-68)
(10-69)
,!iEiii~!~I~~i
Ma..mum forces experienced by a transport airq:aft for an airline on a particular route for 10 flights
are (in units of graviry, g)
1.15,1.23,1.56,1.69,1.71,1.83,1.83,1.85,1.90,1.91.
10-8 Other Interval Estimation Problems 257
The sample average and sample standard deviation are cai.culared to be i """ 1.666 and s : : :. 0.273,
respectively. Rmay be of importance to predict the next maximum force experienced by the <L.."'Crnft.
Since !C.cG!!.9 = 2.262. the 95% prediction interv<u onX! 1 is
I
t"!2"~IJo.273)2(1 + 1 I" XII" 1.666+ t"/2'~1 JO.Z73)'(1-..!..).
i :.666 -
L018"X" ~2.314.
I 10, I W
Reconsider the maximum forces for the t:rar..spon nL"craft in Exru:Lple 10-27. A two-sided tole:ance
interval is desired that would cover 99% of all maximum forces with 95% confidence, From Tabk
XIV (A.ppendix), with I - a = 0.95, q ~ 0.99, and n la, we ftnd that k = 4.433, The sample aver-
age and sample standard deviation were calculated asx= 1,666 ands= 0.273. :espectively. The result-
ing tolerance intervalls then
1.666 ± 4.433(0.273)
or
(0.4 56,2.876).
Therefore, We conclude that we are 95% confident that at leaSt 99% of all maximum forc~s would lie
between 0.456 g and 2,876 g.
258 Chapter 10 Parameter Estimation
I 1 + P X~.4 (10-70)
n=-+-"-'--.
2 l-P 4
Thus, in order to be 95% certain that at least 90% of the population will be included
between the extreme values of the sample, we require a sample of size .
Note that there is a fundamental difference between confidence limits and tolerance
limits, Confidence limits (and thus confidence intervals) are used to estimate a parameter
of a popu1ation~ while tolerance limits (and tolerance intervals) are used to indicate the lim-
it, between which we can expect to find a proportion of a population. As n approaches infin-
ity, the length of a confidence interval approaches zero, while tolerance limits approach the
corresponding quantiles for the population. .
10-9 SlTh:l.MARY
This chapter has introduced the point and interval estimation of unknown pa.rnmeters. A
number of methods of obt.airring point estimators were discussed, including the method of
maximum likelihood and the method of moments, The method of maximum likelihood usu~
ally leads to estimators that have good statistical properties. Confidence intervals were
derived for a variety of parameter estimation problems. These intervals have a frequency
interpretation. The N/o-sided confidence intervals developed in Sections 10-2 and 10-3 are
summarized in Table 10-5. In some instance.'>t one-sided coclidence intervals may be appro-
priate. These may be obtained by setting one confidence limit in the two-sided confidence
interval equal to ttie lower (or upper) limit of a feasible region for the parameter, and using
a instead of cd2 as the probability level on the remaining upper (or lower) contidenco
limit. Confidence intervals using a bootstrapping technique were introduced, Tolerance
intervals were also presented. Approximate confidence intervals: in maximum likelihood
estimation and simultaneous confidence intervals were also briefly introduced.
Table 10~S Summary of Confidence Interval 'Procedures
Point
Problem Typ~ Estimator Two~Slded 100(1 a)% Confidence Interval
dl known
Mean II of a normal distribution, varianl;c X
.~~==--~--~~~~~----~=------
Difference in means of two nonnal distributions XI--X;: XI --Zajl II)
~ -
5; XI--X.1 +z'111
O'~
-+-
0';
)11 and 11.,.. variances O'~ and a~ known "1 11: '" 111
s'
Ratio of the variances ~/a~ oflwo
nomulI dislribulions
.-
s~
0'1
cr; s;
::;.-L;s:iF
S2
fJ.j1,~,-1,,,,-1
~
o
10-10 EXERCISES
10-1. Suppose we have a random sample of size 2n. 10~10. Find the estimator of /_ in the exponential dis~
from a population denoted X. and E(X) <= P and V(X) tribudon by the method of moments, based on a ran-
= cf. Let dom sample of size n,
10~11~ Find moment estL.'1lators of the parameters r
and ). of the gamma distribution. based on a random
sarr.ple of size n.
be two estimators of JL Vlhich is the better estimator
10~12. Let X be a geometric random variable with
of j.1.? E.~plain you:- choice,
parameter p. Find an estimator of p by the method of
10-2. Let Xl' X;., ..• , X1 denote a :-a.'1dom sample from moments, based on a random sample of size n.
a population having mean j.1. and variance <fl. Con- 10~13. Let X be a geometric random variable with
sider the following esti.n.tators of jJ.: parameter p. Fmd the maximum likelihood estimator
8, =~l~~;. _ _C, of p, based Or'! a random sample of size n.
10-14. Let X be a Bernoulli random variable with
parameter p. Find an es:::i:mator of p by the method of
moments, based on a random sample of size n,
Is either est.mator U:lbiased? \\!hich estimator is "bet- 10,,15. Let X be a binor::rial random v:a..-iable with
ter"? In what sense is it better? parameters n (known) and p, Find an estimator of p by
10-3. Suppos.e tha: 8, and &;. are estimators of the the method of moments, based on arandom sample of
paramet", 8. Weknow thatE(e,)~ e,E(fI,) = el2, Vee,) size N.
:= 10, and '\'(8;,) ;:: 4. \Ar'hid:. esti..::::lator is ·'ber-..er"? In 10-16. Let X be a binornia~ random variable with
what sense is. it better? parameters n and p, both unknown, Fmd es:imators of
n and p by the method of momen~, based on a random
104. Suppose tha: lil' ~, and t\ a..-re estimators of e.
sample of size N.
We ","ow lb." E(G,) = E{e.,) = e, E(B,) '" e, Vee,) = 12,
Veil,) 10, and E(i!, - 0' = 6. Compare these three 10-17. Let X be a. binomial random va..';able with
estimators. Whicb do you prefer? \\-by? Pl'Lrameters n (unlo:1own) and p. Find the maximum
likelihood estimator of p, based on a random sample
10,.5. Let three random samples of siz.es n1 := 10, n2 ;;;;;;
of size N.
8, and n~ = 6 be taken from a. populatioa willi mean J.1.
10~18. Set up the likelihood function for a random
and variance a'. Let S~, and S~ be the samp:e vari~
ances. Show t.'1a. sample of size n from a Weibull distribution. What
difficulties would be encountered in Obtaining the
82 _ lOSt . . . 8Sj -+ 6si maximum likelihood estimators of the three parame~
24 ters of the Weibull distribution?
is an unbiased estin:Ia{or of dl, 10~19. Prove that.if {j is an unbiased estimator of e,
10-6. Best Linear Cnbiased Estimators. An e.<:;tima.~ and iflirc,,-.lf'(e);;;;;; 0, then 8is a consistent estimator
e
tor is called a :;mear estim.ator if it is a linear combi~ of e.
nation of the observ-ations in the sample, 8is called a 10~20. Let X be a random ..mabIe with mean jJ. and
best linear unbiased cstL."l1ator if. of all linear func~ varianee d". Given two random samples of sizes n1
dons of the observations. it both is ur.biased and has and n2 with sample means XI and X;;l respe;;tive1y,
m.in.llnum variance. Show that the sample mean X is show ':hat
the bes;: linear unbiased estimator of the population
0< a < 1,
meanp.
:is an unbiased estimator of J1. Assuming X\ and Xz to
10M7. Find L,e t'laX.i.mum likelihood estimator of the
be ffidependent. find the value of a that minimizes the
parameter c of the Poisson distribution, based on a
va...""iance of X,
random sample of size n.
10-21. Suppose that the random variable X has the
10~8. Find the esti:nator of c in the Poisson distribu~
prohability distribution
tion by the method of moments, based on a random
sample of size n. fix) = ({+ 1)x1, o<x< 1,
10~9. Fmd the maximum likelihood e¢-timator of the =0, oth~'ise,
parameter A. in the exponential distribution, based on a Let X;, Xz, "', Xii be a random sample of size fl., Find
random sample of size !L the maximum likeJihood estimator of Yo
10-10 Exercises 261
10-22. Let X have :..'lJ.e truncated (on the left at x) expo~ 10-32. The time be'h'leen failures of a mili.i::ig machine
nential distribution is exponentially distributed with parameter A.. Sup-
fix) ~ A exp[-A(X - x,)]. x> x, > 0, pose we assume an exponential prior 0:1 A with a
mean of 3000 hOllrs. Two machines are observed and
;:: 0, otherwise.
the average t:i.:;;te betvleen failures is i "'" 3135 hou\'s,
Let X" X2 • ,." X" be a random sample of size n. Find Assumiag a squared-error loss, detem:U:le the Bayes
the ~:::\Um likelihood estimator of ).,. estimate of )..
1(}~23. Assr:me that ;t in the previous exercise is
10-33. 'The weight of boxes of candy is noonally dis
}:nmvn but xf is unknown. Obtain the maximum like-
tributed wi'll mean J1. and variance ~. It is reasoaable
I, lihood estimator of At'
to assume a prior density for il that is normal Vtith a
1(1·24. Let X be a random variable whh mean f1 and mean of 1O pounds and a variance of B" De:ermine the
'l2.Iia."lce r;il, and let X:. XZ7 .",X" be a random sample Bayes estir:1ate of jJ. gjven that 11 sample of size 25
of size n from X. Show thaI the estimator G produces i :0.05 po:mds. If boXes ':...1at weigh less
K",~lr .'
~i"'! ,Xi+l - Xi) is unbiased for .an appropriate than 9.95 pounds are defective, what is the probability
that defective boxes Viill be produced?
choice for K. Find the appropriate value for K.
10~25. Let X be a normally distributed random vari· 10-34. The number of defects that occur on a silicon
i able with mean 11 and variance UZ. Assume that rr is wafer used in integra:ed circult manufacturing is
:: known and Jl unknown. The prior density for ,u is known to be a Poisson random variable with patame-
asSUJUed to be normal with mean Jio ani! variance ~. let A. Assume that the prior density for A is eAl1onen~
Determi.."le the posterior density for 11, given a :andom ti.al \Vith a parameter of 0.25, A total of 45 defects
sample of size!t from X. were observed on 10 wafers:. Set up an integral that
defines a 95% Bayes. interva: for .t. "rna:: difficulties
1()"26. LetA,Zbe norm.illy distributed \Vi.th known mean would you er.counter in evaluating &.:is inregral?
Jl a!1d unknown variance al. A.~:''''::le that the prior
dcllSity for 11 cr is a gan:u:na distribution with parame w 10~3S, The random variable X has a density function
ter:s m + 1 and m~, Det:::m:rine the pOsterior density
f(xW)~2;. 0<x<8,
for ller, given a random sample of size n from X. e
10-21. Let X be a geometric random variable \Vith and the prior density for eis
pa.~ter p, Suppose we assume a beta distributon
J'(e)~l.o<e<1.
with parameters a and b as the prior density for p.
Detennine the posterior density for p, given a random eassUttling n: "" 1.
(a) Find th::: poslerior density for
sl!II1ple of size n from X. (b) Find the Bayes estimator for e assuming :he loss.
10~38. "When Xl' X2, ,,'. X~ a;e indepe~dent Poisson and G') = 0.18 fluid ounces for the tv.·o machines,
random variables, each 'With pa.'1UD.eter A, and when n respecti,,-ely. Two random samples of n.j ;:::: 12-'bottles
is relatively large, the sample mean X is approxi- from machine 1 and ~ = 10 bottles from machine 2
mately normal \\-1.th mean Aand variance Nn. are selected. and tie sample mean fill volUJX'.es are Xl
{a) What is the distribution of the statistic ,; 30,87 fluid ounces andJ"2 = 30.68 fluid ounCes,
X-I.? (a) Construct a 90% 'twt.rsided confidence interval on
~}.i;; . the mean difference in fill volume,
(b) Construct a 95% two-sided confidence interval on
(b) Use the results of (al 10 fud a 100(1- a)% confi· tk mean difference il1 fill volume. Compare the
dence interval for ;•. v.idth of this interval to the width of the interval in
10-39. A maD:Jfactu.rer produces piston rings for an part (a).
automobile engjne, It is known :hat ring diameter is (c) Construct a 95% upper-confidence interval on the
approximately normally distributed and bas standard mean difference in fill volume.
deviation Cf= 0.001 mm.A random sample of 15 rings 10~46. The burning rates of tviO different solid-fuel
has a mean diameter ofx;c: 74.036 mm. rocket propellan.ts are bcirtg studied. It is known that
(a) Construct a 99% two-sided confidence iutervalon both propellants have approximately the same standard
the mean piston ring diameter. de'.'iation of burning rate; that is. 0', = OJ::::::: 3 cmIs. Two
(b) Construct a 95% lower-confidence limit on the random samples of n1 :::::: 20 and 1t;;:::::: 20 spe~ens are
mean piston ring diameter. tested, and the sample mean burning rates are Xl ;;;.; 18
cmJs and ~ 24 cm/s. Construct a 99% confidecce
1040. The life in houts oia 75-Wlight bulb is known
interval on the mean difference:u:' burning rate,
to be approximately normally distribQted, "With a stan-
dard deviation CJ = 25 hours. A random sample of 20 10-47. Two different formulations of a lead~free gaso-
bulbs has a mean life of x:::::; :014 hou."'S. line are being tested to study their road octane mlJll w
and pyrotechnics), Bond strengths for 15 such materi~ two-sided confidence interval on the difference in
als ate'shown below, Constxuct a two-sided 95% con- mean Voltage.
fidence interval on the mean bond stre!l.gth.
lO~56. Random samples of SlZC 20 wc.."'C <l:-dwn from
323,312,300.284.283,261,207,183. two independent normal populations. The saxr,p1e
180,179.174.167.167,157. J20. means and standl:rd deviations were XI ':;:;: 22.0, .)1 ;::;
10·50. The wall thlckness of 25 glass 2·liter bottles
1.8. ~ ;. 21.5, and $1 = 1.5. Assuming that 0;, er: : : :
find the following:
was measured by a quality-control engineer, The salll·
(a) A 95% two-sided confidence inter'la: on!1l - ,1)..):
pIe mean was x 4.05 mm and the sample sta.r:.dard
deviation was s 0.08 mm. FInd a 900/0 lower-confi- (b) A 95% upper-confidence ir.tcrval ou Ji; fLl.
dence in!erVa: on the mean wall thickness. (c) A 95% lower-confidence interval 0!1 PI 1-0"
10-51. An industrial er:.gineer is in~erested in estimat- 10..51. The diameter of sreel rods m.a.nufactu.red on
[
ing the mean time required to assemble a printed cir~ two different extrusion machines is being investi-
cuit board, How large a sample is required if the gated. Two random samples of sizes n l ::: 15 alld n., ;;;;
engineer wishes to be 95% confident that the erro!' in 18 are selected, and the samp:e mea.'1S and sample
f
t
estiJnating the mean is less than 0,25 minutes? The variances are :i\ ;:: 8.73, s~ :;:;: 0.30. X1 = 8,680, and $~ :;:;:
standard dev:iat!ou of assem'Oly time is 0.45 minutes. 0.34. respectively, Assuming that oi == a-i. construct a
I 95% two~sided corlictence interval on the difference
I lO~52.- A random sample of size 15 fro:n a norma]
in mean rod diameter.
population has mean x 550 and variance S2:::::: 49,
Find the following; 10~58. Random samples of sizes n: == 15 and "2 10
(a) A 95% two-sided confidence interval on jl. are drawn from two mdependent Donnal populations.
(h) A 95% lower-confidence interval on ,u.
The sample means anc. variances are x; :::
300, if = 16.
~::: 325, s~ =49. Assuming mar o~ ",~. construct a
(c) A 95% upper-confide::.ce intm'3.l o~}1.
95% two-sided confidence interval On 11: - P?,
(d) A 95% two-sided prediction intm'3.l for a single
lO~59. Consider the data in Exercise 10-48. Construct
observation.
the following;
(e) A two~sided tolerance interval that would eover
(a) A 95% two-sided confidence interval on <fl.
90% of all observations with 99% confidence.
(b) A 95% loweN;orJ'idence interval on d.
10~53. An article in Computers in Cardiology (1993,
(c) A 95% upper-coDfictence interval On <f.
p, 317) presents the results of a heart stress test in
which the stress is induced by a particlliar drug. The 10-60. Consider the data in Exercise 10-49. Construct
heart rates (in beats per minute) of nine male patients the following:
after the drug is administered are recorded. The aver- (a) A 99% two-sided confidence int:en'al 00 <f.
age heartrute was found to be x= 102.9 (bpm) with a (b) A99% lower~confidc:lCeinterval on cr-,
sample standare. deviation of s :;:;: 13.9 (bpm), Fmd a
90% eonfidence interval on the mean heart rate after
(c) A 99% upper-confideI:.ee interval on cr.
:he drug is ad.tt1.h'iistered. 10~61. Construct a 95% two-sided confidence interval
on !he variance of the wall thickness data in Exercise
10R54. Two independent random samples of sizes n1 == 10-50.
18 and lIz = 20 are taken from two normal popula~
tions. The sample m.eans are Xl ::;:;: ZOO and Xz = 190. We 10~6.2. In a random sample of 100 light bulbs, the
know that the variances are 0"2 15 and O"~:::::: 12. Fmd samp:e standard deviation of bul'O life was found to be
the folJoW..ng: ' • 12.6 houtS. Compute a 90% -apper-confc.ence interval
on the variance of bulb life.
(a) A 95% rn,'o-sided confidence interva~ on)1: - f.L::..
(b) A 95% lower-eonfidence interval Ot! III - fLl. 10~63. Consider the data in Exercise 10-56. Construct
(e) A 95% upper-coufidenee interval on Jil - J.L;;. 3 95% tvlo-sided eonfidence interval on the rajo of
the population variances ti/~,
lO~S5. The output voltage from two different types of
10~64. Consider the data in Exetcise 10-51. Construct
transformers is being investigated, Ten transfonners
the following:
of each type are seleeted at random and the voltage
measured. The sample z:..eans 12.13 voltS and (a) A 90% two-sided confidence interval on ~/a;.
12.05 volts. We know that the variances of output (b) A 95% two~sided confidence interval on 0-71a;.
voltage fur the two types of transformers are O'~ = 0.7 Compare the width of this interval '1.-ith the 'tvidth
and 0; = 0.8, respectively. ConstrUct a 95% of the interval in part (a).
264 Chapter 10 Parameter Estimation l
(c) A 90% loweN!onfidenceinterval On ~/o;. Subject PSJ FSI
(d) A 90% upper-confidence interval on ~/a;,
25.9 33.4
10-65. Construct a 95% two-sided confidence interval 2 30.2 37.4
on the ratio of the variances ~Jo; using L'1e data in
3 33.7 48.0
Exercise 10-58.
4 27.6 305
10.-(i6. Of 400 randomly selected motorists, 48 were 5 33.3 27.8
found to be uninsured, Construct a 95% two-sided 6 34.6 27.5
confidence interval on the uninsu.red rate fer lnOtonS!$, 7 33.1 36.9
1CJ..67. How large a samp!c would be required in Exer- 8 30.6 31.1
else 10~66 to be 95% confident that ilie error in esti- 9 30.5 27.1
mating the uninsured rate for motorists is less than 10 25.4 38.0
0.037
10M6S. A manufacturer of electronic calculators is Find a 95% confidence interva: on t'1e difference in
interested in estimating the fraction of defective units mean completion times. Is there any indication that
produced. A random sample of 8000 calculators con~ one joystick is preferable?
tains 18 defectives. Compute a 99% upper...confidcnce
10~73~ The manager of a fleet of automobiles is test-
intcrval on ilie fraction defective.
ing two brands of radial tires. He assigns one tire of
10-69. A study .is to be conducted of the percentage of each brand at ~dom to the t\vo rear wheels of eight
homeowners who own at least two television sets. ca.."'S and runs the cars U!1til the tires wear out. The data
How large Ii sample is :required if we wish to be 99% (in kilometers) 3...'"C shown below:
confident that· the error in esti.:nating this quantity is.
less than 0.01?
Car Brand I Brand 2
10*10. A study is conducted to determine whether
there is a significant difference in union membership 1 36.925 34.318
based on sex of the person. A random s.ample of 5000 2 45.300 42.280
factory-employcd men were polled, and of this group, 3 36,240 35,500
785 were members of a union. A random sample of 4 32.100 31.950
3000 fac:ory-employed women were also polled, and 5 37.210 38.015
of this group, 327 were members cf a union, Con- 6 48.360 47.800
struct a 99% confidence interval on the difference in 7 38,200 37,810
proportions PI - P2' 8 33,500 33,215
10~11. The fraction of defective product produced by
two production lines is being analyzed. A random
sample of 1000 units from line 1 has 10 defectives, Find a 95 % confidence i."l.rerval on !he difference in
while a random sample of 1200 units from line 2 has mean mileage. Which brand do you prefer?
25 defeetives. F'md a 99% confidence interval on the 10~ 74. Consider the data in Exercise IQ..50. Find ecn~
differc:lce in fraction defective produced by the two fidenee intervals on J.l and (72 such that we are at least
lines. 90% confident that bofuinter\.'ai~ si.xru.li.taneously lead
to correct conclusions.
10~72. The resultS cf a study on powered wheelchair
driving performance were presented in the Proceed- 10~75. Consider the data in Exercise 10-56. Suppose
tngs of the IEEE 24th Annual Northeast Bfoengineer~ that a ~dom sa:npk of size n:; = 15 is obtained from
ing Conference (1998, p. 130). In this study, the a third nor:wU population. Vtit.'1 Xj::: 20.5 and 5J ::; 1.2.
effects of two types of joysticks, force sensing (FSJ) Find two-sided confidence intervals on #1 - IJ;;. f.ll -
and position sensing (pSJ), on pov,:er wheelchair f.lJ, and iJ...J. - f.lJ such that the probability is at least 0.95
control were investigated, Each of 10 subjects was that all three intervals. simultaneously lead to comet
asked to test both joysticks. One response of inter~"t is c(!!1c1~sicns.
the time (in seconds) to complete 3 predeter:rcined 10-76. A random variable X is normally distributed
couzse. Data typical of this type of experiment ax as with mean f.land variance a1 10. The prior density
follows: for J1. is uniform between 6 and 12. A random sample
10,10 Exercises 265
of size 16 yields x = 8. Construct a 90% Bayes inter~
density for lIcr.If a random sample of size 10 y:elds
va: for f.L Could you reasonably accept the hyPothesis
that J1 = 9? Z(x; - 4)2 =. 4.92, detenr,lne me Bayes estimate of
lJO"': asst.l.!tli.ng a squared-error loss. Set up aa integral
lO~ 77. Let X be a noI'Ill..'illy distributed random Ya.ri~ that defines a 90% Bayes interval for 1/62.
able with mean fJ. =: 5 and unknown variance d2, The
10-78. Prove that if a squared~ecror loss function is
prior dcr.sity for 11(12 is a gamma distribution with
used, the Bayes estimator of eis the lUcan of the pos~ ,
parameters r::;;; 3.and A= 1.0, Determine <he posterior terior distribction for 8.
Chapter 11
Tests of Hypotheses
Many problems requL""e that we decide whether to accept or reject a statement about some
parameter, The statement is usually called a hypothesis, and the decision-making procedure
about the hypothesis is called hypothesis testing. This is one of the most useful aspects of
statistical inference. since many types of decision problems can be formulated as
hypothesis-testing problems. This chapter will develop hypothe.~is~testing procedures for
severa} important situations.
11-1 INTRODUCTION
11-1.1 Statistical Hypotheses
A statistical hypothesis is a statement about the probability distribution of a random vari-
able. Statistical hypotheses often involve one or more parameters of this distribution. For
example, suppose that we axe interested in the mean compressive strength of a particular
ty'Pe of concrete. Specifically, we are interested in deciding whether or not the mean COm~
pressive strength (say /1) is 2500 psi. We may express this fonnally as
Ho' /1 = 2500 psi, (11-1)
HI' /1 ,,2500 psi.
The statemcntHo' /1 = 2500 psi in equation 11-1 is called the null hyporhesis, and the state-
*
ment HI: J1 2500 psi is called the alternative hypothesis. Since the alternative hypothesis
specifies values of /1 that could be either greater than 2500 psi or less thaI: 2500 psi, it is
called a two-sided alternative hypothesis. In some situations, we may wish to formulate a
one~sided alternative hypothesis, as in
266
11-1 Introdcccion 267
We are inrerested in making a decision about the truth or falsity of a hypothesis. A pro~
cedure leading to such a decision is called a test of a hypothesis. Hypothesis~testing proce-
dures rely on using the information in a random sample from the population of interest. If
this information is consistent with the hypothesis., then we would conclude that the hypoth-
esis is true: however, if this information is inconsistent with the hypothesis, we would con-
clude that the hypothesis is false.
To test a hypothesis. we must take a random sample. compute an appropriate test s.ta~
Listie from the sample d~ and then use the information contained in this test statistic to
make a decision, For example, .in testing the null hypothesis concerning the mean com-
pressive strength of concrete in equation 11 ~ 1. suppose that a random sample of 10 concrete
specimens is tested and the sample mean i is used as a test statistic. 1f x> 2550 psi or ifi
< 2450 psi, we will consider the mean compressive strength of this pa..--ticular type of con-
crete to be different from 2500 psi. That is, we would reject the null hypothesis Ho: j1 :
2500. Rejecting Ho implies that the alternative hypothesis, HI' is true. The set of all possi-
ble values of;;; that are either greater than 2550 psi or less than 2450 psi is called the criti-
cal region or rejection region for the test.A1ternatively, if 2450 psi s::x S 2550 psi. then we
would accept the null hypothesis Ho: j1 = 2500. Thus, the interval [2450 psi, 2550 psi] is
called the acceptance region for the test. Note t:hat the boundaries of the critical region,
2450 psi and 2550 psi (often called the critical values of the test statistic), have been deter-
mined somewhat arbitrarily. In subsequent sections we will show how to construct an
appropriate test statistic to determine the critical region for several hypothesis-testing
situations.
Sometimes it is more convenient to work with the power of the test. where
Note that the power of the test is the probability that a false null hypothesis is correctly
rejected. Because the results of a test of a hypothesis a..""e subject to error, we cannot '1Jrove"
or "disprove» a Statistical hypothesis. However, it is possible to design test procedures that
control the error probabilities (t and f3 to s"itably small values.
is True is False
l=eptH, NQ error Type II etro::
RejectH, lYre I error No error
268 C""pter II Tests of Hypotheses
The probability of type I error ais often called the significance level or size of the teSt
In the concrete~testing example, a type 1 error would occur if the sample meanx> 2550 psi
or if;' < 2450 psi when in fact the true mean compressive strength 11- = 2500 psi. Genernlly,
the type I error probability is controlled by the location of the critical region. Thus, it is USu-
ally easy in practice for the analyst to set the type I error probability at (or near) any desited
value. Since the probability of wrongly rejecting Ho is directly controlled by the decision
maker, rejection of Ho is always a strong conclusion. Now suppose that the null hypothesis
Ho: J1 =: 2500 psi is false. That is, the true mean compressive strength J1 is some value other
than 2500 psi. The probability of type IJ error is not a constant but depends on the true mean
compressive strength of the concrete. If jJ. denotes the true mean compressive strength, then
[3(!1-) denotes the type IJ error probability corresponding to J1. The function [3(!1-) is evaluated
by finding the probability that the test statistic (in this case Xl falls in the acceptance region
given a particular value of j.1., We define the operating characteristic curve (or OC curve) of
a tcst as the plot of [3(!1-) against J1.. An example of an operating characterisric curve for the
concrete-testing problem is shown in Fig. 11-1. From this curve, we see that the type II error
probability depends on the extent to which He: J1. = 2500 psiis false. For example, note that
[3(2700) < [3(2600), Thus we can think of the type II error probability as a measure of the
ability of the test procedure to detect a particular deviation from the noll hypothesis He.
Small deviations are harder to detect than large Ones. We also observe that since this is a
two-sided alternative hypothesis, the operating characteristic curve is symmetric; that is,
[3(2400) =[3(2600). Fu:ttherrnore, when 11- =2500 the probability of type II error [3 = I - a,
The probability of type II error is also a function of sample size, as iliustrated in Fig.
11-2. From this figure, we see that for a given value of the type 1 error probability a and a
given value of mean compressive strength the type 1I error probability decrca..~s as the sam~
ple size n increases. That is, a specified deviation of the true mean from the value specified
in the null hypothesis is easier to detect for larger sample size..I) than for smaller ones. The
effect of the type I error probability a on the type IJ error probability fl for a given sample
size n is illustrated in Fig. 11~3. Decreasing acauses f3 to increase, and increasing a causes
[3 to decrease.
Because the type IJ error probability [3 is a function of both the sample size and the
extent to which the null hypothesis Ho is false, it is customary to think of the decision to
accept Ho as a weak conclusion, unless we know that ,8 is acceptably small. Therefore,
rather than saying we <'accept H o'" we prefer the terminolQgy "fail to reject R o." Failing to
reject He implies that we have not found sufficient evidence to reject HOI that is, to make a
strong statement. Thus failing to reject Ro does not necessarily mean that there is a high
.,
'•.00
L~
t
___________ _
I
~(24CO) ~ ~(260C) .[---------- :
I I
~(23GO)~~(2700) ----- ---t---t--..L-- ::===-~
~7Co
Li
0,00 L ...=;;;;;:;;;;~=-----;;;;~-=::::~=_~
2500 /L
Figure 11~2 The effect of sample size on the operating characteristic curve.
1,00
0,00 '---"'---------:=;;-----="--~
2500 ~
Figure llR3 The effect of type I error on the operating characteristic curve.
probability that Ho is true. It may imply that more data are required to reach a strong cOn~
elusion. This can have .important implications for the formulation of hypotheses.
This would imply that the critical region is located in the upper tail of the distribution of the
test statistic. That is. if the decision is to be based on the value of the sample mean XI then
we would reject Ho in equation 11~6 ifi is too large. The operating characteristic curve for
the test for this hypothesis is shown in Fig. 114. along with the operating characteristic
curve for a two-sided test. We obsen'e that when the true mean j.J. exceeds J.ifJ (Le., when the
alternative hypothesis. Hi: f1. > J.i.fJ is true), the one-sided test is superior to the two-sided test
in the sense that it has a steeper operating characteristic curve. When the true mean j.J.:=: J.ifJ.
both the one-sided and two-sided tests are equivalent. However, when the true mean j1Is
less than /4. the two operating characteristic curves differ. If j1 < ,l.l.w the two-sided test has
a higher probability of detecting this departure from f.1o than the one-sided test. This is intu-
itively appealing, as the one~sided test is designed assuming either that fJ. cannot be less than
f.1o or, if IJ. lS less than f.1o, that it is desirable to accept the null hypothesis.
In effect there are two different models that can be used for the one-sided alternative
hypothesis. For the case where the alternative hypothesis is HI: J.l > J1.o, these two models are
H 0: jJ. ~ f.1o,
(11-7)
Hoc> f.1o
and
Ho' IJ. <;, f.1o, (11-8)
H t , IJ. > f.1o.
In equation 11-7, we are assn.ming that jJ. cannot be less than I1rr and the operating charac-
teristic curve is undefi.'1ed for values of j1 < ,Uo. In equation 11-8, we are assuming that fJ. can
be less than f.1o and that in such a situation it would be desirable to accept Ho. Thus for equa-
tion 11-8 the operating characteristic curve is defined for all values of jJ. <;, J.1.o. Specifically,
if IJ." .11u, we have (Jrjl) ~ I af,p), where af,p) is the significance level as a function of J1.
For situations in which the model of equation 11-8 is appropriate, we define the significance
level of the test as the maximum value of the type 1 error probability a; that is, the value of
a at J1. = ikJ. In situations where one~sided alternative hypotheses are appropriate, we will
usually write the null hypothesis with an equality; for example, Ho: jJ. ~ f.1o. This will be
interpreted as including the cases Ho' jJ. <;, f.1o or Ho' .u" .11u, as anropriate.
In problems where one~sided test procedures are indicated. analysts occasionally expe-
rienee difficulty in choosing an appropriate formulation of the alternative hypothesis. For
example, suppose that a soft drink beverage bottler purchases 10-ounce nonreturnable bot-
tles from a glass company. The bottler wants to be sure that the bottles exceed the specifi-
cation on mean internal pressure or bursting stren"cth. which for lO-ounce bottles is 200 psi.
~l
air------_
1.00 ;
OC curve for a
two-sided test
one--siaed test
o.oo'====---l--~2::=="-+-
"" "
Figure 11-4 Operatir.g characteristic CUJ."V'C,5 fortvvo-sided and one-sided tests.
11-2 Tests of HypoL'>eses on a Single Sample 271
The bottler has decided to formulate the decision procedure for a specific lot of bottles as a
hypothesis problem. There are two possible formulations for this problem. either
H,: fl." 200 psi,
(11-9)
Ht : fl.> 200 psi
or
Consider the formulation in equation 11-9. If the null hypothesis is rejected, the bot-
tles will be judged satisfactory~ while if Ho is not rejected, the implication is that the bot~
tles do not conform to specifications and should not be used. Because rejecting Ho is a
strong conclusion, this formulation forces the bottle manufacturer to "demonstrate" that the
mean bursting strength of the bottles exceeds the specification. Now consider the formula-
tion in equation 11- to. In this situation. the bottles will be judged satisfactory unless Ho is
rejected. That is, we would conclude that the bottles are satisfactory unless there is strong
evidence to the contrary.
Which fonnulation is correct, equation 11-9 or equation 11-1O? The answer is "it
deperu:ls;' For equation 11-9, there is some probability that II, will be accepted (i.e .. we
would decide that the bottles are not satisfactory) even though the true mean is slightly
greater than 200 psi. This formulation implies that we want the bottle manufacturer to
demonstrate thar the product meets or exceeds our specifications, Such a formulation could
be appropriate if the manufacturer has experienced difficulty in meeting specifications in the
past. or if product safety considerations force us to hold tightly to the 200 psi specification.
On the other hand, for the formulation of equation 11-10 there is some probability that II0
will he accepted and the bottles judged satisfactory even though the true mean is slighCy less
than 200 psi. We would conclude that the bottles are unsatisfactory only when there is strong
evidence that the mean does not exceed 200 psi; that is, when Ho: fl.2: 200 psi is rejected. This
formulation assumes that we are relatively happy wiD the bottle manufacturer's past per-
formance and that small deviations from the specification of fl.2: 200 psi are not harmful.
In formulating one-sided alternative hypotheses. we should remember that rejecting Ho
is always a Strong conclusion, and consequently, we should put the statement about which
it is important to make a strong conclusion in the alternative hypothesis. Often this will
depend on our point of view and experience with the situation.
A random sample of size n. Xl' X2• •.• , Xll~ is available. Each observation in this sam~
pre bas unknown mean J1 and known variance d'. The test procedure for He: J1 = !10 uses the
test statistic
(11-12)
If the null hypothesis He' J1 = !10 is true, then E(Y:; !10, and it follows that the distribution
of Zo is N(O, I). Consequently, if H,: J1 =!10 is true, the probability is 1 - a; that a value of
the test statistic Zo falls between - Zqfl and Zal2' where Za/2 is the percentage point of the
standard normal distribution such that P{Z;" Zan) = a/2 (i.e., Z"" is the lOO(l-a/2) per-
centage point of the standard normal distribution). The situation is lllustrated in Fig. 11-5 ..
Note ",lilt the probability is a; that a value of the test statistic Zo would fall in She region Zo
>Za!2 or 20 <-Za.'2 whenHo: fJ.::;:;.Uo is true, Clearly, a sample producing a value oftbe test
statistic that falls in the tails of the distribution of Zo would be unusual if Ho' jl = !10 is true;
it is also an indication that Ho is false, Thus, we should reject Ho if either
Zo>Za/2 (1l-l3a)
or
(ll-l3b)
and fail to reject Ho if
(11-14)
Equation 11~14 defines the acceptance region for Ho and equation 11-13 defines the criti-
cal region or rejection region. The type I error probability for this test procedure is ex,
The burning rate of a rocket propellant is being studied. Specifications req~ that the mean burr.in.g
rate must be 40 em/s, Furthermore. suppose that we know that the standard deviation of the burning
rate is approximately 2 cmfs, The experi:nenter decides to spe'cifY a type I errot probability a== 0.05,
and he will base the test 0;1 a random sample of size n == 25. The hypot.heses we wish to test are
He: Jl ~ 40 emfs.
H,:Jl*40cmls.
Twenty-five specimens are tested, and the sample mean burning rate obtained is i;:; 41.25 em/so
The value of u.1.e test statistic in equation 11-12 is
x- JJ.o
Zo:=! (J/~T;;
41.25-40
3.125.
2/.f25
Cr:tcal region
aJ2 ),(!il-----=-----!
Since a= 0.05. thebcundaries of the critical region arc Zo.m L96and~Zo.tT'..5 =-·1.96. and we note
that Zo falls in the critical region. Therefore. He is reJected. and we conclude mat the mea:: burning
rate is not equal to 40 cds,
Similarly, to test
Ho: Jl = 11<1, (11-17)
HI: j.l <!L:r
we would calculate the teststaostic Zo andrejectHo on values of ZO that are too small. That
is, the critical region is in the lower tail of the N(O, 1) distribution, and we reject Ho if
(11-18)
(11-19)
The distribution of the test statistic z" under both the null hypothesis Ho and the alternative
hypotbesisHl is shown in Fig. 11-6, From exam.irring this figure, we note that if Hl is true,
a type II error will be made ooly if - Zan S z" 5 Z<>/2 where :z;, ~ N( 15 .J;;/a, 1). That is, the
probability of the type II error f3 is the probability that z" falls between - Za12. and Z<>/2 given
that Hi is lrue. This probability is shown as the shaced portion of Fig, 11-6. Expressed
mathematically, this probability is
/ 15.[;;1
f3=<11 1 ZaP-a -<II -ZO/2-
r 0..[,;)' (11-20)
, / \ a
where <!>(z) denotes the probabilily to the left of z on the standard normal distribution. Note
that equation 11-20 was obtained by evaluating the probability that Zo falls in the interval
L
274 Cbapter 11 Tests of Hypotheses
Under H 1 : ji.*'1kz;
-z~'2 o
[- Za/2' Zari] 0:1 the distribution of Zo when Hi. is true, These two points were standardized ~'i
to produce equation 11-20. Furthermore, note that equation 11-20 also holds if Ii< 0, due
to the symmetry of the normal distribution.
While equation 11 ~ 20 could be used to evaluate the type II error, it is more eonvenient
to use the operating characteristic curves in Charts VIa and VIb of the Appendix. These
p
curves plot .as calculated from equation 11-20 against a parameter d for various example
sizes n. Curves are provided for both a= 0.05 and a = 0.01. The parameter d is defined as
dJI1- 1101 18 1
(11-21)
(J (J
\Ve have chosen d so that one set of operating characteristie curves can be used for all prob-
lems regardless of the values of f.1.Q and (5. From examining the operating characteristic
curves or equation IlM20 and Fig. 11-6 we note the follOwing:
1. The further the true value of the mean 11 fromllo, the smaller the probability of type
II error pfot a given n and ex. That is, we see that for a specified sample size and a..
large differences in the mean are easier to detect than small ones.
2. For a given 8 and a, the probability of type n error fi decreases as n increases. That
is, to detect a specified difference in the mean 0, we may make the test more pow-
erful by increasing the sample size.
CO!l.Sider the rocket propeBant problem in Example 11-1. Suppose that the analyst is concerned about
the probability of a type II error if the true mean burning rate is J1. = 4-1 em/s. We may use the operat-
ing characteristic curves to find {J. Note that 5=41 40 = 1, n::; 25. 0"= 2. and a= 0.05. Then
and from Chart VIa (Appendix), ?lith n = 25, we find that fJ= 0.30, That is, if the true mean bur.Ling
rate is J1.=41 em/s, the:J. there is approximately a 30% chance that this will nothe detected by the test
w:ithn=25,
;iixipipiiJ~~j
Once again, eo:.mder the rocket propella:ot problem in Example 11-1. Suppose that the analyst would
like to design the test so that if the true mean hurning rate differs from 40 r::tD!s by as much as 1 cr:::Js,
11-2 Tests of Hypotheses on a Sing:e Sample 275
the test will detect this (i.e" reject Ho: JJ. = 40) "''ith a high probability. say 0,90. The operati;:g char~
acteristic curves can be I.:.sed to:tmd the sample size that wi!! give Sl.:.ch a test. Since d;;;;:; lti - .14llkf=
112. a; 0,05. and fJ= 0,10, we find fron Chart VIa (Appendix) tlJat tlJe required sample size is n =
40. approximately,
In general. the operating characteristic curves involve three pru:ameters: j3. 8, and n,
Given any two of these parameters, the value of the thlrd can be determined. There are two
typical applications of these curves,
L For a given n and 8. find j3, This was il!ustratedin Example 11-2, This kindofproh-
lem is often encountered when the analyst is concerned about the sensitivity of an
experiment already performed. or when sample size is restricted by economic or·
other factors,
2. For a given j3 and ii, find n, This was il!ustrated in Example 11-3, This kind ofproh-
lem is usually encountered when the analyst has the opportunity to select the sam-
ple size at the out~et of the experiment
Operating characteristic curves are given in Cha.-ts vlc and VId (Appendix) for the
one-sided alternatives. If the alternative hypothesis is H t : J.L > /10. then the abscissa scale on
these charts is
(11-22)
When the alternative hypothesis is HI: J.L < /10> the corresponding abscissa scale Is
d=J1,~J1.
(5
(11-23)
It is also possible to derive formulas to determine the appropriate sample size to use to
obtain a particular value of j3for a given 0 and a. These formulas are alternatives to using
the operating characteristic curves. For the two-sided alternative hypothesis, we know from
equation 11-20 that
or if 0> 0,
(11-24)
since <p(
-Za/2 - 8.[; j (J) = 0 when 0 is positive. From equation 11-24, We take nor:nal
inverses to obtain
or
n= (11-25)
1
I
276 Chapter 11 Tests of Hypotheses
Returning to the rocket propellant problem of Example l:i.~3, we note that 0'=2,0=41-40;:;: 1, €X=::
0.05, a.'1d /3= 0.10. Since Zan. =2:,.f12S = 1.96 and Zfj =2:,.tO = 1.28, the sample size required to detect
this departure from He: J.l= 40 is found. in equation 11-25 :0 be
which is hJ. close agreement \\'ith the value determined from the operating characteristic curve, Note
that the approximation is good, since <I>(~Z.I2-0.rn/"')=<!>(-!'96-(I)..J4212)=-5.20=O
which is small relative to /3,
Ho: 8= eo>
H"e",eo
will lead to rejection of Ho if and only if 80 is not in the interval [L, UJ. As an illustration,
consider the rocket propellant problem in Example 11·1. The null hypothesis Ho' (1 = 40
was rejected using a= 0.05. The 95% two-sided confidence interval on (1 for these data may
be computed from equation 10·25 as 40.47 :0;(1:542.03. That is, the interval [L, UJ is [40.47,
42.03], and since .110 40 is not included in this interval, the null hypothesis Ho' (1 = 40 is
rejected.
P-Values
Computer software packages are frequently used for statistical hypothesis testing. Most of
these programs calculate and report the probability that the test statistic will take on a value
I
.~
1:-2 TesTS of Hyporbeses 0;1 a Single Sample 277
at least as extreme as the observed value of the statistic when Ho is true, This probabllity is
usually called a P~value. It represents the smallest level of significance that would lead to
rejection of Ho. Thus, if P=O.04 is reported in !he computer output, the null hypothesis Ho
would be rejected at the level a = 0.05 but not at the level a = O.O!. GeMrally, if P is less
than or equal to a, we would reject H0' whereas if P exceeds a we would fail to reject H o'
It is customary to call the test statistic (and the data) significant when the null hypoth~
esis Ho is rejected, so we may think of theP-value.s the smallest level aat which the data
are significant. Once the P-value is known, the decision maker can determine for himself
or herself how significant the data are without the data analyst formally imposing a prese-
lected level of significance.
It is not always easy to compute the exact P-value of a test. However. for the forego-
ing normal distribution tests it is relatively easy. If ~ is the computed value of the test sta~
tistic, then the P-value is
To illustrate, consider the rocket propellant problem in Example 11 -!. The computed value
of the test statistic is Zo = 3,125 and since the alternative hypothesis is two·tailed, the p.
value is
P = 2[1 - <1>(3.125)] =0.0018.
Thus, Ho: JL = 40 would be rejected at any level of significance where a:?: P ~ 0.0018, For
example, Ho would be rejected if a= 0.01, but it would not be rejected if a=O.OOl.
hypothesis tes~. Therefore, care must be taken when interpreting the results of a hypothesis
test when the sample sizes are large.
Statistical Analysis
Suppose that X is a normally distributed random variable with unknown mean f1 and vari~
ance cr.
Vole wish to test the hypothesis that f1 equals a constant 110, Note that this situation
is similar to that treated in Section 11~2.1. except that now both J.1 and if- are unlolOWD.
Assume that a random sample of size nj say Xl' Xl"'" Xfl.' is available. and letX and S2 be
the sample mean and variance, respectively.
Suppose that we wish to test the two-sided alternative
Ho: 11= f1v, (11-27)
H[:I1"".Uo·
The test procedure is based on the statistic
(11-28)
which follows the t distribution with n 1 degrees of freedom if the null hypothesis Ho: 11
: ; :. J.1o is true. To testHo: J.1::: 110 in equation 11 ~27. the test statistic Ie in equation 11-28 is cal-
culated, and Ho is rejected if either
(1I-29a)
or
(Il-Z9b)
where tall, 'I _ I and -tan.. n _ I are the upper aud lower Ct!2 percentage points of the t distri-
bution with n - I degrees of freedom.
For the one-sided alrernative hypothesis
Ho: 11 = /10, (11-30)
H,: 11'> /10,
r
! 11-2 Tests of Hypotheses on a Single Sample 279
The breaking strength of Ii textile fiber is a r.ormally distributed random variable. Specifcations
require that the :t;).can breaking strer.gth shQuld equal 150 psi. The manufact'.1ter would like to detect
any significant departure fro:.;::: this ..'alue. Thus, he wishes to test
no: /1 = 150 psi,
HI:).i;J:.lSOpsi.
A random sample of 15 fiber specimens is $elected and their breaking stre~gths determined, The sam-
ple mean and variance are computed from the sample data as i = 152.18 and 52 16,63. Therefo;:e,
the [est statistic is
'1-/1, 152,18-150
'0 = --,--;::=- "'" -r,=:".=- '= 2.07.
sj.,Jn ,,16.63/15
The type I error is specified as a = 0.05. Therefore t{W2S.14 :=: 2,145 and -toA)'l"s,!4 -2, V:S. and we
would CQucludc that there is not sufficient evidence to reject the hypothesis that JJ. == 150 psi.
G a
= So'
z+S~
= w (11-34)
i
L
280 Chapter 11 Tests of Hypotheses
1
Ii
nancentrality parameter o.JnICi. ~ote that if 0= 0, then the noncentral tdistribution reduces
ta the usual or central t distribution. In any case, the type II error of the two-sided alterna-
tive (far example) would be
/3 =P(- 'ai2•• _ 1S to S; 'ai2.n _110;< 0)
=P(-
where to denOtes the noncentral t random variable. Finding the type II error for the I-test
involves fmding the probability contained between two points on the noncenttal t
distribution.
The operating characteristic curves in Charts VIe, Vlf, VIg, and Vlh (Appendix) plot /3
against a parameter d for various sample sizes n. Curves are provided for both the two-sided
and one-sided alternatives and for 0; = 0.05 or a = 0.01. For the
two-sided alternative in
equation 11-27, the abscissa scale factar d on Charts VIe and VI!is defined as
d
181 (11-35)
For the one-sided alternatives. if rejection is desired, ie., )1 > Jl.rr as in equation II-3D, we
use Charts VIg and Vlh with
(11-36)
i~~tl1jiJ~_i1'.i:
Consider t..'le fiber-testing problem in Ex:ample 11-5. If the b;'eaking strength of this fiber differs from
150 psi by as much as 2.5 psi, the analyst would like to reject the null h)"pothes;s Ho: j.t = 150 psi v,rith
a probability of at least 0,90. Is the sample size n = 15 adequate to ensure that the test is lhis sensitive?
Ifwe usc the sample standard deviation s = --/16.63 =4.08 to est:imare cr, then d = 101/0"= 25/4,08 ='
0.61. By referring to the operatiJ::i.g characteristic curves in Chart VIe, with d = 0.61 and n = 15, we
find ~ = 0.45. Thus, the prabability of rejecting Ho: J.L; 150 psi if the true mean diffurs from this value
by ±2..5 psi is 1 - : 0.45 ~ 0.55, approxi.oJately. and we would conclude that a sample size of
v-! 11-2 Tests of HY;>otheses on a Single SampJe 281
= 15 is not adequate. To find the sample size required to give the desired degree of protection, enter
1:1
the operating characteristic curves in Chart VIe with d == 0.61 and f3 = 0.10, and read the con:espon~
ding sample size as 11. 35, approxbately,
ic
where S' is the sample variance. Now if Ho: d' = ,,; is true, then the test statistic follows
the chi-square distribution with n -I degrees of freedom. Therefore,Ho' d'=,,; would be
rejected if
(11-40a)
or if
;Co<X;-a/2,n-t, (lI-40b)
where X~, n _ 1 and i-
an., ~ _ 1 are the upper and lower ctJ2 percentage points of the chl-
square distribution with n - 1 degrees of freedom.
The same test statistic is used for the one-sided alternatives. For the one-sided
hypothesis
Ho: d' =<7~ (11-41)
H:: if- > O'~.
we would reject Ho if
, 2
Xo> Xa.r.~l· (11-42)
For the other one-sided hypothesis,
Ho' d'=<7:, (I !-43)
Hi: 02< 0';,
we would rejeet Ho if
(11-44)
282 Chapter 11 Tests of Hypotheses
Consider the machine described in Example 10-16, which is used to fill cans with a soft drink bever.
age. If the variance of the fill volume exceeds 0.02 (fluid our.cesf1, then an unacceptably large ~~
centage ofilie cans wi:l be underfilled. The bottler is interesred in testing the hypothesis
H,: d'; 0.02.
H,:<r> 0.02.
A randoCl sample of n.:::::: 20 cans yields a sample variance of il:::::: 0.0225. Thus, the test statistic is
(19)Q0225 21.38
0.Q2
Ifwe choose 0::.:::0.05, we find t.~at .ioll~l~ : : : 30.14, and we would conclude that there is no strong evi~
deuce tha: the variance offill volume e.,,<ceeds 0.02 (fluid ounces)2.
I..;~, (1145)
"0
for various sample sizes n, where O'denotes the true value of the standard deviation. Charts
VIk and \11 are for the one-sided alternative H t : d' > 0-;, while Charts VIm and \'In ardor
the other one-sided alternative HI: a'l < ~. In using these charts. we think of eras the value
of the standard deviation that we want to detect.
In Example 11-7. find the probability of rejecting Ho: a1 = 0.02 if the true va.,"iance is as large as d- "'"
0,Q3, Since 0' .jO,03:;;;O.1732and <To .J0.02:=O,l414~ea.bscissaparameteris
From Chart \'Ik, v.ith;1.= 1.23 and n 20, we find that {:J ""'" 0.60. ThaI is, there is only about a 40%
chance that Ho: d1 = 0.02 will be ::ejected if the variance is really as large as <T2 = 0.03. To reduce p.
a larger sample size must be used. From ':he operating characteristic curve. we note that to reduce f3
to 0,20 a sample size of?5 is necessary.
(11-46)
(11-48)
and we would reject Ho if Zo > Zan or if Zo < - Zan' The same test statistic would be used
for the one-sided alternatives. If we are testing
Ho: cf (l'~ (11-49)
HI: cf > 07"
we would reject Ho if Zo > Za> while if we are testing
Ho: cf = d,;,
( 11-50)
HI: cf < d,;,
we would reject Ho if Zo < -Za-
An injection-mo1ded plastic part is used in a graphics printer. Before agreeing to a lQng~term contract.
the printer manufacturer wants to be sure using a"'" 0.01 t."1at the supplier can produce parts vrith a
standard deviation of length of at most 0.025 m.aL The hypotheses to be tested are
Ho: d' 6.25 X 10-4,
H,' d' < 6.25 X 10-4.
smce (0,025)'2= 0,000625. Arandoro sample of n::: 50 parts is obtained, and t.1C sample standard devi~
ation is s = 0,021 I1t01. The test statistic is
Since ~Zom -2.33 and the observed value of Zo is not smaller than this critical value, Ho is aot
rejected. That is, the evidence from the supplier's process is not strong enough to justify a long-term
contract.
An approximate tesl: based on the nonnal approxi.-nation to the binomial will be given. 11:lis
approximate procedure will be valid as long as p is not extremely close to zero or 1, and if
the sample size is relatively large. Let Xbe the number of observations in a random sample
of size n that belongs to the class associated with p. Then, lithe null hypothesis Ho: P =Po
is true, we have X - N (npo. nPo(l- pf)). approximately. To testHo: p =Po calculate th~ test
statistic
(II-52)
Critical regions for the one-sided al.ternative hypotheses would be located in the usual'
manner.
A semieonductor firm produces logic devices. The contract with their custome!; calls for a fraction
defective of no more ilian 0.05. They wish to test
He' P = 0,05,
H,:p>0,05,
A random sample of 200 devices yields six defectives. The. test statistic is
6-200(0,05)
-1.30
~200(0,05)( 0,95)
Using a= 0.05. we f...1.d !hat ZV.C5:::: 1.645, and so we cannot reject the null h}-pothesis that? = 0.05.
(11-55)
(11-56)
These equations can be solved to find the sample .3ize n. that gives a test ofleve1 a that
has a specified f3 risk. The sample size equations are
(11-57)
For the situation described in Example 11"10. suppose that we wish to find the ,8 error of the test if p
;;; o.m, Using equation 1i-56. the f3 error is
( 0.05-0.07 +1.645.1(0.05)(0.951/200 \
f3=~ " ;
l ~(0.07)(O.93)/?OO )
= <1>(0.30)
= 0.6179.
This type II error probability is not as small as one might like. but ft "'" 200 is not'particularlY large and
0,07 is not very fat from the null value Po =< 0,05. Suppose that we wanl the f3 error to be no larger than
0.10 if:he true value of the fraction defective is as large as p;:; 0.07. The required sample size would
be found from equation 11-58 as
n = [L645~(O.05)(O.95) +L2ll~(O.07)(0.93)J'2
0.07 =0.05
= 1174,
which is a very large sample size, However, notice that we are trying to detect a very small deviation
fron:: the null value p, = 0.05.
286 Chapter 11 Tests of Hypotheses
_ _ It' a~)
Xl-X,-N, il,-iLl,-+-' .
aJ
\ nl llz
Thus, if the null hypothesis Ho: il, =iLl is true, the test statistic
ZO = ~~X"-",
V
J7n:f + <1~"2 (11.60)
follows the N(O, I) distribution. Therefore, the procedure for testing Ho: ill = iLl is to cal-
z;,
culate the test statistic in equation 11-60 and reject the null hypothesis if
(1I-6Ia)
or
Zo>Z.. (11-63)
To test the other one-sided alternative hypothesis,
Ho: il, = .u" (11-64)
HI: J1.1 <fl2.
use the test statistic Zo ill equation 11·60 and reject Ho; il, = iLl if
Zo<-Za- (11-65)
11-3 Tests of Hypotheses on Two Samples 287
...,,,.,S1'.,'-.'
Ex'ii.np'j"ii
{,."",/.",."".,,-,.,
2
The plant mar..ager of an orange juiee canni.....g facility is interested in co:rcparing the performance of
twa different production lines in her plant. AI; line number 1 is relatively new, she suspects that its ont~
put in number of cases per day is greater than the n1.Ullber of cases produced by the older line 2. Ten
xl
days of data are selected at random for each line, for which it is found that ='- 824.9 cases per day
and X::.:;:;; 818.6 cases per day. From experience with operaticg this type of equipmectlt is la:-ovm that
d, = 4() and 0; = 50. We wish :0 test
Ho: J1.t =p,.
H j:J1.j>p,.
The value of the test statistic is
Xj-.i'2 824.9-8,8.6 ~2.10.
20
0; ai
i -;;~. +--;;;
j /40 50
i 10 + 10
"Gsing a:= 0.05 we find that 21:!,os 1.645, and since 4 > 4.1)5' we would reject H J and condude that
t.'le mean number of cases per day produced by the new produetion line is greate! than the mean num-
ber of cases per day produced by the old line.
1111-1121 !81
d= rc;r~ai ~vl+ai' (11~66)
and one must choose equal sample sizes. say n ;;;; fl.l ::::; ~. The one~sided alternative hypothe-
ses requi::e the use of-Charts VIc and V1i For the one-sided alternative HI: 111 > P, in equa-
tian 11-62, the abscissa scale is
d (11-67)
I" 2 '
+ Ci2~(Ji
where n =r.l :;;; nz· The other one-sided alternative hypothesis, HI; Pl < f11., requires that d
be defined as
(11·68)
n- aT. +0'22
(11-69)
-ufj",+u?;,/",'
288 Chapter 11 Tests of Hypotbeses
Ii '" >' ".,. and lbeir values are fixed in advance, lben equation 11-69 is used directly to cal-
culate n. and the operating characteristic curves are entered 'With a specified d to obtain /3.
n,
If we are given d and it is necessary to determine and "., to obtain a specified f3, say f3*,
then one guesses at trial values of n 1 and i1:!, calculates n in equation 11 ~69, enters the curves
with the specified value of d, and finds Ii Ii f3 ~ f3", then lbe trial values of n l and"., are sat-
isfactory.If f3>' f3*. lben adjustments to ", and"., are made and the process is repeated.
Consider the orange juice production line problem in Exampk 11 ~ 12. If the true Cifferenee in ~a:u
production rates were 10 cases per day, find the sample sites required to detect this differ~::ce with a
probability of 0.90. The appropriate value of the abscissa parameter is
d 1',-1'2 10
1.05,
~(jf +O'i .)40+50
It is also possible to derive formulas for the sample size required to obtain a specified
f3 for a given 0 and IX These formulas occasionally are useful supplements to lbe operating
characteristic curves. For the two~sided alternative hypothesis. the sample size n! ; i1:! = n
is
n
(Z<X/2 + zl (<Y? 2
+ <Yn
(11-70)
0
This approximation is valid when <r>( -Za/2 - o~~-I~CJf + ~T) is small compared to {i
For a one-sided alternative, we have nl :::: ~ n. where
(11-71)
The derivations of equations 11-70 and 11-71 closely follow the single-sample case in Sec-
tion 11-2. To illustrate the use of these equations. consider the situation in Example 11-13.
a;
We have a one-sided alternative with a~ 0.05, o~ 10, = 40,0; 50, and f3 = 0.10. Thus
Za 20.05 = 1.645, Zp = 20.10 = 1.28, and the required sample size is found from equation
11-71 to be
develop the test procedure, but moderate departures from normality do not adversely affect
the procedure. There are two different situations that must be treated. In the first case, we
assume that the variances of the two normal distributions are unknown but equal; that is, .i
: : : a;
cr. In the second, we assume that ~ and a~ are unknown and not necessarily equal,-
Case 1: 01: 01: = cr Let X, and X, be two independent normal populations with
unknown means 11: and f.L2, and unknown but equal variances <17;;;: We wish to test a;;;:;: rr.
Ho: f.11 :;:::: P'2~
(11-72)
HI: /11 '" /12.
Suppose thatX w X12• ,.,. XliI) is a random sample ofn} obsex::ati~ns ffOmXl\andX::1. X22•
... , X"" is a random sample of,," observations from X" Let Xl' X" S;, and S; be the sam-
ple me~s and sample variances. respectively. Since both S~ and S; estimate the common
\"a..;ance <i-, we may combine them to yield a single estimate, say
t -
~-x.,•
0- II 1 (11-74)
Sp_I-+---
~ nl ""2
(lH5b)
we reject H~: f.Ll ,~.
The one-sided alternatives are treated si:nilarly. To test
Ho' /11 ~ /12, (11-76)
HI: /11 > /12,
compute the test statistic to in equation 11.-74 and reject Ha: /11 = /12 if
(11-77)
For the other one-sided alternative
Ha: /1, = /12, (11-78)
H ,: /11 < /12,
. Example 1(14
Two catalys':S are being analyzed to dete.."'!!line how t"ey affect the mean yield of a chemical process.
Specifically, cata!yst 1 is cu....-rently in use, but catalyst 2 is acceptable. Since catalyst 2 is eheaper, if i'I'
it does not change ti:e process yield, it should be adopted. S\;,ppose we v.ish to test the hypotheses
HO:!'I=i'?,
H,:!',"i'?.
PLot plant data yields nl =- 8, Xl)::;;; 91.73, s~ =3.89, nl. = 8. ~:= 93,75, and 4.02. From equation
IH3. we find
(7)3.89+ 7(4.02)
3.96.
h8-2
The test statistic is
91.73-93.75
2.03.
'1 1
1.99 1-+-
H S
Using tX= 0.05 we find:hat tM2'i.:4:::: 2.145 and-tC.CZ5, [4 =-2.145, and. conseqr.ently, Ho: #) = 111
cannot be rejected. That is. We do no: have srror.g c\'idence to conclude that catalyst 2 results in a
mean yield that differs fron:: the .rne:an yield when catalyst 1 is used,
a; u;: In some siruations, we cannot reasonably assume that the unknown vari-
Case 2: '#'
ances ~ and 0'; are equal, There is not an exact t statistic available for testing Ho: Pi ~ .'-'2
in this case. However, the statistic
•
tG=~~·~
Xl-X,
1St + si 01-80)
~ ", P..,
is distributed approximately as t with degrees of freedom given by
(St + SfJ'
v lnl n.z 2
(11-81)
(SUn,), (si. In.,J
-'-'-'-"''-- + .l..::..--c:..<_
n: +1 liz +1
if the nUL hypothesis Ho: 11, ~ i'? is true. Therefore, if ~" a;,
the hypotheses of equations
11-72.11-76, and 11~78 are tested as before, except that to is used as the test statistic and
n1 + 1lz - 2 is replaced by v in determining the degrees of freedom for the test. This general
problem is often called the Behren&-Fisher problem.
A manufacrurer ofYideo display units is testing two microcirc;rit designs to determine whether they
produce equivalent current floW. DevelOpment engineering has obtained the following data:
We wish to test
He: 1'1 = fJo,
HI:p'l:;LfJ.-?
where bot.;' populations are assumed to be normal, but we are unwilli."lg to asS\lJne that the un.1caown
variances ci:
and a;
are equal. The test statistie is
Using a;:::, 0.10, we find that t:d7..v:::::: to 05,J6 = 1.746. Since It~1 < t(,05.16' we cannot rejeet Hfi. III :!; /k}.,
d = '---'-_:.1 (11-82)
2a 2a
To use these curves, they must be entered with the sample size n* ::: 2.n - L For the one-
sided alternative hypothesis of equation 11-76, we use Charts VIg and VIh and define
d f.11-J.Lz=J.., (U-83)
2a 2a
whereas for the other one-sided alternative hypothesis of equation 11-78/ we use
!~~~Iil~l;;
Consider the catalyst expc...ment in Example 11-14. Suppose that if catalyst 2 produces a yield that
differs from the yield of catalyst 1 by 3.0% we would lil.--e to reject the null hypothesis .....ith a proba~
bility of at leas! 0,85. What sample size is required'? Using!ip 1.99 as a rO\1gh escrnlte of the
L
292 Chapter 11 Tests of Hypotheses
common standard deviation a, we have d = 1000za= 13.00i/(2)(1.99) = 0.75. From Chart VIe (Appen-
dix) with d;;;;: 0.75 and fJ= 0.15. we find n* =20, approximately. Therefore, since n" =2n - 1.
n .......l 20-!-1
n=-2-=-2-=1O.5 ~ll (say),
11-3.3 ThePairedt-Test
A special case of the two-sample t-tests occurs when the observations on the two popula~
lions of interest are collected in pairs. Each pair of observations, say (X:J• X1.j), is taken
under homogeneous conditions, but these conditions may change from one pair to another.
For example, suppose that we are interested in comparing two different types of tips for a
hardness-testing machine. Ibis machine presses the ti.p into a metal specimen with a known
force. By measuring the deplh oflhe depression caused by lhe tip. the bardness of the spec-
imen can be determined. If several specimens were selected at random, half tested with tip
I, half tested wilh tip 2. and lhe pooled or independent Hestin Section 11·3.2 applied. the
results of the test could be invalid. That is, the metal specimens could have been cut from
bar stock that was produced in different heats, or they may not be homogeneous. which is
another way hardness might be affected; then the obseNed differences betvleen mean hard-
ness readings for the two tip types also include hardness differences between specimens.
The correct: experimental procedure is to collect the dara in pairs; that is, to take two
hardness readings of each specimen, one with each tip. The test procedure would then con~
sist of anal)l'zmg the differences between hardness readings of each specimen. If there is no
difference between tips. then lhe mean of lhe differences should be zero. This test procedure
is called the paired t-test.
Let (Xu. X'l)' (XI2 , Xnl, ... , (X",. X,J be a set of n paired observations. where we
assume lhat XI .:: N(p.l' cr;)
and Xl - NCp., 0) Define lhe differences between each pair of
obsenrations as Dj =Xu - X2i' j = 1, 2, ...• n.
The D] are normally distributed v;.':ith mean
so testing hypolheses about the equality of III and '"'- can be aecomplished by perfonDing a
one~sarnple Hest on }J.D' Specifically, testing Ho: J1.~ = fJ? against HI: J.i.l ;t!; f-12 is equivalent
to testing
15
(ll-86)
where
(11·87)
J
11-3 Tests of Hypotb.eses on Two Samples 293
and
s'-D (11-88)
are the sample mean and variance of the differences. \Ve would reject Ho: /l.v =0 (implying
that Il,,,, 10 if to> Ian., <_lor if to < -1",'2" -1' One-sided alternatives would be treated
similarly.
l;E~~i~'lll!z!
An article in the Joul7Ul1 of Strain Analysis (Vol IS, No.2. 1983) compares several methods fur pre-
dicting the shear strength for steel plate girders. Data fur two of these methods. the Karlsruhe and
Lehigh procedures. when applied to nine specific girders. are shown in Table 11-2. We wish to deter-
mine if there is any difference (on the average) between the tvlo methods,
The sample aYera~e a.nd standard deviation of the differences d, are (1 = 0,2739 and Sd "'" 0.1351.
so the test statistic is - "
0.2739
0.1351/.J9 6,08.
For the twQ~sided alternative HI: 1lfJ;;t 0 and a= 0.1, we would fail to reject only if Itol < t~[},l, 3 """ 1.86.
Since to> to.05 , 5' we conclude that the two strength prediction methods yield d.:.fferent results. Spec:f~
ically. the KarhTUhe method produces, on average, higher stfCngt:!l predictio::tS than does t.!..e Lehigh
method.
Table 11·2 Strength P:edictions fur Nine Steel Plate Girders (Pre-
dicted LoarliObserved Load)
l
294 Chapter 11 Tests of HypotOeses
1
which is compared to tOIl. 211- 2- and of course, the paired t statistic is
15
to=S I"
Di ">In
the numerators of both statistics are identical However, the denominator of the two-sam~
pIe t-test is based on the assc.mption that XI andX2 are independent. In many paired exper-
iments, there is a strong positive correlation bet¥leen Xl and X2• That is,
v(15) = V(Xj - X2 )
= V(X:) + V(X,)-2Cov(X;, X,)
20-2 (1- p)
n
assuming that both populations Xl and Xz ha'!'e identical variances. Furthennoret S ~/n esti-
!llates the variance ofD. Now, whenever there is positive correlation within the pairs, the
denominator for the paired t-test will be smaller than the denominator of the t'No-sample
t-test This can cause the t\Vo-sampJe t-test to considerably understate the significance of the
data if it is incorrectly applied to paired samples.
Although pairing will often lead to a smaller value of the variance of Xl - X;.:, it does
have a disadvantage, Namely. the paired t~test leads 1.0 a loss of n.. - 1 degrees of freedom in
comparison to the two-sample t-test. Generally, we know that increasing the degrees of
freedom of a test increases the power against any fixed alternative values of the parameter.
So how do we decide to conduct the experiment-should we pair the observations or
not? Although there is no general answer to this question, we can give some guidelines
based on the above discussion. They ate as follows:
1~ If the experimental units are relatively homogenous (small a) and the correlation
between pairs is small, the gain in precision due to pairing will be offset by the loss
of degrees of freedom, so an independent-samples experiment should be used,
2. If the experimental uIUts are relatively heterogeneous (large a) and there is large
positive correlation betv.'een pairs, the paired experiment should be used.
The rules still require judgment in their implementation, because (j and p are usually
not \:nm>;ll precisely. Furthermore, if the number of degrees of freedom is large (say 40 or
50). then the loss of n - 1 of them for pairing may not be serious, However, if the number
of degrees offreedom is small (say !O or 20), then losing half of them is potentially serious
ifnot compensated for by an increased precision from pairing.
1
Pi - aj2'''1 -1,n2-1 (11-92)
The same test statistic can be used to test one-sided alternative hypotheses, Since the
notation Xl and X2 is arbitrary, let Xl denote the population that may have the largest vari~
ance. Therefore, the one-sided alternative hypothesis is
Chemical etching is used :0 remove >copper from. printed cir>cuit boards, X and X:t represen~ p::ocess
yields when twO different concentrations are used Supp¢se 6at we wis:::' to test
Ho: a;
Hj; vi:? 0;.
Two samples of sizes n t = liz """ 8 yield 37=3,89
L
296 Chapter 11 Tests of Hypotheses
If a::. 0.05. we find th~t Fo,q.t5. 7, 7"'" 4.99 and FQS75• 7, 7= (FruJ1j •7, 7r! "'" (4.99r I 0.20. Therefore,
0';.
'Ne ca.'1ll()t reject Ho: 0'1"'" and we can conclude mat there is no strong evidence that the variance
of me yield is affected by the concentration,
Choice of Sample Size
Charts v1o, VIp, v1q, and VIr (Appendix) provide operating characteristic curves for the
F-test for a= 0.05, and a= om, assuming that n, = "? = "- Cb,,'ts VIo and VIp are used
with the two-sided alternative of equation 11-89. They plot j3 .gabst the abscissa parameter
A="1 (11-95)
,1.=
0',
By referring to ChartVIo, with p;:::; 0.20 and 1.= 2, we find that a sample size of nl "'" fl;1::' 20, a:pprox~
imately, is necessary.
r,1 1
S 1--+-·
pOj 2n, 2n"
where Sf} is the pooled estimator of the common standard deviation (j, This statistic has an
approxiinate standard normal distribution when ,,; = 0-;. We would reject Ho if .z;, > Zan or
if~ <- Zc:.'Z' Rejection regions for the one~sided alternatives have the same form as in other
two~sample normal teSts.
We will present a large-sample procedure based on the normal approximation to the bino-
mial and then outline one possible approach for small sample sizes.
z PI-P2
(11-99)
If
(11-100)
the null hypothesis is rejected.
Two different types of.:fire control computers are being considered for use by the U.S. Army in six-
gun IDS-rom barteries. The two computer systems are subjected to an operational test in which the
total number of hits of the target are counted. Computer system 1 gave 250 hits out of 300 rounds,
while computer system 2 gave 178 hits out of260 rounds. Is there reason to believe that the two com-
puter systems differ? To answer this question, we test
Ho:PI=PZ,
Hl:Pl *pz·
Note thatp, =250/300 = 0.8333,p, = 178/260= 0.6846, and
250+178
0.7643.
300+260
The value of the test statistic is
If we use a = 0.05, then 20,025 = 1.96 and -ZO.025 = -1.96, and we would reject Ho, concluding that
there is a significant difference in the two computer systems.
298 Ch3j:;ter 11 Tests of Hypotheses
1
I
Choice of Sample Size
The computation of the f3 error for the foregoing test is somewhat more involved than .in the
single-sample case. The problem js that the denominator of z;, js an estimate of the standard
deviation ofpl - pz under the assumption that PI::;;;; P2 =p. \\'hen Ho: PI:::::' P2 is false, the stan~
dard deviation ofPl -pz is
(11-101)
(11-102)
where
and <lj;, _{h is given by equation 11-101. If the alternative hypothesis is H,: P, > P2' then
(lH03)
(11-104)
For a specified pair of values Pl and P2 we can find the sample sizes nl ""2 ~ n
required to give the test of size a; that has specified type II errOr fl. For the two-sided alter-
native the common sample size is approximately
(11-105)
where ql = 1 - PI and q2 = 1 - Pz' For the one~sided alternatives, replace Zd2 in equation
11-105 with2w
11-3 Tests of Hypotllese, on 1\vo Samples 299
HO:PI;p"
B,:p,>p"
Given that Xl + X:;:.;:;:; Y,large values of Xl supportH1• whereas small or moderate values of
Xl support Ho- Therefore, we will reject Ho whenever Xl is sufficiently large.
Since the combined sample of n1 + lLz obse...,ryations contains Xi + Xz : : : Ytotal successes,
if Bo: p, ; p" the successes are no more likely to be concentrated in the first sample than
n,
in the second, That is, all the ways in which the + '" responses can be divided into one
sample of nl responses and a second sample of fl-.2 responses are equally likely_ The number
of ways of selecting Xl successes for the first sample leaving Y - X: successes for the
second is
Because outcomes are equally likcly. the probability of there being exactly Xl successes in
sample I is detennined by the ratio of the number of sample I outcomes having Xl suc-
Cesses to the total number of outcomes. or
:~~§,p,!~~[@i~
InsulatirJ.g cloth used in printed circuit boards is manufactured in large rolls_ The manufacturer is try-
ing to irr;prove the process yield. that is, the number of defect~free rolls produced. A sample of 10 rolls
contains exactly four defect-free rolls. From analysis of the defect types, manufacturing cO$ineering
S:lggests S2VC:a1 changes in the p!"ocess. Fo:!lowing implerne::ttation of these chlUlges. another sample
of 10 rolls yields 8 defect-free rolls. Do the data support the claim thz:t the new process is bet::er tha.'1
the old one, using a:;:;Q,lO?
l
300 Chapter II Te:,'ll; of Hypotheses l
To aIlswer this question. we compute the P-vahle. mour ex.amp1e. nl """ns = 10, Y= 8 "t" 4= 12, and
the observed value of xl = 8, The values of Xi that are more extreme than 8 are 9 and 10, Therefore
12\,'8'
. ( 8 121
p(XJ :8112 successes): ({oy
=0.0750.
,10 J
(12Y81
19;\I;
P(XI =912successes
I ) =~=O.0095.
llO J
. I \
P(XJ :10 12 successes/= (20)
Ga~J :0.0003.
,10
The P-value is p: 0.0750 -;. 0.0095 + 0.0003 = 0.0848. Thus. at the level a: 0.10, ::he nulll:ypothe-
sis is rejected and We conclude that the e:.gineering changes have improved the process yield,
This test procedure i~ sometimes called the Fisher-lrviin test Because the test depends
on the asst::mption that Xl + X2 is :fixed at some value, some statisticians have argued against
use of the test when X, + X, is not actually fixed. Clearly X; + X,is notfixed by the sampling
procedure in our example. However, because there are no other better competing procedures,
the Fisher-lrviin test is often used whether or not Xl + X2 is acrually fixed in advance.
2_ *'
Xo - L,
[=1
(0, -Ed'
Ej
(11-107)
··Eiim·i.ii;2Z·
p ..; •.• .,
'Co .••
A Completely Specified Distribution A computer scientist has developed an algorithm for gener~
ating pseudorandom i...tegerS over :he inter.'al 0-9. He codes the algorithm and genera:es 1000
pseudo::andom digits. 'The data are shown ir. Table 11-3. Is :here evidence th~t the rz:ndom number
genera~or is working correctly'?
If the randoJ:J. number generator is working correctly, then the values 0-9 should follow the dis-
crete uniform distribution, which implies that each of the integers. should occur about 100 ti.D::.es. That
is, the e.xpecredfrequencies Ei = 100, for i "'" 0, 1, ... ,9. Since these expected f:.-eG,uencies ean be deter-
mined v.rifuout estimating any parameters from the sample c.ata, the resul:ing chi-situate goodness-of-
fit test will have k - p - 1 = 10 - 0 ~ 1 9 degrees of freedom.
The observed value of the test statistie is
J'
xl "" I. \o!. :.~
, i
J=; Et
(94-100)' + (93-100)' + ... +",(9_4",-",IO..c0),-'
100 100 100
= 3.72.
Since ioM.? = 16,92 we ~are unable to rejeec the hypothesis tlJ.at the data eome from a discrete uniform
distribution. Therefore, the :random numbet generator seems to be working satisfactorily.
E~~~.ti83
A Discrete Distribution The number of defects in printed cirenit boards ls hypothesized to follow
a Poisson distribution. A random sample of n 60 printed boards have been eollected, and the :mID-
ber of defects observed. The following data result:
H ') c-<'''(075''
,}
o:Pt x ;; .x! "=0,1,2.. ",
The expected frequencies are obtained by multiplying the sample size times the respective pmbabU-
lties. Since the expected frequency in the las~ cell is less than 3, we combine the last two cells:
The test statIstic (which will have k - P - t = 3 - 1 - t :;:; 1 degree of .freedom) becomes
and since X~~,l 3.&4. we cannot reject the hypothesis that the occurrence of defects follows a
Poisson dis~Jlution v,.itlt mean 0.75 defects per board.
J
11·4 Testing for Goodness of Fit 303
are equal Suppose we decide to use k;;; 8 cells. For the standard normal distribution the intervals :hat
divide the scale into e:gh' eq"alIy1£<ely segments are (0, 0.32), [0.32, 0,675), [0.675, 1.15), [Ll5, -l,
and thea: fow: "mir.::or image" intervals on the other side at zero. Denoting :hese standard normal end-
points by Go. at ...., as> it is a simple matter :0 calculate the endpoints that are necessary for the gen-
ernl normal problem at hand; :narnely, we define the new class interval endpoints by the transformation
a;::::::i + sa;; i;:;;;. O. 1, ,,,.8, Por example. the sixth interVal's right endpoint is
12.04 + (0.0&) (0.675) = 12.094.
For each interVal. Pi:::;; t:::::: 0,125. so the expected cell frequencies are El 100 (0,125) = 12.5.
The complete table of observed and expected frequenc:es is given IT. Table
The computed value at the chi-square statistic is
Si:1ce rNO parameters in t.l"e normal distribt.1ion have been estimated. we would compare ,to:::::: L12 to
a cru-s(j.uare disttibutio:l vrith k - P - 1 =: 8, - 2 - 1 : : : 5 degrees of freedom. Using a:::.. 0,10. we see
that X;,I...i = 13.36, and so we conclude that there is no reason to believe that output voltage is not nor-
mally distributed.
Probability Plotting
Graphical methods are also useful when selecting a probability distribution to describe
data, Probability plotting is a graphical method for determining whether the data conform
to a hypothesized distribution based on a subjective visual examination of the data. The
general procedure is very simple and can be performed quickly. Probability plotting
requires special graph paper, known as probability paper, that has been designed for me
hypothesized distribution, Probability paper is widely available for the normal, lognormal,
Weibull, and various chi-square and gamma distributions. To construct a probability plot,
the obsenrations in the sample are first ranked from smallest to largest. That is. the sample
Xl' X2 , .••• X,,,! is arranged as X(li' X(z), ...• X(Il» where X(f> ~ XU+ 1)' The ordered abserva~
tions XC;) are then plotted against their observed cumulative frequency U - O.5)ln on the
appropriate probability paper. If the hypothesized distribution adequately describes the
data, the plotted points will fall approximately along a straight line; if the plotted points
deviate significantly from a straight line, then the hypothesized model is not appropriate.'
Usually, the determination of whether or not the data plot as a straight line is SUbjective.
E~~i!e~~:;Z~(
To illustrate probability plotting, consider the followlllg data:
-{).314, 1.080,0.863, -{).179, -1.390, -{).563, 1.436, LlS3, 0.504, -{).801.
We hypothe:,ize that the$C data are adequately modeled by a normal distribl.!tion. The observations are
arranged in llscending order and their cumulative frequencies U- 0.5)/11, calculated as follows:
j 0.5) I n
-1.390 0.05
2 -{).801 0.15
3 -0563 0.25
4 -{).314 0.35
5 -{).179 0.45
6 0.504 055
7 0.863 0.65
8 L080 0.75
9 Ll53 0,85
10 1.436 0.95
The pall's of values ~)) and U - 0.5)1n are now plotted on normal probability paper; This plot is sbo\Vtl
in Fig. 11-7. Most normal probability ,?aper plots lOO(j - 0.5)/n on the right vertical scale a:::ld
100[1 - U - 05)ln~ on the left vertical scale. with the variable value plotted on tl:.e horizontal scale.
We have chosen to plot X,j) versus IOOU 0,5)1n On the right vertical in Fig. 11-7. A straight line.,
1I 99
2r 98
5~ 95
10 L 90
•
:[ 20 L SO
70 ~
"l
"~
I
:t
50 •
• SO 0
I
50 :::>
""
1:. 60
1:
~
"
~ 70
SO 20
90 10
95 ~ • -5
98~'~ ___ L-._~L- ___ ! I ..~. ~.~,_J2
-2,0 -1.5 -1,0 -0,5 0 0.5 1.0 1,5 2,0
XU:
Figure 11~7 Normal probability plot
J
11-4 Testing for Goodness of Fit 305
chosen subjectively, has been drawn through the plotted points. In drawing the straight line, one
should be influenced more by the points near the middle than the extreme points. Since the points fall
generally near the line, we conclude that a normal distribution describes the data.
We can obtain an estimate of the mean and standard deviation directly from the normal
probability plot. We see from the straight line in Fig. 11-7 that the mean is estimated as the
50th percentile of the sample, orjl = 0.1 0, approximately, and the standard deviation is esti-
mated as the difference between the 84th and 50th percentiles, or 0-= 0.95 - 0.10 = 0.85,
approximately.
A normal probability plot can also be constructed on ordinary graph paper by plotting
the standardized normal scores Zj against XU), where the standardized normal scores satisfy
j XCJ) (j-O.5)/n Zj
1 -1.390 0.05 -1.64
2 -0.801 0.15 -1.04
3 -0.563 0.25 -0.67
4 -0.314 0.35 -0.39
5 -0.179 0.45 -0.13
6 0.504 0.55 0.13
7 0.863 0.65 0.39
8 1.080 0.75 0.67
9 1.153 0.85 1.04
10 1.436 0.95 1.64
Figure 11-8 presents the plot of Zj versus XU). This normal probability plot is equivalent to
the one in Fig. 11-7. Many software packages will construct probability plots for various
distributions. For a 1vf!nitab® example, see Section 11-6.
2.0 , - - - - - - , - - - - - , - - - - - , - - - ,
1.0
-1.0
•
-2.0 :-:c-----,:-:c---;.------;:'-;;-------;;'
-2.0 -1.0 0 1.0 2.0
XU)
l
306 Chapter 11 Tests of Hypotheses
~ E(X-fl')'
, - =7--:-;-;':?f-
~ -,
"'y}Jl 2\3/2-
[<1 )'
14
{3, 5
c 6
~
.0
·c
:i
!: 7;
J!
9'
p,
Figure 11-9 Regions in the P" {3, plane for vmom standard di,ttibutions. (Adapted from G. I.
Hahn and S, S. Shapiro, Statisrical Models in Engineering, John Wlley & Sons, New York, 1961;
used with permission of th¢ publisher and Professor E, S, Pearson. Urri\'eTI)1ty of London.)
1l~5 Con:ingency Table Tests 307
and
, M.
/3 2 ; M"
• 2
where
M·;-
II'(X -X)'
_., j = 1,2,3,4,
J , '
n h:::1
and plot the point~" fJ.,. If this plotted point falls reasonably close to a point, line, or area
that corresponds 10 one of the distributions given in the figure, then this distribution is a log-
ical candidate to model the data.
From inspecting Fig. 11-9 we note that all normal distributions are represented by the
point /31 ; 0 and {3, ; 3. This is reasonable, since all normal distributions have the same
shape, Similarly, the exponential and uniform distributions are represented by a single point
in the /3" {3, plane. The gamma and lognormal distributions are represented by lines,
because their shapes depend on their parameter values. NOte that these lines are close
together. whieh may explain why some data sets are modeled equally well by either distri-
bution, We also observe that there are regions of the /3" {3, plane for which none of the dis-
tributions in Fig. 11-9 is appropriate, Other, more general distributions, such a<; the Johnson
Or Pearson families of distributions, may be required :in these cases, Procedures for fitting
thesefamilies of distributions lmd fignres similar to Fig, 11-9 are given in HaIm and Shapiro
(1967).
Column
Row 2 c
1 i 0:: 0 12__
2 ,0,,,,--+_O~""--t-- 0,
cell, given that the two classifieations are independent. Then Pij = Ui""j. where u/ is the
probability that a randomly selected element falls in row class i and Vj is the probability that
a randocly selected element falls in column class j. Now, assuming independence, the
maximum likelihood estimators of Uj and Vj are
,Ie
uj=-LO/i.
12 I""-l "
• 1 r
Vj =- 2..:0ij' (11·108)
n i=:l
Therefore, assuming independence, the expected number of each cell is
(11·109)
2 +~ (Oij-ES 2
Xo=-:..L E -X(~l)(C-;)' (11·1lO)
1""1/,,,1 ij
A company has to choose among three pension plans. Management wishes to know whether the pref~
crence forplan:s independent of job classification. The opinions of arandom sa:nple of 500 employ-
ees are shown in Table 11-6, We may compute ul ;:: (340/500} ". 0.68. ~ ::: (1601500) ::: 032, Vi ;:::
=
(2001500) 0040. il, = (2001500) 0040. and y) = (100/500) = 0.20. The expecTed frequencies may be
computed from equation ll-l09. For exan:ple, the expected number of salaried workers favoring pet.-
sinn plsn 1 is
11-6 Salnpi" COT.1puter Output 309
Pension Plan
2 3 Total
Saillried
workers 136 136 68 340
Hourly
workers 64 64 32 160
Totals 200 200 100 500
The expected frequencies a...-e shovro in Table 1l~7, The test statistic is computed from equation
11-110 as follom:
Since X~.O$;.! ;;; 5.99, we reject the typotbesis of independence and conclude that the preference for
pension plans is not independent of job classification.
Using the t<.vo-way contingency table to test independence between two variables of
classification in a sample from a single population of interest is only one application of con-
tingency table methods. Another common situation occurs when there are r populations of
interest and each population is divided into the same c categories. A sample is then taken
from the ith population and the counts entered in the appropriate columns of the illl row. In
this situation we want to investigate whether or not the proportions in the c categories are
the same for all populatioos_ The null hypothesis in this problem states that the populations
are homogeneous with respect to the categories. For example, when there are only two cat-
egoriest such as success and failure, defective and nondefective. and so on, then the test for
bomogeneity is really a test of the equality of r binomial parameters. Calculation of
expected frequencies, determination of degrees of freedom, and computation of the
chi-square statistic for the test for homogeneity are identical to the test for independence.
A study was conducted on the tensile strength of a pa.."ticular fiber under various temperr.:.tures. The
results of the study (given in :MFa) are
226,237,272,245,428,298,345,201,327,301,317,395,332,238,367,
L
310 Chapter 11 Tests of Hypotheses
SUpPOSe: it is ofiuteres[ to determ.ine if the mean tensile strength is greater than 250 t1Pa. That is, test
Ho:Jl~250,
HI: Jl > 25Q.
A non:nal probability plot was constructed for the tensile strength and is given in Fig. 11 ~ 10. The no:~
mality assumption appears to be satisfied. The population variance for tensile strength is asswned to
be unknown. a."ld as a result. a single-sample Hest will be used for this problem.
The resulrs from Minitah® for hypothesis testing and confidence interval OZl the mean are
The P-value is reported as 0,004, leading us 'to :eject the null hypothesis and conclude that the mean
tensile strt':ngth is greater tha.'1 250 MFa. The lower one-sided 95% confdence mterval is given as
272<,u...
t:1qlDipJ;li:28"
Reconsider Examp1e 11-11. comparing two methods for predicting the shear strength for steel plate
girders.. The Mlnitab$ output for !he paired t-test using Ct= 0.10 is
Fail.'~d T Karlsl.'uhe -
N Mean Sc:Dev SE Mean
Kar ls:ruhe 9 1. 3401 0.1460 0.0487
Lehigh 9 1. 0662 0.0494 0.0165
Difference 9 0.2739 0.1_351 0.0450
190% CI for mean difference: (0,1901. 0.3576)
IT~Test of mean difference "" o (v5o not '" 0): T-Valc.e 6.08 P-Valc.c ""
~ooo
0,999
I
0.99
0.95
~ ---- -
---
;0. 0.80
:a=0.50 ~
£ 0.20 I •
0.05
0.01
I
'.--.--- .
0,001 ,
I i
200 300 400
Tenslle strength
Figure 11-10 No:mal proh.bilityp!ot fur Example 11-27.
j
11-6 Sample Co:nputer Ouqmt 311
The results of t.ie MinitablPl output are iT.: agreement with the J;eS'Jlts round in Example 11~ 17"
:M£nitab((!l also provides the appropriate confidence interval for:he problem. Using a= 0.10, the level
of confidence is 0,90; the 90% confideot.-'e interval on the differel:ce between the two methods is
(0,1901,0.3576). Since the confidence interval does not coutain Zero, we also conclude Liar there is
a significant difference between l1e two mefuods.
The nc.:rnber of airline flights canceled is recorded for all airlines for each day of service. The Dum-
ber of flights recorded and the Dumber of 'llese flights that were canceled on a single day ir. Match
2001 are provided below for t'NO major airlines.
Is there a significant difference in the proportion of canceled flights for the two airlines? The hypothe-
ses of interest are He: Pl = h, VefS1.:S H!: Pi ¢; P2' A two~sample test 0:< proportions and a twO-sided
confide::.ce interva: on proportions are
Sample X N Sample p
1 115 i128 0.054041
2 49 635 0.07716$
The P"valile is given as 0.04.8, indicating that there is a significant difference between the proportions
of flight... canceled for American and America West airlines at a 5% level of significance. The 95%
confidence interval of (-O.()<i60. -0.0003) indicates thatAmeIica West airlines had a statistically sig-
nificant higher proportion of canceled flights k1an American Airlines for the single day.
,~~~~1!}1:3~
The mea::J. compressive strength for a particular high-strength couercte is hypothesized to be f.J. 20
(MPa). It is knOWD that tile standard deviation of compressive strength :s (j-= 1.3 MPa. A g':Onp of
engineers wants to determ.ioe ':he number of concrete specimens that will be needed in the study to
detect a dec:ease ill the mean compressive strength of two stm:dard deviations. Ii 6c average com-
pressive strength is a;;t>J,ally less than",u - 20', they want to be confident of corre.."tly detecting this sig-
nificant difference. In other words. the test of interest would be HQ: ,il ;;;; 20 versus H;: J1. < 20. For this
study. the significance level is set at a= 0.05 and the power of the test is 1- {J= 0.99, What is :he win-
imam nrunber of conerere specimens ::hat should be used L'1 t1'.is study? For a difference of 2a or 2.6
MFa, a=0.05, and 1- f3= 0.99, ::heminimllm sampie size can be found using}o.1initab~. The result~
ing output is
I,
i
-2.6 6 0.9900 0.9936
L
312 Chapter 11 Tests of Hypotheses
The .rnin.imum number of specimens to be used in the study sholli.d be r. = 6 in order to attain t!:te
desired power and level of significance.
A rr:anuiactu."'et of robber belts wishes to inspect and control the number of nonconforming belts pro.
duced on line. The proportion of noncor':orrning be~ that is acceptable is p =. 0.01, For practical pur-
poses., if thc proportion increases top = 0.035 0(' greatet'. the manufacturer wants to detect this change.
That is, the test of interest would be Ho: p::: 0.01 versus HI: P > 0.01. If the acceptable level of sig-
nificance is a::: 0.05 and the power is 1 {3 = 0.95, how many rubber belts should be selected fat
inspection? For a:::. 0,05 and 1 - f3::: 0,95. the appropriate sample size can be determined using
Minitab@.Theoutputis
Therefore, to adequately detect a significant change in thc proportion of nonconforming n:bbcr belts,
random sac:ples of at least n. """ 348 woU:d be needed.
11-7 SUMll4ARY
This Chapter bas introduced hypothesis testing. Procedures for testing hypotheses on means
and variances are summarized in Table 11-8. The chi-square goodness of fit test was intro~
duced to test the hypothesis that an empirical distribution follows a particular probability
law. Graphical methods are also useful in goodness-of-fit testing. particularly when sample
sizes are small. 1\vo-way contingency tables for testing the hypothesis that two method.;; of
classification of a sample are independent were also introduced. Several eompurer examples
were also presented.
11-8 EXERCISES
11-1. The breaking strength of a fibet used in manu- (b) 'What sample size would be required to detect a
fac:uring cloth is tequited to be at least 160 psi Past true mean yield of 85% with probability 0.95?
experience has indicated that the standard deviation of
11-3_ The diameters of bolts are knowu to have a
bteaking strength is 3 psi. A random sample of four
standa:d deviation of 0.0001 inch. A tandom sample
speci!nens is tested and the average breaking strength
of 10 bolt.,> yields an average diameter of 0.2546 inch.
is found to be 153 psi.
(a) Test the hypotbesis that the true mean diameter of
(a) Should the fiber be judged acceptable with
bolts equals 0.255 inch. using a= 0.05.
a=O.05?
(b) Vihat size sample would be necessary to detect a
(b) \Vh31 is rhe probability of accepting Ho: Ji. ~ 160
true mean bolt diameter of 0.2551 inch with a
iftbe fber has a trne breaking strength of 165 psi?
probability of at least 0"90?
11~2. The yield of a chemical process is being stud-
ied. The variance of yield is known from previous 114. Consider the data in Exercise 10-39.
ex.perience with this process to be 5 (units of a2 ):;:::per~ {a) Test the hypothesis that the mean piston ring
centagel). The past five days of plant operation have diametetis 74.035 rom. Use ct= 0,01.
resulted in the following yields (in percentages): 91,6. (b) 'What san::.ple size is required to detect a troemean
88.75,90.8,89.95,91.3. diameter of 74.030 with a probability of at least
(a) Is there reason to believe chcyield is k:ss than 90%1 0.951
ll-S ExeJX:ises 313
---_
Null Hypothesis
Ho: I' = i4J.
.... --
Test Statistic
x-~
Hypothesis
H,:I'''iJ<;
Rejection
141>
Parameter
d-II' - i4J11a
z,
a' known (II -.In H!:j1.> Pc 4>2. d=(I'-iJ<;)/"
-
H,: 1'- iJ<;,
t·:=~i-,-
X -Po
H,: I' < i4J
H,: i"" .u"
4<-2.
ItJI > lall,n_1
d = (i4J - I'll"
d = II' - iJ<;jia
cf unkJ'1own S/"IlI. H,:I':>iJ<; t(»ta.II_1 d - (I' -101<7
H,:I'<i4J to <-(a,,,-! d - (i4J - I'll"
H,:I', '""', Xl -X1 H,: 1', '" p" Iz,,! :> Za/1 d-I)l,··)l,1 ~O'f +a·~
ZQ= ~ ...=
(j~ fl.'ld ~ knowL Ja 2
a~ Hj:Jl,"> P2 Z. (. ,; r;-:;
d=: J.lJ-).t).l. "';0'1 +O'!
,-L+......t....
iJ "I ": H J:J1.j<.Uz 4 <-Z, d~()J.2 -PI)/.,}(T~ +a~
H,: 1': =,""" H!:j1.;#,Uz Irol > t crn.>1, +,,:-2 d = II': - p,,112a
tc =
a~=~ crunknown HI~J!!>111. to> tCl.,I1, h71-2 d - (1', -/'o)l2er
": H,:I', <p" d=(fl2-J.l.))!2a
lSt-+-1
~.,
s;:
~ ,\2
fl.l 11;.;.1
v= -2
"1),(-'),
\Si.tI! ,S;;r,,!
- - - + ..- - -
It, """ 1 ll:! +1
f :z
Ho:if=~ , (n-l)S' H[:a'",o; %0> Xc!'.!..n-I A;:;:;; GIGo
Xo=·.... _--
ci 2
or
.. ;:
Xa<XI-011.lI_1
o 2
H,:a'>d';, XO>XV,lI-l A= oioe
Hl: a' < Of; X~<X~ _o:.n_l ;.= (flan
He:"~ = 0; Fo=S~lS~ H!:a~;zf: cS Fr:>FtPZ.n:_:.nz._l A= Ci1/U1
or
Fo < FI_all.lIl I,tt;:-,
H;:d>~ Fo>Fl7.nl" .,,7z-1 A=at / l.1:!
11-5~ Consider the data in Exercise 1040, Test the ume. whether or not this yolu.::r:,e is 16.0 ounces, A
hypothesis that the mean life of the light bulbs is 1000 random sample is taken from the outpu: of each
hours. Use Ct= 0.05. machine.
U--6. Consider the data in Exercise 1041. Test the
hypothesis that mean compressive strength equals Machine 1 Machine 2
3500 psi. Gsc a= 0.01.
16.03 16.01 16.02 16.03
11~ 7. Two machines a..~ used for filling plastic bottles 16.04 15.96 15.97 16.04
with a net volume of 16.0 ounces. The filling 16.05 15.9S 15.96 16.02
processes can be assumed normal, with standard devi-
16.05 16.02 16.01 16.01
ations '" =O.oJ5 and '" =oms. Quality engineering
16.02 15.99 15.99 16.00
s~ that both mach.i.net :fill to the same net vol-
314 Chapter 11 Tests of Hypotheses
(a) Do you tbin.k that quality engi!leering is correct? at random fro:n the current production. Assume that
Usea:M5. shelf life is normally distributed,
(b) Assuming equal sample sizes, what sample size 108 days 128 days
=
should be used to assure that f3 0,05 if the true 134 163
difference in means is 0.0751 Asswue that
124 159
a=O.05.
116 134
(c) What is the power of the test in. (a) for a true dif-
ference in .means of 0.075? (a) Is there any evidence that t.":e mean shelf life is
greater than or equal to 125 days?
11~8. The f;1m development department of a local
store :s considering the replacement of its current (b) If it is important to detect a ratio of Clr::r of 1.0
Elm-processing .machine. The time in which it takes with a probability 0.90, is the sample size
the machine to completely process a roU of film is su.fEcient?
impo:"".ant. A random sample of 12 rolls of 24-expo- 11~ 14. The titanium content of an alloy is being stud-
sure color film is selected for processing by the c\.tr- ied in the hope of ultimately increasing the tensile
rent machine. The average processir.g time is 8.1 strength. An analysis of six. recent heats choSen at mc-
minutes, 'filth a sample standard deviation of 1.4 mi:t- dam produces the following ritanium con~ents.
utes. A random sample of 10 rolls of the same type of 8.0% 7.7%
film is se:ected for !esting in the new machine. The
9.9 11.6
average processiltg ti;:ne is 7.3 minutes. wit.~ a sample
9.9 14.6
standard deviation of 0.9 minutes, The local store '>Vill
not pUIChase the new machine unless the processing Is th.et'C any evidence that the mean titanium content is
time is more than 2 minu~es shorter than the current greater than 9.5%1
machine. Based on this information. should they pur- 1l~15. An article in the Journal of Construction
chase the new machine? Engineering and Management (1999, p. 39) presents
11-9. Consider the data in Exercise 1{)"45. Test the some data on the nwuber of work hours lost per day
hypothesis that both machines filllO the same volum!!. on a construction project due to weather-related inci~
Use Ct= 0.10. dents, Over 11 workdays, the following lost work
11.. 10. Consider the data in Exercise IO-4ti Test He: hours were recorded.
j1;::= J.l::. against H! :,U: > JIz, using a;;::; 0.05, 8.8 8.8
ll~ll. Consider the gasoline road octane number 12.5 12.2
data in Exercise 10-47. If fonm.:lation 2 produces a 5.4 133
higher road. octane number than fClfl'4ulation 1, the 12.8 6.9
::r.anufacturer would like to dctect tl".is, Fonnulate and
9.1 2.2
test e.n appropria:-c hypo:hesis, using (X= 0,05.
14.7
11 ~ 12. The la:eral deviation in yards of a certain type
of mortar shell is being investigated. by the propellant Assuming work hours are :r.otma11y distributed. is
manufac:urer. The following data have been observed. there any evidence to conclude that the mean number
of work hours lost per day is greater than 8 hours?
11 ..16. The percentage of scrap produced in a !!leW
Round Deviation Round Deviation
finishing operation is hypothesized to be less than
11.28 6 -9.48 7.5%. Several days were chosen at random and the
2 -10.42 7 6.25 percentages of scrap were calculated.
3 -<l.51 8 IO.ll 5,51% 732%
,[ 1.95 9 -<l.65 6.49 8.81
5 6.47 10 -0.68 6.46 8.56
537 7.46
Tes~ the hypothesis that the mean latetru deviation of (2.) In your opinion. is the true scrap rate less than
these mortar shells is ::ffO. Assume that lateral devia~ 7.5%1
tion is normally distributed. (b) Ifitls important to detect a ratio of (JIG: 1.5 v.'ifu
11~13. The shelf life of a photographic film is of a probability of at least 0.90, what is the :mini~
interest to the manufacturer. The manufacturer mum sample size ';hat can be used?
observes the following shelf life tor eight urits chosen (c) For 5/(1:::.2.0, what is the power of the above test?
1\ "8 Exercises 315
63 82 64 56
Process Yields (0/.;:) 81 68 72 63
1 24.2 26.6 25.7 57 59 83 74
2 21.0 22.1 2\.8 20.9 22.4 22.0 66 75 59 82
82 73 65 82
Ca) Is there reason to believe that process 1 has a
greater mean yield? Use 0: O.OL Assume that (a) Test the hypothesis that the two variances are
bolli variances are equal. equaL Use 0::;;;0.05.
Co) ~l;ssuming that in order to adopt process 1 it must (b) Using the results of {a), test the hypothesis that
produceame:m yield thatls at least5% grea,.-er than the mean buming times are equal.
ti:at of process 2, what are your ;re....--ommendations:?
11 ..22. A new filtering device is insta:led in a cber:ll~
(e) Fl.Ilrl the power of the test in part (a) lithe mean cal unit. Before its i:tstaEation. a random sample
yield of process 1 is 5% greater L'lan that of yielded the fo:Cov.-ing in1:orma:ion about the percent-
process 2. age of impurity: XI = 12.5. si :;: 101.17, and n! =8.
(d) VIhat samp1e size is required for the test in part Aiter installation, a random sample yielded x,::::: 10.2,
1 "
(a) to ensure that the null hypothesis will be sl=94.73,nz=9.
rej~ with • probability of 0.90 if tho mean (a) Can you conclude that the two variances are
yield of process J exceeds tbe mean yield of equal?
process 2 by 5%?
(b) Has the filtering device reduced the percentage of
11~20. An article that appeared in the Proceedings of impurity significantly?
the 1998 Winter Simulation Conferen.ce (1998, p.
11-23~ Suppose that two rando:;). samples were drawn
:079) discusses the concept of validation for traffic
from norn:al populations with equal variances. The
simulation :;).OOels, The stated purpose of the study is
s:unple data yields x\ =: 20,0, nl :::;: 10, L(x!! Xj)l
to design aad modify the facilities (roadways and con-
1480, X, = 15.8, n., = 10, and J:(x,; -x,j' -1425.
trol devices) to optimize efficiency and safety of traf-
fle flow. Pan of the study compares speed observed at (a) Test the hypotl'.esis that the !'No means are equal
various ir.tersections and speed simulated by a model Use 0:=0.01,
being tested, The goal is to determine whether the (b) Fwd the probability that the null hypothesis in (a.;
simulation model is representative of the actual .,-till be rejected if the t.-ue difference in means
observed speed. Field data 1s collected at a particular is 10.
location and then the simulation model is imple~ Ce) What sample size is required to detect a true dif-
mented, Fourteen speeds (ft/sec) are measured at a ference in means of 5 with probability at least 0,80
particular location, Fourteen observations are simu~ if it is known at the starr of the experiment that a
tated using the proposed model The data are; rough estimate of the common variance is 150?
L
316 Chapter 11 Tests of Hypottleses
1 I
11~24. Consider the data in Exercise 10-56. this san:ple size is used and the sample standard devi~
(a) Test the hypothesls that the means of the two nor~ ation s = 0.007, what is your conclusion, using a~
mal distributions are equal. Use a ~ 0.05 and 0,01? Construct a 95% upper-confidence interval for
cr:
assume that ~ 0;. \he true variance,
(b) v.'hat sample size is required to detect a difference 11~29. The manufacturer of a power supply is inter~
in means of2.0 with a probability of at least 0.SS1 ested in the variability of output voltage. He ha.s tested
(c) Test the hypothesis that the: variances of the two 12 units. chosen at random, with the: following results:
distributions ru;e equaL Use (J,;;!;; 0.05. '1
5.34 5.65 4,76
(d) Find the power of the tesl in (c) if the variance of
5.00 5.55 5.54
a population is four times the other. 5.07 5.35 5.44
11-25. Consider the data in Exercise 10-57. Assw:o- 5.25 5.35 4.61
fig that ~:=: 0;. test the hypothesis that the mean rod
diameters do not differ. Use a:::::: 0.05. (a) Test the hypothesis t1at cr =05. Use a=0.05.
11~26. A chemical company produces a certain drug cr
(b) If the true value of = LO. what is the probabil~
whose weight has a standard deviation of 4 mg. A new ity that the hypothesis in (a) v"ill be rejected? ' I·
method of producing this drug has been proposed. 11-30:. For the data in Exercise 11-7, test the hypo£h.. -
although some additioual cost is involved, Manage~ esis that the two variances are equal. using a = 0.01.
ment wpl authorize a change in production technique Does the result of this test influence the manner in
only if the standard deviation of the weight in the new which a test on means would be cor.ducted1 What
process is less than 4 mg. If the standard deviation of sample size is necessary to detect ~j~;;:; 25, 'With a
weight in the new process is as small as 3 mg, the probability of at least 0,901
company would like to switch production methods
11~31. Consider the fOliOVling two sampies, draVt"D.
with a probability of at least 0.90. Assuming weight to
from two normal populations.
be normally distributed and a = 0.05, how many
observations should be taken? Suppose the
resea.."'Chers choose n ;: :; : 10 and obtain the data below, Sample 1 Sa.'11ple 2
Is this a good choice for n? What SI:lOuld be their
----_.- .. ,.
4.34 1.87
decision?
5,00 2.00
16.628 grams 16,630 grams
4.97 2.00
16.622 16.631
4.25 1.85
16.627 16.624
5.55 2.11
16.623 16.622
6.55 2.31
16.618 16.626
6.37 2.28
11~27. A manufacturer of precision measuring instru- 5.55 2.0'7
ments clai:ns that the standzrd deviation in the use of 3.76 1.76
the i=:.sil\;!;).cnt is 0.00002 inch. A.'1 analyst, who is
1.91
u:1aware of the claim, uses the instrument eight times
2.00
and obtain." a sample standard deviation of 0,00005
inch.
(a) Using a= 0,01. is the claimjustiiied? Is there evider.ce to conclude that the variance of pop-
(b) Compute a 99% confidence interval for the true ulation 1 is greater than the variance of population 21
variance. Use a 0.01. Find the probability of detecting
(c) "Vhat is the power of the test if the true standard
if,1c?, ~ 4.0.
de-...1ation equals 0.000047 11-32. Two machines produce :metal parts. The yari~
(d) ·What is the smallest sample size that can be used ancc of the weight of these parts is of interest. The fol-
to detect a true standard deviation of 0,00004 with lowing data have been collected.
a probability at least ofO,95? Use a= 0.01.
11-28. The standard deviation of measu."'ements made Machine 1 Machine 2
by a special thermocouple is supposed to be 0,005
"1 =25 1""'2=30
degree. If the standard deviation is as great as 0.010.
we wish to detect it -with a probability of at least 0,90,
x\ 0.984 x" ~ 0.907
Use a;;;::: 0.01. What sample size should be used? If
s~;;:; 13.46 $;:9.65
11-8 Exercises 317
(a) Test the hypothesis that the var:ances of the twO Do the d&.ta support the designer's theory? Use
macr.rnes are equaL Use a=0.05. a=O.OS.
(b) Test the hypothesis that the two machines proo:.J.ce 11·36. An article in the Internation.al Journal of
parts having the same mean weight. Usc a= 0.05. Fatigue (1998. p. 537) discusses the bending fat~gue
11~33. In a hardness test. a steel ball is press.ed into resistance of ge::\!" teeth when using a pa.."ticular pre-
the material being tested at a standard load, The diam~ stressing or presetting process. Presetting of a gear
eter of the indentation is measured, which is related to tooth is obtained by applying and then removing a
the hardness, Two types of steel balls are available, single overload to the machine element. To detcrmine
and thei:' pe:iormance is compa.red on 10 specimens. significant differences in fatigue resistance due to pre-
Each speci::nen is tested twice. once with each ball. setti:1g. fatigue data were A "preset" tooth and
The results are given below: a "nonpreset" tooth were paired if t.'1.ey were present
on the same gear. Eleven pairs were formed a:ld :he
Ball x 75 46 57 43 58 32 61 56 34 65 fatigue life measured for e:lCrL (The final response of
Bally 52 41 43 47 32 49 52 44 57 60 inrerest is In[(fatigue life) X 10-'n
Test the hypothesis that the two: steel balls give the Pair Preset Tooth Nonpreset Tooth
same expected hardness measurement. Use a= 0.05, 3.813 2.706
1l~34. 1\.'1o:.ypes of exercise equ;p:nent.A and B. for 2 4.025 2.364
handicapped individuals are often !..!Sed to determine 3 3.042 2,773
the effect of the particul:u exercise aD heart rate (in 4 3.831 2558
beats per minute), Seven subjectS participated In a 5 3.320 2.430
study to determine whether the two types of equip- 6 3.080 2.616
ment have the same effect on hem rate, The results
7 2.498 2.765
are given in the table below,
8 2.417 2.486
9 2.462 2.688
Subject A B
10 2236 2.700
1 162 161 11 3,932 2.8:0
2 163 187
3 140 199
Conduct a test of hypothesis to de:ermine whether
4 191 206 presetfulg significantly i::creascs the fatigue life of
5 160 161 gear teeth. Use a=O.10.
6 158 160
11-37. Consider the deta in Exercise 10·66. Test the
7 155 162 hypothesis that the uninsa.w. rate is 10%. Use a:;;; 0.05.
11~3S.. Consider the data in Exercise 10-68, Test the
CQnd~ct an appropriate tCSt of hypo:hesis to deter~ hypothesis t.13t the fraction of defective clLculators
mine whether there is a significant difference in heart produced is 2.5%.
rate due to the type of equipmcnt used.
11~39. Suppose that we . . . -ish to ~est the hypothesis
11·35. An aircraft designer has theoretical e..1dence Ho: PI = fJ.:. against the alternative HI: iJl ;{; 112, where
t.b.at painting lhe airplane reduces its speed at a speci~ botb variances ~ and a; are kno'n'Il. A total of fl.[ + nz
fied power and flap setting. He tests six consecutive = N observations can be taken. How should these
airplanes from the assembly line before and after observations be allocated to the two populations to
painting. The results are shown below. mv:imize the probability thatHowill be rejected if H j
is true. and JlI - P1. : : : b;{; O'?
Top Speed (mph)
11-40. Consider the ~:r:ion membership study
Airplane Painted Not Pairted described in Exercise 10-70, Test the hypothesis that
the proportion of men who belong to a union does not
1 286 289
differ from the proportion of women who belo:tg to a
2 285 286 union, Use a=0.05.
3 279 283
1l~41. Using the data in :a~se 10-71. determine
4 283 288
Whether it is reasonable to conclude that production
5 281 183
line 2 proeuced a higher fraction of defective prcd:lct
6 286 289 t.ian ili:.e 1. Use a = 0.01,
318 Chapter 11 Tests of Hypothe:>es
11-42. Two different types of injectio::-molding (a) It is reaso:::able to conclude that these data come
machines are used to fonn plastic parts. A part is con- from a normal dist."ibution? Use a chi-square
sidered defective if it has excessive shrinkage or is goodncss-of-fit test
discolored. Two random samples, each of size 500, (b) Plot the data on nonnal probability paper. Docs a.1.
are selected, and 32 defective parr.3 are found lj. the assumption of norrn.ality seem justified?
sample from machine 1. while 21 defectl.',le part<; are
11-48. Defects on wafer su.'1'aces in integrated cirelli:
found in the sample from machine 2. Is it reasonable
fabrication are unavoidable. In a particular process the
to co::c1ude that both machines produce tie same me-
follol1ting data were collected.
don of defective parts?
1143. Suppose that we wish to test Ho: PI = fi2 Number of Number of Wafers
agains~ H!: f.lj ;i: il2., where ~ and 0; are known. The Defeetsi with i Defects
total sample size N is fLxed, bu',; the allocation of
observations to tl'lC two populations such that r.( + "''2 0 4
Does it seem reasonable to conclude that presst:..'"e tions and ranges. Would you conclude that deflection
strength is normally distributed? and range are independent?
226,16 psi 211.14 psi
Latera: Deflection
202,20 203,62
219.54 188,12 RaDge (yards) Left Normal Rigt.t
193,73 224.39 0- 1.999 6 14 8
208.15 221.31 2,000 - 5,999 9 II 4
;95.45 204.55 6,000 - 1l.999 8 17 6
;93,71 202.21
200,81 201.63 11-56. A stc:.dy is being made of the failures of an
electronic componer.t There ate four t}"pcs of fail'JIes
11~2. A company operates four machines for three
possible arid two mounting positions for the deviee.
shifts each day. From production records. the follow-
The folloVling data have been taken.
ing data on the number of breakdowns ate collected.
Fellure Type
Mounting Position A B C D
22 46 18 9
2 31 II 9 14 2 4 17 6 12
3 15 17 16 10
Would you conclude that the t:,rpe of failure is inde-
Test the hypothesis that breakdowns are independent pendent of the moor-ting position'?
of the shift.
11-57. An article in Research in Nursing and Health
11-53. Patients in a hospital are classified as surgical (1999, p. 263) summarizes data collected from a pre-
or r:.tedicaL A record is ~ept of the number of tmes vious study (Research in Nursing and Healih, 1998, p.
patients require nu;:sing service during t.~e night and 285) on the rela.tionship between physical activity and
whether these patients are or. Medicare or not. The socio-economic status of 1507 Caucasian women. The
data are as follows: data artl given in the table below.
Patient Category
Physical Activtty
Medicare Surgical Medical
Socia-economic Stat'.!s Inactivc Active
Yes 46 52
Low 216 245
No 36 43
Medium 226 409
Test the hypothesis that caJ.s by surgical-medical H:gh 114 297
patients are independen: of whether the patier.rs are
receiving Medicare. Test the hypothesis that physical activity is bdepend~
11~54. Grades L'1 a statistics course and aD operations ent of socia-economic status.
research course taken simultaneously were as follows 11-58. Fabric is graded into three classifications: A.
for a group of students. E, and C. The results below were obtained from five
looms. Is fabric classification independent of th~
Statistics Operations Research Grade loom?
Grade A B C Other
A 25 6 17 13 ::J~ber of Pieces of Fabric
B 17 16 15 6 in Fabric Classi5catio:::
C 18 4 18 10 Loom A B C
Other 10 8 11 20
185 16 12
f:..:e the grades in statistics and operations research 2 190 24 21
related? 3 170 35 16
4 ,58 22 7
11·55. All experiment with artillery shells yields the
5 185 22 15
following data on the characteristics of laterdl deflec-
320 ChaptCI 11 Tests of Hypotheses
1l~59. An article in the Joumal of Marketing l1~()O. Consider the injection molding process
Research (1970, p. 36) reports a study of the relation~ described in li'Cercise 11-42.
ship between facility conditions at gasoline stations (3) Set up this problem as a 2 X 2 contingency table
and the aggressiveness of their gasoline marketing pol~ and perform the indicated statistical a:'.alysis.
icy, A sample of 441 gasoline stations was investigated (b) State deady the hypothesis being tested. Are you
with the results shown below obtained. Is there e-.i- testing homogeneity or independence?
dence that gasoline pricing strategy and facility CODdl~
(c) Is tb.i.s procedure equivalent to the test procedure
tions a..re independent?
used if. Exercise 1:-42?
Condition
Subst2.ndard Standard Modem
Aggressive 24 52 58
Neutral 15 73 86
Nonaggrcs..<;ive 17 80 36
Chapter 12
30
25
0;
20
~
~
~
S
.c
rn
c
i" 15
0;
.'!!
$
'0;
c
{!!. 10
J
+ I
Graphical interpretation of the data is always a good idea. Box plots show the vari-
ability of the observations within a treatment (factor level) and the variability between treat-
ments. We now show how the data from a single-factor randomized experiment can be
analyzed statistically.
where Yij is the Ui)th observation, J1. is a parameter common to all treatments, called the
overall mean, 't',. is a parameter associated with the ith treatment, called the ith treatment
effect, and €ij is a random error component. Note that Yij represents both the random vari-
able and its realization. We would like to test certain hypotheses about the treatment effects
and to estimate them. For hypothesis testing, the model errors are assumed to be normally
and independently distributed random variables with mean zero and variance 0"2 [abbrevi-
ated NID(O, 0"2)]. The variance 0"2 is assumed constant for all levels of the factor.
The model of equation 12-1 is called the one-way-classification analysis of variance,
because only one factor is investigated. Furthermore, we will require that the observations
be taken in random order so that the environment in which the treatments are used (often
called the experimental units) is as uniform as possible. This is called a completely ran-
domized experimental design. There are two different ways that the a factor levels in the
experiment could have been chosen. First, the a treatments could have been specifically
chosen by the experimenter. In this situation we wish to test hypotheses about the 'fl , and
conclusions will apply only to the factor levels considered in the analysis. The conclusions
cannot be extended to s:imilar treatments that were not considered. Also, we may wish to
estimate the 7:'1' This is called the fixed effects model. Alternatively, the a treatments could
be a random sample from a larger population of treatments. In this situation we would like
to be able to extend the conclusions (whiCh are based on the sample of treatments) to all
treatments in the population, whether they were explicitly considered in the analysis or not.
Here the 'f; are random variables, and knowledge about the particular ones investigated is
relatively useless. Instead, we test hypotheses about the variability of the 7:'; and try to esti-
mate this variability. This is called the random effects, or components of variance, model.
In this section we will develop the analysis of variance for the fixed-effects model. one-.
way classification. In the fixed-effects model, the treatment effects ~ are usually dnfined as
deviations from the overall mean, so that
(12-2)
Let y;, represent the total of the observations under the ith treatment and Y" represent the
average of the observations under the ith treatment Similarly. let y .. represent the gnnd
total of all observations and y.. represent the grand mean of all observations. Expressed
mathematically,
i=1,2, .. "a.
, n (12-3)
y .. = 2,2:>ii' Y··=y··/N,
1=1 i"'i
whereN =: an. is the total number of observations. Thus the "dot" sUbscript notation implies
sum.m.ation over the subscript that it replaces,
We are interested l!J. testing the equality of the a treatment effects. t;sing equation 12-2,
the appropriate hypotheses are
Ho: 't': 1J :;; ... :;; '4<) 0,
(12-4)
H!: ~ ¢ 0 for at least one l.
That is, if the null hypothesis: is true. then each observation is made up of the overall mean
f.L plus a realization of the random error Ell'
The test procedure for the hypotheses in equation 12-4 is called the analysis of vari~
aIlee. The name "analysis of variance" results from partitioning total variability in the data
into its component pa.'1S, The total corrected sum of squares, which is a measure of total
-variability in the data, may be written as
(12-5)
or
(12-6)
a a
+2 2,2,(Yi.- 'y.)(Yi; - Yi.)
~l j=l
Therefore we have
j
(12-7)
12~ 1 The Completely Randomized Single-Factor Experiment 325
Equation 12-7 shows !hat !he total variability in the data, measured by !he total corrected
sum of squares, can be partitioned into a sum of squares of differences betv,,'een treatment
means and the grand mean and a sum of squares of differences of observations within treat-
ments and the treatment mean. Differences bet\Veen obsen-ed treatment meanS and the
grand mean measure the differences between treatments. while differences of observations
within a treatment from the treatment mean can be due only to random error. Therefore. we
write equation 12-7 symbolically as
wheres <v a!1d Q, is chi-square with Vi degrees offree-dom (i= 1, 2 •. ,.,3), Then QH Q;tt
Q: are independent chi-sGuare random variables with Vl , v2 > •• " v, degrees of freedom,
respectively~ if and only if
Using this the-orem, we note that the degrees of freedom for SSt=.tm~n~ and SS£:,' add up
to N - 1, so that SS""',m"Ja" and SSe/a" are independently distributed cbi-square random
variables, Therefore, under the null h)l'othesis, the statistic
(12-8)
follows the Fa _ I• N _ a distribution. The quantities MS<=.:rw:nu and ~y'S£ are rr.ean squares.
The expected values of the mean squares are used to show that Fo in equation 12-8 is
an appropriate test statistic for Ho: 1) =: 0 and to determine the criterion for rejecting this null
hypothesis. Consider
:i
326 Chapter 12 The Analysis of Variance
(12-10)
'r;
"
12~ 1 The Completely Randomized S:.ngle-Factor Exp"'...riment 327
Consider the hardwood concentration experiment described in Section 12-1.1 We can use the analy-
sis of wrian<:e to test '.he hypothesis that different hardwood concentrations do not affect the mean
tensile strength of the paper. The sums of squares for analysis of variance are computed from equa-
dOO$ 12·9, 12-10, and 12-11 as follows:
" _ (383)'
=(7) +(8J-+ ,,+(20t---=512,96,
24
An appropriate estimation criterion is to estimate J1 and 1'; such that the sum of the squares
of the errors or deviations Ei I is a minimum. This n:ethod of parameter estimation is called
the method of least squares: In e.ti.,nating }.l and ~i by least squares, the nonnality assump_
tion on the errors Elf is not needed. To find the least-squares estimators of J1 and 1'!~ we form
the sum of squares of the errors
(12-12)
j!.
and find valnes of }.l and T" say and ii' that minimize L The values p. and ii are the solu-
tions to the a + 1 simultaneous equations
~a}.lL -;
jl,.{
=0
,
Differentiating equation 12-12 with respect to JJ.. and 1'1 and equating to zero, we obtain
a ,
-iLI,(Yirj!.-r,)= 0
i.=:l J=l
and
,
-2I,(Yi rj!.-i',)=O, i::::: 1,2.".,a.
1';;1
=)'t' ,
Equations 12-13 are called the least-squares normal equations. Notice that if we add the
last a normal equations we obtain the first normal equation. Therefore, the normal equations
are not linearly independent, and there are no unique estimates for f.4 't'p1:.z,.,.,'t'". One way
to overcome this difficulty is to impose a constraint on the solution to the normal equations.
There are many ways to c1loose this constraint Since we have defined the treatment effects
as deviations from the OY"erall mean, it seems reasonable to apply the constraint
(12-14)
This solution has considerable intuitive appeal since the overall mean is estimated by the
grand average of the observations and the estimate of any treatment effect is just the differ-
ence bct\lleen the treatment average and the grand average,
This solution is obviously not unique because it depends On the constraint (equation
12-14) that we have chosen. At :first this may seem unfortunate. because two different
experimenters could analyze the same data and obtain different results if they apply differ-
ent constraints. However, certainfimctions of the model parameter are estimated uniquely.
regardless of the constraint. Some examples are 'i-"r which would be estimated by ~,-~ ~
)1. -:ip and f.1 + r i • which would be estimated by it +7:; = <~};.' Since we are usually interested
in differences in the treatment effects rather than their actual values, it causes no concern
that the 7:1 cannot be estimated uniquely. In general, any function of the model parameters
that is a linear combination of the left-hand side of the normal equations can be estimated
lI1liquely, Functions that are lI1liquely esrimated, regardless of which constnlint is used, are
called estimable functions,
Frequently, we would like to construct a confidence interval for the ith treatment mean,
The mean of the ith treatment is
l= 1. 2, .... a.
A point estimator of f.1l would beil} =it + 1i Now t if we assume that the errors are nor~
mally distributed, each 1" is !\'1D(;1,! <i'ln), Thus, if <i' were known, we could use the nQr-
mal distribution to construct a confidence interval for fli' V sing MSE as an estimator of > cr
we can base the confidence interval on the t distribution, Therefore, a 100(1 - a)% confi-
dence interval on the ith treatment mean Pi is
(12-16)
A 100(1- a)% confidence interval on the difference between any two treatment means, say
J11 - /lj. is
~~#l!1~)~-2 .'
We can use the results given previously to estimate the mean tensile s.trengths at different levels of
hardwood concentration for the experiment in Section 12~ 1. L The mean tensile strength estimateS are
Yl' : : :. AI(,:::::' 10.00 psi,
[2L17± (2,086)-I65lj6 ].
[2Ll7 ± 2.17],
330 Chapter 12 The Analysis of Variance
[Yi'- Yj ±tc:.{l.N-Q~2MS5/n ].
[17.oo-1j.67±(2.086)~2(6.51)/6l·
[L33±3.07].
ment number and the fitted value YI' ' These plots do not reveal a:ly model inadequacy or
unusual problem with the assumptions,
:
·•
:•
.
'
•
. •
4 6
Residual value
Figure 12--2 l\~onnal probability plot of residuals from the hardwood co:r.centration experiment.
t
4
2
·1•• •
2 3 4
-2
t
Figure 12~3 Plot of residuals YS. treatment.
332 Chapter 12 T.::le A..'1alysis ofVatiance
2~
!
f
• ••
I
number of observations N:::: k;"'ln l The computational fOIDlulas for SST and SSfftltilY:rJs
+
become
a "
SST 2::2>3
r""l j:<
N
and
J
12~2 Tests on Individual Treattnent Means 333
We might suspect at the outset of the experiment that hardwood concentrations 3 and 4 pro~
duce the same tensile stren~ implying that we would like to test the hypothesis
H,: p, = ,,",,
H,: p, *11"
This hypothesis could be tested by using a linear combination of treatment toWs, say
Y:r-Y,·=O.
If we had suspected that the average of hardwood concentrations 1 and 3 did not differ from
the average of hardwood concentrations. 2 and 4. then the hypothesis would have been
L;
with the restriction that ""tC, O. These linear combinations are called cOntrasts. The sum
of squares for any contrast is
(12-18)
and has a single degree of freedom. If the design is unbalanced, then the comparison of
treatment means requires that 1~ ",1niGo:::: 0, and equation 12~ 18 becomes
(12-19)
A contrast is tested by comparing its sum of squares to the mean square error. The resu1t~
ing statistic would be distributed as F, with 1 and N - a degrees of freedom.
A very important special case of the above procedure is that of orthogonal contrasts.
Two contras:s with coefficients (cJ and (dJ are otthogonal if
For a treatments a set of a-I orthogonal contrasts will partition the sum of squares due to
treatments into a-I independent single-degree-of~freedom components. Thus. tests per~
formed on orthogonal contrasts are independent.
There are many ways to choose the orthogonal contrast coefficients for a set of treat-
ments. Usually, something in the nature of the experiment should suggest which compar-
isons will be of interest. For example, if there are a=::3 treatments, with treatment 1 a
"control" and treatments 2 and 3 actua11evels of the facto! of interest to the experimenter.
then appropriate orthogonal contrasts might be as follows:
Note that contrast 1 with Ci =-2, 1, 1 compares the average effect of the factor \Vith the con-
trol while contrast 2 with d i = 0, - 1, 1 compares the two levels of the factor of interest.
Contrast coefficients must be chosen prior to running the experiment, for if these com·
parisons are selected after examining the data. mOSt experimenters would construct tests
that compare large observed differences in means. These large differences could be due to
the presence of real effects or they could be due to random error, If experimenters always
pick the largest differences to compare, they will inflate the type I error of the test, since it
is likely that in an unusually high percentage of the comparisons selected the observed dif-
ferences will be due to error.
Consider the hardwood concentration experiment. There an:: four levels of hardwood concentration.
and the poss:.Ole sets of comparisons between these means and the associated orthogonal comparisons
are
Ha: f,i; ..... f,i;, =P-t -1-'1, C1 t;;;,YI' -Y2' -Yy + Y4"
H,,: 3f,i( + ,LL:.,:;;; P-t + 3fJ. 4t C2 = - 3y!' -)'2' + YJ- - 3Y4' •
H,: 1', + 3", ~ 3", + ",. C3=-YI·+3Y2' 3Y3'+Y4'
Notice that the contrast constants are orthogonaL Using the data from Table 12-1. we find the numer-
ical values of the contrasts and the sums of squares as follows:
(-9)' •
C1 ~60-94-102+127~-9. SSe. ~ 6(4) ~ 3.,8,
These contrast sums of squares C-Ontpletely partition the treatment sum of squares; that is. SS_G::=
SSe l + SSC~ - SSe:; == 382.79. These tests on the contrasts are usually incorporated into the analysis of
variance, such as shown in Table 12-6. From this analysis, We conclude that there are sig:n.ificant dif~
ferences between bardwood concentrations 1,2 vs. 3, 4. but that the average of 1 and 4 does not dif~
fer from the average of 2 and 3. nor does the average of 1 and 3 differ from the average of 2 and 4.
J
12-2 Tests on Individual Treatment Means 335
wbere y~ is the largest sample mean and y""" is the smallest sample mean out of p sample
means. Let qa (a,f) repre.<:;ent the upper a percentage point of q, where a is the number of
treatments andfis the number of degrees of freedom for error. Two means) y;. and Yj. (i *'
j), are considered significantly different if
Thble XlI (Appendix) contains values of q. (a,fJ for a= 0.05 and 0.01 and a selection of
values for a and! Thkey's procedure has the property that the overall significance level is
exactly a for equal sample sizes and at most a for unequal sample sizes.
~~,!,!~'g4:
We will apply Tukey's test to the hardwood concentration experiment, Rec::ill that there are a = 4
rnea.1S, n:::: 6) and MSc :::: 6,51. The treatment means axe
Y" = 10.00 psi. y,. = 15.67 psi. y, = 17.00 psi, Y.., =: 21.17 psi.
Fro", Table XII (Appendix). with a= 0,05. a =4. and!= 20. we find qOM(4, 20) = 3.96.
336 Chapter 12 The Analysis of Variance
when sample sizes are equal. This expression represents a 100(1 a)% simultaneous con-
fidence interval on aU pairs of means Jl; - j..l;.
If the sample sizes are unequal, the 100(1 - 0:)% simultaneous confidence interval On
all pairs of means JJ.; - ~ is given by
~.
I I
I I I
10.0 12.0 14.0 16,0 1M 20.0
i 1,2"",,a,
Yij=/1+'r i '€ij { ' - (12·21)
J -1,2,"",n,
where 1; and €;, are independent random variables. Note that the model is identical in struc-
ture to the fixed-effects case, but the parameters have a different interpretation. If t.'1e vari-
ance of 'T, is ~, then the variance of any observa:ion is
(12-22)
still holds. That is~ we partition the total 'Vmiability in the observations into a component
that measures variation between treatments (SSI.."tlI;mI:1lJ and a component that measures vari-
ation within treatments (S5E)' However, instead of testing hypotheses about individual treat-
ment effects) we test the hypotheses
If 0, all treatments are identical. blo1 if U; > 0, t.'len there is variability between treat-
ments, 'The quantity SSE/a' is distributed as chi-square with N - a degrees of freedom, and
under the null hypothesis, SS~enl$/cf is distributed as chi-square with a 1 degrees of
freedom, Further, the random variables are independent of each other. Thus, under the null
hypothesis, the ratio
SS"""""",/( a - I) (12-23)
SSE/eN -a)
L
338 Chapter 12 The Analysis of Variance
If we square and take the expectation of the quantities in brackets; we see that terms involv-
ing -r: are replaced by d!. as E(r;) O. Also, terms involving I7", 1~' L~.. lL;", I€~' and
L;.IL;~ I ~ are replaced by nO". 000". and an 0;, respectively. Finally, all cross-product
tenns involving ~i and €ijhave zero expectation. This leads to
or
E(MS=Il,:) = (J1 + nd;. (12-24)
A similar approach will show that
E(MS,) = (Jr, (12-25)
From the expected mean squares. we see that if Ho is true, both the numerator and the
denominator of the test statistic, equation 12-23, are unbiased estimators of 0-2; whereas if
HI is true, the expected value of the numerator is greater than the expected value of the
denominator. Therefore, we should rejectHo for values of Fo that are too large. This implies
an upper~tail, one-tail critical region, so we reject Ho if Fo > Fa. a_ I,N_,:'
The computational procedure and analysis-of-variance table for the random-effects
model are identical to the fixed~effects case. The conclusions., however, are quite different
because they apply to the entire population of treatments.
We usually need to estimate the \'ariance components (0-2 and ~) in the model. The
procedure used to estimate 0-2 and a! is called the "analysis..of-variance method," because
it uses the lines in the analysis-of-variance table. It does not require the normality assump-
tion on the observations. The procedure consists of equating the expected mean squares to
their observed values in the analysis-of-variance table and solving for the variance compo-
nents. When equating observed and expected mean squares in the one-way-classification
random-effects model, we obtain
MSu<;wnctJ.I1 = 0-2 + nO-;
and
(12-27)
1
no=-
a ~n21 ,
Ln,-.c;-1
~ i
r
a-I 1-
'-1
;
L
L
kcl
11.
,
Sometimes the analysis-of-variance method produces a negative estimate of a variance
component Sinee variance components are by defmition nonnegative, a negative estimate
of a variance component is unsettling. One course of action is to accept the estimate and use
it as evidence that the true value of the variance component is zero, assuming that sampling
variation led to the negative estimate. \\'!rile this has intuitive appeal, it will disturb the sta~
tistical properties of other estimates, Another alternative is to reestimate the negative vari~
3..1.ce component with a method that always yields nonnegative estimates. Still anodier
possibility is to consider the negative eS"Jmate as evidence that the assumed linear model is
incorrect, requiring that a study of the model and its assumptions be made to find a more
appropriate model.
;J,l~~ii;)l#:~
In his book Design aruiAnalysis afExperiments (2001). D. C. Montgo:nery describes a single-factor
experiment involving the random-effects model. A textile manufacturing company weaves a fabric on
a la.,'"gc number of loows. The company is interested in loom-to-loom. variability in tensile stre::1gth,
To investigate this, a IruUlufacturing engineer selects four looms at FaJldom and makes fOtlf st."'\'::ngth
determinations on fabric samples chosen at random for each loom. The data are shown in Table 12-7,
and the analysis of variance is summarized in Table 12-8.
From the analysis of variance, we conclude :hat the looms :.n the p:ar.t differ significantly in their
ability to produce fabric of unifonn strength. The variance components are estimated by &::;:::: 1.90 and
., _ 29.73-1.90 6%
vr 4 . .
= 6.96 + 1.90
8.86.
Most of this variability is attributable to dill'erences between looms.
Observations
Loom 1 2 3 4 Totals Averages
1 98 97 99 96 390 97.5
2 91 90 93 92 366 91.5
3 96 95 97 95 383 95.8
4 95 96 99 98 388 97.0
1527 9SA
l
34(l Chapter 12 The Analysis of Variance
Process
fallout
(a)
110 '.
pSI
(b)
Figure 12.-6 The distribution of fabric strengdl, (a) Current process. (0) Improved process.
12-4 The Randomized Block Design 341
IY1D
y"
he
L
342 Chapter 12 The Analy'sis of Variance
Let Yi. be the total of all observations taken under treatment i,let Y'} be the total of all
observations in blockj, let y .• be the grand total of all observations, and let N = ab be the
total number of observations. Similarly, y;. is the average of the observations taken under
treatment i, yO) is the average of the observations in blockj, and y•. is the grand average of
all observations. The total corrected sum of squares is
Expanding the rigbt-hand side of equation 12-29 and applying algebraic elbow grease
yields
(J b ., (1 b .,
2,2,(Yij - y.,f =b2, (y,.-; ..)' +a2, (;.j - y.)"
i=1 J"'-l i=1 /,=1
o b 2
(12-30)
+ 2,2,(Y,j-Y"-Y'j+)'.')
i=1 j~l
or. symbolically,
SSr= SSumrr,cn~ + SSblocu'l- SS£, (12-31)
The null hypothesis of no treatment effects (H,: "" = 0) is tested by the F ratio,
MS~,,,IMS£. The analysis of variance is summarized in Table 12-9. Computing formu,
las for the sums of squares are also shown in this table. The same test procedure is used in
cases where treatments and/or blocks are random.
Ex..nJ~~eX2~6
Au experiment was performed ::0 determine the effcct of four different chemica:s on the stre';lgth of a
fabric. These chemicals a.--e used as part of the penna.nent-press finishing process. Five fabric samples
were selected, and a randomized block design was run by testing each chemica! type once in random
order on each fabric s:unple. The data are shown in Table 12-10.
The S~ of squares for the analysis of variance are corcputed as follows:
Treatments a-I
a-I
S~l"'.'~
Blocks h-I 0-1
ss
Error SSE (oy subtraction) (a -1)(h -1) "
(a-l)(o-l)
Tota! ab-I
124 The Randomized Block Design 343
Row Row
Chemica! Fabric Sawe1e Tota~s, Averages,
Type 2 3 4 5 y,. Y.·
1 1.3 1.6 0.5 1.2 1.1 5.7 1.14
2 2.2 2,4 OA 2.0 1.8 8.8 1.76
3 1.8 1.7 0.6 1.5 1.3 6.9 1.38
4 3.9 4.4 2.0 4.1 3.4 17.8 3.56
CollUli.n
totals, Y'J 9.2 10.1 3.5 8.S 7,6 39.2 1.96
Colu;r.n
averages. Y'j 2.30 2.53 0.88 2.20 1.90 (y .. ) (]i.. )
SS
b,od:s
; -.& ZL_ iab
£,..., a
/",1
The analysis of variance is summar'.z.ed in Table 12-11. We would conclude that there is a signiiicant
difference in 11e chemical types as far as their effect on fabric strength is oo:l.cernoo,
·-T-·_··_··_·
2
•
",-
•
•
~ 0
••
1il •
E ••
0
Z
-, ,•
•
•
-2 •
0.50 0.25 a 0.25 0.50
Res;d:tal value
Figure 12-9 Normal probability plot of residuals from the randomi2ed block design.
Chapter 12 The Analysis of Variance
4, which provides the greatest strength, also has somewhat more mability in strength, Fol,
low-up experiments may be necessary to confirm these findings if they are potentially
important,
•
-0.5
el/+
, I
0.5 ;-
! r
01- t5
t
.~
2 3
1
-O.5~
I
Figure 12-11 Residuals by block.
eat
°l· •
. ••
'
I
• •
•
°1 ..•• •
2 4 6
-O'5~
Figure 12-12 Residuals versus YI}"
j
r l2-5 Determining Sample Size in Single-Factor Experiments
EXPERIMENTS
In any experimental design problem the choice of the sample size or number of replicates
to use is important. Operating characteristic curves can be used to provide guidance in mak-
this selection. Recall that the operating characteristic curve is a plot of !.he type II (fJ)
error for various sample sizes against a measure of the difference in means that it is impor-
tant to detect. Thus, if the experimenter knows how large a difference in means is of poten~
rial importance, the operating characteristic curves can be used to determine how many
replicates are required to give adequate sensitivity.
We first consider sample size determination in a fixed-effects model for the case of
equal sample size in each treatment. The pnwer (1- (3) of the test is
1- {3= P(Reject HoIH, is false} (12·34)
=P{Fo > F", ,.,."••iH, is false I·
To e\'aluate this probabili~ statement, we need to know the distribution of the test statistic
Fo if the null hypothesis is false. It can be shown tllat if He is false, the statistic Fo =
MS=_.IMSE is distributed as a noncentral F random variable, with a I and N - a
degrees of freedom and a noncentra1ity parameter o. If Ii ~ O. then the noncentral F distri·
bution becomes the usual centroi F distribution.
The operating characteristic curves in Chart vn of the Appendix are used to calculate
the pnwer of the test for the fixed-effects model. These curves plot the probability of type
II error ({3) against <1>. where
,
nL'!f
'--,-.
{=1
(12-35)
aIr
o.
The parameter tt;2- is .related to the noncentrali::y parameter Curves arc available for a::::
0.05 and a I::: 0.01 and for several values of degrees of freedom for the numerator and
nt~?
<I>
:1 ;"'1n(5j'=n..
=--,-=- ,
QIJ 5
and the openting chal;a.cteristic curve for a -1 = 5 -1 =4 and N - a = a(n-l) = 5(11 -1) euor degrees
offreedomis shown in Chart VII (Appendix). As a fut guess, try n =4 replicates. This yields <Ilz 4, =
<1)=2, and5(3)= 15 crrordegrees offrcedom. Consequently, from ChartVlI. we find that i3~ 0.38.
348 Chapter 12 TbeAnalysis of Variance
Therefore, the PO?'ef of the test is approxllnately 1 - P 1 - 0.38 ;;;: 0.62. whicn is less than the
required 0.90, and so we conclude that It:::::: 4 replicates are not sufficient. Proceeding in a similar man-
ner, we can constru:ct the following display.
Thus, at least 11 = 6 replicates must be run in order to obtain a test with the required power.
(12-37)
In the randomized block design, replace n with b, the number ofblocks. Since u' is usually
unknown, we may either use a prior estimate or define the value of cr! that we are interested
in detecting in terms of the ratio d!/~.
Consider a completely randomized design with five treatments selected at random, with six observa w
a-;
tions per treatment and a=O.05. We \¥ish to determine the power of the testlf is equal to cT. Since
a;
a 5,11 = 6, and = 02, we oay compute
A~~1+(6)1 ~2.646.
From the operating characteristic curve with a-l :::::4, tV -a = 25 degrees of freedom and ct= 0,05,
we find that
{J = 0.20.
Therefore. the power is approrimately 0.80.
~~~ .....- - ..
12·6 SAMPLE CO:\1PUTER OUTPUT
Many computer packages can be implemented to carry out the analysis of variance for the
situations presented in this chapter. In this section, computer output from 1-linitab® is
presented.
j
12-6 Sa'JOple Computer Output 349
The analysis of variance results are identical to those presented in Section 12-1.2. 11initab®
also provides 95% confidence intervals for the means of each level of hardwood concen-
tration using a pOOled estimate of the standard deviation, Interpretation of the confidence
intervals is straightforward. Factor levels with confidence lnterYals that do not overlap are
said to be significantly different.
A hetter indicator of significant differences is provided by confidence intervals based
on Tukey's test on pairwise differences, an option in MinitabV , The output provided is
The (simultaneous) confidence intervals are easily interpreted, For example, the 95% con-
fidence intenra} for the difference in mean tensile strength between 5% hardwood concen-
tration and 10% hardwood concentration is (-9.791, -1.542), Since this confidence interval
does not contain the value 0, we conclude there is a significant difference between 5% and
10% hardwood concentrations. The remaining confidence intervals are interpreted simi~
larly, The results provided by M.in.itab' are identicil to those found in Section 12-2.2,
350 Chapter 12 The Analysis afYanancc
12-7 SGMMARY
This chapter has introduced design and analysis methods for experiments with a single fac-
tor. The importance of randomization in sing1e~factor experiments was emphasized. In a
completely randomized experiment, all runs are made in random order to balance out the
effects of unknown nuisance varia.bles. If a known nuisance variable can be contrOlled,
blocking can be used as a design alternative. The fixed-effects and random~effects models
of analysis of variance were presented. The primary difference bet\\'een the two models is
the inference space.1n the fixed-effects model inferences are valid only about thefactor lev-
els specifically considered in the analysis) while in the random~effects model the conclu~
sions may be extended to the population of factor levels. Orthogonal contrasts and Tukey's
test were suggested for making comparisons between factor level means in the .:fixed~effects
experiment. A procedure was also given for estimating the variance components in a ran-
dom-effects model. Residual analysis was introduced for checking the underlying assump-
tions of the analysis of variance.
12-8 EXERCISES
12--1. A study is conducted to determine the effect of Observations
cuttir.g speed on the life (in hO":lrs) of a particular
?".acbine tooL Four leveis of cutting speed are selected 2 3 4 5 6
for the study with the following results: 125 2.7 4,6 2.6 3.0 3.2 3.8
160 4.9 4.6 5,0 4.2 3,6 4.2
Cuttir.g 200 4.6 3.4 2.9 3.5 4.1 5.1
Tool Life
41 43 33 39 36 40 (a) Does C~6 flow rate affect e~ch unifon:.rity1 Con-
2 42 36 34 45 40 39 struct box. plots to compare the fac~or levels and
34 38 34 34 36 33 perform the analysis of variance,
3
4 36 37 36 38 35 35 (b) Do the residuals indicate any problems v.i.th the
underlying assun:ptions?
tz..3. The compressive strength of concrete is being
(a) Does cutting speed affect tool life? Draw compar-
studied, Fom: different mi.:x.ing techniques are being
ative box piots a.id perform an analysis of investigated. The following data have been collected:
variance.
(b) Plot average tool life agai.."lst cutting speed and Mixing
interpret the results. TechrJque Compressive Stren..,c-th (psi)
(c) Use Tukey's tese to investigate Cifferences 3129 3000 2865 2890
between the individual levels of cutting speed. 2 3200 3300 2975 3150
Interp:et the results. 3 2800 2900 2985 3050
(d) Find the residuals and examine them for model 4 2600 2700 2600 2765
inadequacy.
Ca) Test the hypothesis that mixing techniques affect
J2...2. I!l "Orthogonal Design for Process Optimiza-
the strength of the concrete. 1;se a=' 0.05.
tion and Its Application to Plasma Etching'; (Solid
Sfate Technology, May 1987), G. Z. Ym:md D, W. lil- Cb) Use Thkey's test to make comparisoos bem'eetl
lie describe (1,'1 experiment to determine the effeet of pairs of means. Estimate tl::.e treatment effects.
q,) flow rate on the ur..iformity of the etch on a sili- 12-4. A textile mill has a la..rge nurr:.ber of looms. Each
con wafer used in integrated circuit manufacturing, loom is supposed to provide the sam.e output of cloth
Three flow rates are used in the experiment., and the per minute. To investigat!! this assumption. fiv!! looms
resulting unifomrity (in percent) for six :eplicmes is as are chosen at random and the!! output measured at dif~
follows: ferent times. The following data.are obtained:
•• .....:...J
r
I 12-8 Exercises 351
Output (lb i miJ.<) {c) Compute a 95% interval estimate of the mean for
LoOIC
coating type 1. Compute a 99% interval estimate
4.0 4.1 4.2 4,0 4,1 of the mean difference bern'een coating types 1
2 3,9 3.8 3.9 4,0 4,0 and 4,
3 4,1 4,2 4,1 4,0 3,9
(d) Test all pairs of means using Tukey's test, with
4 3,6 3,8 4,0 3,9 3,7 a= 0,05,
5 3,8 3.6 3.9 3,8 4,0
(e Assu:ni.."1g t.l)ar. coating type 4 is currently in use,
what are your recommendations to the manufac-
(a) Is this a:5xed- or randoIll~effects experiment? Ate ~"'Cf? We wish to minimize conductivity.
the looms similar in output?
12-7. The response time in milliseconds was deter-
('J) Estimate the .wability between looms. mined fot three different types of clrccirs t:.sed in an
(c) Estimate the experimental error variance. electronic calculator. The results a..-e recorded here:
(d) What is the probability of accepting Hr; if ~ is
four times the experimental error variance? Circuit Type Response Time
(e) Analyze the residuals from this experiment and 1 19 22 20 18 25
check for model itmd£quacy, 2 20 21 33 27 4{)
3 16 15 18 26 17
12-5. 1m experiment was run to determine whether
four specific :5ring tempetatu:res affect the density of a
(a) Test the hypothesis that the three circuit types
cer.ain type of brick. The experiment led to the fol~
have the same response time.
lowiDgda~:
(b) Lse Tuley's test to compare pairs of treatment
Temperatu.·..c means.
_-"C-=F)-'-_ _ _ _ _ _D=en=,=ityL...................... _ _ (c) Construct a set of Ort.1og0r..al contrasts. assuming
that at the outset of the experiment you suspected
100 21.8 21.9 21.7 21.6 21.7 21.5 21.8
the response time of circuit type 2 to be different
125 21.7 21.4 21.5 21.5
from the other two.
150 21.9 21.8 21.8 21.6 21.5 -
(d) What is the power of this test for detecting
21.7 21.8 21.7 21.6 21.8 -
-r;
L~., ler = 3.0?
{e) Analyze the residuals from this experiment.
(a) Does dle fixing temperature affect the density of
the bricks? 12·S.In "The Effect of Nozzle Design on the Stability
and Performance of Turbulent Water Jets" (Fire Safety
(b) Estirnate the components in the model,
JOUT1'.aI. VoL 4, Angus! 1981), C. Theobald describes
(c) Analyze the residuals from the expe~ent. an experiment in which a. shape fac:or was determined
12-6.An electro::tics eogineer is interested in the effect for several different nozzle designs at different lev>!ls
on tube conductivity of .five different types of coating of jet efflux velocity. Interest in this experimer.(
for cathode ray tubes used in a telecommunications focuses primarily on nwil.e design, and velocit} is a
system display device, The follov.i.ng conductivity nuisance factor. The data are shown below:
data are obtained:
:43 141 150 146 1 0.78 0,80 0,81 0.75 0.77 0,78
2 152 149 137 143 2 0,85 0,85 0,92 0,86 0,81 0,83
3 134 133 132 127 3 0,93 0,92 0,95 0,89 0,89 0,33
4 129 127 132 129 4 1.14 0,97 0,98 0,88 0,86 0,83
5 147 148 1'4 142 5 0,97 0.86 0,78 0,76 0.76 0,75
(a) Is there any difference in conductivity due to coat~ (a) Does noz:ile type affect shape factor? Compare the
ing type? Use <X= 0,05, noZ<.1es using box plots and the analysis of variance.
(b) Estimate the overall mean and the treatment {b) Use Tukey's test to determine specific differences
effects. between the nozzles. Does a graph of average (or
352 Chapter 12 The Analysis of Variance
standard deviation) of shape factor versus nozzle 12~11. Suppose :hat five uonnal populations have
type assist -.vith the conclusions? coUll!J.On variance IT =100 and means Ji.: = 175~ J.i: =:
{c) Analyze:he residuals from this experiment. 190,)1., 160,)1., = 200, and)1., 215. How = y
observatiollS per population must be taken so :hat the
12-9. In his bookDesigrt andAnalysis ajExperimen.ts probability of rejecting the hypothesis of equ.allry of
(2001), D. C. Montgomery describes an experiment to means is at least 0.951 Use a= 0.01.
determine the effect of four chemical agents on the
strength of a particular type of clo:h. Due to possible 1,2..12. Consider testing the equality of ilie means of
variability from cloth to cloth. bolts of cloth are con- two normal populations where t.'1e variances are
sidered blocks. Five bolts are sclec:ed and all four uokuo\\'n but assumed equal. The appropriate test
chemicals in random order are applied to each bolt. procedure is the two-sample Hest Show that ::he
The resulting tensile streng"..hs are two-sample t-test is equivalent to tl:e one-way~
classification analysis of variance.
Bolt 12-13. Show that the variance of the linear combioa~
Chemical 2 3 4 5
tion Z:~"[c;y,, is 02Z:;d!ni c7,
12··J4. In a fixed-effects moru;l. suppose that there ue
1 73 68 74 71 67 n observations for each of four treatments, Let~, ~,
2 73 67 75 72 70 and a; be single-degree~of-freedom components for
3 75 68 78 73 68 r.h~ ortl;ogO;.1al cootrasts. Prove :hat SSl(tt!!l!Ienu = Q~ +
4 73 71 15 75 69 Q;+ 12;.
12-15. Consider the data shown in Exercise 12-7.
(a) Is there any dIfference 1."1 tensile sjre:tgth between (a) Write out the :east squares normal equations for
the chemicals? this problem., and solve them forfl and1;., making
(b) Use Tukcy's test to investigate sped..£ic differ~ the usual coostraint (Z:~'" Ji,-;; 0). Estimate 'C) - ; .
ences between the chemicals. (b) Solve the equations in (a) using the constraint~:::;;
(c) Analyze the residuals itom L.":ris experin1ent. O. Are the estimators f, and fi the same as you
found in (a)? Why? Now estim.ate 1'[ ~ 't; and com-
12~10. Suppose that four normal populations have
pare your answer with (a). '\\t'hat statement can
COUll!J.On variance <f =25 a.."l.d means ,Il! = 50, ~ ::60.
you make about estimating contrasts in the !)
iLl;;;:: 50, and Ji.4 """ 60. How ma.1Y observations should
be taken on each population so that the probability of (c) Estimate jl+ 1';. 21'] -11- 1";: and jl + 'C( + 12 using
rejectiug the hypothesis of equality of means is at least the two solutions to the normal equations. Com-
0,901 Use" = 0.05. pa..-e L1e results obtained !n each case.
Chapter 13
Design of Experiments
with Several Factors
An experiment is just a test or a series of tests. Experiments are performed in all scientific
and engineering disciplines and are a major part of the discovery and learning process. The
conclusions that can be drawn from an experiment will depend. in part, on how the exper-
iment was conducted and so the design of the experiment plays a major role in problem
solution, This chapter introduces experimental design techniques useful when several fa:~
tors are involved,
.~.:t.ailii(i~l:fr
A Characterization Experiment A development engineer is working on a new process for sol-
dering electronic components to printed circuit hoards. Specifically. he is working with a r.ew type of
flow solder machine that be hopes will reduce the number of defective solder joints. (A flow solder
machine preheats printed circuit boards and then rr..oves them into contact with a wave of liquid sol-
der. This machi:1e makes all the electrical and most of me mechanical COllt.ections of the components
to the pr. ."1ted circuit board, Solder defects require touehup or rework. which adds cost and often dam~
.ages the boards.) The flow solder machine has several variables that the engineer can controI. They
are as follows:
1. Solder temperature
2.. Prebeat temperature
3. Conveyor speed
4. Flux type
5. Flu., specific gravity
6. Solder wave depth
7. Conveyor angle
In addition to t.iese controllable factors, there are several factors that cannot be easily controlled
once the machine enters routine manufacturing, ineluding the following:
1. Thickness of the printed circuit board
2. Types of components used on the board
3. Layout of the components of the board
4. Operator
5. En"irorunental factors
6. Production rate
Sometimes we call tl:Ie uncontrollable factors noise factors, A schematic representation of the
process is shOVf"U in Fig. lJ., L
353
354 Chapter 13 Design of Experiments with Several Factors
Controllable factors
Input
11 r Process Outpl..1:
(printed circuit boards) (flow solder machine} (defects, y)
. J
ll-~
Z1 2'2 Zq
In this sit'.lscion the engineer is interested in characterizing the flow solder m:lClrine; that is, he is inter-
ested in detemrining which factors (borb controllable and uncontrollable) affect the occurrence of
defects on the p.r:inted circuit boards, To accomplish this he can design an experiment that v;:ill enable
him to estimate the magnitude and direction of the factor effects, Sometimes we call an experiment
such as this a screening experiment The information from this characterization study Or screening
experiment can be used to identify the critical factors, to determine the direction of adjustr.::Jent for
these factors to reduce the n:nnber of defects, and to assist in determining which factors should be ca.re.
fully controlled dcring manufacturing to prevent high defect levels and erratic process perfo.rmance.
;Ekli~~:i3:~.J
An Optimiza.tion Experiment In a characterization experiment, we are interested in determining
which factors affect the response. A logical next step is to determine the region in the important fac-
tors that leads to an optimum response, For example. if the response i.,. yield, we would look for a
region of maxim1.ll:l yield, and if the response is cost, we would look for a region of minimum cost.
As an illustration. suppose that the yield of a chemical process is influenced by the operating
temperature and the reaction time. We are currently operating the process at 155°F and 1.7 hours of
reaction time and experiencing yields around 75%. Figure 13-2 shows a view of the time-tempera-
ture space from above. In this graph we have connected points of constant yield with lines. These lines
are called contours. and we have shown the contours at 60%, 70%, 80%, 90%, and 95% yield. To
locate the optimU1l1, it is necessary to design an experiment that varies reaction time and temperature
together. This design is illu:strated in Fig. 13-2. The responses observed at the four points:in the exper-
iment (145'F, 1.2 br), 045°F, 2.2 br), (16S'F. 1.2 br), MId (165'1', 2.2 br) indicate that we should
move in the general direction of increased temperature and lower reaction time to increase yield. A
few additional IUIlS could be performed in this direction to locate the region of maximum yield.
These examples illustrate only two potential applications of experimental design meth-
ods. In the engineering environment~ experimental design applications are numerous. Some
potential areas of use are as follows:
1. Process troubleshootiI:g
2. Process development and optimization
3. Evaluation of :material alternatives
4. Reliability and life testing
5. Performance testing
13-2 Factorial Experiments 355
200
CUrrent operating
160 conditions
150
140
Figure 13M2 Contour plot of yield as a function of reaction time and reaction temperature, illustrat-
ing an optimization experiment.
Factor B
Factor .4 B,
10 20
30 4()
That is. changing factor A from levell to level 2 causes an average response increase of 20
units. Similarly, the main effect of B is
B= 20+40 _10+30 =10.
2 2
In some experiments. the difference in response between the levels of one factor is not
the same at all levels of the other factors. V/hen this occurs. there is an interaction bet\Veen
the factors. For example, consider the data in Table 13-2. At the first level of factor B, the
A effect is
A =30-10 = 20,
and at the second level of factor B. the A effect is
A=O-20=-20.
Since the effect of A depends On the leVel chosen for factor B, there is interaction between
A andB.
Vt'hen an interaction is large, the corresponding main effects have little meaning. For
example, by using the data in Table B-2, we find the main effect of A to be
A 30+0_10+20 =0
22'
and we would be tempted to conclude that there is no A effect. However. when we exam~
ined the effects of A at different levels offactor B. we saw that this was not the case. The
effect of factor A depends on the levels of factor B. Thus, knowledge of the AB interaction
is more useful than knowledge of the main effects. A significant interaction can mask the
significance of main effects.
The concept of interaction can be illustrated grnphically. Figure 13-3 plats De data in
Table 13-1 against the levels of A for both levels of B. Note that the B, and B,lines are
roughly parallel, indicating that factors A and B do not interact significantly. Figure 13-4
plots the data in Table 13-2.ln this grnpb, the B, and B, lines are not parallel, indicating the
interaction between factors A and B. Sucb graphical displays are often useful in presenting
the results of experiments.
An alternative to the factorial design that is (unfortunately) used in practice is to change
the factors one at a time rntberthan to vary them simultaneously. To illustrate this one~factor~
at~a-time procedure, consider the optimization experiment described in Example 13-2, The
Factor B
Faztor A B,
A, 10 10
A, 30 o
,
-~
:3-2 Factorial Experiments 357
:~
i 30~
\D 20 i-
~ 10~
O~
i ................ L .. ~ ___ _
A,
Factor A Figure 13-3 Factorial experiment, no interaction.
50 -
§ 4O~
<
~ 30~ 81
~ 20~ 8,
8 10~ 8,
o~ 8,
I
Figure 13..4 Factoria: experiment, with
A, A, interaction,
engineer is interested in finding the values of temperature and reaction time that maximize
yield, Suppose that we fix temperature at 155'F (the current operating level) and perform
five runs at different levels of time, say 0.5 hour, La hour, 1.5 hours, 2.0 hours, and 2.5
hours. The results of this series of runs are shown in Fig. 13-5. This figure indicates that
maximum yield is achieved at about 1.7 hours of reaction time. To optimize temperature, the
engineer fixes time ar 1.7 hours (rhe apparent optimum) and perfonns five runs at different
temperatures, say l40'F, 150'F, 160'F, 170'F, and 180'F. The results of this ser of runs are
plotted in Fig. 13-6. Maximum yield occurs at about 155°F. Therefore, we would conclude
that running the process at 155'F and 1.7 hours is the bestset of operating conditions, result-
ingin yields around 75%,
BO
70 •
~ I •
.",
:.; 60
•
>-
• •
50
70
~
:Q
m
:;: eo •
Temperature (!JA
Figure 13~6 Yleld versus temperature with reaction time constant at 1.7 hr"
Figure 13-7 displays the contour plot of yield as a function of temperature and time
with the one-factor-at-a~time experiment shown on the contours. Clearly the one-factor-at-
a-time design has failed dramatically here, as the true optimum is at least 20 yield points
higher and occurs at much lower reaction times andbigher temperatures. The failure to dis-
cover the shorter reaction times is particularly important as this could have significant
impact on production volume or capacity, production planning, manufacturing cost. and
total productivity.
The one-factor-at-a-time method has failed here because it fails to detect the interac-
tion between temperature and time, Factorial experiments are the only way to detect inter-
[ 18D!
~
~m
0.
E
m
t-
150l
14G~
'-_;;,',;:" .--~~.:--.~.---.-.-L.----L......-.-
Oll 1.0 1.5 2.0 2.5
Time (hr)
Figure 1J..7 Optimization experiment using the onc-:factor-at~a~time method.
13-3 Two-Factor Factorial Experiments 359
(13-1)
where J.1 is the overall mean effect.. -t: is the effect of the ith level of factor A,~, is the effect
of the jth level of factor S, (1:!llij is the effect of the interaction between A and S, and is 'il'
a NID(O, a') (norma! and independently distributed) random error component. We are
interested in testing the hypotheses of no significant factor A effect, no significant factor B
effect, and no significant AB interaction. As with the single-factor experiments of Chapter
12, the analysis of variance will be llsed to test these hypotheses, Since there are two fac-
tors under study, the procedure used is called the two-way analysis of variance.
a ""J!:l:!
360 Chapter 13 Design of Ex-periments with Several Factors
observations. Define Yf ... YT' YIj" and y... as the correspondlng row. colllIIlD.. cell, and
grand averages. That is,
.1
i::::: 1.2•... ~a,
= LLLfCVI.-Y .. )+(Yj.
r"'l 1::::]·k=-1 •
-yJ
+ (YiJ. -Yio. -Y.I +Y... )+(Y,jk-Yi}.jf
a 2 b 2 (13-3)
=bnL(Yi.. -Y.) +anL(Y:! -Y:)
1",1 1::::1
ab "aon, 2
+nLL(YJ}. -y, -Y.j+Y.,J + LLLlY,!k -Yij) .
1""1 j=-l i=1 1",1 k><ol
Thus. the total sum of squares is partitioned into a sum of squares due to "rows," or factor
A (SSA)' a sum of squares due to "columns," or factor B (SS.), a sum of squares due to the
interaction between A and B (SSM). and a sum of squares due to error (SSE)' Notice that
there must be at least tv.'o replicates to obtain a nonzero error sum of squares.
The sum of squares identity in equation 13-3 may be written symbolically as
(13-4)
There are abn - 1 total degrees of freedom. The main effect<; A and B have a - I and
&-1 degrees of freedom, while the interaction effectAB has (a - 1)(b-1) degrees cffree-
dom. Within each of the ah cells in Table 13-3, there are n -I degrees of freedom between
the n replicates, and observations .in the same cell can differ only due to random error.
Therefore, there are ahen -1) degrees of freedom for error. The ratio of each sum of squares
on the right~hand side of equation 13-4 to its degrees of freedom is a mean square.
Assuming that factors A and B are fixed, the expected values of the mean squares are
13-3 Two-Factor F2.ctorial Experimcnt:5 361
a b
, nL2JrJ3)~j
E(MS ) ~ E( SS,s J ~ ,,2 + i=[ j=1
All (a-l)(b-I) (a-l)(b-I) .
and
Therefore} to test He: 't;::;:; 0 (no row factor effects), Ho: /3,- -;::; 0 (no column factor effects), and
H,: (~fi),! ~ 0 (no interaction effects), we would divide'the corresponding mean square by
the mean square error, Each of these ratios will follow an F distribution with numerator
degrees of freedom equal to the number of degrees of freedom for the numerator mean
square and ab(n-I) denominator degrees of freedom, and the critical region will be located
in the upper tail, The test procedure is arranged in an analysis-of-variance table. such as is
shown in Table 13-4.
Computational fonnulas for the sums of squares in equation 13-4 are obtained easily,
The total sum of squares is computed from
a b r.,., y2
SSy~ LLLYi;, -~"'"
1=1 j=l aim .(:",1'
(13-5;
(13-6)
and
b 2 2
SSB -~.L,;
'\' Y,j. y",
--. (13-7)
}=1 an abn
We usually calculate the SS" in two steps. First, we compute the sum of squares between
the ab cell totals, called the sum of squares due to Hsubtotals":
IJ b y~.
SSS\lbtO".,;:Us;; LL..JL:..-
nan
1",1 J""1.
b
Table 134 The .Analysis~of- Variance Table for the Two- Way~aassification Fixed~Effects Mode:
l
i
!
362 Chapter 13 Design of Experi;;nentS with Several Factors
'This sum -of squares also contains 5Sft and SSs' Therefore, the second step is to compute
SSAS as
. E~pl.13,3·.
Aircraft primer paints are applied to alUI1li.nilI!l. surfaces by two n:.ethods: dipping and spraying, The
purpose of the primer is to improve paint adhesion. Some par...s can be primed using either applica-
don method and engineering is interested i'11eamiug whether th...--ee different primers differ in their
adhesio::::. properties. A factorial experiment is performed to investigate the effect of paint primer type
and application method on paint adhesion. 'Three speci.-.ner.s are painted with each primer using each
application method. a ilnish palnt applied, and the adhesion force measured. The data from the
experiment are shown in Table 13-5. The circIed nUDlbers in the cells are the cell totals Y,r . The sums
of squares required to perfow the analysis of variance are computed as follows:
Ss. "' ,
= ~YL_~
,)'?'!5 ~ or. abn.
(89.8)'
4.58- 4.91 = 0.24.
3 18
and
The analysis of variance is summarized in Table 13~6, Sinee F'J,(r.;,i,l'2 =- 3,S9 and FC.O,..,11 = 4.75. we
..:onclude that the main effects of primer type and application method affect adhesion force. Fu..'tber M
more, since 1.5 <: F O.05 ,2,12' there is no indication of interaction between these factors.
13-3 Two-Factor Fac:or'.J1l Experiments 363
A graph of t.':1e cell adhesion force averages Yir versus the levels of primer type for each application
method is shown in Fig. 13--8. The absence ofintcraction is evident by the parallelism of the two lines.
Furthermore. since a large response indicates greater adhesion force, we conclude that spraying is a
superior application method and that primer type 2 is most effeetive.
Tests 011 Individual Mean. When borh factors are fixed, comparisons between rhe indi-
vidual means of either factor may be made using Tukey's test. When there is no interaction;
these comparisons may be made using either the rOw averages Yi" or the column averages
Y.).. However. when interaction is significant, comparisons between the means of one factor
(say A) may be obscured by rhe AS interaction. In rhls case, we may apply Tukey's test to
the means of factor A, 'With factor B set at a particular level.
Yi}·t
7,0-
6.0 rf ~ng
5,Or
4.0~
~Ping
a,oc
2 Figure 13-& Graph of average adhesion force versus
Prlmertypa primer types for Example 13-3,
364 Chapter 13 Design of Experiments with Several Facton:
That is~ the residuals are JUSt the difference between the observations and the corresponding
cell averages.
Table 13-7 presents the residuals for the aircraft primer paint data in Example 13-3.
The normal probability plot of these residuals is shown in Fig. 13-9. This plot has tails tbat
do not fall exacdy along a straight line passing through the center of the plot, indicating
some potential problems with the normality assumption, but the deviation from normality
does not appear severe. Figures 13-10 and 13-11 plot the residuals versus the levels of
primer types and application methods, respectively, There is some indication that primer
type 3 results in slightly lower variability in adhesion force than the other two primers. The
graph of residuals versus fitted values Yl/k;;;;; YU' in Fig. 13-12 reveals no unusual or diag~
nostic pattern.
Table 13M7 Residuals for the Aircraft Primer Paint Experiment in Example 13-3
Application Met.'-lod
Primer Type Dippi.'g Spraying
I -iJ.27, 0.23, 0.03 0.10, -OAO. 0.30
2 0.30, -0.4{), 0.10 -0.27, 0.03, 0.23
3 -0.03, -0.13, 0.17 0.33, -iJ.17, -0.17
2.0.
1J ••
•
•
•
ZI oJ ••
••
-1.0 •
•
-2.0 • ! ..J
-0.5 -0.3 -0.1 +0.1 +0.3
SI}X' residual
Figure 13-9 Nonnal probability plot of the residuals from Example B-3.
13-3 Two~Factor Factorial Experiments 365
+0.5
Or-------~~------~~~~-· ~
D S Application method
••
-0.5 c
•
•
••
0 • •
.4 5 6 A
Yijk
• ••
• •
••
-0.5
in the analysis-of-variance model as there are observations, and the error degrees offree~
dam is zero. Thus, it is not possible to test a hypothesis about the main effects and interac~
tians unless some additional assumptions are made. The usual assumption is to ignore the
interaction effect and use the interaction mean square as an error mean square. Thus the
analysis is equivalent to the analysis used in the randomized block design. This
366 Chapter 13 Design of Ex.periments with Several Factors
no~interaction assumption can be dangerous, and the experimenter should carefully exam~
ine the data and the residuals for indications that there really is interaction present. For more
details, see Montgomery (200 1),
where the parameters 1:i' pj • (~f5)ii' and ti" are random variables. Specifically, we assume
that~i is NIO(O, a;), p) is NIO(O, ~), (o/!I} is NIO(O, 0".,), and e,jk is NIO(O, 0"), The vari·
ance of any observation is
V(Yij,) ~ a; + ~ + a;p + 0",
and 0;.. ~ o;.~~ and (52 are called variance components. The hypothcses that we are inter~
ested in testing are H,: ~ ~ G, Ho: ~ ~ G, and Ho: 0"" ~ O. Notice the similarity to the one·
way classification random-effects model,
The basic analysis of variance remains unchanged; that is, 8S".• SSe. SSw SSI~ and SSE
are all calculated as in the fixed-effects case, To construct the test statistics. we mUIit exam-
ine the expected mean squares. They are
E(MS,) ~ 0" + no"., + bna;,
E(MS,) ~ 0" + na:." + ana;,
E(MSA,) ~ 0" + na:p, (13·11)
and
E(MS,J = <1'.
~ote from the expected mean squares that the appropriate statistic for testing H;;:
0"",: 0 is
(13·12)
since under Ho both the numerator and denominator of Fo have expectation <f', and only if
He is false is E(MSAIl ) greater than E(MS,), The ratio Fo is distributed as F,,_ ""_ !),,~".IJ'
Similarly. for testing Ho: 0-; ::: 0, we would use
which is distributed as F £>""1,(0-:)(1;-1)' These are all upper-tail, one. .tail tests, Notice that these
test Statistics are not the sarne as those used if both factors A and B are fixed, The expected
mean squares are always used as a guide to test statistic construction.
13-3 Two-Factor Factorial E}."Periments 367
The variance components may be estimated by equating the observed mean squares to
the::r expected values and solving for the variance components, This yields
-2
(j = MS"
(13-15)
Suppose that in Example 13-3, a large number of primers and several application methods could \:Ie
used. 'Three pri..'"Uers, say 1. 2. anc 3, were se:ected at rancom, as were :he two application methods,
Tne analysis of va.--iance assuming the rando:n effects :;node! is sho'W'D in Table 13-8.
Nerice that the first four colurr.ns.in the analysis ofvanance table are exactly as in Exa.'1lple 13-
3, Xow. however. the F ratios are computed according to equations 13~12 through 13~14. Since
F&,~,1.l2::::: 3.89, we conclude that interaction is not sigrlficant. Also, since FI)IJ!,;';;::::: 19_0 and FM).I.? =
18,5, we conc!ude that both types and application. methods significantly affect adhesion force,
although primer type is just ba:ely significant at (j. = 0,05. The '1-ariance components may be o!Stimated
using equation 13-15 as follows:
,,' =0,08,
0,12-0,08 00133
v"
-2
= 3 "
_, 2,29-0,12
"r,~ 6 ",0,36,
Clearly, :he tw'Q largest variance components are for prirr.er typeS (6';::::: 0.36) and application meth-
ods 0.53),
In this model, 'li is a fixed effect defined such that 'L~ '" I 't, = 0, ~ Is a random effect, the inter~
action term ('f/3)ij is a random effect, and el;~ is a NID(O, a2) random error. Ie is also cus-
tomary to assurne that Ilj is NlD(O, <i,,)
and that the interaction elements (rfJ!ij are nonnaJ
random variables "ith mean zero and variance l(a - 1)la]d'fJ' The interaction elements are
not all independent,
The expected mean squares in this case are
a
bnI ~?
E(.M:SA)=a2+/ra;jJ+ h:
a-I
(13-17)
E( MS. ) ~(f 2 +an(fp,
2
and
(13-18)
whicb is distributed as F.. _1,(c-I)(I1- n' For testing He: cr1 = 0, the test statistic is
MSB
Fo="--, (13-19)
MSe
which is distributed as Fb_l,ab(n_ 1)' Finally. for testing Ho:cr;p = 0, we would use
(13-20)
and
(13-21)
This general approach can be used to estimate the variance components in any mixed
model. After eliminating the mean squares containing :fixed factors, there will always be a
set of equations remaining that can be solved for the variance components, Table 13-9
summarizes the analysis of variance for the two~factor mixed model.
J
13-4 General Factorial Experb.en.ts 369
(13-24)
(13-25)
c 2 2
SSC _- "" Y..k, Y....
I £ . . , - - - - -.
.1:=1 abn. abcn
(13-26)
l
310 Chapter 13 Design of Experiments with Several Factors
To compute the two-factor interaction sums of squares, the totals for theA x B,A x C, and
B x C cells are needed. It may be helpful to collapse the original data table into three two-
way tables in order to compute these totals. The sums of squares are
rr
11. b
;""1 j=l
,2
Yij.
en
"2
c_L',-SSA -SSo
abcn (13-27)
= SSw.hroroh(AlT; - SSA - SSB'
SSAC = rr
11. C 2 2
Yd. -~-SSA -SSe
i=l k""l bn abcn (13-28)
;; SSwblotalS{AC) - SSA - SSe>
and
SSnc = rr
b c
J=l K""t
2 2
y ,j>' -~-SSB -SSe
an abcn (13-29)
= SS'.'''WCBC) - SSe -SSc·
The three-factor interaction sum of squares is computed from the three"way cell totals Yijt-
as
Table 13-10 The Analysls~of-Variance Table for the Three-Factor Fixed-Effects Model
, MS;.
SS, a-I
(1z + bcrir."t:
A MS,
a-t MS,
) acnI.j3; MSn
B SSe b-I MS. CI +~-
b-I MS,
'nl; , MSc;_
(12+atJ It
C SSe e-I MS c
c-I MS,
01 + brIE{'ry)~~ MS AC
AC SSAC (a-I)(e-I) MS;.c (a-l)(c-l) MSc
;
.J
13-4 Gen::ral Factorial Experiments 371
(13-30b)
The error sum of squares may be found by subtracting the sum of squares for each main
effect and interaction from the tOtal sum of squares, or by
(13-31)
a ,,2 y'1
SSA=I:!i.:.:..:...-~
[ .. I ben abcn
a b;( 2
SSAii ::::: L: 2: Yuen" - Men
;"'! j .. j
y:". -SSA - SSs
=0.0625,
b c 2 2
SSlJC -_""Y.jk. Y.... SSe- 5Sc
..t......t....------
j .. :.t»o: an abcn
+148\' (J77')'
='138"I +(44)' _ (4")'
" I . 10.5625-3.0625
4 16
L
372 Chapter 13 Design of Experiments \'lith Several Factors
1M analysis of ..'afiance is summarized in Table 13-12. Feed rate has a significant effect on su.rface
finish (a: < 0,01). as does the depth of cut (0.05 < a < 0.10). There is some indication of a mild inter-
action betweet. these factors, as the Ftest for the AJ3 interaction is just less than the 10% critical value.
Obviously factorial experiments wit'" three or more factOTS are complicated and require
many runs, particularly if some of the factors have several (more than two) levels. This '
leads us to consider a class of factorial designs with all factOrs at two levels. These designs
are extremely easy to set up and analyze, and as we will see, it is possible to greatly reduce
the number of expe~ental runs through the tecbnique of fractional replication.
High b
(+) ,....-----------tab
Low
(-) (1) a
Low A High
(-) (+) Figure 13·13 The 2' lilctorial desigr..
374 Chapter 13 Design of Experiments with Severa:. Fac:t.ms
in that treatment combination; if it is absent, the faetor is run at its low level. For example.
treat:raent combination a indicates that factor A is at the high level and faetor B is at the low
leveL The treatment eombination with both factors at the low level is denoted (1), 'This nota-
tion is used throughout the 2K. design series. For example. the treatment combination in a 24
design with A and C at the high level and B and D at the low level is denoted ac.
The effects of interest in the 22 design are the main effects A and B and the two~factor
interaction AB. Let (I), a, b, and ab also represent the totals of all n observations taken at
these design points. It is easy to estimate the effects of these factors. To estimate the main
effect of A we would average the obsenrations on the right side of the square, where A is at
the high level. and subtract from this the average of the observations on the left side of the
square, where A is at the low level, or
A= a+ab _ b+(l)
2n 2n
(13-32)
J...[a +ab -b -(1)].
2n
Similarly, the main effect of B is found by averaging the observations on the top of the
square, where B is at the high level. and subtracting the average of the observations on the
bottom of the square, where B is at the low level:
B= b~aiJ._ a+(l)
2n 2n
(13-33)
= I- [b+ab-a-(I),..
2n '
Finally, the AB interaction is estimated by taking the difference in the diagonal averages in
Fig. l3~13i or
(13-34)
The quantities in brackets in equations 13-32, 13-33, and 13-34 are called contrasts.
For example, the A contrast is
Contrast, ~ a - ab - b - (1).
mthese equations, the contrast coefficients are always either +1 Or -L A table of plus and
minus signs, such as Table 13-13, can be used to determine the sign of each treatment com~
bination for a particular contrast. The column headings for Thble 13MB are the main effects
A and B, AB interaction. and I, which represents the total. The row headi:r.:.gs are the treat~
ment combinations, Note that the signs 1:1 the AS column are the products of signs from
(1) + +
a + +
b + +
ab + + ~ +
13·5 The 2' Factorial Design 375
colu."llOS A and B. To generate a contrast from this table, multiply the signs in the apprQpri~
ate column of Table 13-13 by the treatment combinations listed in the rows and add.
1b obtain the sums of squares for A, B, and AB, we can use equation 12-18; which
expresses the relationship between a sing1e-degree-of~freedom contrast and its sum of
squares:
SSA
40
SS. =
4n
SSAB =
40
The analysis of variance is completed by computing the total sum of squares SST (~ith
40 - 1 degrees offreedom) as usual, and obtaining the error sum of squares SSE [with 4(0
- 1) degrees of freedom] by subtraction.
§;;§gI~',l~~
An article in the AT&T Technical Jo"""u (MarcblApril, 1986, Vol. 65, p. 39) describes the applica-
tion of two-level experimental designs to integrated circuit manufacturing. A basic processing step in
t.'1.is industry is to gr:ow an epitaxial layer on polished silicon wafers. The wafers are mounted On a sus-
ceptor and positioned inside a bell jrtr, Chemical vapors are introduced rr..rough nOl.7..les near the top
of the jar, The susceptor is rotated and heat js applied. These conditions are maintained until the eplR
taxiallayer is thick enough.
Table 13-14 presen!'$ theresuJts of a 2' factorial design \\-1.thn=4replicates using the factors A ""
deposition time and B:::: arsenic flow rate. The two leve~s of deposition time are-= short and,.;. == long,
and the two levels of arsenic flow rate are -:=; 55% and - = 59%. The response variable is epitaxial
layer thickness (urn). We may find the estimates of the effects using eq'Jations 13 32, 13-33, and
R
13-34 as follows:
A = ;n [a+ah-b-(ll]
l
376 Chapter 13 Design of Experiments with Several Factors
B=..!..[b+ab-a-(I)]
2n
1 [55.686+59.156-59.299-56.081
=2(4) • ] =-{).O67,
AB=..!..[ab+(I)-a-b]
2n
1
= 2(4) [59.156+56.081 -59.299-55.686]=0.032.
The numerical estimates of the effects indicate that the effect of deposition time is large and has a pos-
itiYe direction (increasing deposition time increases thickness), since changing deposition time from
low to high changes the mean epitaxial layer thickness by 0,836 pm. The effects of arsenic flow rate
(B) and the AB interaction appear small.
The magnitude of these effects may be eonfi'llled with the analysis of variance, The sums of
squares for A, E, andA,B are computed using equation 13-35:
(Contrast)'
SS=-- -,
n·4
ra+ab-b-(I)l'
SS = <
A 16
= [6.688]2
16
~2.7956,
•
SS.=-
,[b_+_ab_-_"_-l-'(llCL.l'
<
16
[-{).5381'
16
= 0.0181,
[ab+(I)-a-bj'
SSAB 16
_ [O.256f
- 16
=0.0040.
The analysis of variance is summarized in Table 13-15. This confirms our conclusions obtained by
examining the magnitude and direction of the effects; deposition time affects epitaxial layer tb.ickness,
and from the direction of the effect estimates we know that longer deposition ti.:n.es lead to thicker epi~
taxia11aYer5,
ResidualAnalysis It 1s easy to obtain the residuals from a 2;; design by fitting a regression
mode! to the data. For the epitaxial process experiment, the regression model is
y; 130 + 13,x, + E,
since the only active variable is deposition time, which is represented by Xl' The low and
high levels of deposition time are assigned the vaiues Xl =-1 and Xl =+1, respectively. The
fitted model is
y 14.389+lro.836ixl>
2 /
where the intercept A, is the grand average of all 16 obServatiODS (Y) and the slope $, is one·
half the effect estimate for deposition time. The reason the regression coefficient is one-half
the effect estimate is because regression coefficients measure the effect of a unit change in
Xl on the mean of y, and the effect estimate is based on a two-unit change (from -1 to +1),
This model can be used to obtain the predicted values at the four points in the design.
For example, consider the point with low deposition time (x: = -1) and low arsenic flow
rate. The predicted value is
ji 14.389 + (0.836
2 J
IH): 13.971 pm,
+T
ZJ
+1
o~ .. .••.
•
•
1
-t
-2~
•
••
O~------~L~O-W------~H~i9~h----~
•
ol~:------~IL~O-W----~H~i~9h---- ...
s~O.077
(+) b
points. Most of this difference is attributable to the unusually low thickness measurement
associated with YIS' The standard deviation of the fout' observations -with A and B at t.'1e low
levels is also somewhat larger than the standard deviations at the remaining two runs. This
could be an indication that there are other process variables not included in this experiment
that affect the variability in epitaxial layer thickness. Another experiment to study this pos-
sibility, involving other process variables, could be designed and conducted (indeed, the
original paper shows that there are two additional faetors. uneonsidered in this example,
that affect process variability).
Be
e r''---T-,
,--:err-
I
I
I
I
bi __
~~_ ------ ab
..,- ... ..,-' , +1
-1 :... ( 1 ) " " ' - - - - -. . /
-/a
-1
A
Figure 13·18 The 2' desigu.
level, and sUbtracting from that quantity the average of the four treatment combinations on
the left side of the cube, where A is at the low level. This gives
I
A = --[a ~ ab+ac + abe- b -e- bc- (I)]. (13-36)
4"
In a similar manner the effect of B is the average difference of the four treatment combina-
tions in the back face of the cube and the four in the front, or
1 •
C=--lc~ac+be+abc-a-b-ab-\I
, )1 ,
(13-38)
4n '
~ow consider the tv/o-factor interaction AB. \¥hen C is at the low level, AB is just the
average difference in the A effect at the t\Vo levels of E, or
The ABC interaction effect is the average difference between theAB interaction at the two
levels of C. Thus
ABC = :n ([abc~bcl-[ac-cl~[ab~bl+[a~{I)]}
(13-42)
= 2-[ abc -
4n
bc~ac+ c~ab + b + a - (I)l..
The quantities in brackets in equations 13-36 through 13-42 are contrasts in the eight
treatment combinations. These contrasts can be obtained from a table of plus and minus
signs for the 2' design, shown in Table 13-16. Signs for thc main effects (colu.mns A, B, and
C) are obtained by associating a plus with the high level of the faetor and a minus with the
low level. Once the signs for the main effeets have been established. the signs for the
remaining eolumns are found by multiplying the appropriate preceding columns row by
row. For example. the signs in column AB are the produet of the signs in columns A and B.
Table 13-16 has severnl interesting properties.
1. Except for the identity column l, each eolumn has an equal number of plus and
minus signs.
2~ The SUJr. of products of signs in any two columns is zero; that is, the columns in the
table arc orth..ogonal,
3. Multiplicating any column by column [leaves the column unchanged; that is, I is an
identity element.
4. The product of any two columns yields a column in the table; for example, A x B =
AS and AB x ABC =A'B'C C, since any colu.mn multiplied by itself is the iden-
tity column.
The estimate of any main effect or interaction is determined by multiplying the treat-
ment combinations in the first column of the table by the signs in the corresponding main
cffect or interaction colUlllll. adding the result to produce a contrast, and then dividing the
contrast by one-half the tolal number of runs in the expctimcnt. Expressed mathematically,
(Contrast)2
SS - (13·44)
n2k
F!¥~il(>!~11~J
Consider the surface--roughncss experiment described originally in Example 13-5. This is a: 23 facto-
rial design in the factors feed rate (.4)j deptl::. of cut (B), and tool angle 'G), 'Yri.tb 10 =2 replicates. Table
13-17 presents the observed surface~rougbness data.
The main effects may be estimated using equations 13-36 through 13-42. The effect of A is, for
example.
1
=-[27]=3375,
8
and the sum of squares for A is found using equation 13-44:
(ContrastS
(21)'
=--=455625
2(8) ~ .
C= 0.875.
AB 1.375,
AC=O.I25.
BC=-O.625.
ABC= 1.125.
From. examining the magrrimde of the effects. clearly fced rate (factor A) is dom.i."lant. followed by
depth of cut (8) and the AB interaction. although the interaction effect is relatively small. The analy-
sis of \':It....ance is summarized in Table 13-181 and it confirms our interpretation of the effect estimates.
Total 92.9375 15
Other Methods for Judgfng Significance af Effects The analysis of variance is a formal
way to detenni:1e which effects are nonzerQ. There are t\\'O other methods that are useful.
In the first method, we can calculate thc standard errors of the effects a.~ compare: the mag-
nitudes of the effects to their standard errors. The second method uses normal probability
plou> to assess the importance of the effects,
The standard error of an effect is easy to find, If we assume that there are ft replicates
at each of the 26 runs in the design, and if Yil' yc~' , .. , Yill are: the observations at the ith run
(design point), then
Where we have obviously assumed equal variances for each design point. This is also the
variance estimate gNen by the mean square error from the analysis of variance procedure.
Each effect estimate has variance given by
V{Effect) = v[contrast]
n2k 1
1
-:r
-'-'::""'T,
(n2 k
V(Contrast).
Each contrast is a linear combination of 27. treatment totals, and each total consists of n
ohservations. Therefore.
V(Contrast) = n2'<i'
and the variance of an effect is
The estimated standard error of an effect would be found by replacing rr with its estimate
s' and taking the square root of equation 13-46,
To illustrate for the surface-roughness experiment, we find that S' = 2.4375 and the
standard error of each estimated effect is
s,e,(Effect) = ~;;
= ,} 2, ~3-2 (2.4375)
=0,78,
Therefore two standard deviation limits on the effect estimates are
A: 3.375 ± 1.56,
B: 1.625± 1.56,
C: 0,875 ± 1.56,
AB: 1.375 ± 156,
AC: 0,125 ± 1.56,
BC: -0,625 ± 156,
ABC: Ll25 ± 1.56,
These intervals are approximate 95% confidence intervals. They indicate that the wo main
effects. A and 8, are important, but that the other effects are not, since the intervals for all
effects except A and B include zero.
Normal probability plots can also be used to judge the significance of effects, We will
illustrate that method in the next section.
Projection o/2k Designs .Any zi< design will collapse or project into another i:. design in
fewer variables if One or more of the original factors are dropped Sometimes this can pro-
vide additional insight into the remaining factors. For example, consider the surface-rough~
ness experiment. Since factor C and all its interactions are negligibl~ we could eliminate
factor C from the design. The result is to collapse the cube in Fig. 13~ 18 into a square in the
A - B plane-however, each of the four runs io the new design has four replicates, In gen-
era!, if we delete h factors so that r = k - h factors remain, the origioalZ' design with n repli-
cates will project into a 2r design with n2h replicates.
Residual Analysis We may obtain the residuals from a 2' design by usiog the method
demonstrated earlier for the 22 design. As an example, consider the surface~roughness
experiment. The three largest effects areA, B, and theA.Binteractiou. The regression model
used to obtain the predicted values is
y = Po + p,x, + Ax, + /3,,,,,,,,,,,
where Xj represents factor A~.l1 represents factor B, and x!A:z represents the AD interaction.
The regression coefficients,Bl,A. andft!~ are estimated by one·balfthe corresponding effect
esti:clates andfJo is the grand average, Thus
,
y =.
110625'~ ,(3.3751
- - IX 1 + (1.6251 (1.375\
- 0-)X, + --)x}xz.
\ 2 / '" 2
13-5 The 2< Factorial Design 385
and the predicted values would be obtained by substituting the low and high levels of A and
B into this equation, To illustrate, at the treatment combination where A, B, and C are all at
the low level, the predicted value is
=9.25.
The observed values at thls run are 9 and 7, so the residuals are 9 - 9.25 = -0.25 and
7 - 9.25 ~ -2.25. Residuals for the other seven runs are obtained similarly.
A normal probability plot of the residuals is shown in Fig. 13-19. Since the residuals
lie approximately along a straight line. we do not suspect any severe nonnormaJiry in the
data. There are no indications of severe outliers. It would also be helpful to plot the resid·
uals versus the predicted values and against each of the factors A, 8, and C.
Yates' Algorithm for the 2" Instead of using the table of plus and minus signs to obtain the
contrasts for the effect estimates and the sums of squares, a simple tabular algorithm
de'\1.sed by Yates can be employed, To use Yates' algorithm. construct a table with the treat-
ment combinations and the corresponding treatment totals recorded in standard order. By
standard order, we mean that each factor is introduced one at a time by combining it with
all factor levels above it. Thus for a 22, the standard order 1s (1), a, b. 00, while for a 23 it is
(1), a, b, ab, c, ac, be, abc, and for a 2' it is (1), a, b, ab, c, ac, be, abc, d, ad, bd, abd, cd,
acd, bcd, abed. Then follow thls four-step procedure:
L Label the adjacent column [IJ. Compute the entries in the top half of thls column by
adding the observations in adjacent pairs. Compute the entries in the bottom half of
this column by changing the sign of the first entry in each pair of the original obser~
vation. and adding the adjacent pairs.
2. Label the adjacent column [2J. Construct column (2) using the entries in column
(1). Follow the same procedure employed to generate colurr,n (1]. Continue this
process until k columns have been constructed, Colunm [k] contains the contrasts
designated in the rows.
3. Calculate the sums of squares for the effects by squaring the entries in column [k]
and dividing ey ni:'.
4. Calculate the effect estimates by dividing the entries in column [k) by n2'·'.
+2! I
+1 ~
•
~ 0-
:~
-2~~2~5~OO~--~~--~~~--~~~--~~~--1~.O~a~3~3----7i.~75~OO
J
Residual, ej
Figure 13-19 Normal probability plot of residuals fro", the surface·roughness experimen:
1
386 Chap,er 13 Design of E.."'Cperiments v.ith Several Factors
Exa ';'/lie13-8.
Consider the surfacc-rougr..ness experiment in Example 13-7, This is a 23 design with n;= 2 replicates.
The analysis of this data using Ya,es' algorithm is illustrated in Table 13~19. Note that the sums of
squares computed from Yates' algorithm agree with the resultS obtained in Example 13-7,
.~~ ... ~~-
.Exam;,l" 1:>:9
An article in Solid State Technology ("Orthogonal Design for Process Optim.i:z.ation and its Application
in Plasma Etchir.g," May 1987, p. 127) describes the application of factorial designs in developing a
nitride etc]:: process ou a single-wafer plasma etcher. The process uses y,., as the reactant gas. Iris pos-
sible to 'vary the gaz flow, the power applied to the cathode, the pressure in the reactor chamber, and the
spacing berwee!l the ::mode azd the cathode (gap). Several response variables would usually be of i:.ter-
cst in this process, but in this exa.:mple we will concentrate on etch rate for silicon ni::ride.
We will use a single replicate of a r design to i.;:.vestigate this process. Since it is un!ikcly that
the three-factor and four-factor intera.ctions are significant, we will tentatively plan to combine diem
as an estim.ate of error. The factor levels nsed in the design are shown here:
Design Factor
A B C D
Gap Pressure c,F,Flow Power
Level (em) (roThrr) (SeeM) (w)
J
13-5 The 21 Factorial Design. 387
Table 13-20 presents the data from the 16 runs of the 24 design. Table 13~21 is the table of plus
and minus signs for the 24 design. The signs in the columns of this table can be used to estimate the
factor effects. To illustrate, the estimate of factor A is
A B C D Etch Rate
(Gap) (Pressure) (C,F, Flow) (power)
-1 -1 -I -1 550
-1 -1 -I 669
-I 1 -I -1 604
1 1 -1 -1 650
-1 -1 -1 633
1 -J -1 642
-1 ! -1 601
1 1 1 -1 635
-1 -1 -1 1 1037
-1 -1 1 749
-1 -1 1052
1 -1 868
-1 -1 1075
1 -1 860
-1 1 1 1063
1 729
388 Chapter 13 Design ofExperi.."Pents with Several Factors
significant, then the estimates will behave like a random sample drawn from a normal dis~
tribution with zero mean, and the plotted effects will lie apProximately along a straight line,
Those effects that do not plot on the line are significant factors:.
The normal probability plot of effect estimates from the plasma etch experiment is
shown in Fig, 13-20, Clearly the main effects of A and D and the AD interaction are signif-
icant, as they fall far from the line passing through the other points, The analysis ofvmance
summarized in Table 13-22 confirms these findings. Notice that in the analysis of variance
we have pooled the three- and four~factor interactions to form the error mean square. If !.he
nonnal probability plot had indicated that any of these interactions were important, they
then should not be included in the error term.
Since A =-101.625, the effect of increasing the gap benveen the cathode and anode is
to decrease the etch rate. However. D 306.125, so applying higher power levels \Vill
increase the etch rate, Figure 13-21 is a plot of the AD interaction, This plot indicates IIlat
the effect of changing the gap width at low power settings is smaU. but that increasing the
+2,~,------,~----~----'------r-----r------Dnil
~
l
+1
Zj O~.
gap at high power settings dramatically reduces the etch rate. High etch rates are obtained
at high power settings and narrow gap widths.
The residuals from the experiment can be obtained from the regression model
1400
1200
c
1000
]
800 - - - - - - - - - - - Dh;gh = 325w
~
~ 600 _-------
.. D10w = 275w
u
W
400
200
0
Low (0.80 em) High (1.20 em)
A (Gap)
Figure 13~21 AD interaction from the plasma etch experiment
390 Chapter 13 Design of Experlments with Several Factors I
•
•
•
.. •
•
•
• •
Block 1 Block 2
[IJ
(1)
Block 1 Block 2
ab
(1) ~ 8C , c
all
~ be [abc
Figure 13<-23 The 2: design in two blocks, Figure 13~24 The 2} design in two blocks,ABC
confounded.
The reason for this is apparent from the table of plus and Illlnus signs for the 2' design
(Table 13-13), From this tablel we see that all treatment combinations that bave a plus on
AB are assigned to block 1. while all treatment combinations t..1at have a minus sign on AB
are assigned to block 2.
This scheme can be used to confound any 2k design in two blocks. As a second exar.I-
ple. consider a 2' design. run in two blocks. Suppose we wish to confound the three-factor
interaction ABC witb blocks, From the table of plus and minus signs for the 2' design (Table
13-16), we assign the treatment combinations that are minus on ABC to block 1 and those
that are plus onABCto block 2. The resulting design is shown in Fig. 13-24.
There is a more general method of constructing the blocks. The method employs a
defining contrast, say
(1347)
where Xi is the le'vel of the ith facto:: appearing in a treatr:lent combination and (Xi is the
exponent appearing on the ith factor in the effect to be confounded. For the 2k system. we
have either lX, = 0 Or lj and either x, = 0 (low level) or Xi =: 1 (high level). Treatment combi-
natious that produce tbe sarne value af L (modulus 2) will be placed in tbe sarne block.
Since the only possible values of L (mod 2) are 0 and I, this ,,'ill assign tbe 2' treatment
combinations to exactly wo blocks.
ill an example consider a 2 3 design with ABC confounded with blocks. Here Xl corre-
sponds to A. X Z to E, X] to C, and lX, :::::: ~ :::: OJ = L Thus. the defining contrast for ABC is
L=x1 +X1 +X1•
To assign the treatment combinations to the wo blocks, we substitute the treatment com-
binations i:Jto the defi.n.ing contrast as follows:
A ~hortcut method is useful in constructing these designs. The block containing the
treaunent combination (I) is called the principal block. Any element [except (I)l in the
principal block may be generated by multiplying two other elements in the principal block
modulus 2. For example, consider the principal block of the 2' design with ABC con-
founded, shown in Fig. 13-24. Note that
ab . ac :::; a?bc bC j
ab· bc=ab"J.c=ac,
a.c • be = abC' = abo
Treatment combinations in the other block (or blocks) may be generated by multiplying one
element in the Dew block by each element in the principal block modulus 2. For the 2' with
ABC confounded, since the principal block is (I), ab, ac, and be, we know that b is in the
other block. Thus, the elements of this second block are
b'(I)=b,
b·ab=ab'=a,
b 'ac =abc,
b . be = b'c = e.
Jt'xample13;:lO
An experiment is performed to investigate the effects of foW" factors on the terminal miss distar.ce of
a shoulder-fired ground~to-air missile.
The four faclors are target type CAl, seeker type (8). tatget altitude (C). and target range (D),
Each. factor may bc conveniently ron at two levels, and the optical tracking system will allow termi~
nal miss dist.'Ulce to be measured to the nearest foot. Two differet.t gunners are used in the flighl test,
and since there may be differences between individuals, it was decided to conduct the 24 design in two
blocks with ABeD confounded. T!;.us, the defining contrast is
L=x~ +.:tz+x,+x4-
The experimental design and the resulting data are
Block 1 Block 2
(1).3 ,·7
ab ... 7 b .5
ae =6 c 6
be ~8 d.4
ad"" 10 abe = 6
bd.4 bed = 7
cd 8;lid acd =9
: abed I t 9 abd ~ 12
The anal:ysis of the dcsign by "rates' algorithm is shown in Table 13-23. A noJ;ID.al. probability plot of
the effects would reveal A (ta.'"get type), D (target nL1.ge), and AD to have large effects. A confuming
an.alysis of variance. using three~factor interactions as error, is shown in Table 13-2J..
It is possible to confound the 21:. design in four blocks of 2}:;·-1 observations each. To con~
struct the design. twO effects are chosen to confound with. blocks and their defullng con-
trasts obtained. A third effect, the generalized interaction of the tvlo initially chosen, is also
:.3-6 Confoun.Cing in the 2k Design 393
confounded with blocks. The generalized interaction of two effects is found by multiplying
their respective columns.
For example, consider the 24 design in four blocks. If AC and BD are confounded with
blocks, their generalized interaction is (AC)(BD) = ABCD. The design is constructed by
using the defining contrasts for AC and BD:
L} =X1 +X:J,
L,. ~ x, + "•.
394 Chapter 13 Design of Experi:nents with Several Factors
~
I
iabed I bed : cd
--
- '
This general procedure can be extended to confounding the 2k design in 21' blocks.
where p < k. Select p effects to be confounded, such that no effect chosen is a generalized
interaction of the others. The blocks can be constructed from the p defIning contrasts L:, L"
... ~ Lp associated with these effects. 10 addition, exactly 2! - P - 1 other effects are COD-
founded with blocks, these being the generalized in:eraction of the origina! p effects cho-
sen. Care should be taken so as not to eonfound effects of potential interest.
For more information on confounding refer to Montgomery (2001, Chapter 7). That
book contains guidelines for selecting factors to confound with blocks so that main effects
and low-order interactions are Dot confounded. In particular, the book contains a table of
suggested confounding schemes for designs with up to seven factors and a range of block
sizes, some as small as two roos,
Table 13-25 Plus and Minus Signs for the 21 Factorial Design
treatment combinations are shown in !he top half of Table 13-25. We will use bo!h the COD-
ventional notation (a, hI c, ... ) and the plus and minus notation for the treatment comblll.3.~
tions. The equivalence between the two notations is as follows:
Notation 1 Notation 2
a +--
b -+-
c --+
abc +++
Notice that the 23 -1 design is formed by selecting only those treatment combinations
!hat yield a plus an !he ABC effect, Thus ABC is called !he generator of this particular frac-
tion. Furthermore,. the identity element I is also plus for the four runs, so we call
[=ABC
the defining relation for the design.
The treatment combinations in the 2:!< -: designs yield three degrees of freedom associ-
ated with !he main effects, From Table 13-25, we obtain !he estimates of the rna;:'! effects as
1
A =z:[a-b-c+abc],
I
B=-[-a+ b -c+ abc],
2
I
C=-[-a-b+c+abc].
2
It is also easy to verify that the estimates of the two-factor interactions are
I
BC =z:[a-b -c+abe],
I
AC= 2[-a+b-c+abc],
AB= 1 -b-c+abc].
2
e
Thus, !he linear combination of observations in column A, say A , estimates A + BC
e e
Similarly, E estimate.B .,.AC, and c estimates C ~ AB. Twa or mare effects !hat have this
proper:y are called aliases. In our 2'-1 design, A and BC are aliases, B andAC are aliases,
and C and.4.B are aliases. iiliasing is !he direct result of fractional replication. In many prac-
tical situations, it will be possible to select the fraction so that !he main effectS ""d low-
order interactions of interest will be aliased with high~order interactions (which are
probably negligible).
The alias structure for this design is found by using the defining relation I = ABC, Mul-
tiplying any effect by the defining relation yields !he aliases for !hat effect. In OUr example,
the alias of A
A=A . ABC =A'BC = BC,
since A . I = A and A' = 1. The aliases of B and C are
B=B·.4.BC=AB'C=AC
396 Chapter 13 Design of Expe:irnents with Several Factors
and
C=C·ABC=ABC'=AB.
Now suppose that we had chosen the other one-half fraction. that is" the treatment com~
binations in Table 13-25 associated with minus on ABC. The defining relation for this
design is I =--ABC. The aliases are A =-BC. B =-AC, and C ~ --AB. Thus the estimates of
.4.B, and C with this fraction really estimateA- BC, B -AC, and C -AB. In practice, it usu-
ally does not matter which one·half fraction we select. The fraction with the plus sign in the
defining relation is usually called the principal fraction, and the other fraction is usually
called the altemate fraction.
Sometimes we use sequences of fractional factorial desig::lS to estimate effects. For
example, suppose we had run the principal fraction of the 23 -! design. From this design we
have the following effect estimates:
e,.=A+BC.
C,=B+AC.
ec=c+AB.
Suppose that we are willing to assume at this point that the t:\J,Io-factor interactions are neg~
ligible. If tile}' are, then tbe 23 - J design has produced estimates of the three main effects, A,
B, and C, However, :if after running the principal fraction we are uncertain about the inter~
actions. it is possible to estimate them by running the alternate fraction. The alternate frac~
tien produces the following effect estimates:
e;=.4-BC.
e~ B-AC,
e;~C-AC.
If we combine the estimates from the two fractions t we obtain the following:
Effect i
I
;j(A+BC+A-BC)=A ~A +BC-(A-BC)] =BC
1
1(B+AC+B-AC)cB }[B+AC-(B-AC)] =AC
~(C+AB+C-AB)=C
1
I=C ~[C+AB- (C -AB)) =AB
Thus by combining a sequence of two fractional factorial designs we can isolate both tile
main effects and the two~factor interactions. This property makes the fractional factorial
deSign highly tL>;eful in experimental problems a') we can run sequences of small. efficient
j
experiments, combine infonnation across several experiments, and take advantage of learn-
ing about the process we are experimenting with as we go along.
A 2;:-1 design may be constructed by \\-TIting down the treatment combinations for a
full factorial with k - I factOrs and then adding the kth factor by identifying irs plus and
minus levels with the plus and minus signs afthe highest-order interaction ±ABC,,· (K-
I). Therefore, a 2'-1 fractional factOrial is obtained by writing down the full 2' factorial and
then equating factor C to the ±AB interaction, Thust to obtain the principal fraction. we
would use C =+ AB as follows:
r
l
i Full 2'
13-7 Fractional Replication of the 21 Design 397
A B A B C=AB
+
+ +
+ +
+ + + +
To obtain the alternate fraction we would equate the last column to C = -AB.
To illustrate the use of ZI one-half fraction, consider the plasma etch experiment described in Exa"n-
pIe 13-9, Suppose that we decide to use a 24 - 1 design with 1::; ABeD to investigate the foU!' factors.
gap (A). p::essure (B), C<F6 flow rate (C), and power setting (D). This design would ~e consl'l.lc[ed by
writing down a 2.) in the factors A. B, and C and then setting D = ABC. T'r.e design ond the resulting
etch rates are shov,rn in Table 13-26.
In this design, the main effects a.."'C aliased with the three-factor interactions: note that the a:ias
of A is
A·!=A·ABCD,
;;;;;.A 2BCD,
= BCD,
and sirnila::ly
B=ACD,
C=.4BD,
D=ABC.
The two~factor interactior..s are aliased with each otheL For ex.nnple, the alias of AB is CD:
AB·!=AB·ABCD,
=A'B'CD,
CD.
Tb.e other aliases axe
Treatr:1ent Etch
A B C D=ABC Combinations Rilte
(1) 550
+ + ad 749
T T bd 1052
+ + ab 650
+ + cd 1075
+ + ac 642
+ + be 601
+ + + + abed n9
398 Chapter 13 Design of Experiments with Several Factors
AC=BD,
AD=BC
The estimates of the mal.!;. effects and their aliases are found using the four colUl.Il:ls of signs in
Table 13~26. For example, from column A we obtain
t, =A + BCD =t(-550 + 749-1052+ 650 -1075 +642 -601 + 729)
;=-127.00.
The other columns produce
t 8 =B+ACD:4.00,
tc=C+ABD=lLSO,
and
tD=D+ABC 290.50.
e
Clearly .e" a....d b are large, and if we believe that me tlL."'ee-factor interactions are negligible. then. the
main effects A (gap) and D (power setting) significantly affect etch rate.
The interactions are estimated by formi::!g the AE, AC, and AD colu::n.1.s and adding them to the
table. The signs in theAB column are +,-, -. +, +, -. -, +, aDd this column prod..:ces the estimate
t" =AB + CD =-j:(550-749-1052 + 650 + 1075 - 642- 601 + 729)
=-10.00.
From theAC and AD columns we find
e;,c=AC+BD ""- 25.50,
£Ao=AD+BC=- :97.50.
The tAD estimate is large; the most straightforward interptetacion of the results is that this is the AD
inreractio!].. Thus, the results obtained from the 2d ~ I design agree with the full factorial results in
Example 13.-9.
Normality Probability Plots and Residuals The normal probability plot is very useful in
assessing the significance of effects from a fractional factorial. 'This is particularly true
when there are many effects to be estimated. Residuals can be obtained from a fractional
factorial by the regression model method shown previously. These residuals should be plot-
ted against the predicted values~ against the levels of the factors. and on normal probability
paper. as we have discussed before, both to assess the validity of the underlying model
asSl.!IDptions and to gain additional insight into the experimental situation.
Projection of the 2k - 1 Design If one or core factors from a one-half fraction of a 210 can
be dropped, the design will project into a full factorial design. For example, Fig. 13-25 pres-
ents a t·- 1 design, Notice that this design will project into a full factorial in any t'No of the
three original factors. Thus. if we think that at most two of the three factors are important
the 2:} -1 design is an excellent design for identifying the significant factors. Sometimes we
call experiments to identify a relatively few significant factors from a larger number of fac~
tors screening experiments, This projection property is highly useful in factor screening. as
it allows negligible factors to be elim.i:mted, resulting in a stronger experiment in the active
factors that remain.
j
13-7 f<ractional Replication of the 2~ Design 399
I I
I /
/ /
I I
/ I
b l I
I I
I I I
abc /
I /
II I A
tf-- - a
I
, II
"I I I
I I I
I I
I I I I
I •
C
~ V
Figure 13-25 Projection of a 2 3 - 1 design iIlto three 2). designs.
In the 2'-1 design used in the plasma etch experiment in Example 13-11, we found that
two of the four factors (B and C) could be dropped. If we eliminate these two factors, t':te
remaining columns in Table 13·26 form a 2' design in the factors A and D, with two reFli·
cates. This design is shown in Fig. 13-26,
Design Resolution The concept of design resolution is a useful way to catalog fractional
factorial designs according to the alias patterns they produce. Designs of resolution III, IV-,
and V are particularly imporraut The cefioitions of these terms and an example of each
follow:
1. Resolution III Designs. These are designs in which no main effects are aliased with
any other main effect, but main effects are aliased with two~factor interactions, and
two·factor interactions may be aJiased with each other. The 2' -, design with I ~
ABC is of resolution m. We usually employ a subscript Roman numeral to indicate
design resolution; thus this one-half fraction is a t'm 1 design,
2~ Resolution IV Designs. These are designs in which no main effect:is aliased with
any other main effect or two-factor interaction. but two-factor interactions are
(1052.1075) [749,729)
+1~------------------'
1(550,601) (650,642)
-1~~~~--------~
-1 +1
A (Gap)
Figure 13-26 The 22 design obtained by dropping factors B and C from the plasma etch
experiment
400 Chapter 13 Design of Experiments with Severa!. Factors
1
aliased "ith each other, The 2'-1 design Mth I =ABCD used io Example 13-11 is
of resolution N (21; 1),
3. Resolution V Designs. These are designs in which no main effect Or two-factor
interaction is aliased with any other main effect or two-factor interaction, but two~
factor interactions are aliased with three~factor interactions. A 25 -! design with I ~
ABCDEis ofresolution V (2',,1).
Resolution ill and N designs are particularlY useful io factor screening experiments,
A resolution IV design provides very good information about main effects and will provide
some information about two-factor interactions.
To find the alias of any effect, simply multiply the effect by each word io the foregoiog
deflning relation. The complete alias strUct1lre is
A = BCE= CDF=ABDEF,
=
B =ACE DEF ABCDF, =
C =ABE = ADF =BCDEF,
D=ACF=BEF=ABCDE.
E=ABC=BDF=ACDEF.
F=ACD=BDE=ABCEF,
AB= CE= BCDF=ADEF,
AC=BE=DF=ABCDEF,
liD = CF = BCDE = ABEF,
r 13~ 7
AE=BC= CDEF=ABDF,
AF=CD=BCEF=ABDE,
Fractional Replic.ation of the 21,- Design 401
BD=EF=ACDE=ABCF,
BF=DE= ABCD = ACEF
ABF=CEF=BCD=ADE,
CDE=ABD=AEF= CBF.
);otice that this is a resolution IV design; main effects are aliased with three~fac~or and
higher interactions, and two~factor interactions are aliased with each other. TIlls design
would provide very good information on the main effects and give some idea about the
strength of the two-factor interactions. For example, if the AD interaction appears signi:fi~
can~ either AD andior CF are significant, If A andior D are significant main effectS, but C
and F are not, the experimenter may reasonably and tentatively attribute the significance to
the AD interaction. The construction of the design is shov.'1l in Table 13-27.
The same pr.nciples can be applied to obtain even smaller fractions. Suppose we wish
to investigate seven factors in 16 runs. This is a 27 - 3 design (a 118 fraction). This design is
constructed by writing daVID. a 24 design in the factors A, B, C, and D and then adding three
new columns. Reasonable choices for the three generators required ace I = AliCE. I =
BCDF, and I = ACDG, Therefore, the new columns are formed by setting E = ABC, F =
BCD, and G =ACD. The complete defining relation is found by multiplying the generators
together two at a time and then three at a time, resulting in
I =ABCE = BCDF = ACDG = ADEF = BDEG =ABFG = CEFG.
Notice that every main effect in this design will be aliased with three-factor and higher
interactions and that two--factor interactions will be aliased with each other. Thus this is a
resolution IV design,
For seven factors. we can reduce the number of runs even further. The 27 -.; design is
an eight-run experiment accommodating seven variables, This is a 1116 fraction and is
Tabl,13·:.7 Construction of the 26 - 2 Design with Generators I "" ABCE and I =ACDF
A B C D E=ABC F=ACD
(1)
+ ~
+ dej
.;. be
+ + + obf
~ ~
+ eif
+ + at:
+ + + bcf
+ + + abce
+ .;. dj
+ ade
+ .;.
+ + bdef
+ + + obd
+ + + ede
+ + + oedf
+ + bed
+ + + + obcdef
402 Chapter 13 Design of Experiments with Several }actors
obtained by first writing down a 2' design in the factors A, B, and C, and then forming the
four new columns, from 1= ABD, I = ACE, 1= BCF, and 1= ABCG. The design is shown
in Table 13-28.
The complete defining relation is found by IDuetiplying the generators together two,
three, and :finally four at a time, producing
I =ABD =ACE =BCF =ABCG = BCDE =ACDF = CDG =ABEF
=BEG=AFG=DEF=ADEG=CEFG=BDFG=ABCDEFG.
The alias of any main effect is found by multiplying that effect througb each term in the
defining relation. For example, the afuu; of A is
A=BD CE=ABCF=BCG=ABCDE=CDF=ACDG
=BEF =ABEG = FG =ADEF = DEG = ACEFG =ABDFG
= BCDEFG.
This design is of resolution rn~ since the main effect is aliased with two-factor interactions.
If we assume that all three-factor and higher interactions are negligible, the aliases of the
seven main effects are
e,=A+BD+ CE+ FG,
e,=B+AD+CF+FG,
fc=C+AE+BF+DG,
CD=D+AB+CG+EF,
C,=E+AC+BG+DF
C,=F+BC+AG+DE,
eo G+CD+BE+AF.
This 27:i!4 design is called a saturated fractional factorial, because all of the available
degrees of freedom are used to estimate main effects. It is possible to combine sequences
of these resolution ill fractional factorials to separate the main effects from the two-factor
interactions. The procedure is illustrated in Montgomery (2001, Chapter 8).
In constrUcting a fractional factOrial design it is important to select the best set. of
design generators. Montgomery (2001) presents a table of optimum design generators for
designs 'With up to 10 factors. The generators in this table will produce designs of max-
imum resolution for any specified combination of k and p. For more than 10 factors, a
+
+ + +
+ +
+ + +
+ + +
+ + +
+ + +
+ + + + + + +
13~8 Sample Computer Output 403
resolution ill design is recommended. These designs may be constructed by using the same
method illustrated earlier for the 211«4 design. For example, to investigate up to 15 factors
in 16 runs, v.'rite down a 24 design in the factorS A, B, C. andD, and then generate 11 new
columns by taking the productS of the original four columns two at a time, three at a time,
and four at a time. The resulting design is a 21~!- n fractional factoriaL These designs, along
witit other useful fractional factorials, are discussed by Montgomery (2001, Chapter 8).
The Minitab® results are in agreement with the results given in Table 13-6.
The ou:put from Minitab'" is slightly different from the results given in Example 13-7. Hests
on the main effects and interactions are provided in addition to tite analysis of variance on
the significance of main effects. two~factor interactions, and three-factor intera(.i.ions. The
404 Chapter 13 Design of Experiments with Several Factors
Al~OVA results indicate that at least one of the main effects is significant, whereas no two-
factor or three-factor interaction is significant.
13-9 SUM1VlARY
This chap:er has introduced the design and analysis of experiments with several factors j
concenrrating on factorial designs and fractional factorials. Fixed, randoIr., and mixed mod~
els were considered, The F~tests for main effects and interactions in these designs depend
on whether the factors are fixed or random.
The 2k. factorial designs were also introduced. These are very useful designs in which
all k factors appear at two levels. They have a greatly simplified method of statistical analy-
sis. In situations where the design cannot be run under homogeneous cQnditiQns~ the 2~
design can be confounded easily in 2 P blocks. This requires that certain interactions be con-
founded with blocks. The 2' design also lends itself to fractional replication. in which only
a parJcular subset of the 2" treatment combinations are run. In fractional replication~ each
effect is aliased with one or n:ore other effects. The general idea is to alias main effects and
low-order interactions with higher-order interactions. This chapter discussed methods for
construction of the 2~-P fractional fac:orial designs, that is, a If2P fraction of the 2* design.
These designs are particularly useful in industrial experimentation,
13-10 EXERCISES
13-1. An article in the Journal of Materials Process- 25, and 30 minutes-andrandornly chooses two types
ing Technology (2000. p. 113) presents results from an of paint from several that axe available, He conducts
experiment involving tool wear estimation in m.ilEng. an experiment and obtains t.'1e data shown here. Ana~
The objective is to minimize tool wear. 1\vo facto;s of lyze the data and draw conclusions, Estimate the vari-
~nterest in ~e study were cutting speed (mlmin) and ance components.
d:;pth of cut {mm). One response of interest is tool
flank wear (mm). Three levels of ea.cll factor were
selected and a factorial experiment with three repli- Drying Time (min)
cates is run. Analyze the data and draw conclusions. Paint 20 25 30
74 73 78
Depth of Cut
64 61 35
2 3 50 44 92
0.170 0.198 0.217 92 98 66
12 0.IS5 0.210 0.241 86 73 45
0.110 0.232 0.223 68 88 85
0.178 0.215 . 0.260
15 0.210 0.243 0.289
13~3. Suppose that in Exercise 13-2 paint types wc:e
0.250 0.292 0.320 fixed effects. Compute a 95% interval estimate of the
0.212 0.250 0.285 mean difference between the responses for paint type
18.75 0.238 0.282 0.325 1 and painttype 2.
0.267 0.321 0.354
13-4. The factors that influence the breaking strength
of cloth are being stu.died. Four machines and three
13~2..A:1 engineer suspectS that the surface finish of a ope{ators axe chosen at random and an experiment is
meta: part is lnI1uenced by the :ype of paint used and run using cloth from the same one~y.ard segment The
the dr)'ing time, He selects three d..")ing times-20, results are as follows:
]3·10 Exe,cises 405
13..12. Four factors are thought to possibly influence Experimental Statistics (No. 91, 1963) involves
the taste of a soft drink be¥"erage: type of sweetener tlame*testing fabrics after applying fIre-retardan,
~4), rario of syrup to water (B), carbonation level (C). treatments. There are four factors: type of fabric (A),
and temperature (D), Each factor can be run at two type of firc-reta."Iiant treatment (B), laundering condi~
levels. produci.Ilg a 24 design. At each run in the ton (C-the low level is no laundering. the hlgh level
design, samples of the beve:.-age are given to a test is a.Fter one laundering). and the method of conducting
panel consisting of 20 people. Each tester assigr.s a rhe flame test (D). All factors are ron at two levels. and
point score from 1 to 10 to the beverage. Total score is the response ..-a.riable is the inches of fabric burned on
the response variable, and the objective is to futd a a standard size test sample. The data are
formulation that m.aximizes tOW sCOre, Two replicates
of this design are run, and the results shown here. (1) 42 d ...,,40
Analyze the data and draw conclusions, a=;: 31 ad = 30
b =45 hd 50
Tteatment ~lica:e Treatment Replicate:
Combinations I II Combinations T II ab = 29 abd = 25
e = 39 cd =40
(1) 190 193 d 198 195
a 174 178 ad 172 176 DC =28 aed "'" 25
b 181 185 bd 187 183 be =46 bed = 50
ab 183 180 aba 185 186 abc = 32 abed = 23
c 177 178 cd 199 190
ac 181 180 aed 179 175 (a) Esti.matc the effects and prepare a nonnal proba-
be 188 182 bed 187 184 bility plot of the effects.
abc 173 170 abed 180 180 (b) Construct a nonnal probability plot of thc residu-
als and comment on the results"
13~13. Consider the experiment in Exercisc 13~12. (c) Construct an analysis of variance table assuming
Plot the residoals against the levels of factors A, B, C. that three~ and four-factor interactions are
and D. Also construct a normal probability plot of the negligible.
residuals. Comment on these plots.
13-17. Consider the data from the first replicate of
13-14. Fmd the standard error of the effects for £he Exercise 13-10, Suppose rhat these o!:;servations could
cxperit:.lcnt in Exercise 13~ 12. Using the standard not all !:;e ron under the same conditions. Set up a
errors as a guide. what factors appear significant? design to run these observations in two blocks of four
13-15. The data sho~'D here represent a single repli~ observations, each withABC confounded. Analyze Lie
cate of a 25 design that is used in an experiment to data.
study the compressive strength of concrete. The fac- 13~18. Consider the data from the firSt replicate of
tors are nUx (A). time {B). laboratory (C), temperature E."l:ercise 13~ 12. Construct a desig=. with two blocks of
(D), and drying time (E). Analyze the data, assuming eight observations each. with ABCD confounded.
that three-factor and higher interactions are negligible. Analyze the data.
Use a nom'.a1 probability plot to assess the effects.
13 ..19. Repeat Exercise 13-18 assuming that four
(1) = 700 d=,OOO de =1900 blocks are required. Confound ABD and ABC (and
e= 800
consequemJy CD) wi'" blocks.
a 900 ad=ll00 ae: 1200 ade=1500
13~20. Construct a 2$ design in four blocks. Select the
b= 3400 hd = 3000 be =3500 bde=4000
effects to be confounded so rhat we confound the
ab= 5500 abd=6100 abe =6200 abd< =6500 highest possible interactions with blacks.
e= 600 cd:::::: SOO ce = 600 cde =1500
13·21. An article in Industrial and Engineering
ae= 1000 acd=ll00 ace""" 1200 aede =2000 Chemistry ("Factorial Experiments in Pilot Plant
be= 3000 bed =3300 oC$=3006 bede=3400 Studies," 1951, p. 1300) repo::ts on an experiment to
abc = 5300 abcd=6000 abee = 5500 abed< =6300
investigate the effects of temperature (A), gas through-
put (B). and concentration (C) on the strength mprod-
uct solution in a recirculation unit TWo blocks were
13~16. An experiment described by ~L G. Natrella in used with ABC confounded, and the .experiment was
the National Bureau of Standards Handbook of replicated twice. The data are as follows:
r 13·10 Exercises 407
13-25~ An article in Cement and Concrete Research (a) 'What is the generator for dlls fraction?
(2001. p.1213) describes an experimenttQ investigate (b) Analyze the data. What factors influence mean
the effects of four metal oxides on several cement bulk density?
properties. The four factors are aU run at two levels (e) Analyze the residuals from this experiment and
and one response of i:r.terest is the mean bc.lk density comment on your findings.
(g!cm~). The four factors and their le'l.'els are
13-.26. Consider the 2,,-2 design in Table 13-27. Sup-
pose that aftet analyzing the original data, we find that
Low Level High Level factors C and E can be dropped. Vlhat type of 2t
Factor (-1) (+1) design is left in the remaining variables?
ll.: %Fe,G, 0 30 13-27. Consider the 2'-' design in Table 13·27. Sup.
15
pose that after the original data analysis. we find that
8: % 7.J10 0
factors D and F can be dropped, ""'hat type"'of 2k
c: %PbO 0 2.5
design is left in the rero.aining variables? Compare the
D: %0:,0, 0 2.5
results with Exercise 13-26. Can you explain why the
answers are different?
Typical results from this type of experiment are given 13~28. Suppose that in Exercise 13-12 it was possible
in the follO'Ning !able: to nm only a one-half fracdon of the Z' design. Con-
strUCt the design and perform the statistical analysis,
Run ll. B C D Density use the data from replicate 1
-I -1 -I I 2.001 13-29. Suppose that in Exercise 13-15 only a one·half
fraction of the 2~ design could be run, Construct the
2 I -1 -I -1 2.062
design and perform the analysis.
3 -1 -1 -1 2.019
4 1 -1 2.059 13-30. Consider the data in Exercise 13~15, Suppose
that only a one-quarter fraction of the 2;5 design could
5 -1 -1 -1 1.990
be run. Construct the design and analyze the data.
6 1 -1 1 2.076
13,,31~ Construct a 2t\t:~ fractional factorial design.
7 -1 1 1 2.038
Write do'Wll the :a.:iases, assuming that only main
8 1 1 1 -1 2,118
effects aLd two-factor interactions are of inteCCSL
J
r chap_te_r_l_4____________________
est Chapter 15 will present the case involving more than one regressor variable.
where the intercept Pc and the slope /31 are unknown constants. We assume that each obser-
vation~ y~ can be described by the model
409
410 Chapter 14 Simple Linear Reg:re!)sio~ and Correlation
and the sum of squares of the deviations of the observations from the true regression line is
n n
L =2>1 =l., (Yi - Po - Plx,t
(14-4)
(14-5)
j=l 1=1
n n
Pol., X, + P,l.,X;
(14-6)
Equations 14-6 are called the least~squares normal equations. The solution to the normal
equations is
(14-7)
( 14-8)
,y,
where y= (l/niE;. andX'= (lln)L;.lx" Therefore, equations 14·7 and 14-8 are the least-
squares estimators of the intercept and slope, respectively. The fitted simple linear regres-
sion model is
(14-9)
14-1 Simple Lbear Regression 411
(14-10)
and
n
S", = :2>,(x, (14-11)
i=l
We callSp; the corrected sum of squares of.x and Sr; the corrected sun: of cross products of
x and y. The extreme right-hand sides of equations 14-10 and 14-11 are the usual computa-
tional form.u1as, Using this new notation, the least~squares estimator of the slope is
- S
f3.. =-'L. (14-12)
S'"
A chemical engineer is investigating the effect of process operating terr.perature on product jield, The
study results irl the following data:
Tempernture. 'C (xl 100 110 120 130 140 150 160 170 180 190
YIeld, % (y) 45 51 54 61 66 70 74 78 85 89
These pairs of points are plotted irl Fig. 14--1. Such a display is called a scatter diagram. Examination
of this scatter diagramirldicaLes that there is a s'JOng relationship between yield anc temperature, and
the tentative assumption of the straight~li.:le model y = Pc + PIX - € appears to be reasonable, The fol-
lollti."l.g quantities IDay be computed:
S =
n
~ x:
,L,.
1"',
-1.l'~" yI
1O,L,'
,'",j)
=218 500- (:450)' =8250
' 10
and
.ty ,L..;"
10
S = "" ",y.--I "" x·1
lO:.L,;' 1"'-
1 (!O i[""y I10
j
(1450)(6731
,=101570-,,""""",""-' -3985 .
I' 10
/..1 \.;",1,; ;",,1 )
, S", 3985
/3,, =-=--=0.483
S," 8250
412 Chapter 14 Simple Linear Regression and Correlation
l
100
9C •
80
•
•
70
•
60 •
,,-"
ID 5D • •
;;: •
40
30
20
10
! 1 ! ! i !
0
100 110 120 130 140 15D 160 170 160 160
Temperature, x
Figure 14.1 Scatter diagram of yield versus temperature.
and
Po ~ y - fi ii =67.3 (0.483)145 =- 2.739.
The fitted simple linear regression model is
y= - 2.739 + 0.481<.
since '2.7= !(xl-X) = 0, r..;. ;7:'(X1 ;;;;; Sx,x. and by assurnptionE(e,) = O.Thus'PI is an unbi-
ased estimator of the true slope PI' Now consider the variance of Pt , Siuce we have
assumed that VeE) = 0-', it follows that V(y); 0-', and
(14-13)
The random variables {y,) are uncorrelared because the {€,} are uncorrelated, Therefore, the
variance of the stlmm equation 14-13 is just the sum of the variances, and the variance of
each term in the sum, say VIY;(XI is VZ(x; - i)2. Thus.
S'"
Using a similar approach~ we can show that
and (14-15)
ftc
:Note ~t is an ::nbiased estimator of Po. The covariance of Po and 13: is not zero; iu fact.
Cov({3e' f3,) ;-rr"iIS""
It is usually necessary to obtain an estimate of <fl. The difference bet\Veen the obser-
vation 1, and the corresponding predicted value y" say e} = y, - y" is called a residual, The
sum of the squares of the residuals, or the error sum of squares, would be
(14-16)
, )2 '
y,
A more couyeni",nt computing formula for SSe may be found by substiruting the fitted
model {30 - i3,x, into equation 14-16 and simplifying, The result is
n
SSE=2:Y7
1=1
and if we let :Lf""tyl-ny'2 =i(Yi _y)2 =: Syy. theu we may 'Write SSfi as
j",l
(14.17)
414 Chapter 14 Simple Linear Regression and Correlation
The expected value of the error sum of squares SSe is E(SSE) = (n - 2)(1'. Therefore,
(14-18)
is an unbiased estimator of oz.
Regression analysis is widely used and frequently misused. There are several common
abuses of regression that should be briefly mentioned. Care should be talren in selecting vari-
ables with which to construct regression models and in determining the form of the approxi~
mating function. It is quite possible to develop statistical relationships among variables that
are completely unrelated in a practical sense. For example, one might attempt to relate the
shear stren.,"1h of spot welds with the nn.'Uber of boxes of computer paper used by the data
processing department. A straight line may even appear to provide a good fit to the data, but
the relationship is an unreasonable one On which to rely. A strong observed association
benveen variables does not necessarily imply that a causal relationship exists beween those
variables. Designed experiments are the only way to detennine causal relationships.
Regression relationships are valid only for values of the independent variable \'vithin
the range of the original data. The linear relationship that we have tentatively assumed may
be valid over the original range of XI but it may be unlikely to remain so as we encounter x
values beyond that range, In other words, as we move beyond the range of values of x for
which data were collected, we become less certain about the validity of the assumed modeL
Regression models are not necessarily valid for extrapolation purposes,
Finally, one occasionally feels that the model y = .8x "". is appropriate. The onission of
the intercept from this model implies, of course, that y =0 when x = O. This is a very strong
assumption that often is unjustified. Even when t'.VO variables, such as the height and weight
of men, would seem to qualify for the use of this model, we would usually obtain a better fit
by including the intercept, because of the limited range of data on the independent variable,
(14-20)
14-2 Hypothesis Te}''ting in Simple Linear Regression 415
foDows the tdistribution with n - 2 degrees of freedom underHo' p, P,.,. We would reject
Ho: p, ~ P;. 0 if
Itol > t"","~" (14-21)
where to is computed from equation 14-20.
A similar procedure can be used to test hypotheses about the intercept To test
Ho: p, P,.o'
(14-22)
H,' Pn " Pn.o,
we wood use the statistic
(14-23)
H" /3,,, O.
This hypothesis relates to the significance of regression. Failing to reject Ho: PI :.:: 0 is
equivalent to concluding that there is no linear relationship between.x and y. This situation
is illusrrated in Fig. 14-2. Note that this may imply either tbatx is of little value in explain-
ing the variation in y and that the best estimator of y for any x is y = Y (Fig. 14-2.a) or that
the true relationship between" and y is not linear (Fig. 14-2b). Alternatively, if Iio: {3, = 0
is rejected, this implies that ):is of value in explaining the variability in y. This is illustrated
in Fig. 14-3. However, rejecting HO=fJl :::::: 0 could mean either that the straight-line model is
adequate (Fig. 14-3a) or that even though there is a linear effect ob, better results cnuld be
obtained with the addition of higher-order polynomial terms in x (Fig. 14-3b).
The test procedure for Iio:/3, 0 may be developed from two approaches. The frst
approach starts with the following partitioning of the total corrected sum of squares for y:
(14-25)
., . , .
, "
•
• •
•
.. , • • • • •
•
...
• • •
(a) (b)
x
(a) (b)
The two components of Sy! measure, respectively. the amount of variability in the Yi
accounted for by the regression line, and the residual variation left unexplained by the
regression line. We usually call SSE:=: ICY! - Yr)'l the error sum of squares and SSp;::::
I.;=,(Yi - Y)'l the regression sum of squares. Thus, equation 14-25 may be v.rritten
(14-26)
Comparing equation 14-f 6 with equation 14-17. we note that the regression sum of squares
SSE'S
(14-27)
S" has" - 1 degrees of freedom, and SS, and SSE have 1 and n - 2 degrees of freedom,
re·spectively.
We may show thatE[SS,,!(r. 2)] ~ a'- and E(SS;;J ~ a'- + f3;S_ and that SSE and SS,
are independent. Thus, if Ho: f3, = 0 is true, the statistic
Note thatt; in equation 14-30 is identical to Fain equation 14-28. It is true, in general, that
the square of a t random variable with! degrees of freedom is an F random variable, with
One and!degrees of freedom in the numerator and denominator, respectively. Thus, the test
using", is equivalent to the test based on Fo.
We will test the model developed in Example 14-1 for significance of regression. The fitted model is
y:::- 2,739 ..;..0.483.x, and Syy is computed as
:47225- (673)'
, :0
=1932.10.
1932.10 -1924.87
7.23.
The analysis of variance for testing Ho: f3! 0 is summarized in Table 14-2. Noting that Fo 2138.74
t":";
and
418 Chapter 14 Simple Linear Regression and Correlation
are both <listributed as t wilb n -2 degrees of freedom. Therefore, a 100(1- a)% confidence
interval on Ibe slope /3, is given by
(14-31)
,
/30-la12
I'
I I MS;; -+-J5./3o
",-2 n S
[ 1 ",' :
S./3o+ta'2'-2~MSE -+-J'
n S
, !,
: [1 j"l (14-32)
,xx xx
We will find a 95% confidence interval on Ll:le slope of the regression line using the data in Exa:rnplc
14-1. Rec.all that ~, ~ 0.483, S~= 8250, and MS,= 0.90 (see Table 14-2). Then, from equation 14-31
we fine.
or
TIris simplifies to
0.45% Il,'; 0,507.
A confidence interval may be constructed for the mean response at a specified x. say
x". This is a confidence interval about E(YIxJ and is often called a confidence interval about
Ibe regression line. Since E(ylx,) ~ /3, + /31x", we may obtain a point estimate of E(YIxJ from
Ibe fitted model as
E~) ""y,=i3,+P,xo'
Now Yo is an unbiased point estimator of E(YIxJ, since Aand p, are unbiased estimators of
/30and /3" The variance ofYo is
V(i'o) =O"'~.!.+ ("" - x)'],
Ln S~
A
and Yo is normally distributed, as and PI are normally distributed. Therefore, a 100(1 -
a)% confidence interval abom the true regression line at x '\'0 may be computed from
•
Yo -ta{l,n-2 _MSEl-+
/1 (1 (xo-xl"J
~ n Sa (14-33)
,----,-----,-,.
'"1
"E (YixO)';Yo+'aJ2~-2> 1MSEI' -+
(x,-x)'1 .
\I n Sa
I' ,
r 14-3 btcrval Estimation in Simple Linear Regression 419
Tne width of the confidence interval for E(y1xJ is a function of xc' The interval width is a
minimum for Xu =::X and 'Widens as ~o - Xl increases, This widening is one reason why using
regression to extrapolate is ill-advised,
We will construct a 95% confidence interval about the regression line for the data in Example 14~ 1.
The fi!ted model is Yi; = - 2.739 ..;.. 0.483x,., and the 95% confidence interval on E(yix,,) is found from
equation 14-33 as
_ f r
1 (xo -14,\ :
Yc ::2.30\!0.90 Iii--' &250 '
-21]I'
1 \ ,
The fitted values Yo and the coa:espomii."lg 95% cor.fldence limits for the poi::ts i= 1, 2. ""
10, a:-e displayed in Table 14-3. To illustrate the use of this table, we may find the confidence
interval on the true mean process yield at Xc = 140"C (say) as
Table 14-3 Confidence Interval about the Regression Line, Example 14-4
100 110 120 130 140 150 160 :70 1&0 190
45.56 50.39 55.22 60.05 64.&8 69.72 7455 79.38 114.2[ 89.04
95% confider.ce
limits ±L30 ±LlO ±Q.93 ±Q.79 ±Q,71 ±Q.7t ±Q,79 ±Q.93 ±LlO ±1.30
100
90 /~
~ ... /
80
~~
~-; ...
70
"- 60
~
:;: 50
40
30
20
10
0' I I !-~
100 110 120 130 140 150 160 170 180 190
Temperature. x
Figure 144 A 95% confidence interval about the :regression line for E;r;ampie 14-4.
420 Chapter 14 Si.rcple Linear Regres~10n and Correlation
because Yo is independent ofy,> Thus, the 100(1 - a)% prediction interval on a future obser,
vations ar Xc is
(14,35)
(14-36)
14-5 Measuring the Adequacy of the Regression Model 421
To illustra::e the cons!ruction of a prediction interval, suppose we use the data in Exam-
ple 14-1 and find a 95% prediction .inter,fal on the next observation .on the process yield at
Xo == 160°C, Gsing equation 14~351 we find that the prediction interval is
which simplifies to
L
Figure14.5 patterns for residual plots, [al Satisfactory, (b) funnel, (e) double bow, (d) nonlinear,
[Adapted from Montgomery, Peck, and Vming (2001),J
Xi that look like Fig. 14-5c also indicate inequality of variance, Residual plots that look like
Fig. 14.5d indicate model inadequacy; that is, higher-order terms should be added to the
model.
l{xam"i,,14~5'
The residuals for the regression model in Example 14-1 are computed as follows:
e, =45.00 - 45.56=- 0,56, e, =70,00 - 69.72 = 0.28,
e, ~ 51.00 - 50.39 = 0,61, .,=74,00 -74.55 =-0.55,
e~ 54.00 - 55.22 =-1.22, <, = 78,00 -79.38 1.38,
e, = 61.00-60,05 = 0,95, e, = 85,00 - 84.21 = 0.79,
e, = 66,00 - 64.88 = Ll2, e" = 89,00 -89.04=- 0,04.
These residuals Me plotted on normal probability paper in Fig. 14--6. Since the residuals fall approx~
imately along a straight line in Fig. 14-6, we conclude that there is no severe departure from nonnal-
ity. The residuals are also plotted agai.'1st y! in Fig. 14-7a and against Xi in Fig, lA-7b. These plots
indicate no serious model inadequacies.
2 98
5 • 95
10 90
•
20 80
• 70
30
.~ 40 60
j3
m 50 50
.0
e
c.. 60 40
70 30
BO 20
•
90 -: 10
95 ~ • -is
9B~··-L· I I I 2
-1.5 1.0 -0.5 0.0 0.5 1.0 1.5 2.0
Residuals
Figure 14-6 Nonnal probability plot of residuals.
e,t
2.00~
1,00~ • •
0.00 c •
-1.00 ~
-2.00 l ------~-----------------~- ·
40 50 60 70 80
$,
2.00
----------~---------------
:.00 • •
0.00 . •
• •
-1.00
---~----------------------
.
-2.00
.L ! ! ! !! )
100 110 120 130 ~40 :50 160 170 180 190 x,
Figure 14-7 Residual plots for Example 14-5. (0) Plot agalnsty" (b) plot against.tr
'The danger of using a regression model that is a poor approximation of the true func~
tional relationship is illustrated in Fig. 14·8. Obviously, a polynomial of degree two or
greater should have been used in this situation.
424 C1:.apter 14 Simple Linear Regression and Correlation
l,
• •
•
• • •
• •
• •
• •
• •
x
Figure 14-8 A regression model displaying llck of fit.
We present a test for the "goodness of fit" of the regression model. Specifically, the
hypotheses we wish to test are
Ho: The model adequately fits the data
H,: The model does not fit the data
The test involves partitioning the error or residual sum of squares into the NO components
SSE = SS" + SSWF
where SSi'E is the sum of squares attributable to "pureH error and SSWF is the sum of squares
attributable to the lack of fit of the model. To compute SS" we must have repeated obser-
vations OIl y for at least one level of x, Suppose thar We have n total observations such that
repeated observations at XI'
repeated observations at~,
(14-37)
The total sum of squares for pure error would be obtained by summing equation 14-37 over
all levels of x as
~ ,
SSpE = I,L(Y,,, _y,)2 (14-38)
There are n~ = :L:! (n; - 1) ;:::: n m degrees of freedom associated with the pure-error
sum of squares. The sum of squares for lack of fit is simply
(14-39)
wid, n - 2 - n, = m - 2 degrees of freedom. The test statistic for lack of fit would then be
~ _ SSwd(m-2)
TQ - (14-40)
SSPE!(n-m)
J
14-5 Measuring the Adequacy of the Regression ~lodel 425
This test procedure may be easily introduced into the analysis of va....-fance conducted
for the significance of regression. If the null hypothesis of model adequacy is rejected, then
the model must be abandoned and attempt!) made to find a more appropriate model. If Ho is
Dot rejected, then there is DO apparent reason to doubt the adequacy of the model, and MSPE,
and MSWF are often combined to estimate rfl.
i;:Xi#~re 1ft;.
SuppQse we have the follov,,1ng data:
We may compute St)';;; 10.97, S,ry:= 13.20, S;u: = 52.53~ y:;; 2.847, ru;d x;;; 4.382. The regression model
isy,;:::: 1.703 + Q,26Qx, a.nd the,regressionsum of squares is SSr:= f3IS1:)' = (0.260)(13.62) = 3.541, The
pure-error SUr:l of squares is computed as follows:
The analysis of variance is summarized in Table 14-4. Since Foz, a, 1 L 70. we cannot reject the
hypothesis that the tenta;tve model adequatelY describes the data. \Ve ",ill poor la¢k~of-fit and pure~
error mean squares to fonn the denominator mean square in the test for significance of regression.
Also, since =4.54, we conclude that{1!"# O.
(I4-4J)
is called the coefficient of determination and is often used to judge the adequacy of a
regression model. (VIe will see subsequently that in the case where x and y are jointly dis-
tributed random variables, R2 is the square of the correlation coefficient between x and y.)
Clearly 0 ,; R' ~ I, We often refer loosely to R' as the amount of variability in the data
explained or accounted for by the regression model. For the data in Example 14·1, we have
R' = SSFIS" = 1924,87/1932,10 = 0,9963; that is, 99,63% of the variability in the data is
accounted for by the model
The statistic Rl should be used with caution. since it is al'W"&ys possible to makeR2 unity
simply by adding enougb tenns to the model For example, we can obtain a "perfect" fit to
n data points with a POlplOmiaI of degree n - 1. Also. Rl will always increase if we add a
variable to the model, but this does not necessarily mean the new model is superior to the
old one. Unless the error Sum of squares in the new model is reduced by an amount equal
to the original error mean square, the new model will have a larger error mean square than
the old one, because of the loss of one degree of freedom. Thus the new model will actu~
ally be worse than the old one,
There are several misconceptions about R2, In general, R? does not measure the mag-
nitude of the slope of me regression line, A large value of R' does not imply asleep slope,
Furthermore. Rl does not measure the appropriateness of the model, since it can be artifi-
cially inflated by adding higber-order polynomial terms, Even if y and x are related in a non·
linear fashion, J?l will often be large. For example, R2 for the regression equation in
Fig. 14-3b will be relatively large, even though the linear approximation is poor. Finally,
even though R2 is large, this does not necessarily imply that the regression model will pro~
vide accurate predictions of future observations.
J
14.7 Correlation 427
This transformation requires that the transformed error ten:ns 1n E be nor:maEy and inOO·
pendently distributed with mean 0 and variance a?.
Another intrinsically linear funetion is
(11
y;/30+/3'l-I+€'
x/
14-7 CORRELATION
Our development of regression analysis thus far has assumed that x is a mathematical vari~
able. measured with negligible error, and that y is a random variable. Many applications of
regression analysis invulve situations where both;; and y are random variables. In these sit~
uations, it is usually assumed that the obser"llations (Yi' XI)' i = 1. 2, "', n, are jointly dis-
tributed random variables obtained from the distribution/(y, x). For example, suppose we
wish to develop a regression model relating the shear strength of spot welds to the weld
diameter. In this example, weld diameter cannot be controlled. We would randomly select
n spot welds and observe a diameter (Xi) and a shear strength (y.) for each. Therefore, (y, x)
are jointly distributed random variables.
We usually assume that the joint distribution ofy! and Xi is the bivariate normal distri-
bution. That is,
(14-42)
0: a;
where J.i.l and are the mean and va..--iance of Y. J.4 and are the mean and variance of x,
and p is the correlation coefficient between y and x, Recall from Chapter 4 that the corre-
lation coefficient is defined as
(14-43)
where
(14-44a)
(14-44b)
and
--' ' ,
"]2 = 0;(1 - {r), (l4-44c)
That is~ the conditional distribution of y given x is normal with mean
E(ylx) = Po + P,X (14-45)
and variance O"~i' Note that the mean of the conditional distribution of y given x is a
straight~line regression modeL Furthermore, there is a relationship between the correlation
coefficient p and the slope PI' From equation 14-44b we see that if p=(), then P, 0, which
implies that there is no :regression of y On x. That is, knowledge of x does not assist us in pre-
dicting y.
The method of maximum likelihood may be used to estimate the parameters 130 and {JI-
It may be show'n that the maximum likelihood estimators of these parameters are
{Jo=y-/J"i (14400)
and
i-I
(l4-46b)
We note that the estimators of the intercept and slope in equaden 1446 are identical to
those given by the metl::od of least squares in the case where x was assumed to be a mathe~
matical va..~able, That is, the regression model with y and x jointly normally distributed is
equivalent to the model with x considered as a mathematical variable. This follows because
the random variables y given x are independently and normally distributed with mean Po +
PIX and constant variance 0':,. These results will also hold for any joint distribution of y and
x such that tl::e conditional distribution of y given x is noxmaL
It is possible to draw inferences about the correlation coefficient p in this model. The
es.timator of p is the sample correlation coefficient
(14-47)
J
14-7 Correlation 429
Note that
'S ,,1/,2
s j'
III, ~.,YY
\ .a
r, (14-48)
so the slope PI is JUSt the sample correlation coefficient r multiplied by a scale factor that
is the square root of the "spread" of the y values divided by the "spread" of the x values,
Thus il,
and r are closely related, although they provide somewhat different information,
The sample correlation coefficient r mea!."1U'es the linear association between y and x. while
il,measures the predicted change in the mean of y for a mtit change in x. In the case of a
mathematical variable x, r has no meaning because the magnitude of r depends on the
choice of spacing for x. We may also write) from equation 14-48,
which we recognize from equation 1441 as the coefficient of determination. That is, the
coefficient of determination RZ is just the square of the sample correlation coefficient
between y and x.
It is often useful to test the hypothesis
H,:p=O,
(14-49)
H,:p,eO,
The appropriate test statistic for this hypothesis is
(14-50)
which follows the t distribution with n - 2 degrees of freedom if Ho: P = 0 is true, There-
fore, we would reject the null hypothesis if It,l > 1"',"_1' This test is equivalentto the test of
the hypothesis He: Il, = 0 given in Section 14-2, This equivalence follows directly from
equation 1448,
The test procedure for the hypothesis
H,: P=Pa'
(14-51)
H,:p",po,
where P,,e 0, is somewhat more complicated, For moderately large samples (say n;' 25) the
statistic
1 l~r
Z = arctanh r = -l:n-'- (14-52)
2 l-r
is approximately normally distributed with mean
1 I~p
Ilz
,
=arctanh P =-l:n--
2 I-p
430 Chapter 14 Simple Linear Regression and Correlation
and variance
<r. =(n-3t:.
Therefore, to test the hypothesis Ho: P = Po' we may compute the statistic
Zo =(arctanh T - arctanh pOl(n - 3)ln (14-53)
and reject H,: P = Po if IZo! > Z.~.
It is also possible to construct a 100(1- a)% confidence interval for P using the trans-
formation in equation 14-52. The 100(1 - a)% eonfidence intelval is
~~~@1Efi
Montgomery, Peck, and VUJing (2001) descn.be an application of regression analysis in which an
engineer at a soft-drlnk bottler is investigating the product distribution and route service operations
for vending machines. She suspects that the time required to load and service a machine is related to
the number of cases of produ..-t delivered. A random sample of 25 (etail outlets having vending
machines is selected. and tOe in-outlet delivery time (in minutes) and volume of product delivered (in
cases) is observed for each outlet. The data-are shown in Table ~4-5. We assume that delivery time and
volume of product delivered ate jointly normally distributed.
Using the data in Table 14-5. we may calculate
S" = 6105.9447, s~ = 698.5600, S.,= 2027.7132.
The regression model is
y= 5.1145+ 2.9027<.
The sample correlation coefficient between x and y is computed from equation 14-47 as
S;ry 2027.7132
T 0.9818.
l
S:n;SYJ t [(698.5600)(6105.9447)1'"
9.95 2 14 11.66 2
2 24.45 8 15 21.65 4
3 31.75 11 16 17,89 4
4 35.00 10 17 69.00 20
5 25.02 8 18 10.30 1
6 16.86 4 19 34.93 !O
7 14.38 2 20 46.59 15
8 9.60 2 21 44.88 15
9 24.35 9 22 54.12 16
10 27.50 8 23 56.63 17
11 17.08 4 24 22.13 6
12 37.00 11 25 21.15 5
13 41.95 12
i
J
r
14-9 Summary 431
Note that R'l ~ (0.9818)2 0,9640, or that approximately 96.400/0 of the variability in delivery time is
explained by the linear relationship with delivery volume, To test the hypothesis
I Iio: p~ 0,
Ii,: p" 0,
we can compute the test statistic of equation 14~50 as follows:
~
_ '~n-2
=
0,9818,,23 2480
to - ,......-- • ,
.,jl_,2 .,jl-0,9640
Since to.Ol.'i,1:! = 2.069, we reject Ho and conclude fuat the correlation coefficient p;t O. Fi:::ally, we may
construct an approximate 95% confidence .interval on p from equation 14-54, Since arctanh r ;:; ;
arctmh 0,9818 = 2.3452, equation 14-54 becot:1<S
\ ~22 \ -fii)
which reduces to
0.9585 s P s 0,9921.
The regression equation is pro'tided along with the results from the Hests on the individual
coefficients. The P-values indicate that the intercept does not appear to be significant
(P-value = 0.114) while the regressor variable, temperarure, is statistically significant
(P-va!ue = 0), The analysis of variance is also testing the hypothesis that Ii,: /31 = 0 and can
be rejected (P-value = 0)_ Note also that r = 46.17 for temperature, and f = (46,17)'
2131.67 = F. Aside from rounding, the computer results are in agreement with those found
earlier in the chapter,
14-9 SUMMARY
This chapter has introduced the simple linear regression model and shown how least-
squares estimates of the model parameters may be obtained. Hypothesis ..testing procedures
432 Chapter 14 Simple Linear Regression and Correlation
and confidence interval estimates of the model parameters have also been developed. Tests
of hypotheses and confidence intervals requrre the assumption that the observations y are
nonnally and independently distnouted random ...'3Iiables. Procedures for testing model
adequacy, including a lack-of-fit test and residual analysis, were presented. The correlation
model was also introduced to deal with the case where x and yare jointly normally distrib-
uted. The equivalence of the regression model parameter estimation problem for the case
where x and y are jointly nonnal to the case where x is a mathematical variable was also dis-
cussed, Procedures for obtaining point and interval estimates of the correlation coefficient
and for testing hypotheses about the correlation coefficient were developed.
14-10 EXERCISES
14-1. Montgomery, Peck. and Vwing (2001) present (a) Fit a linear regression model relating games won
data concerning the performance of the 28 National to yards gained by an opponent
Football League teams in 1976. It is suspected that (b) Test for significance of regression.
the number of games won (y) is related to the number (e) Find a 95% confidence interval for me slope.
of yards gained:.'U.Sh.ing by an opponer:.t (x). The data (d) '\Vh.a: percentage of total variability is explained
are shown below, by the lIlodel?
(e) Find::he residuals and prepare appropriate resid-
Yards Rushing by ual plots.
Teams Games Won (y) Opponent (x) 14-2. Suppose We would like to use the model devel~
oped in Exercise 14-1 to prediet ::he number of games
Washington 10 2205
a team will ..........n if it can limit the opponents to 1800
Minnesota 11 2096 yards rushing. Find a poin~ estimate of the number of
New Englaod 11 1847 games won if the opponents gain only 1800 yards
Oakland 13 1903 rushing. Find a 95% prediction interval on the number
Pittsbcr:gh 10 4151 of games WOn.
Baltimore 11 1848 14-3.. Motor Trend magazine frequently presents per-
Los Angeles 10 1564 formance data for automobiles. The table below pres-
Dallas 1! 1321 ents data from the 1975 volume of Motor Trend
Atlanta 4 2577 concern.iog the gasoline milage perfonnance and the
Buffalo 2 2476 engine displacement for 15 automobiles.
Chicago 7 1984
10 1917 :'files I Displacement
Cinc!"1.Ilati
Automobile Gallon (y) (Cubic (x)
Cleveland 9 1761
Denver 9 1709 Apollo 18.90 350
Detroit 6 1901 Omega 17.00 350
Green Bay 5 2288 t\ova 20.00 250
Houston 5 2072 Monarch 18.25 351
Kansas Cit)' 5 2861 Duster 20.01 225
Miami 6 2411 Jensen ConY. 11.20 "40
~ew Orleans 4 2289 Skybawk 22.12 231
New York Gia:ats 3 2203 Monza 21.47 262
'S'ew York Jets 3 2592 Corolla SR-5 30.40 96.9
Philadelphia 4 2053 Carnaro 16.50 350
St Louis 10 1979 Eldorado 14.39 500
San Diego 6 2048 Trans Am 16.59 400
San. Francisco 8 1786 Charger SE 19.73 318
Seatt::c 2 2876 Cougar 13.90 351
Tampa Bay 0 2560 Co,rvette 16.50 350
J
,4-10 Exexoises 433
(a) Fit a regression model relating mileage perform- (a) Fit aregression model relating sales price to taxes
ance to engine displacemenL paid.
(b) Test far significance of regression. (b) Test for significance of regression,
(c) 'What percentage of total ,,"arizbility in mileage is (c) W"hat percentage of the variability in selling price
explained by the model? is explained by the taxes paid?
Cd) Find the residuals for this model. Construe: 11 nor-
(d) Fi:r:.d a 90% confidence inten'a! On the mean
mal probability plot for the residuals. Plot the
mileage if the engine displacement is 275 cubic
residuals versus y and versus x. Does the model
inches,
seem satisfactory?
144. Suppose that we wish to predict the gasoli:r:.c 14-7. The stre:r:gth of paper used ill the manufacture of
mileage from l! car v.ith a 275 cubic inch displacement cardboard boxes (y) is related to the percentage of
engine, Find a point estimate. using the model devel- hardwood concentration in the original. pulp (x). UnCer
oped in Exercise 14-3. ar.d an appropriate 90% interval controlled cor.ditions, a pilot plant manufaetu.'>'CS 16
estimate. Compare this interval to the Onc obra:ned in samples. each from a different batch of pu~p, and
Exercise l4-3d, 'Which one is wider. and why? measures the tensi::e strength. The data are shown here.
14~5. Fmd the rcsidl;als from the moder in Exercise
14.-3. Prepare appropriate residua: plors anC commer.t : :~~4 (:7;411:~/ 11~~211~~911:60911~8 1~:9
on model adequacy.
y ! 11.3 123.0 12':),1 145.2 1343 ,,<.:.4.5 J43.'1 146.9
14~6. An article in TechnomeIrics hy S. C. Narula and
J. E Wellington ("Prediction, Linear Reg:ession, and
a :Mi.nin::n.ntJ. SUDl of Relative Errors:' VoL 19, 1977)
x • zlZSl28 I 2$ • 3.0 • 3.0 • 3.2 I B
(a) Fit a simple linear regression model to the data,
presents data On the selling price and annual taxes for
27 houses. The data are shown below. (b) Test for lack of fit and significa.'1ce of regression.
(c) Construct a 90% confidence interval on the slope
l
434 Chapter 14 Simple Linear Regression and Correlation
shown in the following table. Assume that weight and Is the es-:ir.1ator of t.'1e slope in the simple linear
blood pressure are jointly normally distributed. regression model unbiased?
19 215 150
, , ,
20 195 163 /loLwixl-P:Lwixr= Lw/x;'Y;.
t=l ;"'! i",:
21 180 156
22 143 124
23 :240 170 14 20. Consider the data shOWll below, Suppose that
M
.J
Chapter 15
Multiple Regression
Many regression problems involve more than one regressor variable. Such models are
called multiple regression models. Multiple regression is one of the most widely used sta-
tistical techniques. This chapter presents the basic techniques of pardII1eter estimation, con-
fidence interval estimation, and model adequacy checking for multiple regression. We also
introduce some of the special problems often encountered in the practical use of multiple
regression including model building and variable selection. autocorrelation in the errors t
j
437
438 Chapter 15 Multiple Regression
Models that are more complex in appearance than equation 15-2 may often still be ana~
lyzed by multiple linear regression techniques. For example, consider the cubic polynomial .I
model in one independent variable, i
y Po + P,x+ pzX' + f3i' + e. (15-3)
If we let x, =x, A1 ::;;r, andxJ =;:3, then equation 15-3 can be written
y = Po + PIXI + {3.,x,+ {3.,x) - e, (15-4)
which is a multiple linear regression model wil.h three regressor variables. Models that
include interaction effects may also be analyzed by multiple linear regressiOTIlllethods, For
example, suppose that the model is
y= Po + PIX, + f3,x, + Pi',xlx" + s. (15-5)
If we let x) = XIx" and p, Pll' then equation 15-5 can be written
y = Po + P,x l + f3,x, + {3.,x, + e, ( 15-6)
which is a linear regression model In general, any regression model that is linear in the
paramJJrers (111e {3's) is. linear regression model, regardless of the shape of the surface 111.t
it generates.
(15-8)
Xu
J
15-2 Estimation of the Parameters 439
The function L is to be minimized with respect to A" /3" ... , /31- The least-squares estiroa-
tors of /30' /3" _", /31 must satisfy
(\5-9.)
and
• • ., •
+;3k 2. xilxi/;; ;::: LXilYi'
- • 2
ffi· L,x.
v ,. + /31 L, Xi! + ffi2 L, X il X
', +.,
r""l i""l 1::.1 /."'i i:1 (15-10)
Note that there are p;::: k + 1 normal equations, one for each of the unknown regression coef~
ficients, The solution to the no.rrnaI equations will be the least-squares estimators of the
regression coefficients /10, r /Jll ... , /J
1ti5 simpler to solve the normal equations if they are expresse<.iin matrix notation, We
now give a matrix development of the normal equations that parallels the development of
equation 15-10. The model in terms of the observations, equation 15-7. may be written in
matrix notation,
y=XIHe,
where
:1 x 11 x" ., .
y=
.. '
-".2 x""
aLI
- =-2X'y+2X'X~=O, '
a~~
which simpliftes to
X'X~ X'y. (15-12)
Equations 15-12 are the least-squares normal equations. They are identical to equations
15-10. To solve the normal equations, multiply both sides of equation 15-12 by the inverse
of X'X. Thus, the least-squares estimator of ~ is
p= (X'Xt'X'y. (15-13)
It is easy to see that the matrix form of the normal equations is identical to the scalar
form. Writing out equation 15-12 in detail we obtain
n
n
,
l>i!
ie::
, n
n
LX'2
t=l
,
n
LXi.<
,:;;;;1
cpo l ;
n
LV"
n
tti.l
, , n , n
If the indicated matrix multiplication is performed, the scalar form of the normal equations
(that is, equation 15-10) will result. In this fonn it is easy to see that X'X is a (p X p) sym-
metric matrix and X'y is a (p x 1) column vector. Note the special structure of the X'X
matrix. The diagonal elements of X'X are the sums of squares of the elements in the
columns QfX, and the off-diagonal elements are the sums of cross products of the elements
in the columns ofX. Furthermore, note that the elements ofXfy are the sums of cross prod-
ucts afthe columns of X and the observations {yJ.
The fitted regression model is
Y=Xp. (15-14)
In scalar notation, the fitted model is
k
y, Po + LP j );.,
j:::1
The difference between the observationy, and the fitted valuey; is a residual, say e;:::: Yt -Yi'
The (n xl) vector of residuals is denoted
o=y-y. (IS-IS)
15~2 Estimation of the Parameters 441
iu.l article in the Journal of Agricultural Engineering and Research (2001, p. 275) describes the use
of a regression model to relate the damage susceptibility of peaches to the height at which they are
dropped (drop height. measured in rom) <C!d the density of the peach (measured in glcmJ ). One gocl
of the ana]ysis is to provide a predictive model for peach damage to serve as a guideline for harvest-
ing and postharvesting operations. Data typical of this type of experiment is given in 1'able 15-2.
We will fit the multiple linear reg:::-e.....sion model
y= i>o+ pjx 1 + f3J.~ +"
to these data. The X matr:..x and y vector for this model are
303.7 O·90i 3.62
I:!1 366.7
336.8
304.5
U14
1.01
0.95
7.27
2.66
1.53
;1 346.8 0.98 4.91
il 600.0 1.04 10.36
:1 369.0 0.96 5.26
1 418.0 1.00 6.09
1 269.0 1.01 6.57
1 323.0 0.94 4.24
x= 562.2 1.01 '
y~
s.o4
284.2 0.97 3,46
558.6 1.03 8.50
.1 415.0 1.01 9.34
I 349,5 1.04 5.55
1 462.8 1.02 8.11
1 333.1 1.05 7.32
502.1 1.10 12.58.
311.4 0.91 0.15
351.4 0.96J 5,23 J
The X'X matrix is
,,:, l'
303.7
[ 20
7767.8
3201646
19.93 J
= 7767.8 7791.878 ,
19.93 7791.878 19.9077
~ = (X'Xt'X'y,
or
[ ~lj=
~Ol [ 20 7767.8 19.93 T'C 120.79]
7767.8 3201646 7791.878J' l'S1l29.17
~, 19.93 7791.878 19.9077 122.70
-33,831]
= 0.01314 .
[
34.890
J
15-2 Estimation of the Parameters 443
Table 15-3 Observations, Fitted Values, and Resid'J:a1s for Example 15-1
Observation Number y,
1 3.62 1.56 2.06
2 7:17 7.27 0.00
3 2.66 5.83 -3.11
4 1.53 3.31 -l.7S
5 4.91 4.92 -0.01
6 10.36 10.34 0.02
7 5.26 451 0.75
8 6.09 6.55 -0.46
9 651 4.94 1.63
10 4.24 3.21 1.03
11 8.04 8.79 -0.75
12 3.46 3.75 -0.29
13 8.50 9.44 -0.94
14 9.34 6.86 2.48
15 555 7.05 -1.50
16 8.11 7.84 0.27
17 1.32 7.18 0.14
18 12.58 11.14 1.44
19 0.15 2.01 -1.86
20 5.23 4.28 0.95
•
:2A
/.1
: :;: e'e.
l
444 Chapter 15 Multiple Regression
MS,= (15-17)
n-p
It can be shown that the expected value of MS£, is 0-2; thus an unbiased estimator of (f'l is
given by
&'=MS£, (15-18)
t"".,:pi?f5~
We \Vill estim..a~ t.'le error va:ciance <:f for the multiple regression problem in Exan:.p1e 15-1, Using the
dara in Table 15~2, we find
20
y'y= I,0 =904.60
1",1
and
120.79 1
~'X"Y=[-33.831 0.01314 34.890J 51129.17J
[
122,70
= 866.39.
Iberefore, the e:ctor sum of squares is
SSe = y'y - ~'X'Y
904.60 - 866.39
=38.2[.
The estimate of IT is
'2 _ SSe _ 38.21 _ 2 04-
" .. - - - - - - ._ i_
n- p 20-3
Ii
bination of the observations, it follows that is nonnally distributed with mean vector II and
covariance matrix u'(X'Xt'. Then each of the qUlllltities
P P
j - j
j O,l•...• k, (15-19)
1~2
'lCf ejj '
is distributed as twith n - p degrees oifreedom, where CJJ is tbei/th element oitbe (X'Xt!
matrix and if' is the estimate of the error variance, obtained from equation 15-13, Therefore,
a 100(1 a)% confidence interval for the regression coefficient Py j = 0, 1, ... , k, is
(15-20)
We will construct a 95% confidence interval on the parameter f31 in Example 15-1. Note that:he point
est:i:n.ate of PL is fil "'" 0.01314, and me diagonal element of (X'X)-! corresponding to fJl is en =
0,0000077, The estimate of <f was obtained in Example 15-2 as :.247. and til.o:.~.11 = 2.110. Therefo~
the 95% confu:1ence interval on [3, is computed from equation 15-20 as
which reduces to
0,00436"; /3, ,,; 0,0219,
We may also obtain a confidence interval on the mean response at a particular point, say
(X 01 ' Xw "., xoJ· To estimate the mean response at this point define the vector
This estimatoris unbiased, since EGo) = E(x;,~) = x;,11 = EVe)' and the variance afY, is
(15-22)
Therefore, a 100(1- a)% confidence interval on the mean response at the point (x,., x"' ""
..tJt)is
Equation 15~23 is a confidence interval about the regreSSion hyperplane. It is tbe multiple
regression generalization of equation 14~33.
446 Chapt:r 15 M",tiple Regression l I
The scientists conducting the experiment on damaged peact.es in Example 15·1 would like to OOn~
struct a 95% confidence interval on the mean damage for a peach dropped from:a height of Xl'" 325
~ if its density is Xi = 0.98 glcm? Therefore,
The estimated mean response at this point is found from equation 15~21 to be
-33,831]
325 0.98] 0.01314 = 4,63.
[ 34.890
: 24,63666 0,005321
-26.74679]~ 1 ]
O"x,(X'Xr'xc =2,247[1 325 0.98]1 000;>321 0.0000077 -(l.008353 325
c-26.14679 -(l,008353 30,096389lo,98
=2.247(0,0718) =0,1613,
Therefore, a 95% confidence interval On the me3l1 damage at this point is found from equation 15-23
to be
which reGuces to
3.78 S E(y,) :;; 5,48,
A
YC-tctj2,I1-p1./cJ
Cl(I'
+:<0 (X",,)-l Xc )
.A.
(15-25)
"' /',c-,(;---,-,-X--X-)-:-t---C'
sYo -Yo +t"12,._p~d'- I+xc\ "0i'
TIlls prediction interval is a generalization of the prediction inteI'\la! for a future observation
in simple linearregressio~ equation 14--35.
In predicting new observations and in estimating the mean response at a given point
(xc;. Xw ••• , X Ok )' one must be careful about extrapolating beyond the region containing the
original observations. It is very possible that a model that fits well in the region of the orig·
inal data will no longer fit well outside that region. In multiple regres>ion it is oiren easy to
inadvertently extrapolate, since the levels of the variables (XI1 • xa.> ... Xi*>' i;;;; 1, 2, , .. , n)
jointly define the region comaining the data. As an example, consider Fig. 15-1, which illus-
15~5 Hyporb.esis Testing in Multiple Linear Regmision 447
: XOj
L Onginal
, rangs for x1
trates the region containing the observations for a two-variable regression model. Note that
the point (.:tol' xO") lies within the ranges of both independent variables Xl and ~~ but it is out-
side the region of the original observations, Thus, either predicting the value of a new
observation or estimating the mean response at this point is an extrapolation of the original
regression modeL
Suppose mat the scientists in Example lSffl wish to construct a 95% prediction interval on the dam-
age on a peach that is dropped from a heignt of-XI = 325 IIlIl1 and has a density of.;s =0.98 glcmJ • Note
that x:: = [1325 0.98], and the point estimate of the damageisYn =x~~ =4.63 rom. Also. inE.~ample
15-4 we calculated x~(X'Xrlx.:: = 0.0718, Therefore, from equation 15-25 we have
Rejection of He: /3) == 0 implies that at least one of the independent variables XI' x?, .... xk
contributes significantly to the model. The test procedure is a generalization of the proce~
dure used in simple linear regression. The total sum of squares SfY is partitioned into a sum
of squares due to regression and a sum of squares due to error, say
Syy=SS,+SS,.
and if H,: /3, = 0 is !rue, then S5.10" - xi, where the number of degrees of freedom for the
x' is equal to the number of regressor variables in the model. Also, we can show that S5,10"
- X;,'," and SSe and SS, are independent. The test procedure for H,: J3, = a is to compute
_. SS,lk _ MS R
Ee- (15-27)
SSE!(n-k-l) MSE
and to reject Ho if FI) > F a.k,lI-i.'-I' The procedure is usually summarized in an analysis of vari~
ance table such as Table 15-4.
A computational formula for SS, may be found easily. We have derived a computa·
tion~ formula for SSE in equation 15-16, that is,
SSE=y'y- ~'X'y.
Now since SyJ :;; l,~",y~ - O:';"'ly'-y1/n = y'y - CZ;"'ly,lln. we may rev.rite the foregoing
equation
or
(15·28)
(15·29)
(15·30)
15~5 Hypothesis Testing in Multiple Lir.ear Reg;r<>...ssion 449
'~~j~~~~
We v.:ill test for significance of regression using the damaged peaches data from Example 15~ L Some
of the numerical quantities required are calculated in Example J5~2. Note that
I
Syy=y'y-- .L,y,.
r fi ',2
n \i=! )
= 904.60 (120.79)'
20
=175.089,
SSR~~'X'Y-.!.( :~.>,r
n\l:! )
(,20.79)'
=866.39
20
~ [36.88.
SS, =Syy - 5S,
~y'Y-~'X'Y
=38.21.
The analysis of variance is shown in Table 15-5, To testR(J: f31 = A =0, we calcula[e the statistic
Since Fo > Fo,IJ~;)',.1 """ 3.59, peach damage is related to drop height, froitdensity, or both. However, we
note that this does not necessarily imply that the relationship found is an apprOpriate oae for predict-
ing damage as a function of drop height or fruit density. Further tests of model adequacy are required.
where A.
ejj is the diagonal element of (X'xyt corresponding to The null hypothesis
He: ~= 0 is rejected if Itol > 1"",.,_ ,. Note that this is really a partial or marginal test,
p;
because the regression coefficient depends on all the other regressor variables xli ¢;) that
are in the modeL To illustrate the ;se of this test. consider the data in Example 15-1, and
suppose that we want to test
He: f3, = 0,
H,: {3," O.
The main diagonal element of (X'Xt' corresponding to jj, is en ~ 30.096, so the r statistic
in equation 15-32 is
.. 34.89 _ 4.24 .
.)(2.247)(30.096)
Since to"" 22 = 2.110, we reject Ho: /3, =' 0 and conclude that the variable x, (density) con-
tributes sigiuficantly to the modeL Note that this test measures the marginal or partial con~
tribution of X1 given that Xl is in the model.
We may also examine the contribution to the regression sum of squares of a variable,
say x" given that other variables x, (i ¢ J) are included in the model. The procedure used to
do this is called the general regression significance test, or the "extra sum of squares"
method. This procedure can also be used to investigate the contribution of a subset of the
regressor variables to the model. Consider the regression model with k regressor variables
y~Xll+£,
whereyis (nX I), Xis (nxp), II is (px I),. is (nx I), andp = k+ 1. We would like to
determine whether the subset of regressor variables Xt~ X:l~ .. "' xr (r < k) contributes
J
15-5 Hypo~esis Testing in Multip:e Lineru; Regression 451
significantly to the regression model. Let the vector of regression coefficients be pa.'1itioc.ed
as follows:
Ho: Pl =0,
(l5-33)
H,: p, ;cO.
The model may be \.VIitten
(15-34)
where Xl represents the columns of X ass.ociated with PI and ~ represents the columns of
X associated ",itb ~"
Fortbeft,II1110del (including both 13, and ~,), we know that ~ = (X'Xr'X'y. Also, the
regression sum of squares for all var..ables including the intercept is
(p degrees of freedom)
and
y'y -~'X'y
MS" = --'-""""""'.
~ n- p
SS,(I3) is called the regression sum of squares due to 13, To find the contribution of the terms
in 13, to the regression, fit the model assuming the null bypothesis Ho: 13, =0 to be true, Tbe
reduced model is found from equation 15-34 to be
sSR(i3d~2)lr
(15-33)
If Fo > Fa. f, "_p we reject HCr< concluding that at least one of the parameters in 1">1 is not zero
and, consequently, at least one of the variables XI' Xz... " x!' in Xj contributes significantly
to the regression model Some authors call the test in equation 15-33 a partial F-test
The partial F-test is very useful. We can use it to measure the contribution of xJ as if it
were the last var.able added to the model by computing
SSiflip", p" ... , Pi-I' /5;." ... , pJ,
452 Chapter 15 Multiple Regression
This is the inerease in the regression sum of squares due to adding X; to a model that already
includes XI' , ..• Xj_I' ;:,"H' ..•• xk' Note that the partial F-test on a single variable Xj is equiva-
lent to the t-test in equation 15-32. However, the partial F-tes! is a more general procedure
in that we can measure the effect of sets of variables. In Section 15-11 we will show how
the partial F-test plays a major role in model building, that is, in searebing for the best set
of regressor variables to use in the model.
Consider the damaged peaches data in Example 15-1. We will investigate the contribution of the vari-
able x,. (density) to tte model. That is, we wish to test
He: fl, = 0.
H,: /3," 0,
To test this hypothesis. we need the extra sum of squares due to /3,.. or
SS,(/3,I/3,. flc) = SS,cf3" fl" f3;J - SS,(/3,. I30l
= SS/J31' f3,I/3;J - sS,(/3,lf3,).
In Example 15-6 we calc'Jlated
(2 degrees of freedom).
The:[efore. we have
sS'/'/3,I/3" flu) = 136,88 -96.21
=40.67 (1 degree of freedom),
This is the increase in the regression sum of squares attributable to adtfng x,. to a model aJJ:eady con-
taining XI' To testHo: fl.1 ;: 0, [onn the test statistic
Note that theMS£; from thefull model. usi::g both x: ar.d x,.. is used in the denominator of the test sta~
t:.suc. Since F). Cd,I,;7::::: 4.45, we reject Hi; 132 =: 0 and conclude :.bat density (x2) contributes significantly
to the model.
Since this partial F~tes[ involves a single variable, it is equivalent to the Mest To see this, recall
°
that the t-test on Ho: .B: ;;; resulted in the tes: statistic tc :;:;; 4.24. Furt:b.crmore, recall that :he square
of a trmdom variable with v degrees of freedom is an Frandom variable with one and v degrees of
freedom. and we note that (4.24)' =: 17,98::.: FfJ ,
tant part of the multiple regression model building process. A good paper on this sUbject is
Snee (1977) (see also Montgomery, Peck, and VIning, 2001).
(15-39)
The coefficient of multiple determination for the regression model estimated in Example 15~ 1 is
That is, about 78.2% of the variability in damage y is explained when the two regressor variables, drop
height (XL) and fruit density (.:s) are used. The model relating density to Xl ouly was developed. The
value of R2 for this model turns out to be Rl == 0.549. Therefore, adding the variable X? to the model
has increased R2 from 0.549 to 0.782. -
AdjustedR'
Some practitioners prefer to use the adjusted coefficient a/multiple determination, adjusted
R2, defined as
(15-40)
The value Sy-/(n - 1) will be constant regardless of the number of variables in the model.
SSE1(n -p) is the mean square for error, which will change with the addition or removal of
tenns (new regressor variables, interaction terms, higher~order terms) from the model.
Therefore, R ~dj will increase only if the addition of a new term significantly reduces the
mean square for error. In other words, the R!dj will penalize adding terms to the model that
are not significant in modeling the response. Interpretation of the adjusted coefficient of
multiple detemrination is identical to that of R2.
454 Chapter 15 Multiple Regression
We can. calculate K~j for the model fit in Example 15-1. Prom Example 15-6, we found that SSE;;:;:
38.21 and Srr::= 175.09. The estimate R~: is then
38.21/(20-3) _1_ 2.247
~)=I
175.09/(20 I) 9.215
=0.756.
The adjusted R' will playa significant role in variable selection and model building later in
this chapter.
The residt:.aL.... for the n:odel estimated in Example 15-1 are shown.in Table 15-3. These residuals are
plotted 00:1 a normal probability plot in Fig. 15-2. No severe deviatiollS from r.ocrnaliry are obvious,
although the smallest residual (e) = -3.11) does not fall near the remaining residuals. The standard-
ized residual. -3.l7/~j2,247 :::: -2,11, appears to be large and could indicate an unUSt:a1 observation.
The residuals are plotted against y in Fig. 15~3, and againstx j and.A1 in Figs. 15-4 and ;5-5, respec-
tive;y_ In Fig. 15-4, there is some i.'1dication that the assumption of constant variance may not be sat-
isfied. Removal ofth.e unusual observation may improve the model fit, but there is no indication of
error in data colleetion. Therefore, the point 'kill be retained, We will see subsequently (Example 15-
16) iliat!:\Vo other regressor variables are required to adequately model these data,
2
•
•
.•• . .
• •
~
0
0
m
ro 0
E
~ ~ -1
,
• •
••
•
•
1-;
•
I •,
-2 ....
-3 -2 -1 0 2 3
Figure 15-2 Normal probability plot of residuals for E""ample 15-10.
r
15-6 ~!easures of Model Adequacy 455
3~ " i
2- " . " j
~ ~ -----~-"-~"-~------
i
:
[ .
... -"-------+---- .
o:~:~" . '. I
~~l~----------~----------~
2 7
Fitted value
12
3~----------------------------------,
2 •
•
•
-2 r
i
'.
~'
--~~--~~----~;~----~
300 400 500 600
x,
,Figure 154 Plot of residuals against x ( :or Example 15-10.
E~pl~;~s~ii
Sidewall panels for the interior of an airplane are formed i::.l a 1500~ton press. The unit manufactur~
ing cost varies \\1:th the production lot size. The data shown below give :he average cost per unit (in
hundreds of dollars) for this product (y) and :he production lot size (x), The scatter diagram. shown
in F1g. 15-6, indicates that a second-order polynomiul may be appropriate,
y LSI 1,70 [.65 [,55 lAS lAO 1.30 1.26 1.24 1.21 1.20 1.18
x 20 25 30 35 40 50 60 65 70 75 80 90
"90~
1.80 •
1.70 •
~
" i.50
"~
0 •
"- 1.50
0 •u
1.40
m
'"
~
1,30
~
«
1.20
• •
• .
1.10
1,00
, , I .. !~.~
20 30 40 50 60 70 8() 90
Lot size, x
Figure 15-6 D.ta for Example 15-11.
f 15-7 Polyr.omial Regression 457
I j:.811 1 20 400
; 1.70 25 625
1.65 I 30 900
1.55 I 35 1225
L48 1 40 1600
1.40 1 50 2500
y= 1.30 ' x= 60 3600'
1.26 I 65 4225
L24 : 70 4900
1,21 1 75 5625
1.20 ; 80 6400
cLl8j 1 90 8100
Solving the normal equations X"'Xtl = X'y gives the fitted model
, ; 2.1983 - 0.0225, + 0.0001251:<".
The test fo~sign.ificance of:egrcssioIl is shoVt"O in Table :5-6. Since Fo =2171.07 is significant at 1%,
we conclude iliat at least O!:.e of the parameters fJ! and /311 is not zero. Furth.ermore, the sta.1.da!d tests
for model adequacy reveal no unusucl behavior,
In fitting polynomials, we generally like to use the lowest-degree model consistent with
the data. In this example, it would seem logical to investigate dropping the quadratic term
from the modeL That is, we would like to test
H,: 13,,: 0,
H,: 13" "' O.
The general regression significance test can be used to test this hypothesis. We need to
determine the "extra sum of squares" due to {111' or
SS,(f3ulf3;,f3J = SSR(fi" f3 li lf3,) - SS,(f3,If3J·
The sum of squares 5S"(f3,, f3::1f3;J = 0.5254, from Table 15-6. To find SSR(fi,lf30), we fit a
simple linear regression model to the original data, yielding
y 1.9004 0.0091".
It can be easUy verified that the regression sum of squares for this model is
58,(13,113,) 0.4942.
Table 15-6 Test for Significance of Regression for the Second~~er Model in Example 15-11
Source of Sum of Degrees of Mean
Variation Squares Freedom Square F,
Table 15-7 Analysis ofVariaz;ce of Example 15~ 11 Showing the Test for He: [31, 0
Source of $\l."U. of Degree of Mean
Variation Squaces Freedom Square
Reg:-ession SS,([3" {J"i[3,) ~ 0.5:254 2 0.2627 2171.07
Lin_ SS,([3,I[3,) ~ 0.4942 1 0.4942 4084.30
Quadratic SS,(,8"1[3,, .8,) = 0.0312 I 0.0312 257.85
Error 0.0011 9 0.000121
Thtal 0.5265 11
Therefore, the extra sum of squares due to Pll' given that /31 and Pc are in the model, is
SS/fi"lA" /3) SSk(/3" /3,,113,) - SSifiJ/3~
~ 0.5254 - 0.4942
~0.0312,
The analysis of variance, with the test of H,: /311 ~ 0 incorporated into the procedure, is dis-
played in Table 15-7, Note that the quedratic term contributes significantly to the model.
x, X,
0 0 if the observation is fmm operator 1,
1 0 if the observation is from operator 2,
0 1 if the obsertation is from operator 3.
Indicator variables are also referred to as dummy variables. The following example illus~
trates some of the uses of indicator variables. For other applications, see Montgomery.
PeCk, and Vining (200 1).
15-8 Indicator Variables 459
E~pl~(?::i2',
(Adapted from Montgomery. Peck. and Vmiog, 2001), A mechanlcal engineer is investigating the sur~
face finish of metal parts ptoduced on a lathe and its relationship to the speed (in RPM) of the lathe,
The data are shown in Table 15-8. Note that the data have been collected usiGg two different typCs of
cutti.:lg tools. Since it is likely that the type of cutting tool affects tb.e surl'ace fiIrish, we will fit the
model
y ~ /3, + fil~' + Ph + E,
where y is the surface finL'>h, XI is the lathe speed in RPM, and X: is a.'1 indicator va.--!ahle denoting the
type of ct;.tting tool used; 'ilia: is,
x -{
ofor tool type 302,
, - 1 for tool type 416,
The parameters if. ttis model may be easily interpreted. If x,! ;;;; 0, then the model becomes
Y"" Pc + ftjX t + e,
which is :a straight-line model with slope Pi and intercept f3y However, if X. ;;;; 1. then the model
becomes
y ~ 13, + fi,x, + 13,(l) + E; p, - /3, + f3,x, + E,
which is a st:raight~Jine model v.ith slope PI and intercept f3ij + /32' Taus, the model y;; Po + f31X+ f3-zXz
+ e implies that surface finish is linearly related to lathe speed and that the slope f31 does Dot depend
on the type of cutting tool used, However, the type of cutting rool does affect rbe intereept,. and /32 indi~
cates the ehange in the intercept associated with a change in too) type from 302 to 416,
216 d L32.29 J
where.:s is the inmcam! variable. Now if too) type 302 is used,.:s "'" O. and the model is
y= {J,x,+e.
lftool rype 416 is used• .:s;;;;: 1, and the model becomes
y ~ {Jo - {J,", + A+ f3:,x, + e
= (fJ" + fiJ + (/3, + /3;JX, + S-
Note that A is the change in the intercept. and A is the change in slope produced by a change in
tool type,
Another method of analyzing these data set is to fit separate regression models to the data for
each tool type. However, the indicator variable approach has several advantages, rmt, only one
regression model must be estimated. Second, by pooling the data on both :001 types, more degrees of
freedom for error are obtained. Th.lrd. tests of both hypotheses on the parameters 13: end /33 a:e just
special cases of the general regression signific3IWe test
where
•
Skj= I,(xik-x.)(xg-Xj), k,j 1,2. (15-47)
i=l
and note that r" = r" = 1. Then the correlation form of the XX matrix, equation 15-46, is
The quantity r t2 is the sample correlation bet\.veenx[ and~. We may also define the sam-
ple correlation between Aj and y as
j=1,2, (15·50)
l
462 Chapter 15 Multiple Regression
where
x·· -XI
z11;:::: 1~~'2" j= 1,2.
B
The relationship between the parameters b , and b2 in the new model, equation 15-52, and
the paramerers {J" {JI' and A in the original model, equation 1.5-43, is as follows:
(15-53)
{J, (15-54)
(15-55)
The least-squares normal equations for the transformed model, equation 15-52, are
(15-56)
or
(15-57a)
rZy - liz'll'
(J5-5ib)
1-1.~
The regression coefficients, equations 15-57, are usually called staruiardized regression
coefficients. Many multiple regression computer programs use this transfor.:nation to reduce
round~off errors in the (X'X)-l mati"" These round-off errors may be very serious if the
15~9 The Correlation Matrix 463
original variables differ considerably in magnitude, Some of these computer programs also
display both the original regression coefficients and the standardized coefficients, The
standardized regression coefficients are dimensionless, and this may make it easier to com~
pare regression coefficients in situations where the original variables Xj differ considerably
in their units of measuremenL In interpreting these standardized regression coefficients,
hovvever\ we must remember that they a...-.oe still partial regression coefficients (i.e., b, s..'1ows
the effect of z) glven iliat other Zi> i are in the model). Fur",hermore, the hJ are attected
by the spacing of the levels of the x)' Consequently, we should not use the magnitude of the
bj as a measure of the importance of the regressor variables.
While we have explicitly treated only the case of two regressor variables. the results
generaIiz.e. If there are kreg:ressor variables X,I ~> •••• xk• one may vrrite the X'X matrix in
correlation form.
'12 '13 '"
"Jk
r'12 1 rn ,.,
r",
R=,r13 1 (15·58)
TZ3
''''
!
L'ik r2k r,. , " 1
w:tere r.j S/(S;,Sl))Lf2 is dIe sample correlation between x, and x-' and Sij;';;; I.;:",(X"i - Xi)
(X'I - x,). The correlations between Xj and y are
!'ly
: r2'j
g (15·59)
lrky j
"
where "Y =1:;"', (x., - x,)(y" - Y), The vector of standardized regression coefficients b' = £h"
b" ''', b,] :s
0= R·1g, (15·60)
The relationship between the standardized regression coefficients and the original regres-
sion coefficients is ~
~~lil~'131
For the data in Example 15·1, we find
S", = 175,089, S" = 184710.16,
SI,V;;4215.372, S" = 0.047755,
S" = 2,33, Sll ;;;51.2873.
Therefore.
51.2873
0.5460,
~(18471O.16)(0,047755)
464 Chapter 15 Multiple Regression
42\5,372
1(18471O.16)(6i089) - 0,7412,
S"
r2v :::: - - ,.... ljl ",;;l
2.33
••••• - 0.8060.
, (S"S,) "JrO,047755)(175,089)
'. - f)
0.54601
I,
J
From equation 15~56. the normal equations in terms of the standardized regression coefficients are
Ic 1
0.5460
0.5460l'~t]=[0'7412],
1 L'" 0,8060
r~,'J 1 0.5460]",'0,74121
~ b,J L0,5460 1 L0,8060,
1.424737 -<},77791J[0.7412]
= [ ·.{j,77791 1.424737. 0,8060
0.429022]
= [ 0.571754 .
These standardized regression coefficients could also have been computed directly from
elfuer equation 15,57 or equation 15·61. Note that alfuough b, > b" we should be cautious
about ooncluding that the fruit density X, is more important fuan drop height (x,)' smce h,
and b2 are still partial regression coefficients.
15-10.1 Multicollinearity
In most multiple regression problems. the independent or regressor variables Xi are inter-
correlated, In situations whieh this intercorrelation is very large. we say that multi:
collinearity exists. lv1ulticollinearity ean have serious effects on the estimates of the
regression coefficients and on fue genernl applicability of fue estimated model.
The effects of multicollinearity may be easily demonstrated. Consider a regression
model with two regressor variables Xl and x,., and suppose that'xi and ~ have been ;~stan~
dardized.'" as in Section 15~9, so that the X'X matrix is in correlation fo:rm, as in equation
15-49, The model is
,
J
r The (X'Xr' matrix for this model is
15-10 Problems in Multiple Regression 465
C=(X'Xf' =
. [)I(l-";)'
/"'-2
-r,,/(I- rall
-,:,/(I-r,,) lj(l-fl~) J
and the estimators of the parameters are
", -_~i.Y-'12X2Y
P 2'
. 1- r12
,
{3, = X'Y"
2 - '12 x'y
,
l-'l~ ,
where T;2 is the sample correlation between x; and.:tj, and x;y and x~y are the elements of
the X'y vector.
Now, if multicolli.'learity is presen~ x, and.<, are highly correlated, and If,,1 -> 1. In
such a situation) the variances and covariances of the regression coefficients become very
large, since v(li) = ejF' -> ~ as If,,1 -> 1, and COy (A, Ii)
= C"d' -> ± = depending on
whether r 12 ~ ± 1. The large vari,ances for A, imply that the regre.';sion coefficient. . are very
poorly estimated. Note that the effect of multicollinearity is to introduce a ''neat' linear
dependency in the columns ufthe X matrix, As r" ->± 1. this linear dependency becomes
exact. Furthermore, if we assume that x;y -? x~ as Irnl ~ ± 1. then the estimates of the
regression coefficients become equal in magnitude btll opposite in sign; that is = Ii, -A,
regardless of the true values of p, and p,.
Similar problems occur when multicollinearity Lt; present and there are more than t..vo
regressor variables, In general, the diagonal elements of the matrix C (X'X)-J can be
;:.!
written
the variance inflation factor for ~j' ~ote that these factors are the main diagonal elements
of the inverse of the correlation matrix. They are an important measure of the extent to
which multicollinearity is present.
Although the estimates of the regression coefficients are very imprecise when multi-
collinearity is present, the estimated equation may still be useful For example, suppose we
v.1sh to predict new observations. If these predictioQ.'i are required in the region of the
x~space where the multicollinearity is in effect, then often satisfactory results will be
obtained because while individual f3jmay be poorly estimated, the function 'L;',f3;:;, may be
estimated quite well. On the other band. if the prediction of new observations requires
extrapolation, then generally we would expect to obtain poor results. Exlrdpolation usually
requires good estimates of the individual model parameters.
466 Chapter 15 Multiple Regressio"
Multicollinearity arises for several reasons, It will occur when the analyst collects
the data such that. constraint of the form L~, ojX] ~ 0 holds among the colnnms of the X
matrix (the a, are constants, not all zero), For example. if four regressor varia.bles are the
components ~f a mixture, then such a constraint will always exist because the sum of
the cOClponents is always constant. Usually. these constraints do not hold exactly, and the
analyst does not know that they exist.
There are several ways to detect the presence of multicollinearity. Some of the more
important of these are briefly discussed.
1. The ·variance ioilation factors, defined in equation 15-63, are very useful measures
of multicollinearity. The larger the variance inflation factor. the more severe the
multicollinearity. Some authors have suggested that if any variance inflation factors
exceed 10 then multicollinearity is a problem. Other authors CQusiderthis value too
j
liberal and suggest that the variance inflation factors should not exceed 4 or S.
2. The determinant of the correlation matrix may also be used as a measure of multi-
collinearity. The value of this determinant can range between 0 and 1. When the
value of the determinant is I, the columns of the X matrix are orthogonal (i.e" there
is no intercorre1ation between the regression variables). and when the value is O.
there is an exact linear dependency among the columns of X. The smaller the value
of the determinant. the greater the degree of multicollinearity.
3. The eigenvalues or characteristic roots of the correla. .:ion matrix provide a measure
of multicollinearity, If X'X is in correlation form. then the eigenvalues of XX are
the roots of the equation
IX'X - All = o.
One or more eigenvalues near zero implies that multiCOllinearity is present. If Am:»;
and Aml'lC denote the largest and smallest eigenvalues of X'X, then the ratio Amru!Amin
can also be used as a measure of multicollinearity. The larger the value of this ratio,
the greater the degree of multicollinearity. Generally, if the ratio VAm" is less than
10, there is little problem with multicollinearity.
4. SOr:letimes inspection of the individual elements of the correlation matrix can be
helpful in detecting multicollinearity. If an element h·1
is close to 1. then Xi and
X, may be strongly multicollinear. However, when more than tvlO regressor vari-
ables are involved in a multicollinear fashion, the individual Tlj are not necessarily
large. Thus, this method will not always enable us to detect the presence of
multicollinearity.
5. [f the F-test for significance of regression is significant but testt; on the individual
regression coefficients are not significant, then multicollicearity may be present.
Several remedial measures have been pr<!p<Jsed for resolving the problem of multi-
collinearity. Augmenting the data with new observations specmcally designed to break up
the approximate linear dependencies that currently exist is often suggested. However. some-
times this is impossible for economic reasons, or because of the physical constraints that
relate the X,. Another possibility is to delete certain variables from the model 'This suffers
from the disadvantage that one must discard the information contained in the deleted
variables,
Since multicollinearity primarily affects the stability of the regression coefficients. it
would seem. that estimating these parameters by some method that is less sensitive to mul-
tieollinearity than ordinary least squares would be helpful. Several methods bave beeD sug-
gested for this. Hoed and Kennard (1970., b) bave proposed ridge regression as an
15·] 0 Problems in Multiple Regression 467
alternative to ordinary least squares. In ridge regression, the parameter estimates are
obtained by solving
11'(1)= (X'X + Ifr'X'y, (15·64)
where I> 0 is a constant. Generally, mues of 1in the interval 0 :s; I'; I a.-e appropriate. The
ridge estimator 11'(1) is not an unbiased estimator of ]:I, as is the ordinary least·squares esti-
mator ~, but the mean square error of 11'(1) will be smaller than the mean square error of~.
Thus ridge regression seeks to find a set of regression coefficients that is more "stable,lt in
the sense of having a small mean square error. Since mUlticollinearity usually results in
ordinary least-squares estimators that may have extremely large variances, ridge regression
is suitable for situations where the multicollinearity problem exists.
To obtain the ridge regression estimator from equation 15~64, one must specify a value
for the constant I. Of course, there is an "optimum" I for any problem, but the simplest
approach is to solve equation 15-64 for several values of I in the inren'lll 0 :s; I ;; L Then a
plot of the values of Il'(l) against 1is constructed. This display is called the ridge trace. The
appropriate value of 1is chosen subjectively by inspection of the ridge trace. TypicallYI a
value for I is chosen such that relatively stable parameter estimates are obtained. In general l
the variance of 11'(1) is a decreasing function of I, while the squared bias [1l-Il'(I)J' is an
increasing function of l, Choosing the value of [involves trading off these two properties of
13'(1)·
A good discussion of the practical use of ridge regression is in Marquardt and Snee
(1975). Also, there are several other biased estimation techniques that have been proposed
for dealing with multicollinearity. Several of these are discussed in Montgomery, Peck, and
Vrning (2001).
it,,·mp~15-14
(Based on an example in Rald, 1952.) The heat geoerated in calories per gxa.,\ for a particular type of
cement as a function of the qua."1tities of foll::' additives (ZI' z;>.' z,. andz,;) is shown in Table 15-10. We
wish to fit a :nultiple linear reg:res;,lon model to these da:a.
r.:
where Sjj = t (Zij - ;i! is the corrected sum of squares of the levels of <;" The coded data are shown
in Table 15-11. This tran..<;fonnation makes the intercept orthogonal to the o~r regression coeffi-
cients, since the:first column of the X matrix consists of ones. Therefore, the intercept in this model
will always be estimated by y. The (4 x 4) X"X matrix for the foUl' coded variables is the correlation
ma:rix
The ...'ariance inflation factors are the main diagonal clements of this matrix, Note 'J:lat three of t!:!e
variance inflation factors exceed 10, a good indication that multicollinearity is present. The eigenval-
ues of X'X are }'1 := 3.657, i-;::= 0.2679, A, = 0.07127. and }"4 ':;::: 0.004014. Two of the eigenvalues,.:t.
and A4> are re1atively close to zero. Also, the ratio of the: largest to the smallest eige:r:.value is
?.~ = 3.652....=911.06,
Amill 0,004014
which is considerably larger than 10. Therefore, since examination of the variance inflation factors
and the eigenvalues indicates ?otential problems with multicollinearity, wewill use ridge regression
:0 estiro.ate the model parameters.
We solved Equation 15-64 for 'various values of l, and the results are su:m..mar:ized in Tabie 15-12. The
ridge trace is shown in Fig. 15-7, The instabLity of the least-squares estimates tJ;U = 0) is evident
from inspectio:1 of the ridge t"ace. It is often difficult to choose a value of 1from the ridge trace e::.3r
simultaneously stabilizes the estimates of all regression coefficients. We will choose l "'" 0.064. which
implies that the regression model is
y = 25,53 -18.0,566x1 "'r' 17.2202xz + 3{}.0743;:;) + 4.7242J:~,
using A:: y:: 25.53. Coavening the model to the original variables Zt we ha,,'C
y = 2.9913 - 0.8920z, + 0.3483" + 3.3209" - 0.0623".
Pi(l)f
70
60
Ie
o P4 (I}
-10
P;(I)
-2C
-30
-40
-50
--<l°o O.OS 0.:0 0.15 0.20 0.25 0.30 0.35 0.40 D.45 0.50 0,55
Figure 15~7 Ridge trace for Ex.ample 15-14,
1
470 Chap"" 15 Multiple Regression
Clearly if the it.!? point is influential, its removal will result in ~"1 changing considerably
from me value p. Thus a large value of D, implies that the ith point is influential. The sta-
tiseic Di is acu:ally computed using
(15·65)
Xl!
XI':'. ,-~---------------!
, Region containing 1
all obServattons I
except the ith :
I. 1
I
I
I
y~x~
= X(X'Xr'X'y
=Hy.
Thus H is a projection matrix that transforms the obsen'ed values of y into a set of fitted
values 'Y.
From equation 15-65 we note that D, is made up of a component that reflects how well
the model fits the ith observatioll Yi [the quantity eJ~MSE(I- hii ) is called a Studentized
residual, and it is a method of scaling residuals so that they have unit variance] and a com~
ponent that measures how far that point is from the rest of the data ["lil(1 - h,) is the dis-
tance of the ith point from the centroid of the remaining n - 1 points]. A value of DI > 1
woold indicate that the point is influential. Either component of D, (or both) may contribute
to a large value,
[~[':!II~te!5Ii5;
Table !5-13lists the values of D; for the c!a::naged peaches data in Example 15~ 1. 'Ib illusnte the ca1~
culations. CODSider the first observation:
Di
Table 15..13 I:rfb.1cnce Diagnostics for the Damaged Peaches Data in Example 15~15
[2.06N2.247(1-O.249)]'. 0.249
3 (1-0.249)
~ 0.277.
The values in Table 15~13 were calculated using :Minitab", The Cook distance measu..--e Di does not
identify any potentially influential observations in the data, as no value of Di exceeds unity.
15·10.3 Autocorrelation
The regression models developed thus far have assumed that the model error components
S; are uncorrelated random variables. Many applications of regression analysis involve data
for which this assumption may be inappropriate. In regression problems where the depend-
em and independent variables are time oriented or are time-series data. the assumption of
uncorrelated errors is often untenable. For example, suppose we regressed the quarterly
sales of a product against the quarterly point~of~sa1e advertising expenditures. Both vari~
abIes are time series, and if they are positively correlated with other factors such as dispos-
able income and population which are not included in the model, then it is likely that
the error terms in the regression model are positively correlated over time, Variables that
exhibit correlation over time are referred to as autocorrel.ated variables. 11any regression
problems in economics, business, and agriculture involve autocorrelated errorS.
The occurrence of positively autocorrelated errors has several potentially serious con~
sequences, The ordinary least-squares estimators of the parameters are affected in that they
are no longer minimum variance estimator51 although they are still unbiased. Furthermore,
the mean square error !l;fSJ;; may underestimate the error variance cr, Also, confidence inter-
vals and tests of hypotheses. which are developed assuming uncorrelated errors, are not
valid if autocorrelation is present.
There are several statistical procedures that Can be used to determine whether the error
terms in the model are uncorrelated. We ,Yin describe one of these, the Durbin~ WatSOn test
This test assumes that the data are generated by the first-order au.toregressive model
where t is the index of time and the error terms are generated according to the process
{15-67}
where Ipi < 1 is an unknown parameter and 0, is a NlD(O, <f'} random vanable. Equation
15-66 is a sjmple linear regression model, except for the errors, which are generated from
equation 15~67, The parameter p in equation 15~67 is the autocorrelation coefficient. The
Durbin-Watson test can be applied to the hypotheses
He: p=O,
(l5-6S)
H1:p>O.
Note that if Ho: P = 0 is not rejected, we are implying that there is no autocorrelation in the
errors. and the ordinary linear regression model is appropriate.
15-10 Problems in Multiple Regression 473
To test II,: p = 0. first fit the regression model by ordinary least squares. Then, ealcu-
late the Duibin-Watson test statistic
(15-69)
where e, is the rth residual For a suitable value of a, obtain the critical values D a. u and Dq,J.
from Table 15-14.lf D > D~", do not reject II,: p =0; but if D < D .. t • reject II,: p= 0 and
conclude that the errorS are pOSitively autocorrelated. If D13... L ~ D !£ D 1Z, lj' the test is
inconclusive, \vnen the test is inconclusive, the implication is that more data must be col-
lected. In many problems this is difficult to do.
To test for n.egative autocorrelation, that is) if the alternative hypothesis in equation
15-68 is HI: P < 0, then use lY =4 - D as the test statistic, where D is defined in equation
15-69, If a two-sided alternative is specified, then use both of the one~sided procedures. not-
ing that the type I error for the two-sided test is 2a, where a is the type I error for the one-
sided tests.
The only effective remedial measure when autocorrelation is present is to build a model
that accounts explicitly for the autocorrelative structure of the errors. For an introductory
treatment of these methods, refer 10 Montgomery, Peck, and Vining (2001).
An Possible Regressions This approach requires that the analyst fit all the regression
equations involving one candidate va.,;able, all regression equations involving two candi~
date variables, and so on. Then these equations are evaluated according to some suitable cri-
teria to select the "'best" regression modeL If there are k candidate variables, there are 2k
total equations to be examined, For example, if k 4, there are 24:::;: 16 possible regression
equations, while if k 10, there are 2" 1024 possible regression equations. Hence. the
number of equations to be examined increases rapidly as the number of candidate variables
increases.
15-11 Selection of Variables in Multiple Regression 475
There are a number of criteria that may be used for evaluating and comparing the dif·
ferent regression models obtained, Perhaps the most commonly used cri:erion is based on
Ii!
!.he coefficient of multiple detennination. Let denote the coefficient of determination for
a regression model with p terms, that is, p - 1 candidate variables and an intercept ~eml
(note that p !5 k + 1). Computationally, we have
= 1 [(b.)2 .
las + vanance
(j2
"J-
o
Figure 15-9 Plot of ~ against p,
476 Chapter 15 Multiple Regression
Minimum MSE(o)
L -_ _ _ _ _ _ _ _ _~~----
\Ve use the mean square error from the full k + 1 term model as an estimate of a'l; that is,
IT = MSs(k + I). An estimator of r,is
(15-71)
If the p-tel1I!. model has negligible bias, men it can be shown that
E( C,lzero bias) = p.
Therefore, the values of Cp for each regression model under consideration should be plot-
ted against p. The regression equations that have negligible bias will have values of Cp that
fall near the line Cp = P. while those with significant bias will have values of Cp that pwt
above this line. One then chooses as the "best" regression equation either a model with min~
imum Cp or a model with a sJJghdy larger CD that contains less bias than the m.i.nimum.
R;
Another criterion is based on a modification of that accounts for the number of vari-
ables in the model. We presented this statistic in Section 15-6.1. the adjusted R' for the
model fit in Example 15-L This statistic is called the adjusted R~ defined as
'()
RoO n-l("
p =1--- 1-11:). (15-72)
7 n- p P
Note that R:"j(P) may decrease as p increases if the decrease in (n - l){l - R; is not CDW-
pensated for by the loss of one degree of freedom in n - p. The experimenter would usually
select the regression model that has the caximum value of R~(P). However, note that this
is equivalent to the model that minimizes MS.(P). since
j
15-11 Seleetion of Variables in Multiple Rcgra""ssion 477
The data in Tab;e 15~ 15 are an expanded set of data fo:- the damaged peach data in Example 15~1.
Tnere are now five candidate variables, drop height (Xl)' fruit density (.x::;, fruit height at impact point
(x:J. fruit pulp thickness (x4 ), and potential energy of me fruit before me impact (x5 ),
Tab~e 15-16 presents the rescl~ of rur.uing all possible regressions (except the tivial modei wim
ody an intercept) on these data. The values of R~, R~lp), MS/.p), and ep are given in fr.e table. A plot
R!
of the maximum for each subset of size p is shown in Fig. 15~ 1 L Based on this plot there does not
R!
appear to be much gain in z.dding the fift:h variable. The value of does Dot seem to bcrease sigrif-
R!
icantly 'J.'!th the addition of x, over the four-variable model with the highest value. A plot of the
minimum lrfS:/.,p) for each subset of size p is sholA.'J1 in FIg. 15~12, The best two-variable model is
either (x;. X:/) or (..tz. X:/); the best th...'"U:-variable model is (xp.x::, x:J; the besl four-w.u::iablc model is
either (x:,.."t:J,.:s. xJ or (xl' 4' x 3• X5)' There are several models with relatively small values of MSi,p).
but either the three-variable model (xl' 4'xJ or me four~variable model (x,.~, Xz, xJ would be supe~
rior to the other models based on the MSdJ;) criterioll. Further investigation ?oill be necessary.
A Cp p:ot is shown in Fig. 15-13, Only 'fue fiv~variable model has a Cp '5. P (specifically Cp =-
6.0), but the Cp value for the four-variable model (Xl' x.~.~, x,J is Cft:;;: 6.1732. There appears to be
insufficer.t gain in the Cp value [Q justify including x5' To illustrate the calculations, for this equation
(for the rrwdel including (X" x" x,. xJ] we wou:d find
C = SS£(p) -n-2p
p 02
18.29715 20+2'5)=6.1 732.
1.13132 \
•
•
0.7
5 6
P
Figure 15·11 11:0 maximwn R; plot for Example 15·16.
oJ"
:§;
•
~
w
0.8 r
J, 2 3
•
4
.. ·
j
5
t:
6
P
Figure 15·12 The MS,(P) plot for &le 15·16.
35 F'
u~
25,
15 ~
i
5::::L_ _.
• •j I
2 3 4 5 6
P
Figure 15-13 The C, plot for Exa::nple 15·16.
noting tha:(f1= 1.13132 is obtained from the full equation {xt>;S, x 3• x4• xj ). Slnee all other models
(with the exclusion of the five-variable model) contain substantial bias, we would conclude on the
basis of the C", criterion that the best subset of the regressor variables is (x:. X:!,.:s, x,J. Since this model
also results in rela::ively small MSs(p) and a relatively high R!.
we would selecr it as the "best"
regression equation, The final model is
jk-19.9 + 0.0 123x, + 27.3x, - 0.0655", -0.196<,.,
Keep in mind. though, that fi.t.r.her analysis should be conducted on this model as well as other pos-
sible candidar:e models. VV1th additional investigation,. it is possible to discover an even bcttcr.-fitting
model. We v.ill discuss this in mOre detail later in this chapter.
15~ 11 Selection of Variables in Multiple Regression 479
Stepwise Regression This is probably the most widely used variable selection technique.
The procedure iteratively constructs a sequence of regression models by adding or
Table 15-16 Al1 Pos.sible Regressions for the Data in Example 15~ 16
?':.u:nber in Variables in
Modelp-I Ii! if""cp) Gp MS,fp) Model
removing variables at each step. The criterion for adding or removing a variable at any step
is usually expressed in terms of a pattial F·tes<. Let F i• be the value of the F statistic for
adding a variable to the model, and letFoot be the value of the Fstatistic for removing a vari~
able from the model. We must have Fin ~ Fout' and usually F.~ ;;: F 00\'
Stepwise regression begir:s by forming a oneMvariable model using the regressor vari~
able that has the highest correlation with the response va..'iable y. This will also be the vari~
able producing the largest F statistic. !fno F statistic exceeds Fin' the procedure terrnina.tes,
For example, suppose that at this step XI is selected. At the second step the rema.:ining k - 1
candidate variables are examined, and the variable for which the statistic
is a maximum is added 10 the equation, pro,ided that F, > F". In equation 15-73, MS/"xr x,)
denotes the mean square for error for the model containing both XI andx? Suppose that this
procedure now indicates that:s should be added to the model. Now the stepwise regression
algorithm de1lermines whether the variable x; added at the first step should be removed. This
is done by calculating the F statistic
(15·74)
~;m;~~~'i~;4t
We w'Jlapply stepwise regression to the damaged peaches data in Thble 15-15, Miuita.b2i O!!tput is pnr
vided in Fig. 15-14. From this figure, we see that variables XI';S, and.x;, are sigcifica..'lt; this is because
the last column contains entries for only ,xl'~? and.t:J. Figure 15-15 provides the SAS computer output
that '.vil.l support the computatioas ro be calculated ne."tt. Instead of speciff.!lg u~ca1 values of F ~
andFQ<,# we use an advertised type I ¢ITOr of a'"",O.10. The first Step consists ofbcild.ing a sirnp:e linear
regression model using the variable that gives the largest F statistic. This is X-z, and since
S 1.84 1. SO 1.17
:R-Sg 65.3C 7S.0S 87.56
R-Sq (adj) 63.37 75.47 S5.23
C-p 37.7 20.0 7.3
sS.(lldIl2.llc) 46.45691
MSE(x,.x;)
2.26063 - 20.55 > F;, =F,.".,.i1 =3.03,
x. is added to ':he model. :;ow the procedure evaluates whet.'ier or not Xz should be retained, given that
x ~ is in the modeL rus mvol yes calculating
TherefOie.t: should be retained. Step 2 terminates 'IN:ith bot'i x; and X:: in t'ie model.
The :bird step finds the next variable fer entry as;S. Since
:l:; is added to the model. Partial F~te.sts on Xz (given Xl and.x~ and Xl (given ~ and xJ) indicate that
these variables sholtld be retained. Therefore. the third step coucludes with the variables XI' x 2' and Xj
in the model.
At 'the fourth step. neither of the remainir.g terms, x.. or x~. is significant en.ough to be included
in the model. Therefore. the stepwise proced\4'"e ~tes.
The stepwise regression proce~""e would conehlde that the bes~ model includes XI' )'.,.. and-'S.
The usual checks of medel adequacy, such as residual analysis and Cp plots. should be applied to the
equation. These results are si.mi.l.ar to those found by all possible regressions, with. the exception that
Xi was also considered a possible significant variable with all possible regressions,
482 Chapter 15 Multiple Regression
Forward Selection This variable selection procedure is based on the principle that vari-
ables should be added to the model one at a time until no re:mai::ring candidate variables
produce a significant increase in the regression sum of squares. That is. variables a...~ added
one at a time as long as F> Fin' Forward selection is a simplification of stepwise regression
that omits the partial F-test for deleting variables from the model that have been added at
previous steps. TIlls is a potential weakness of forward selection; the procedure does not
explore the effect that adding a variable at the current step has on variables added at ear~
lier steps.
No other variable met the 0.1000 significance level for entry into the.. model,
Summary of Forward Se:e..ction
Variab":e Number Partial Model
Step Entered va.:rs In R-Square R-Square C(p) F Value Pr > P
1 x2 1 0.6530 0,6530 37.7047 33.87 <.0001
2 xl 2 0.1275 0.7805 19.9697 9.88 0.0059
3 x3 3 0.0951 0.8756 7,2532 12.23- 0.0030
Figure 15-15 SAS output for stepwise regression in Example l5~ 17.
15-: 1 Selection of Variables in Multiple Regression 483
Ex:unplelS-18 .
AppE.carion of the forward selection algoriilim to the damaged peach data ir. Tabre 15~ 15 would begin
by adding X-z to the :node1. The::l th~ variable that induces the Wb'eSt partial F~test. given tha! X 2 is in
the model, is added-this is variable XI' The third step enters X:" which produces tIre Largest pa.,"tial F
statistic given 1f<.at x t and x: are in the modeL Since the partial F Statlsr:cs for XJ al'l.d;; are not signif~
iea.I!t, the procedure terminates. The SAS outputfo:- forward selection is given in Fig. 15-16. Note <±tat
forward selection leads to the same:final model as stepwise regression, This is not always the case,
Backward Elimino.tion This algorithm begins with all k candldate variables in the modeL
Then the variable \Vith tbe smallest partial F statistic is deleted if this F statistic is insignif~
kant, that.is, ifF < Foul' Next, the model with k - 1 \'3riables is estim.ated~ and the next vari-
able for potential elimination is found. The algorithm terminates when no further variables
can be deleted.
To apply backward elimination to ilie dara in Table 15~15. we begi."'l by estimating t.'1e fur: model in
a!: five variables. This model is y ~-20.89732 + 0.01 lO2x, + 27.37046x, - 0.06929", - 0.25695x, +
O.Ol66&X,.
The SAS COffiputer output is given in Fig. 15~17. The partial F~tests for eaeh variable are as
follows:
SS,(Pl;3,. /3" p, ,p,. Po) 13.65360
1'j 12m.
MSe 1.13132
SS, I,;3, !f3" {3,. p,. P,. /30 'J 21.73153 ~ 19.2J.
F,
MS z 1.13132
miD.ated. T.:e th.."'Ce~variable model (x!, Xz. Xv bas all variables significant according to the partial
F~test criterion. Note that backward cl.i..::::.Unation has resulted in the same model that was found by
forward selection and stepwise regression, This r::w:y not always happen,
Some Comments on Final )'Jodel Selection livre have illustrated several different
approaches to the seleetion of variables in multiple linear reg..""ession. The final model
obtained from any model-building procedure should be subjected to the usual adequacy
checks. such as residual analysis and examination of the effects of outermost points. The
analyst may also consider augmenting the original set of candidate variables with cross
products, polynomial terms, or other transformations of the original variables that might
improve the model
484 Chap'", 15 Multiple Regression
All variables left k" the model are significant: at the 0.1000 level,
Summary of .Bac1cNard Eli.'l'Iination
Variable Nu'1'1ber Partial Model
Step Entered Vaxs In R-Square. R-Square C(p) F Value PO? > F
1 x5 4 0.0140 0.8955 6.1732 2.17 0.1626
2 x4 3 0, :)199 0.8756 7.2532 2.86 0.:1:7
Figure 15~ 16 SAS output for forward selection in Example 15-18.
15-11 Selection of Variables in Multiple Regression 485
AI: variables lef"t in "the model are signific~~t a~ ~he O.lCOO leveL.
Ke other variable met the 0.1000 significance level for entry into the model.
Summary of St:epwise Select~on
Va:r:"able Variable Number Partial Mode:
Step Entered Removed Vars In .a-Squar€ R-Square Cep) F Value Pr > F
x2 1 0.6530 0.6530 37.7047 33.87 <.0001
2 xl 2 0.1275 0.7805 19.9697 9.88 C,0059
3 ,,3 3 0.0951 0.8756 7.253-2 12.23 0,C030
Figure 15~17 SAS output for bacbvard elimination ill Example 15~19.
A major criticism of variable selection methods. such as stepwise regression, is that the
analyst may conclude that there is one "bestH regression equation. This generally is not the
case, because there are often several equally good regression models that can be used. One
486 Chapter 15 Multiple Regression
way to avoid this problem is to use several different model~building techniques and see if
different models result. For example, we have found the same model for the damaged peach
data by using stepwise regression. forward selection, and backward elim.ination. This. is a
good indication that the threewvariable model is the best regression equation, Furthennore,
there are variable selection techniques that are designe.d to find the best one-variable model.
the best two~variable model, and so forth. For a discussion of these methods, and the van..
able selection problem in general, see Montgomery~ Peck., and Vining (200l),
If the number of candidate regressors is not too large, the all~possib:e-regressions
method is recommended. It is not distorted by multicollinearity among the regressors, as
stepwise-type methods are.
15·12 SU~1{
This chapter has introduced multiple linear regression, including least-squares estimation of
the parameters, interval estimation, prediction of new observations, and methods for
hypothesis testing, Various tests of model adequacy. including residual plots. have been dis-
cussed. It was shown that polynomial regression models can be handled by the usual mul-
tiple linear regression methods, Indicator variables were introduced for dealing with
qualitative variables. It also was observed that the problem of multicollinearity, or inter-
correlation between the regressor variables. can greatly complicate the regression problem
and often leads to a regression model that may not predict new observations well. Several
causes: and remedial measures of this problem, inc1udi.'lg biased estimation techniques,
were discussed, Finally, the variable selection problem in multiple regression was intro-
duced. A number ofmodel~bui1ding procedures, including all possible regressions, stepwise
regression. forward selection. and backward e1imination, were illustrated.
15·13 EXERCISES
lS~l, Consider th.e damaged peach data in Table lS~4. Using the results of Exercise 15~2, find a 95%
15-15, confidence interval on j3~.
(a) Fit a regression model using XI (drop height) and 15~5. The data in the table at the top of page 487 are
x.;. (fruit pulp thickness) to these data. the 1976 team performance statistics for the teams in
(h) Test for significance of regression. the National Football League (Source: The Sparring
(c) Compute the residuals from this modeL Analyze ly'ews),
these residuals using the methods discussed in (a) Fit a multiple regression :nodel relating the num-
this chapter. ber of games won to the teams' passing yardage
Cd) How does this twO-\Mable r:1ode~ compare Veith {x..), the percentage of rushing plays (x,), and the
the two-variable model using x: and .:tz from opponentS' yards rosblng (x,),
Example 15- j ? ' (b) Construct the appropriate residual plots a:.d com-
15-2. Consider the damaged pea.en data in Table ment on model adequacy,
15-15, (e) Test the significance of each variable to the
(a) Fit a regression ::nodel using XI {.:top height),.:tz model, u:sing either the Nest Or the p;trtial F~test
(fruit density), and ~ (fruit height at impact point) 1S-6. The table at the top of page 488 presents gaso-
to cese data. line rr.ileage perforrr,.a.::ce for 25 at.tomobiles (Source:
(b) Test for sigrlficance of regression. Motor Trend, 1975),
(c) Compute the residcals from this model. Analyze (a) Fit a multiple regression :::::lodel relating gasoline
these residuals using the methods discussed in mileage to engine displacement (Xl) and number
"'Us chapter. of carburetor ba:c-els (xJ..
15-3. Using the results of Exercise 15-1, find a 95% (b) Analyze the residuals and comment on mood
confidenee interval on A. adequacy.
15·13 Exercises 487
(c) What is the value of addirtg ,46 to a model aut amo:cn! temperature (xa. the number of days in jre
already contains,X;? lIIDnth (:r;), the average product purity (x,), and :he
tOllS of proouct produced (.xJ. The past year's hlston-
15M'7. The electric power COJ1SllD1ed each month by a cal data is available and is pres:ented in the table at the
chemieal plant is thought to be related to the average bottom of page 488.
488 Chapter 15 ~lu!tipl< Rogression
Automobile y
Apollo 18.90 350 165 260 8.0:1 2.56:1 4 3 200.3 69.9 3910 A
Nova 20.00 250 105 185 8.25:1 2.73:1 3 196.7 72.2 3510 A
Mo::>arCh 18.25 351 143 255 8.0:1 3.00:1 2 3 199.9 74.0 3890 A
Duster 20.07 225 95 170 8.4:1 2.76:1 1 3 194.1 71.8 3365 M
Jenson Cony. 11.20 440 215 330 8.2:1 2.88:1 4 3 1845 69.0 4215 A
Skyhawk 22.12 231 110 175 8.0:1 2.56:1 2 3 179.3 65.4 3020 A
34.70 89.7 70 81 8.2:1 3.90:1 2 155.7
Scirocco
Corolla SR~5 30040 96.9 75 83 9.0:1 4.30:1 2 "5 165.2
64.0
65.0
1905
2320
M
M
Camara 16.50 350 155 250 8.5:1 3.08:1 4 3 195.4 74.4 3885 A
DatsunB210 36.50 85.3 80 83 8.5:1 3.89:1 2 4 160.6 62.2 2009 M
Capri II 21.50 171 109 146 8.2:1 3.22:1 2 17004 66.9 2655 M
"
Pacer 19.70 258 110 195 8.0:1 3.08:1 I 3 171.5 no 3375 A
Granada 17.80 302 129 220 8.0:1 3.00:1 2 3 199.9 74.0 3890 A
Eldorado 14.39 500 190 360 85:1 2.73:1 4 3 224.1 79.8 5290 A
Imperial 14.89 440 2[5 330 8.2:1 2.71:1 4- 3 231.0 79.7 5185 A
NovaLN 17.80 350 155 250 8.5:1 3.08:1 4 3 196.7 72.2 3910 A
Starlirc 23.54 231 no 175 8.0:1 256:1 2 3 179.3 6504 3050 A
Cordoba 21.47 360 180 290 8.4:1 2,45:1 2 3 214.2 76.3 4250 A
Trans Am 16.59 400 185 NA 7.6:1 3.08:1 4 3 196 73.0 3850 A
Corolla E~5 31.90 96.9 75 83 9.0:1 4.30:1 2 5 165.2 61.8 2275 }'1
M"klV 13.27 460 223 366 8.0:1 3.00:1 4 3 228 79.8 5430 A
Celica GT 23.90 133.6 96 120 8.4:1 3.91:1 2 5 171.5 63.4 2535 M
Charger SE 19.73 318 140 255 8.5:1 2.71:1 2 3 215.3 76.3 4370 A
Cougar 13.90 351 148 243 8.0: 1 3.25:1 2 3 215.5 78.5 4540 A
Corvette 16.50 350 165 255 8.5:1 2.73:1 4 3 185.2 69.0 3660 A
~-----------------
y. Miles I gallon.
XL: Displacement (cub~c in.),
x,: Horsepower (ft~lb) .
.'f): Torque (ft~lb).
XA: Compression noo,
x,: Rear a:tle ratio.
;r~: C'WJrctor (barrels).
Xl: No. of tra.-.s:nlssion speeds.
x~; Overall length (in.),
x.;: \Yidth {in.).
X10: Weigl:t Oos).
XI i: Type of tnl.n$mlssion (A-au:omatic. M-manU3l).
j y x, x,
240 25 24 91 100 300 80 25 87 97
236 31 21 90 95 296 84 25 86 96
290 45 24 88 110 267 75 24 88 110
274 60 25 87 88 276 60 25 91 105
301 65 25 91 94 288 50 25 90 100
316 72 26 94 99 261 38 23 89 98
(a; Fit a multiple regression model to these data. (c) Test the hypothesis thar Pn == O.
(b) Test for sigrllficance of regression. (d) Compute the residuals and test for model
( c) Use partial F statistics to test H,: {J, = 0 and H,: adequacy.
(J, =0. 15~10. Consider the follow.Jlg data. which result from
(d) Compute the residuals from this model. .Analyze .an experiment to derer.:nine the effect of x = test time
the res~du.als using the merhods discussed in this in hours at a particular temperature on y change in
chapter. oil viscosity.
15-8. Hald (1952) reports data on the heat evolved in y -4.42 -1.39 -1.55 -1.89 _2.43 -3.15 -4.05 -5.15 -6,43 -7.89
calories per gram of cement (y) for various amounts of
four ingredients: (x!, x... x,> xJ. x O.2S 0,50 0.75 :.00 1.25 L50 1.75 2.00 2.25 2,50
different slopes and intercepts) are required to ade- 15-22. Use the National Football League Team Per~
quately model the data. fonnance data in Exercise 15-5 to build regression
models using rhe following techniques:
15.. 16. Piecewise Linear Regression (I). Suppose
(a) All possible regressions.
that y is piecewise linearly related to x. Thar is. differ·
ent linear rela.tionships are appropriate over the inter- (b) Stepwise regression.
vals - o:x> < X ~:l 3lld x' < x < «l, Show how indicator (c) Fonvard selection.
variables em.;. be used to fit such a piecewise linear (d) Backward elimination.
regression model, assuming that point x' is kno'WU. (e) COIl:.lm.ell.t on the various models obtained.
15~17. Piecewise Linear Regression (II). Consider 15~23. Use the gasoline mileage data in Exercise 15-6
the piecewise linear regression mode3 described in to build regression models using the following tech~
Exercise 15-16. Suppose that at point x' a discontinu- niques:
ity occurs in the .regression function. Show how indi~ (a) All possible regressors,
cator variables can be us~ to iIlcorporate the
(b) Stepwise regression.
discontinuity into the modeI.
(e) Forward selection.
15~18. Piecewise Linear Regression (ill). Consider (d) Backward _atio"
the piecewise linear regression model described in
(e) Comment on the various models obtained,
Exercise 15-16. Suppose that point x' .is not known
15~24. Consider the Rald cement data in Exercise
with certainty aud must be estimated. Develop an
approach that could be used to fit the piecewise linear 15-8. Build regression models for the data using me
regression model following techniques:
(a) i~l1 possible regressions,
1s..19. C'..alcu1ate the standardized regression coeffi-
(b) Stepwise regression.
cients for the regression model developed in Exercise
lS-L (c) FoIW'dl'd selection.
Cd) Backward elimination.
15~20. Calculate the standardized regressioo coeffi-
cients for the regression model developed in Exercise 15~25. Consider the Hald cement data in Exercise
15-2. 15-8. Fit a regression model involving all four :reg...~s
sors and find t.~ variance inflation factors. Is multi-
15~21. Find the variance inflation factors for the collinearity a problem i:1 this model? Use ridge
regression model developed L1 Example 15-1. Do they regression to estimate the coefficients in this model.
ir.dicate t.~at multicollinearity is a prOblem in this Compare the ridge model to the models obtal.!1ed in
model? E..'Cercise 15~25 using variable selection methods.
Chapter 16
Nonparametric Statistics
16·1 INTRODUCTION
Most of the hypothesis testing and eonfidenee interval proeedures in previous chapters are
based on the assumption that we are working with random samples from normal popula-
tions. Fortunately, most of these proeedures are relatively insensitive to slight departures
from nonnality, In general. the t- and F-tests and t eonfidenee intervals will have aetuallev ...
ets of signifieance or eonfidenee levels that differ from the nominal or advertised levels
chosen hy the experimenter, although the difference between the actual and advertised lev-
els is usually fairly small when the underlying population is not too different from the nor-
mal distribution. Traditionally. we have called these procedures parametric methods
beeause they are based on a partieular parametric family of distributions-in this case, the
nonna!. Alternatively, sometimes we say that these procedures are not distnbution free
because they depend on the assumption of normality.
In this chapter we describe procedures ealled nonparametric or distributionwfree meth-
ods and usually make no assumptions about the distribution of the underlying population;
other than that itis eontinuous. These procedures have actual levels of s:ignifieanee a or con-
fidence levels 100(1- a)% for many different types of distributions. These procedures also
have considerable appeal. One of their advantages is that the data need not be quantitative; it
could be categorical (such a..,.;; yes orno~ defective or nondefective. etc.) or rank data. Another
advantage is that l10nparametrie procedures are usually very quick and easy to perform.
The procedures deseribed in this chapter are competitors of the parametric t- and
F-procedUIes describ~d earlier. Consequently~ it is important to compare the performance
ofbotb parametric and nonparametric methods under the assumptions of both normal and
nonnormal populations. In general, nonparametrie procedUIes do not utilize all the infor-
mation provided by the sample) and as a result a nonparametric procedure will be less effi-
cient than the corresponding parametric procedure when the underlying population is
norma1. This loss of efficiency usually is reflected by a requirement for a larger sample size
for the nonparametric procedure than would be required by the parametric procedure in
order to achieve the same probability of type II error. On the other band. this loss of effi·
ciency is usually not large, and often the difference in sample size is very smalL When the
underlying distributions are not nonnal, then nonp,arametric methods have much to offer.
They often provide considerable improvement over the nOrmal-theory parametric methods.
491
492 Cbapter 16 Nonpa..'"3J:uetric Statistics
probability is 0.5 that an observed value of X is greater than or equal to the median. That is.
P(X s,{l) = P(X?;fl) = 0.5.
Since the normal distribution is symmetric, the mean of a normal distribution equals
the median. Therefore the sign test can be used to test hypotheses about the mean of a nor~
mal distribution. This is the same problem for which we used the I-test in Chapter 11. We
will discuss the relative merits of the two procedures iu Section 16-2.4. Note that while the
t~test was designed for samples from a normal distribution. the sign test is appropriate for
samples from any continuous distribution. Thus, the sign test is a nonparametric procedure.
Suppose that the hypotheses are
Ho: {l =/10.
(16-1)
H,: {l "/10.
The test procedure is as follows. Suppose that Xl' X2• •••• Xtl is a random sample of n. obser-
vations from the population of mterest. Form the differences (Xi p,;J, i = 1, 2, .. ,' n. Now
if Ho: {l=/1o is true, any differenceX,-/1o is equally likely to be positive or negative. There-
fore let R' denote the number of these differences (X, - p,;; that are positive and let K denote
the number of these differences that are negative. where R;;: min CRt, R-).
When the null bypothesis is true, R has a binomial distribution with parameters n and
p = 0.5. Therefore. we would find a critical value. say R:
from the binomial distribution that
ensures that P(type I erroD = P(reject Ho whenHo is true) = a. A table of these critical val-
uesR;is given in the Appendix. Table X. If the test statistic R:f R:. then the null hypothe-
sis Ho: P. = /10 should be rejeeted.
~1ontgo:mery, Peck. and Vining (2001) report on a study in which a rocker: motor is formed by bind-
L"lg an igniter propellant and -a sustainer propellant together inside -a metal housing. The shear strength
of the bond between j}e two propellant types is an important characteristic. Res'..llts of testing 20 ran~
comly selected motors are shown in Table 16-1. We would like to test the hypothesis that me median
shear strength i'l 2000 psi.
Tne formal statement of the hypotheses of interest is
H,: fl = 2000,
,a
esis of ::=. 2000 psi.
r
I
16-2 The Sigr. Test 493
:r.ble16-1 Propellant Shear Streng&. Da:a
Observation (i) Shear Strength (X,) Differences (X, - 2000) Sign
2158.70 +158,70
2 1678.15 -32L85
3 2316,00 +316,00
4 2061.30 +{;L30
5 2207,50 +207,50 +
6 1708.30 -291.70
7 1784,70 -215.30
8 2575,10 +575.00 +
9 2357.90 +357,90 +
10 2256.70 +256,70 +
11 2165.20 +165.20 +
12 2399.55 +399.55
13 1779.80 -220.20
14 2336,75 +336,75
15 1765.30 ··234,70
16 2053.50 +53.50 +
17 2414.40 +414.40
18 2200,50 +20050 .,.
19 2654,20 +654.20 +
20 1753,70 -246.30
Exact Significance Levels When a test statistic has a discrete distribution 1 such as R does
R;
.in the sign test, it may be impossible to choose a critical value that has a level of signif-
icance exactly equal to a. The usual approach is to choose R:to yield an a as close to the
advertised level a as possible.
Ties in the Sign Test Since the underlying population is assumed to be continuous, it is
theoretically impossible to find a " that is, a value of XI exactly equal to Po- However)
this may sometimes .happen in practice because of the way the dzta are collected. \Vhen
ties occur, they should be set aside aJ:,d the sign test applied to the remaining data.
One~Sided Alternative Hypotheses We can also use the sign test wben a one-sided alter-
native bypothesis is appropriate. If the alternative is H,: p > p", then reject Ho: p= p" if
R- <R~ if the alternative is H,: p <p", then rejectHo: P = p" if R" <R~ Tne level of signif-
icance of a one-sided test is one-half the value sbo,," in the Appendix, Table X, It is also
possible to calculate. P-value from the binomial distribution for the one-sided case.
The Normal Approximation When p 0.5, the binomial distribution is well approxi-
mated by a norma! distribution when n is at least 10. Thus, since the mean of the binomial
is np a.'1d the vari""ce is np{l- p), the distribution of R is approximately normal, with mean
O.5n and variance 0.25n wbenever n is moderately large. Therefore in these cases the null
l
The two-sided alternative would be rejected if 1201:> Zoo' and the critical regions of the one-
sided alternative would be chosen to reflect the sense of the alternative (if the alternative is
H,: f1 :>f1c, rejectH, if 20 > Z", for example).
Exampk16·2
An autor.o.otive engincer is investigating two different types of metering devices for an electronic :bel
injection system to determine if they differ in their fuel.mi.!eage performance. The system is installed
on 12 different cars, and a tes£is run with each meterslg system on each car, The observed fuel mileage
performance data, corresponding differences. ar.d their signs are show!:. in Table 16~2. Note that K:::: 8
andR-;4. Thm::foreR = min (K,R1 = min (8, 4) =4. From the Appendix, Table X, with n; 12, we
:find the criticlli va.:'.1e for a= 0.05 is~.O$ = 2. SinceR is not less than the critical value ~.(jj' we cannot
reject the null hypothesis that the two metering devices produce the same fuel mileage performa.nee.
tiveness is the value of Pfor departures that are important. A small value of Pimplies an
effective test procedure.
In determining p, it is important to realize that not only must a particular value of Jl, say
A:, + Ll., be used, also the form of the underlying distribution will affect the calculations, To
illustrate, suppose that the underlying distribution is nonnal with 0"= I and we are testing
the hypothesis that {1 = 2 (since P. = f.1 in the normal distribution this is equivalent to testing
that the mean equals 2), It is important to detect a departure from{1= 2 to{1= 3, The situa-
tion is illustrated graphically in Fig, 16-10, ,,'hen the alternative hypothesis is !rue (H,: {1 =
3), the probability that the random variable X exceeds the value 2 is
p =P(X> 2) = P(Z>-I) = 1- 4>(-1) = 0.8413.
Suppose we have taken samples of size 12, At the a = 0,05 level, Appendix Table X indi-
cates that we would reject Ho: {1 = 2 if R ,; R;,,, = 2. Therefore, the Perror is the probabil-
ity that we do not reject No: {1 = 2 when in fact {1 = 3, or
P
'
'2 1 12
1- ~l x (0.1 587)"(O.8413)lt- = 0,2944,
x
J
If the distribution of X has been exponential rather than nonnal, then the situation
would be as shown in Fig. 16-1b, and the probability that the random variable X exceeds the
value.x : :;:; 2 when fl 3 (note when the median of an exponential distribution is 3 the mean
is 433) is
1 2
p=P(X>2) r~_1_;4,33X dx=;<,33 =0.6301.
h 4.33
-1 0 2 3 4 5 6
Under He: il =2 Under ~;;:;'=3
(a)
J(,=2 J.L-2.89
Figure 1&.1 Calcu!ation of {3for the sign test (a) Normal distributions, (b) exponential
distributions.
496 Chzpter 16 Nonparametric Statistics
x={)\ ..
l1ms, the /3 error for the sign test depends not only on the alternative value of {l but on
the area to the right of the value specified in the null hypo thesis under the population prob-
ability distribution. This area is highly dependent on the sbape of that particular probabil·
ity distribution.
2F;iiUli~1~~~;1:
To illustrate the Wucoxon signed rank test. consider the propellant shear strength data presented in
Table 16-1. The signed ranks are
16-3 The WIlcoxon Signed Rank Test 497
The sum of the positive:ranks is R" (-1 +2 + 3 +4 + 5 + 6...:.11 + 13.,.. 15 + 16 +:7 + 18 + 19-
20) =: 150 and the sum of the nega:ive ranks is lC::::. (7 + 8 + 9 + 10 + 12 + 14) = 60. Therefore R:::.
nilil (P:', R') = nilil (150, 60) = 60. Proll: Appendix :able Xl, with n = 20 and a= 0.05, we find the
critical value R~.J1 = 52. Since R exceeds R;, we cannot reject the null uypothesls that the mean (or
median. since the popUlations are assumed to be symmetric) shear strength is 2000 psi,
R n(n+ 1)/4
Zo == ---;===:':;=.==::.00-. (16-3)
..,jn(n + 1)(2n + 1)/24
An appropriate critical region can be chosen from a table of the standard normal
distribution.
Consider th~ fuel metering device data examined in Example 16~2. The signed :an.ks are shown
below.
7 -0.2 -1
12 0.3 2
8 0.4 3
6 0.5 4
2 -0.6 -5
4 0.7 6.5
5 -0.7 -6.5
1 0.8 8
9 0.9 9
10 -1.0 -10
11 !.l !1
3 1.3 12
Note that K= 55.5 and It =: 22.5; therefore.R=min (Ir"t10 = min (55.5. 22.5) 22.5. FromAppcn-
dix Table Xl. with It =: 12 and (1,= 0,05, we find::he critical value ~,iJ) = 13. Since R exceeds (,1)$' we
cannot reject tbe null hypothesis that the two metering devices produce the same mileage
performance.
164 The Wilcoxon Rank-Sum Test 499
p.,. The distributions of X, and X, have the same shape and spread and differ only (possibly)
in their means. The Wilcoxon rank-sur:! test can be used to test the hypothesis flo: 11, = f.L:.
Sometimes this procedure is called the Mann-Whitney test, although the Mallll-Vihitney
test statistic is usually expressed in a different form.
R:
Appendix Table IX contains the critical values of the rank sums for 0: = 0.05 and a
0.0 I. Refer to Appendix Table IX, with the appropriate sample sizes n, and Ilz. The null
hypothesis Ho: Ill;;;;.~ is rejected in favor of HI: J1 1-t: J.L;.if either R: or R'{. is less than or equal
to the tabulated critical value R~
The procedure can also be used for one-sided alcematives. If the alternative is H j : J.lt <
/1), then reject He if R] :;; R~ while for He: #1 > fJ..z1 reject 8 0 if Rz s: R~ For these one~sided
tests the tabulated critical values R;correspond to levels of significance of a::::: 0.025 and
0:'" 0.005.
Exartlple.16-S
The mean axial stress in tensile members used in an aircraft st:ucture is bei:1g studied. Two. alloys are
being investigated. Alloy I is a traditio.nal material and a:loy 2 is a new aluminum-lithium tilloy that
is much lighter than the standard materiaL Ten specimens o.f each alloy type are tested, and the axial
stress meas'..lX""...d. The sample data are assembled in the following table:
Alloy 1 Alloy 2
3238 psi 3254 psi 3261 psi 3248 psi
319-5 3229 3187 3215
3246 3225 3209 3226
3190 3217 3212 3240
3204 3241 3258 3234
2 3187 psi
3190 2
3195 3
3204 4
2 3209 5
2 3212 6
2 3215 7
3217 g
! 3225 9
2 3226 10
3229 11
2 3234 :2
1 3238 13
2 3240 14
3241 15
3246 :6
2 3248 17
3254 18
2 3258 19
2 3261 20
r 16-5 Nor-parametric Methods in the Analysis ofVa.'iauce 501
and \'ariance
, n,":1(n,+,,:!+I)
OR!;;:: 12 .
In this model the error tenns E., are assumed to be normally and independently distributed
with mean zero and variance '0". The assumption of normality led directly to the F-test
described in Chapter 12. The Kruskal-Wallis test is a nonparametric alternative to the
F-test; it requires only that the Eli have the same continuous distribution for all treatments
j= I, 2, .. " D,
Suppose that N .:;;: L~ln! is the total number of observations. Rank all N observations
from smallest to largest and assign the smallest observation rank 1, the next smallest rank.
2, , .. , and the largest observation ranklY. If the null hypothesis
Ho: 1': = }l, = ... 1',
is true. theN obsen'ations come from the same distribution, and all possible assignments of
the 11' ranl:'s to the a s"'!Iple, are equally likely, then we would expect the ranl:'s I, 2, ... , 11'
to be mi'ICed throughout the a samples, If, however, the null hypothesis Ho is false, then
some samples will consist of observations having predominantly small ranks while other
samples will consist of observations having predominantly large ranks, Let R'j be the rank
of observation Y;" and let R;+ and denote the total and average of the ni ranks in the ith
treatment. When'the null hypothesis is true, then
and
The Kruskal-Wallis test statistic lI1.eaSuteS the degree to which the actual observed average
ranks RI , differ from their expected value (N + 1)/2.lfthis difference is large, then the oull
hypothesis Ho is rejected. The test statistic is
12 " (- 11'+1)'
K
N(N+1)~ n,\"R - -2 - '
(16-7)
K= 12 ±n,
N(N + I) i~,
11,2 -3(11'+1). (16-8)
We would usually prefer equation 16-8 to equation 16-7, as it involves the rank totals rather
than the averages.
The null hypothesis He should be rejected if the sample data generate a large value fOr
K. The null distribution for K has been obtained by using the fact that under He each possi-
ble assignment of mlks to the a treatments is equaliy likely, Thus we could enumerate all
possible assignments and count the number of times each value of K occurs. This has led to
tables of the critical values of K, although most tables are restricted to small sample sizes
nf' In practice. We usually employ the following large-sample approximation: Vlhenever He
is true and ei ther
a= 3 and n,2:6 for i = 1, 2, 3
or
for i:::: 1, 2, ... , G,
16-5 Nonparametric Met.100s IT. the Analysis of Variance 503
then K has approximately a chi-square distribution with a-I degrees of freedom. Since
large values of K imply that Ho is false, we would reject He if
K?!i~'_l'
The test has approximate significance level a.
Tu;s in the Kruskal- ..Vallis Test When observations are tied, assign an average rank to
each of the tied observations. \\'1len there are ties, we shQuld replace the test statistic in
equation l6~8 with
(16-9)
where ni is the number of observations in the itb treatment, N is the total number of obser~
vatioM.and
52 = _.I_
C
Note that :f is just the variance of the ranks. When the number of ties is moderate, there
will be little difference between equations 16-8 and 16-9, and the simpler form (equation
16-8) may be used.
In Design and Analysis oj Experiments, 5t.'l Edition (John Weey & SODS, 2(01), 0, C, Montgomery
presents data from an experiment in which five different levels of cotton content in a syn~etic Eber
were tested to daermine if eottOn content has any effect on fiber tensile strength. The sa."llple data and
ranks. from this experimer.t are shown in Table 1&-3. Since :here is a fairly large number of ties, we
use equation 16~9 as the test statistic. From equation 16~ 10 we fud
S2 =~f-iiili~ _N(N+l)']
f; -l~i"'l J""l 4
= 1..[5497.79- 25(26)'
24 4 j
1
~53.03.
Table 16-3 Data and Ran.<.s for :he Tensile Testing Experiment
Percentage of Cotton
IS 20 25 30 35
=_l_r5245.0- 25(26)']
53.03, 4
=19.25.
Si...."lce K> x~,c,'.; ::::;; 13.28, we would reject the nullhypoiliesis andcondude that treatnents differ. This
is the same conclusion given by the usual analysis of variance Fwtest.
16-6 S~1ARY
16-7 EXERCISES
16~1. Ten samples were taken Emma plating bath used the sig:c test to determine whether or not the CWO tips
in an electronics manufacturing process and the bath produce equivalent hardness readings.
pH determined. The sample pH values are given
below: Coupon Tip 1 Tip 2
16 ..3. The distribution time between a..-'rivals in a 4.00 (mm), 4,02, 4.03. 4.01, 4.00, 4.03,
telecommt."Dicauon system is ex-ponential, and the sys- 4.()4, 4.02, 4.03, 4.03.
tem man.ager wishes to test the hypothesis that HG: p.;::.. (a) Suppose that p is the probability that observation
3.5lIlin versus HI: p. > 3.5 min, Xi+5 exceeds observation X" If t.1.ere is no upward
(a) ' ' hat is the value of the mean of the exponential or downward trend. then X;"'5 is no more or Iess
likely to exceed XI or lie below Xt VVhat is the
distribution under Hij: fl.:::. 3.5?
value ofp?
(0) Suppose that we have taken a sample of It:::::. 10
observations and we observe R-;:;; 3, Would the (0) Let V be :he number of values of i for which
sig:1 test reject Ho at a= 0,05? Xi+, > Xr If there is no upward or dov,llward trend
in the measurements, what is the probability dis-
(c) What is the type n error of this test ifjl =4,57
tribution of V?
164. Suppose that we take a sample of n;:;; 10 meas- (c) Use t!:.e data above and the results of parts (a) and
urementS from a normal distributio!l with (1 == 1, We
°
wish to test Ho: J1. "'" against HI: J1. > O. The no:rrn.al
(b) to test He: there is :10 trend versus HI: there is
i!pwatd trend. Use a"",O.05.
test statistic is ~ =(x - #0)/ a/-rn , and we decide Note that this test is a modification of the sign test. It
to use a critical region of 1.96 (:.hat is, rejectRc if4 2: Was dc·...eloped by Cox and S:uart.
1.96). 16..7. Consider the Wilcoxon signed rank test, and
(a) 'What is ex for this test? suppose :hat n == 5. Assume ~t Ho: j.J. = ,l1t.J is trJe.
(b) What is j1fortlllstestif,u= I? (a) How many different sequences of signed r;mks
(c) If a sign tes~ is used, specify the critical region are pOSsible? Errt!rnerate these sequences.
that gives an a value consiste::lt with a for the nor- (b) How ma.1.Y different 'values of R" are there? Find
:nal test the probability tlSsociated with each value of R".
(d) What is the t> value for the sign test if fJ. == 1? {c) Suppose that we define the critical region of the
Compare this with the :esult ob~ed in part (b). test to be R:
such that we would reject if R"'> R:,
16--5. TWo different types of tips can be used in a and R~;::::: 13, What is the approximate a level of
Rockwell hardness tester. Eight COUP0!lS from test this test?
ingots of a nickel-based alloy are selected, and each (d) Can you see from this exerc:se how the critical
coupon is tested twice, once with each tip. The Rock- 'values for the Vlilcoxon signed xank test were
well C-sca1e hardness readings are shown next Use deve,oped? Explain.
506 Chapter 16 Nonparametric Statistics
16..&. Consider:he data in Exercise 16-1. and asstllXte Unit I 25, 27, 29, 31, 30, 26, 24, 32, 33, 38.
that the distribution of pH is symmetric and continu~
ous, Use the Wilcoxon signed rank test to test the
Unit 2 31, 33, 32, 35, 34, 29, 38, 35, 37, 30,
hyp<:1thesis Ho: ,u == 7 against H!; /1 '# 7.
16-16. In Design and h..alysis 0/ Experiments, 5th
16~9. Consider the data in Exercise 16-2, Suppose Edition (John Wiley & Sons, 2001), D. C. Mont-
that the distribution of titanium content is symmetric gomery presents the results of an experiment to com-
and cont:ir.uous. Use the Wilcoxon signed rank test to pare four different :n;.,iting tech.rUques: on the tensile
test the hypotheses Ho: J.l;:;::! 8.5 versus H;: J.l #; &.5. strength of portland cement, The results are sho\Vn
16~10" Consider t.1e data in Ex.ercise 16-2. Use the below. Is t.~ere any indication that rnixi:ng !echnique
large-sample approximation for the Wilcoxon signed affects t..lJ.e st:ength?
rank: test to test :he hypotheses Ho: J1::::: 8.5 Versus H,;
J1 ~ 8,5. Assume that :he distribution of titanium con- :Mixing Technique Tensile St:'ength (lb i In.')
tent is continuo'.)s and symmetric,
Hi-ll. For the large-sample approximation to the 1 3129 3000 2865 2890
Wilcoxon signed rank test, derive the mean and stan- 2 3200 3000 2975 3150
dard deviation of the test statistic used in t!1e 3 2800 2900 2985 3050
procedtlte. 4 2600 2700 2600 2765
16~12. Consider the Rockwell hardness test data in
Exercise 16-5. Assume Li.a: bot."1 distributions are 16~17. An a.rticle in the Quality Control Handbook,
continuous and use ll-)e \Vilcoxon signed rank test to Srd Edition (McGraw~Hill. 1962) presents the results
rest that the mear:. difference iL hardr,ess readings of an experiment perfonned to investigate t!le effect of
between the two tips is zero, th..'"ee differer.t conditioning methods on the breal:ir.g
16-13. An electrical engineer must design a circuit to strength of cement briquettes. Tne data are shown
below, Is there any indication that conditioning
deliver the r.:l:aximum amount of current to a display
rube to achieve sufficient image brightness. Within method affects breaking strength?
rJs allowable design constraints. he has developed
t'NO candidate circuits and tests prototypes of each. Conditioning
The resulting data (in microamperes) is shown below: Method Brealdng (Ib i in,')
Circuit 1: 25:, 255, 258, 257. 250, 251, 254, 2S0, 248 1 553 550 568 541 537
Circuit 2: 250, 253, 249, 256, 259, 252, 250, 251 2 553 599 579 545 540
3 492 530 528 510 571
, Use ;he Wilcoxon r.m.\<;-sum test 1.0 test Ho: J1; = ,u1
against the alternative HI; J11 > ,u1>
16~1S. In Statistics/or Research (John Wiley & Sons,
16-14. A consultant frequently travels from Phoenix, 1983), S. Dowdy and $, We3rden present the results of
Arizona, to Los Angeles. California, He will use one an experiment to measure stress resulting from oper-
of two airlir:es, Uni:ed or Southwest. The number ating hand-held chain saws. The experimenters meas-
of minutes that his flight arrived late for the last ured the kickback angle through which the saw is
six trips on each airlbe is sbo\Vll belOW. Is there evi- deflected when it begins to cut a 3-inch stock syn-
dence that either airli,'le has superior on-time arrival thetic board. Shown below are deflection angles for
perfon:l3!lce'1 five saws chosen at random from each of four differ~
eut manufacturers. Is there any evidence that the man-
United 2 19, -2 8 0 (minutes late)
ufacturers' products Ciffcr with respect to kickback
Southwest 20 4 8 8 -3 5 (minu:es late) ar:gle?
The quality of the products and services used by our society has become a major consumer
decision factor in many, if not most, businesses today. Regardless of whether the consumer
is an individual, a corporation. a military defense program, or a retail store, the consumer
is likely to consider quality of equal importance to cost and schedule. Consequently, qual-
ity improvement has become a major concern of many U.S. corporations. This chapter is
about statistieal quality control and reliability engineering methods, two sets of tools that
are essential in quality~improvement activities.
507
508 Chapter 17 Statistical Quality Control and Reliability Engineering
Statistical methods playa vital role in quality improvement. Some applications include
the following:
1. In product design and development, statistical methods, including designed exper-
iments, can be used to compare d.i.f.te:rent materials and different components or
ingredients1 and to help in both system and component tolerance determination.
This can significantly lower development costs and reduce development time.
2. Statistical methods can be used to determine the capability of a manufacturing
process. Statistical process control can be used to systematically improve a process
by reduction of variability,
3. Experiment design methods can be used to investigate improvements in the process,
These improvements can lead to higher yields and lower manufacturing costs.
4. Life testing provides reliability and other performance data about the product. This
can lead to new and improved designs and products that have longer useful lives and
lower operating and maintenance costs.
Some of these applications have been illustrated in earlier chapters of this book. It is essen-
tial that engineers and managers have an in-depth understanding of these statistical tools in
any industry or business that wants to be the bigh-quality, low-cost producer, In this chap.
ter we give an introduction to the basic methods of statistical quality control and reliability
engineering that, along with experimental design, form the basis of a successful quality~
improvement effolt..
of operating with little variability around the target .or nominal dimension. Online statisti ~
cal process controls are powerful tools useful in achieving process stability and improving
capability through the reduction of variability.
It is customary to think of statistical process control (SPC) as a set of problem-solving
tools that may be applied to any process. The major IDols of SPC are the following:
1. Histogram
2. Pareto chart
3. Cause-and-effect diagram
4. Defect~concentration diagram
5. Control chart
6. Scatter diagram
7. Check sheet
While these tools are an important part of SPC, they really constitute only the technical
aspect of the subject. SPC is an attitude-a desire of all individuals in the organization for
continuous improvement in quality and productivity by the systematic reduction of vari~
ability. The control chart is the most powerful of the SPC tools. We now give an introduc-
tion to several basic types of control charts.
ISoo::.etimes contnum c:J.use is used.insre:ld of "random" or '"'chAnce cause." and special cause is used ifIstead of
"assignab:e cause,"
510 Chapter 17 Statistical QUality Control and Reliability Engineering
We distinguisb between control charts for measurements and control charts for attrib-
utes, depending on whether the observations on the quality characteristic are measurements
or enumeration data. For example, we may choose to measure the diameter of a shaft, say
with a micrometer, and utilize these data in conjunction with a control chart for measure-
ments. On the oilier hand, we may judge each unit of product as either defective or nonde-
fe;::tive and use the fraction of defective units found or the total number of defects in
conjunction with a control chart for attributes. Obviously. certain products and quality char-
acteristics lend themselves to analysis by either method. and a clear-cut choice between the
two methods may be difficult.
A control chart, whetberfor measurements or attributes, consists of a cemerline, corre-
sponding to the average quallty at which the process should perform when statistical "ontrol
is exhibited, and two control U",jts. called the upper and lower control limits (UCL and
LCL). A typical control chartls shown in Fig. 17-1. The control limits are chosen so that val-
ues falling between them can he attributed to chance variation, while values falling beyond
them can be taken to indicate a lack of statistical conrroL The general approach consists of
periodically taking a random sample from the process, computing some appropriate quan-
tity, and plotting that quantity on the control chart. When a sample value falls outside the
control limits, we search for some assignable cause of variation. However, even if a sample
value falls between the conrrollimits~ a trend Or some other systematic pattern may indicate
that some action is necessary, usually to avoid more serious trouble. The samples should be
selected in such a way that each sample is as homogeneous as possible and at the same time
maximizes the oppo:rturUty for variation due to an assignable cause to be present. This is usu-
ally called the ratiolllll subgroup concept. Order of production and source (if more than one
source exists) are commonly used bases for-obtaining rational subgroups.
The ability to interpret control charts accurately is usually acquired with experience. It
is necessary that the user he thoroughly familiar with both the statistical foundation of con-
trol charts and the nature of the production process itself.
Sample number
standard deviation chart, depending on how the population standard deviation is estImated.
We will discuss only the R chart.
Suppose that the process mean and standard deviation, say p. and a. are known, and,
furthermore, that we can assume that the quality characteristic follows the normal distribu-
tion. LetXbe the sample mean based on a random sample of size n from this process. Then
the probability is 1 - IX that the mean of such random samples w,n fall between
f.l+Za/2{a/.Ji;) and I1-Z<l/2(a/..;;;). Therefore. we could use these two values as the
upper and lower control limits, respectively. However, we usually do not know f.1 and a and
they must be estimated. In addition, we may not be able to make the normality assumption.
For these reasons, the probability limit 1 - a is seldom used in practice. Usually Zan is
replaCed by 3, and "3.sigma" control limits are used.
'Wben J1. and a are unknown. we often estimate them On the basis of preliminary sam~
pIes. taken when the process is thought to be in control. We recommend the use of at least
20--25 preliminary samples. Suppose k preliminary samples are a;"Ililable, each of size n.
Typically, n will be 4,5, 'Or 6; these relatively small samp:ie sizes are widely used and often
arise from the C'Onstruction of rational subgroups. Let the sample mean for the ith sample
be Xi. Then we estimate the mean of the population, j)., by the grand mean
- l'
X IL.X,.
1'=;
(17-1)
UCL = X
= ~-······F
3- R.
d.2~n .
(17-4)
3
LCL=X--:-;=J:(
and we would use as the upper and lower control limits on the R chart
UCL R+ 30"w. R
d2 '
(17-6)
LCL=R- 30"w R.
d,
Setting D3 = 1- 30"w1d2 and D4 := 1 + 30"w1d2' we may rewrite equation 17-6 as
UCL =D4R, (17-7)
LCL=D,R,
where D, and D4 are tabulated in Table XIII of the Appendix.
When preliminary samples are used to construct limits for control charts, it is cus~
tomary to treat these limits as trial values. Therefore, the k sample means and ranges
should be plotted on the appropriate charts, and any points that exceed the control limits
should be investigated. If assignable causes for these points are discovered, they should be
eliminated and new limits for the·contrel charts determined. fn this way, the process may
eventually be brought into statistical control and its inherent capabilities assessed. Other
changes in process centering and dispersion may then be contemplated,
Exaxnpie17,1.
A component pa..rt for a jet aircraft engine is manufactured by ax:. investment casting process, The vane
opening on this casting is an impor.ant functiOIlal Parar::le:er of the part, We will illustrate the use of
X&.,'ld R control cha."tS to assess the statistical stability of this process. Table 17-1 presents 20 samples
of five parts each. The values given in the table have been coded by using the last three digits of the
dimension; that is. 31.6 should be 0.50316 inch.
The quantities X:= 33.33 a:;;.d R "'" 5.85 are shown at the foot of Table 17-L Notice that even
though X, X, R, and R are now realizations of random variables, we have still \V:.i.tten them as
17·3 Statistica: Process Control 513
X=33.33 R=5,85
uppercase letters, This is the usual convention in quality control. and it will always be clear from the
context what the notation implies, The trial concrollimits are, for the Xcbart.
or
UCL = 36,70.
LCL=29,96,
For the R chart, the trial co:r:trollimits are
UCL =D,R = (2.115)(5,85) = 12.37,
LCL =D,R = (0)(5,85) = 0,
The X aodR cootrol charts .....1tb. these trial cootrollimit.,> are sho'NU in Fig, 17~2, Notice that sa.'":lples
6, S, 11. and 19 are out of cOlltrol on the X chart, and that sample 9" is out of control on the R chart.
Suppose that all of these assigr.ablc causes can be traced to a defective tool io the wax-molding area,
We should discard these !lve samples and recompute the limits for the X and R charnL These new
revised limits are, fur the X chart,
UCL= X+A,J< =32,90 + (0.577)(5.313) =35,96.
LCL = X 32,90 - (0,577)(5.313) = 29,84.,
LCL = (0)(5,067) = 0,
514 Chapter 17 Statistical Quality ContrQI and Reliability Engineering
c
iii 3S:-
I-"~-""- UCL= 36.70
E
o
Mean = 33.33
! J-----*--
io, ~1;;-O------;2~O
ILCL
29,96
Sample number
, - - - - - - - .... _ - - -
1S
UCL; 12.37
m
0> 10
C
~
.!!l
c.
E
M S
(j)
O~ LCL=O
10 20
Sample number
The revised control charts are shO'WTl in Fig. 17-3, };otice that we have treated the firs..: 20 preliminary
samples $S estimation. data with which to establish control limits. These limits can now be used. to
judge the statistical control offuturc production.. As each l1eW sample becomes available, the values
ofX' and R should be cotUputed and plotted on t.~e cOIltrol charts. It may be desirable to revise the lim-
hs periodically, even if the process rCJllJlins stable. The limits should alv.rays be revised when process
improvements are made.
§ 40r
1-
i UCL~35.96
S 35~
w
""g . L ~
"'~
f
c. "-
"
m
<I)
VS LCL=29.84
I \
0 10 20
Subgroup
15~
m
~
UCL; 10.71
"e 10 ~
'"
c. R;5.C67
:~
E
'"
(J)
LCL;C
0 10 20
Subgroup
Figure 17~3 X a.,.d R control charts fOT vane openi.1g, revised limits,
USL;40
40
•
I
"- 30
35
i
.
0
ID
C
ID
>
--\Nominal •
dirrension ;:: 30
25
LSL
20
Sample r:umber
Figure 17 ~4 Tolerance diagram of vane openings.
516 Chapter 17 Statistical Quality Contro: and Raliability Engineerillg
is appropriate to plot the specification limits on the tolerance chart, since it is a chart of indi~
vidual measurements, It is never appropriate to plot specijicationlimits on a control chart,
Or 10 use the specifications in determining the control limits, Specification limits and con~
trollimits are unrelated. FInally, note from Fig. 174 that the process is running off center
from the nonrillal dimension of 0.5030 inch.
The histogram for the vane opening measurements is shown in Fig. 17-5. The obser-
vations from samples 6, 8, 9, II, and 19 have been deleted from !his histogram. The gen·
era} impression from examining this histogram is that the process is capable of meeting the
specifications. but that it is running off center.
Another way to express process capability is in terms of the process capability ratiQ
(PCR), defined as
(J R =~=2.06.
d, 2.326
Therefore. an estimator for the peR is
USL-LSL
peR
60'
40-20
= 6(2.06)
1.62.
20,------ - - - - --------,
15
o ~__,__rH_r'Ty.,-'-r-H_rry.,-'-r-,,_,-rl,
18 20 22. 24 26 28 30 32 34 36 38 40 42 44
LSL t USL
Ncmina! dimension
Vane opening
17-3 Statistical Process Contral 517
The PCR has a natural intetpretation; (I!PCR)IOO is just the percentage of the toler-
ance band used by the process. Thus) the vane opening process uses approximately
(111.62)100 = 61.7% of the tolerance band.
Figure 17-6a shows a process for which the peR exceeds unity. Since the process nat-
ural tolerance limits lie inside the specifications. very few defective or nonconforming units
will be produced. If PCR= I, as shown in Fig. 17-6b, more nonconforming units result. In
fact, for a normally distributed process, if peR = 1, the fraction nonconforming Is 0.27%.
or 2700 parts per million. Finally, when the PCR is less than mity, as in Fig. 17-6<, the
process is very yield sensitive and a large number of nonconforming U.1.l3 will be produced.
The definition of the peR given in equation 17-8 implicitly assumes that the process
is centered at the nominal dimension, If th.e process is running off center, itS actual capa~
biliry will be less than that indicated by the peR. It is convenient to think of PCR as a meas-
ure of potential capability, that is, capability with a centered process. If the process is not
centered, then a measure of actUal capability is giVe!1 by
D • rUSL-X X-LSL1
P C"k = romi! 3<1 'J
3<1 " (17-9)
(al
PCR
NonconformIng Nonconforming
units ( unilS
LSL USL
(b)
Nonconforming
units
lSL
"
3<T_...;1~-3u-"'"
(e)
Figure 17.-6 Process fallout and the process capability ratio (peR).
518 Chapter 17 StatistiC!! Q~a1ity Cootrol and Reliability Engineering
lI
In effect, peRk is. a one-sided process capability ratio that is calculated relative to the spec-
ification limit nearest to the process mean. For the vane opening process, we find that
Note that if PCR = PCR, the process is centered at the nominal dimension. Since PCR,
1.1 0 for the vane opening process, and PCR = 1.62, the process is obviously running off
center, as was first noted in Figs. 17-4 and 17-5. This off~center operation was ultimately
traced to an oversized wax tOoL Changing the tooling resulted in a substantial improvement
in the process,
Montgomery (2001) provides guidelines on appropriate values of the peR and a table
relating fallout for a normally distributed process in statistical control as a function of peR.
Many U.S. companies USe PCR 1.33 as a minimum acceptable target and PCR = 1.66 as
a mini"llUIn target for strength. safety, or critical characteristics. Also, some U.s, compa M
rues, particularly the au;omobile industry, have adopted the Japanese terminology Cp =PCR
and Cpk ::;;: peRk.- As Cp has another meaning in statistics (in multiple regression; see Chap-
ter 15), we prefer the traditional notation PCR and PCR,.
For example, for m observations, m - I moving ranges are calculated as MR, lx" - XII,
MR3 = lx, - x,1, ... , MRm = Ixm x",-ll. Simultaneous control chans can be established on
the individual observations and on the moving range.
The control limits for the individuals control chart are calculated as
MR
UCL x+3
d2 '
Centerline:::: X, (17-10)
LCL - - ,MR
-x-'-::i;'
where MR is the saJ::1ple mean of the MR i•
17~3 Statisticul Process Control 519
If a moving range of size n = 2 is used, then d2 = 1.128 from Table XIII of the Appen-
dix. The control limits for the moving range control chart are
UCL = D4MR,
Centerline = MR, (17-11)
LCL=D,MR.
Batches of a particular chemical product are se:ected from a pro:::ess ar.d the purity measured on each,
Da:a for 15 successive bat:::hes have been collected mld are given in Table 17-2, The moving ranges
of size n = 2:are also displayed in Table 17~2,
To Set up the control chart for individuals. we ~t need the sample average of the 15 purity
measurements. This average is found to bex == 0357. The average of the moving ranges of two obser-
vations is MR ;::; 0,046. The control limits for the individuals cha...-rt with moving ;:angcs of size 2 using
me linritsin equation 17*lQ are
The control limits for:he moving.range cha.-rt a..--e found using the Emits giver. in Equation 17-:11:
UCL = 3.267(0.046) = 0.150,
Centerline = 0.046,
LCL = 0(0.046) O.
0.9r-======================lUCL=0.879
:·:t Mea.,::::: 0,757
L======::::::=======;::======~LCL=
0.6 0 .. 5 18 15 0.635
Subgroup
(a)
Figure 17~7 Control charts for (a) the individual observations and (b) the moving range ap pu..""ity.
The control charts for individual observations and for the moving range are provided in Fig. 11-1,
Since there are no pointz beyond the controlli..rojts, the process appears to be in statistical canuoI.
. The individuals chart can be interpreted much like the X control chan, An out~of...;;:ontrol situa~
rion would be indicated by either a poi.'u (or points) plotting beyond the controllimi-;s or a pattern
such as a run on one side of the centerline.
The moving range dwt cannot be i::tterpreted in the same way. Although a point (orp0lnts) plot~
ling beyond the control limits would likely indicate a.:l out-of-controlsituation.. a par.em or run on one
side of the centerline is not necessarily an indica:ion that the process is out of conn:oL This is due to
the fact that the moving ranges are correlated, and this cone1ation.lll3.Y natu1'3..lly cause patterns or
trends on the chart
SUFPose D is the number of defective units in a random sample of size n. We assume that
D is a binomial random variable with unkno'wn parameter p. Now the sample fraction
defective is an estimator of p, that is
D
(17-12)
n
Furthennore; the variance of the statistic p is
p(l- p)
n
so we may estimate a; as
';(1- P)
(17-13)
n
The centerline and control limits for the fraction~defective control chart may now be
easily determined. Suppose k preliminary samples are available, each of size n, andD; is the
number of defectives in the ith sample. Then we may take
(17-14)
as the upper and lower control limits. respectively. These control limits are based On the
normal approximation to the binomial distribution. "When p is sman. the normal approxi-
mation may not always be adequate. In such cases. it is best to use control lim.lts obtained
directly from a table ofbinornial probabilities or. perhaps; from the Poisson approximation
to the binomial distribution. If p is small, the lower control limit may be a negative number.
If this should occur, it is customazy to consider zero as the lower control limit.
;.~~I1I~iZ:~
Suppose we wish to construct a £raction~defectiye control char:: for a ceramic substrate pro-
duction line. We have 20 preliminary samples, each of size 100; the number of defectives
in each sample are shown in Table 17~3, Assume that the samples are numbered in the
= =
sequence of production. Note that Ii 80012000 OAO, and therefore the trial parameters
for the control chart are
Centerline =: 0.395
LCL=O.395_3!(0.395)(O.605) 0.2483.
"'~ 100
522 Chapter 17 Statistical QUality Control and Reliability Engineering
The control chart is shoVr'TI in Fig. 17~8. All samples are in control. If they were not, we
would search for assignable causes of variation and revise the limits accordingly,
Although this process exhibits statistical conlrol. its capability (,0 0.395) is very
poor, We should take appropriate steps to investigate the process to detenr.Jne why such a
large number of defective units are being produced. Defective units should be analyzed to
determine the specific types of defects present. Once the defeet types are known. process
changes should be investigated to determine their impact on defeet levels. Designed exper~
iments may be useful in this regard.
..E~!lI~i7~4
Attributes Versus Measuremenl. Control Charts
Tae advantage of measurement control charts :relarlve to the p chart 'With respect to size of sample may
be easily illustrated" Suppose that a DOnnally distributed quality characteristic has a sta:ndard devia-
tion of 4 a.."\d specification limits of 52 and 68. The process is centered at 60, which results iri. a frnc~
tiOD defective of 0.0454. Let the process mean shift to 56, ~ow the fraction defective is 0.1601. If the
~~ 0.35
~
"' 0.25 LCL= 0,2483
o 10 20
Sample number
probability of d::tecting the shi..~ on ';he first sample following the mift is to be 0.50. ';hen the sample
size mus~ be such tl::at ';he lower 3~sigma limit will be at 56. This implies
whose sOlution is It:::; 9, For a p chart, using the no~ approximation to the binomial, we must have
0,0454+3~(0,0454~0,9546) 0.160L
whose solution is n:::; 30. Thus. unless the cost of measurement inspection is more than three times ali
costly as the attributes inspection, the measurement control chart is cheaper to operate.
and
UCL;;;;;c+W. (17-17)
LCL;;;;;c -3-«
are the upper and lower control limits, respectively.
Printed circuit boards are assembled by a combina~on of manual assembly and automation. A flow
sold~r machine is used to make the mechanical and elecrrical connections of the leaded components
to the board. The boards are run thtough the flow solder process almost continuously. and every bour
five boards are selected a.,d i..'lSpectcd for process-control purposes, The number of defects .l:_ each
sample of five boards is noted. Results for 20 samples are shown in Table:'7-4. Now c = 160120;;;;; 8.
and t.tlerefore
UCL~8+3;S ~16,41!4.
LCL ~ S- 3-./8 < 0, set to 0,
F:om lite control chart i:o Fig. 17~9, we see that the process is in controL However, eight defects per
group of five printed circuit boards is too many (about 8/5;;;;; 1.6 defectslboord), and the process needs
improvement. An investigation needs to be made of the specific types of defects found on the printed
circuit boards. This will usually suggest POteJltial ave::mes for process improvement
524 Chapter 17 Statistical Quality Control and Reliability Engineering
20
UCL~ 16.484
c
r1
J£
w
"'0 10
;;;
.0
E
~
z
Sample number
Figure 17-9 The c chart for defeCts in samples of five prir.red circuit boards.
(l7-18)
(17-19)
17-3 Statistical Process Con""l 525
A u chart may be constructed for the printed cl.rcuir board defect data in Examp~e 17~5. Since each
sample contains r. = 5 printed circuit boards, the values of u for eacb sample may be calculated as
shown in the following display:
w
1:>
ill
"
1 r
!
i
0
0 10 20
Sample number
Figure 17,.10 The u chart of defects pe~UDit on printed circuit boards. Example 17 o.
w
Alternatives to Shewhart eontrol charts include the cumulative sum control chart and
the exponentially weighted moving average control chart. These control charts are more
sensitive to small shifts in the process because they incorporate i..'1formation from current
and recent past observations,
against the sample i. The quantity C; is the cumulative sum up to and including the ith sam~
pIe. As long as the process is in control at the target value flo, then Ci in equation 17-20 rep-
resents a random walk with a mean of zero. On the other hand. if the process shifts away
from the target mean, then either an upward or downward drift in Ci. will be evident. By
incorporating information from a sequence of observations, the CUSUM chart is able to
detect a small shi.fi: in the process more quickly than a standard Shewhart chan, The
CUSUM eharts Can be easily implemented for both subgroup data and individual observa-
tions, We will present the tabular CUSUM for indiyidual observations.
The tabular CUSUM involves tvlo statistics. c;- and C;, which are the accumulation of
deviations above and below the target mean., respectively. C; is called the one-sided upper
CUS'(rM and c:
is called the one-sided lower CUSUM, The
statistics are eomputed as follows:
(17-21)
(17-22)
'Nith initial values of C~ :;;:: CO =O. The eonstant, K. is referred to as the reference value and
is often chosen approximately halfway between the target mean, ,'4> and the out-of-control
17~3 Statistical Process Control 527
mean that we are interested in detecting, denoted Jl.l' In other words. K is half oftbe mag-
nitude of the shift from 110 10 )1.1' or
The Sllltistics given in equations 17-21 and 17·22 accumulate the deviations from target that
are larger than K and reset to zero when either qm.u:.tity becomes negative. The Cr..:SUM
control chart plots the values of C; and C; for each sample. If either statistic plots beyond
the decision interval. H, the process is considered out of controL We will discuss the choice
of H later in this ch,,!,!er, but a good rule of thumb is often II = 5a.
A study presented in Food Control {200t p. 119) gives the results of measuring the dry-matter cor-tent
in buttercream from a batch process, One goal of t.1e study is to monitor the amount of Cry matter from
batch to batch. Table 17-5 displays some data ':hatrnay be t']pical of this type of process. Th::: reported
values. Xi' are percentage of dry-ll.la!:ter content examinee after mixing. The target amount of d..ry~mat
ter cor-tent is 45% a:ld assnme that a= 0.&4%, Let us also assume t.1at we.are interested in detect:i:r.g a
shift in the process of mean of at least 10; that is, J1.1 =IJ;;+ 10'=45 + 1(0.84) =45.84%. We will use the
recommended decision i!'lterval of H = 5(j~ 5(0.84) = 4.2. The reference valu~ K. is found to be
The va!ues of C; and c-; are given ir. Table 17-5. To illustrate the calculations, consider the first two
sample batches, Recall that C;::::::. CO = 0, and using equations 17-21 and 17~22 ",itb. K:::::; 0.42 and
J.lo : : :; 45, we have
c: ~=(O,,,,-(45 M2)+C;J
",,-maxlO,xj ~45,A.2't"C;]
and
C: =max[O, (45 -0.42) -xl + C;i
=max(O,4458-x, +CQJ,
Fcrbatth l,x\ =46,21,
=max[O, 4<5.21-45,42 +OJ
~ maz[O, 0,79]
=0,79,
and
C; =max[O, 4458 - 4<5.21 + OJ
= max[O, -L63J
=0,
For batch 2,-<" = 45.73,
c; = l.l3.X[O, 45.73 - 45.42 + 0.79]
= max[O, 1.10]
= 1.10,
and
C; = max[O, 44.58 - 45.73+ 0]
:: max[O, -1.15J
=0.
The CUSL"'M calcJlations give::. in Table 17~5 indicate rhat the upper-sided CUSUM for batch 17 is
C77 = 4.46, which CAceeds the dedsion value of H 4,2. Therefore, the process appears to have
shifted. out of concrol. The CUSUM status chart created using ~1initab® withH =4.2 is given in Fig.
17-1 L The out-of-control control sh:uation is also evident On this cha.'t at batch 17.
The CliSUM control chart is a powerl'uJ quality tool lor detecting a process that has
shifted from the target process mean, The correct choices of H and K can greatly improve
the sensitivity of the control chart while protecting against t.'1e occurrence of false alarms
(the process is actually in control, but the control chart signals out of con~ol). Design rec~
ommendations for the CliSUM will be provided later in this chapter whcn the concept of
average run length is introduced,
17~3 St<ltistical Process Control 529
6~ Upper CUSUM
:~----------------------~r-----+-+---
3~
~ 2~
.r
.~
~
1;-
0
,
"'--~-""-~-"-"";o;;,.----.......
B -1~ "J.~""~~;/'~s..v../ ~'~~ :!",~ .. .,j.~~
-2 ~ ' I:
t
-3 '-
~4 ! Lower CUSU;v,
! ! I
1-42
o 5 10 15 20 25
Subgrouo nurr:ber
We bave presented the upper and lower CUSti'M control charts for situations in which
a shift in either direction away froe the process target is of interest. There are many
instances w ben we may be interested in a shift IT. only one direction, either upward or
dov.nward, One-sided CUSUM charts can be constructed for these situations. For a thor-
ough development of these charts and more details, see Montgomery (2001),
(17-23)
where Ais a weight, 0 < A';; 1. The procedure to be presented is initialized withZo = Po, the
process target mean, If a target mean is unknown, then the average of preliminary data. x,
is used as the initial value of the EWlYfA. The definitior: given :in equation 17-23 demon-
strates that information from past observations is incorporated into the current value of z/.
The value Zi is a weighted average of all previous sample means. To illustrate, we can
replace Zl~l on the right-hand side of equation 17~23 to obtain
i-I
Z; = A2::(1- Ai X;_J +(1- A)i,,".
530 Chapter 17 Statistical Quality Control a=.d Reliability Engineering
The EWMA can be thought of as a weighted average of all past and current observations.
Note that the weights decrease geometrically with the age of the observation. giving less
weight to observations that occurred early in the process. The E\v:MA is often used in fore-
casting. but the EWMA control chart bas been used extensively for monitoring many types
of processes.
!fme observations are independent random variables with variance <f, the variance of
the E'\V"NIA, Zi' is
Given a target mean, Po, and the variance of the EVlMA, the upper control limit! centerline,
and lower control limit for the Em1.A control chart are
UCL
Centerline = il<J,
I(
LCL=.Uo-L<1~ 2~;y-(1-At',
A '[ "'J
where L is the width of the comrollimits. Note that the term 1 (1 - ).)2.i approaches 1 as i
increases. Therefore, as the process continues running. the control limits for the E\V.M.1\.
approach the steady srate values
(17-24)
Although the control limits given in equation 17 -24 provide good approximations, it is rec-
ommended that the exact limits be used for smalJ values of i.
F!x:lrnp", i '7-8
We will now imp!cmenr the EVlMA control chart with 1;::< 0.2 and L;;: 2.7 for the dry-matter content
data provided in Table 17-5. Recall that the target mean.is Ji.J;; 45% and the process standard devia-
tion is assumed to be <5= 0.&4%. Tae EWMA calculations are provided in Thble 17-6. To demonsmtte
some of the calculations, consider the first observ:ation with XI "'" 46,21. We find
=45.34.
17,3 Statistical Process Control 531
Tablel7•• EWMA Calculations for Example 17-8
Batch, i XI UCL LCL
46.21 45.24 45.45 44.55
2 45.73 45.34 45.58 44.42
3 44.37 45.15 45.65 44.35
4 44.19 4.95 45.69 44.31
5 43.73 44.71 45.71 44.29
6 45.66 44.90 45.73 44.27
7 44.24 44.77 45.74 44.26
8 44.48 44.71 45.75 44.25
9 46.04 4.98 45.75 44.25
10 44.04 44.79 45.75 4.25
11 42.96 44.42 45.75 44.25
12 46.02 44.74 45.75 44.25
13 44.83 44.76 45.75 44.25
14 45.02 4.81 45.76 44.24
15 45.77 45.00 45.76 44.24
16 47.40 45.48 45.76 44.24
17 47.55 45.90 45.76 44.24
18 46.64 46.04 45.76 44.24
19 46.31 46.10 45.76 44.24
20 44.82 45.84 45.76 44.24
21 45.39 45.75 45.76 44.24
22 47.80 46.16 45.76 44.24
23 46.69 46.27 45.76 44.24
24 46.99 46.41 45.76 44.24
25 44.53 45.04 45.76 44.24
The EW1v1A values 3I'e plotted on a control chart along with the upper and lower control J.i:::;ito given by
=45.45,
rr""X--'" 1 2(11 ..
LCL = Po -LO'\(Z_i.J[.-(H.) 1
=44.55.
532 Chapter 17 Statistical QUality Control and Reliability Engineering
UCL=45.76
45,5
"::;
~
45.Q ~l;ea.n:: 45
44.5
LCL=44.24
44.0
0 5 10 15 20 25
Sample number
Figure 17·12 EWMAcha.",forE:wnple17·8.
The reIl)aining controllim,its are calculated similarly and plotted on the control chart giver:. in Fig. 17~
12, The conrrollimits ::end to' increase as i increases, but then rend to die steady state values given by
equations in 17-24:
.-)-
UCL= 1'0 +La;12~ A
=45+2.7(0.84») 0.2
12~0.2
'-A-
LCL=uo -La 1_ _•
." ~2-1.
=45-2.7(0.84) 1-0. 2
'12-0.2
= 4424.
The E\V'I\1A, control chart signals at observation 17, indicating that the process is O\;t of concr.l.
The sensitivity of the EWMA control chart for a particular process will depeod on the
choices of L and A. Various choices of iliese parameters will be presented later in this chap-
ter, when the concept of the average run length is introduced. For more details and deyel~
opments regarding the EWMA, see Crowder (1987), Lucas and Saccucci (1990), and
Montgomery (2001).
The ARL is the expected number of samples taken before a control chart signals out of COn~
trOL In general, the ARL is
I
ARL=:-.
p
where p is the probability of any point exceeding the controillmlts. If the process is in con~
trol and the control chart signals out of control,. then we say that afalse alarm has occurred
To illustrate, consider the Xcontrol chart with the standard 30-limits. For this situation, p =
0.0027 is the probability that a single point falls outside the limits when the process is in
controL The in~control ARL for the X CO!ltrol chart is
From this, 1:- fJ is the probability that a shift in the process is detected on the first sample
after the shift has occurred. That is, the process has shifted and a point exceeds the control
limits,--signaling the process is out of contro1. Therefore, ARLl is the expected number of
samples observed before a shift is detected.
The ARLs have been used to evaluate and design control charts for variables and for
attributes. For more discussion on the use of ARLs for these chms, see Montgomery
(2001),
Shift in M=
of 0) 0 0.25 0.50 0.75 1.00 LSD 2.00 2.50 3.00 4.00
ARL 465 139 38,0 17,0 10,4 5,75 4,01 3.11 2,57 2,01
Table 17-9 ARl...$ for Various E"'lMA Control Schemes {Lucas and Saccucci 1990}
Shift iL Mean L= 3.054 L""2.998 L=2.962 L=2.814 L=2.615
(multiple of d) }.;;:;;0.40 ).=0.25 1=0.20 A=O.1O ).=0.05
in the process mean, it is found that small values of I. should be used. Note that for L = 3.0
and I. = LO, the EWMA reduces to the standard Shewhart control chart with 3-sigma l:m-
its.
75
64
diagram is widely used to clearly display the various potential causes of defects in products
and their interrelationships, It is useful in summari7ing knowledge about the process.
As a result of the brainsto:rming session, the team tentatively identifies the following
variables as potentially influential in creating solder defects:
A statistically designed experiment could be used to investigate the effect of these seven
variables on solder defects. Also, the team constructed a defect concentration diagram for
the product. A defect concentration diagram is just a sketch or drawing of the product, with
the most frequently occurring defe<:ts shown on the part. This diagram is used to determine
whether defects occur in the same location on the part The defect concentration diagram
for the printed circuit board is shown in Fig. 17-15. This diagram indicates that most of the
insufficient solder defects are near the front edge of the board, where it makes initial con-
tact with the solder wave. Further investigation showed that one of the pallets used to carry
the boards across the wave was bent, causing the front edge of the board to make poor con-
tact with the solder wave.
17 -4 Reli~bility Engineering 537
Amount
Wf:Ne height
Exhau~ Conveyorsoeed Specifio rav:
Contact time
Conve or an Ie
Wave fluidity
Maintenance
Alignment of pallet
Orler.tation
Temperature
Contaminated lead
Figure 17~14 Cause--and-effect diagram for the print!:d circuit board flow solder process.
Sack
Figure 17~ 15 Defect concentration diagram for a printed cireui: board.
When the defective pallet was replaced, a designed experiment was used to investigate
the seven variables discussed earlier. The results of this experiment indicated that several of
these factors were influential and could be adjusted to reduce solder defects. After the
results of the experiment were implemented, the percentage of solder joints requiring
rework was redueedOOm 1% to under 100 parts per million (0.01 %).
I
538 Chapter 17 Sutistical Quality Control and Reliability Engineering
area of consi.uner product manufacture, high reliability has come to be e.xpected as much as
confonnance to other important quality characteristics.
Reliability engineering encompasses several activities, one of which is reliability mod~
eling. Essentially, the system survival probability is expressed as a function of a subsystem
of component reliabilities (survival probabilities). Usually, these models are fune depend-
ent, but there are some situations where this is not the case. A second important activity is
that of life testing and reliability esfunation.
(17-25)
Since for many components (or systems) the initial failures or time zero failures are
removed during testing, the random variable Tis conditioned on the event that T> 0, so that
the failure density is
g(t)
j(t) = i - p{O)' t>o, (17-26)
=0 otherwise.
Thus, in tenns of/; the reliability function, R, is
The term interval failure rate denotes the rate of failure on a particular interval of time
[r i , t 2J and the terms failure rate, iflstantaneousjailure rate, and hazc.rd \\till be used syn-
onymously as a limiting form of thc interval failure rate as 12 ~ t 1- The interval fajlure rate
FR(tt, t,) is as follows:
(17-28)
g(l)
~.-----.:::-..~----.:::--+;
o
Figure 17~16 A composite:faiJu:re distribution.
17-4 Reliability EngiLeering 539
The second term is for the dimensional characteristic> so that we may express the condi-
tional probability of equation 17-29 on a per-unit time basis.
We will develop the instantaneous failure rate (as a function of I). Let h(l) be !.'Ie haz-
ard function. Then
h{ll~ lim R(t)-R(I+ill) 1
• ", ....0 R(rl ill
=-lim R(I+ill)-R(I) ._1_
"' .... 0 ill R(t)'
or
= =
sinceR(t) l-F(r) and -K(t) J(I). A typical hazard function is shown in Fig. 17-17. Note
that h(t) . dl might be thought of as the instantaneous probability of failure at I, given sur-
vival to t.
A useful result is that the reliability functionR may be easily expressed in tenns of has
h(t)
Early ag~'
h
!
failures Wear Oll:
and Random failures _ _-t-faiiures and_
rancom random
fai!ures
, 1allures
Ja
r
r h(x)dx =- 0 R'«X)"ilx
R x)
=-lnR(X}',!o
so that
(17-32)
Most complex system modeling assumes that only random component failures need be
considered. This is equivalent to stating that the time~to~failure distribution is exponential,
that is,
t<: 0,
othenvise.
so that
i() ;l,,-"
h(t)=-' =--=,1.
R(I) e-"
is a constant. \\tnen all early-age failures have been removed by bum in., and the time to
occurrence of wearolit failures is very great (as with electronic parts), then this assumption
is reasonable.
The normal distribution is. most generally used to mode! wearout failure or stress fail-
ure (where the random variable under study is StreSS level). In situations where most failures
are due to wear, the nonnal distribution may very well be appropriate.
The lognormal distribntion has been found to be applicable in describing time to fail-
ure for some types of components; and the literature seems to indicate an increased
utilization of this density for this purpose.
The Weibull distribution has been extensively used to represent time to failure. and its
nature is such that it may be made to approximate closely the observ-ed phenomena. Vt'hen
a system is composed of a number of components and failure is due to the most serious of
a large number of defects or possible defects, the Weibull distribution seems to do particu-
larly well as a model.
The gll1Iltlla distribution frequently resulfll from modeling standby redundancy where
components have an exponential ti.me~rQ-failure distribution. We will investigate standby
redundancy in Section 17-4.5.
17-4 Reliability Engineering 541
t;' 0,
;::;0, otherwise, (17-33)
(17-36)
fit)
hit)
Figure 17~lS Density, reUabwr:y ft.'Uctio::;., and hazard function for the exponential failure model.
542 Chapter 17 Statistical Quality Control and Reliability Engineering
Emnple 17:9
A diode used on a printed c!.reui~ board has a fated failure rate of 2.3 x lo-e failures per hour. How-
eve:, under an increased tempera~e stress, it is felt that the :ate is about :t5 x 10-5 failures per hour.
The time to failure is exponentially distributed, so that we have
ft.t) "'" (1.5 X 1O-5)e-<lJh<!o-~)r, t~O,
~O, otherwise,
RU) ::::: e-< L5:«11)-5)I, t~O,
=0, othetwise,
and
h(f) ~ L5x 10-', !~O.
=0, otherwise.
To determine the reliability at t = 10 and t "" 1iY, we evaluate R( 10') =- e...iJ,15 ::.:: 0.861, and R( 10 5)
4
:::::-
[15;0.223.
Example17.10.
Three components must all iu..'1ction [or a simple system to funetion. The random variables T;. T2 , and
TJ representing t.i.r!J.e to failure for the components are independent with the following distributions:
4
Tr-N(2XIO J , 4X10 ),
I
T,-WcibUl\y=O, 0=1, '71)-'
T,-lOgnormal(1l = 10, <1'=4).
1: follows that
so that
~ [0.175::0.0498J[0.876)
=0.0076.
For the simple serial system, system reliability may be calculated using the product of
the component reliability functions as demonstrated; however, when all components have
an exponential distribution, the calculations are greatly simplified, since
R(t) ::;:;; e-).;l . e-.t:f ... e-J"I = e-;).I + AZ+ " . . . ..1.,)/,
or
(17-38)
where A.f = L:~).j represent>; the system failure rate. \Ve also note that the system reliability
function is of the same fonn as the component reliability functions. The system failure rate
is simply the sum of the component failure rates, and this makes application very easy.
Consider an electronic cii-c'Jit with three integrated ciTC".rit devices, 12 silicon diodes. 8 ceramic capac-
itors, and 15 composition resistors. Suppose under given Stress levels of <;eroperatt:re, shock, and so
on that each component has fail:rre rates as shown 1:;. the following table. and the component faillh-eS
~ independent
Therefore,
A, ~3(O.013 10-') + 12(1.7 x 10- ) + 8(1.2 x 10-7) + 15(0.61
X
'
x 10-7 )
=3.9189 X 10"',
and
544 Chapter 17 Statistical Quality Control and Reliability Engineering
(17·39)
Three identical components are arranged in active redundancy. operating independently. In order for
the assembly to fun~tion, at least two of t.'e components must function (k = 2), The reliability func~
don for the system is thus
When only one of the n components is required, as is often the case, and the components
are not identical, we obtain
•
R(t) = 1- I1[I- Rj(t)1 (17-40)
i""l
The product is the probability that all components fail, and, obviously, if they do not fail the
system survives. When the components are identical and only one is requited, equation
17-40 reduces to
R(t) = 1 - [I r(tlJ", (17-41)
where r(t) = R/t),j = 1, 2, ... , "-
'When the components have exponential failure laws, we will consider two cases. First,
when the components are identical with failure rare ;.t and at least k compo:cents are required
for the assembly to operate, equation 17-39 becomes
(17-42)
The second case is considered for the situation where the components have identical expo-
nential failure densities and where only one component must function for the assembly to
function. Using equation 17-41. we get
R(tl 1 - [1 - e-'1". (17-43)
In Example 17-12, where th::ee identical components were arranged in an active redundancy, and at
:east !:\Vo were required for system operation. we found
R(t) = (r(t)]'[3 - 2r(tl],
If the component reliability functions are
r(tl = e"",
then
R(t) = .-"'[3 - 2'-'1
=3e-2M _ 2e-3.l.:,
If tv.'o eomponents arc ammgcd in an active redu.Ldancy as described. and only one must function for
the assembly to function, and. furthermore, if the time-to-failurc densities arc exponential with fail-
ure rate 1, tb,en from equation 17-42. we obtci:n
R(t) 1- [1 e-42 = 2e-"-
Standby units are not subject to failure until activated. The time to failure for the assem~
bly is
T = T1 + T'}. + ... + Til!
where T! is the time to failure for the ith component and T1, T2, •• ,. Tn are independent ran-
dom variables. The most common value for n in practice is two, so the Central Limit The-
orem is of little value, However, we know from the property of linear combinations that
"
E(T]: 2,£(li)
1=1
and
" V(li).
Y(T] = 2,
i=l
We must know the distributions of the random variables Ti in oroer to find the distribu-
tion of T. The most common case occurs when the components are identical and the time-to-
failure distributions a..re assumed to be exponential. In this case. Thas a gamma distribution
f(t)=~(;urle-", 1>0,
, (n-l)!
=0 otherwise,
so that the reliability function is
"-1
R(t): 2,e-" (At)k /k!, t>O. (17-44)
k.O
The parameter A. is the component failure rate; that is, E(TJ = ItA.. The mean time to fail-
ure and varia"lce are
MTIF = E[11 = nI A. (17-45)
and
(17-46)
respectively,
1 17-4 Reliability Engineering 547
Example i7:14
Two identical compOllents are assembled in a standby redundant configuration with perfect switch-
ing. The cornponer.t lives are identically distributed. independer.t rand()m variables having an expo-
lOu
ner.tial distribu::ion with failure rate l , The mean time to failure is
(17-48)
.:E"".tipi~17,~~
One hundred units are placed on life test, and the test is run for 1000 Mum. There are two failures dur-
ing test. soft:;;;;; 0.02, ar.d ~~(1000) =0.98. Using a table oft.'e binomial Gistribution, a 95% uppc('con-
fidence limit on pis 0.06, so that a lower limit on R(lOOO) is given by 0.94.
In recent years, t.~ere has been much work on the analysis of failure-u.'Ue data, .:nclud-
ing plotting methods for identification of appropriate failure~time models and paramet:er
estimation. For a good summary oftbis work. refer to Elsayed (1996).
S48 Chapter 17 Statistical Quality Control an.d Reliability Engineering
There are four situations to consider. according to whether the test is stopped after a preas-
signed time or after a preassigned number of failures and whether failed items are replaced
or not replaced during test.
For the replacement test, the total test time in unit hours is Q nt", and for the nonre-
placement test
,
Q= 2> +(n-r)t', (17-51)
If items ace censored (withdrawn items that have not failed), and if failures are replaced
whlle censored items are not replaced. then
c
Q= 2:>! +(n-c)" (17-52)
)"'1
r 17-4 Reliability Engineering 549
where c represents the number of censored items and tj is the time of the jrh censorship. If
neither censored items nor failed items are replaced, then
, c
L= TIf(ti)'TIW)
(17-54)
Then
l~lnL~-rlne-ekti- (n-rt
)" l~ /8
(""1
Two...sided lL'11its
The results are, however. dependent on the assumption that the distribution is exponential,
It is sometimes necessary to estimate the time tR for which the reliability will be R, For
the exponential model, this estimate is
(17-'>7)
,l:;~~~';17',!6.
Twenty items are placed on a replacemenr test that is to be O?Cf3ted until 10 failures occur. The tenth
failure occurs at 80 hours, ar.d the reliability engineer wishes to estimate :he (':lean time '.:0 failure,
95% two-sided limits on 8, R(10D), and 95% two-sided limits on ROOO). Finally, she wishes ':0 esti-
mate the time for whkh the :eliability wil: be 0,8 with point and 9.5% two-sided confdence interval
estimates.
According to equation 17~56 and the results presented in Tables 17~lO and 17-11,
Also,
The two-sided 95% confidence limit is determined from Table 17-11 a.<:;
r
L 34,17 9.591"
1
2(1600)(0,223), 2(1600)(0.223) ~ [20,9,74.45],
: 2Qln(l/R) -
Lower, ol1e~sided limit
" '-J
I %;;'.21'-1
~
T 17-6 Ex=ises 551
17-5 SL"MMARY
This chapter has presented several widely used methods for statistical quality control. Con-
trol charts were -introduced and their use as process surveillance deviees discussed. The X
and R control cha."tS are used for measurement data. ¥t'ben the quality characteristic is an
attribute, either the p chart for fraction defective or the cor u chart for defects may be used.
The use of probability as a modeling technique in reliability analyse, was also dis-
cussed. The exponential distribution is widely used as the distribution of time to failure,
although other plausible models inelude the normal, lognormal, Weibull, and gamma dis-
tributions, System reliability analysis methods were presented for serial systems, as weB as
for systems having active or standby redundaney. Life testing and reliability estimation
were also briefly introduced.
17-6 EXERCISES
17~1. An extrusion die is used to produce aluminum 17..2.. S'I!ppose a process is in control, and 3-sigma
rods. The diameter of the rods is a critical quality controllimlts are in use on the X chart. Let the :mean
eharacteristic. Below are shown X and R values for 20 shift by 1.50. ""'ba! is the probability that t1is shift
samples of five rods eaeh. Specifications on the rods will remain undetected for three consec:.ttive samples?
are 0.5035 ± 0,0010 inch. The values given are the V/hat would this probability be if2-sigma control. lim-
la.st three digits of the measurements; that is. 34.2 is its are used? The sa."Ilple size is 4,
read as 0.50342.
11~3. Suppose that an X chart is used ~o control a :'.01.'-
mall)' distributed process, and that samples of size n
Sample X R Sample X R
are taken every h hours and plotted on the chart, which
1 34.2 3 11 35.4 8 has k sigma limits.
2 31.6 4 12 34.0 6 (a) Find the ex.pected number of samples that will be
3 31.8 4 13 36.0 4 Laken until a false action signal is generated. This
4 33.4 5 14 37,2 7 is called the in-co:::trol average run length CARL).
5 35.0 4 15 35.2 3
(0) S'I!ppose tha~ the ptocess srJfts to an out-of-
6 32.1 2 16 33.4 10 eontrol state, Find the expected number of Sacl-
7 32.6 7 17 35.0 4
pies that will be taken until a false action is
8 33.8 9 18 34,4 7 generated. This is the out-of-cootrol ARL.
9 34.8 10 19 33.9 8 (e) Evaluate the in-comrolARL fork= 3. How does
10 38.6 4 20 34.0 4 this change if k::; 27 What do you thi:lk about the
use of 2-sigma limits in practice?
(a) Set up the X wd R cha..''ts, revising the trial control
(d) Evaluate the out-of-controlARL for a shift of one
limits if neeessary. assuming assignable causes
sigma, give:;', that n = 5.
cau be found,
(0) CalctUa>e peR and PCRk' Interpret these ratios. 11-4~ Twenty-five samples of size 5 are dra'Wn from a
(c) What pereentage of defectives is being produced process at regular intervals, and the following da!a a(e
by this prOl..---ess? obtained:
552 Chapter 17 Statistical Quality Control and Reliability Engineering
!
Does the no:rrnality assumption appear to be
satisfied? standard deviation of this quality characteristic is 5.
(b) Set up an. individuals control chart for the \Vbe:re ~hou1d we een~er the process to minimize the
weights. Interpret the results, fraction: defective produced? Now suppOse the mean
sl'.ifts to 105 ar.d we atc using a sample size of 4 on an
17-9. The follo~1ng are the numbet of defective sol- X chan:, Wh2.t is the probability that such a sb.ift wit!
der joints found during successive samples of 500 sol· be detw.ed on the first sou:.ple following t."ic shift:
derjol."lts, What sample size would be needed on a p chart to
obtain a similar degree of protection?
Day No. of Defectives Day No. of Defectives 17-14. Suppose the following fraction defective had
been found L'1 successive samples of size 100 (read
106 11 42 down):
2 116 12 37
3 164 13 25 0.09 0.03 0.12
4 89 14 88 0.10 0.05 0.14
5 99 15 101 0.13 0.13 0.06
6 40 16 64 0.08 0.10 0.05
7 \\2 17 51 0.14 0.14 0.14
8 36 18 74 0.09 0.D7 0.11
9 69 19 71 0.10 0.06 0,09
10 74 20 43 0.15 0.09 0.13
21 80 0.13 0.08 0.12
0.06 O.ll 0.09
Construct a fraction-defective control chart. Is the
process in control? Is the process in control with respect to its fraction
defective?
17~10. A process is controlled by a p chart using sam-
ples of size 100. The centerline on the chart is 0.05. 17~15. The foUov.iug- represent the nu.-r.ber of solder
What is the probability th2.t the control chart detccts a defects observed on 24 samples of five printed circuit
shift to 0.08 on the first sample f&lowing the shift? boards: 7. 6.S. 10.24. 6.5,4,8. 11.15. 8,4. ,6.
II,
What is the probability t.hat the shift is detected by at 12, 8,6,5,9,7, 14,8,21. Can we conclude that t.'e
least the third sample following the shift? process is in control using a c chart? If not, asst.::;ne
assignabie causes can be found and revise the control
17~11. Suppose a p :::hart with cC'I}terline at p with k limitS.
sigma units is used to control a process. There is a
17-16. The following represent t.'e number of de:fects
critical fraction defective p< t3at must be detected with
per 1000 feet ir:. rubbcr-cov:red wire: I, 1,3,7,8, 10.
probability 0.50 on the firs: sample followe:g the shift
5,13,0,19,24,6,9, 11, 15,8,3,6.7,4,9,20,11,7,
to dUs state. Derive a general fonuula for the sample
size that snould be used on this enart.
18, 10.6,4.0,9,7,3, I, 8, 12. Do the data come from
a control1ed process?
17~12. A normally distributed process uses 66.7% of 17-17. Suppose t.'e nt::mber of de:ects in a unit is
the specification band, It is ce:.1tered at the nominal knOVI'l1to be 8, If the number of defects ma unit shifts
dimension, located halfway between the upper and to 16, what it !he probability that it will be detected by
lower specifica~on limits. the c chart on the :Erst sample follo'Ning the s..lift?
(a) What is t."ie process capability ratio peR? 17~lS. Suppose we are inspectir:.g disk drives for
(b) What fallout level (fraction defective) is pro- defects per unit, and it is h:o-wn that there is an aver-
duced? age of two defects per urit. We decided to make our
(c) Suppose the mean shifts. to a distance exactly 3 inspection unit for the c chart five disk drives. and we
standard deviations below the upper 1>1lecification control the tOtal number of defects per inspection unit.
llinil What is the value of PCRk? How has peR Describe the new control chart,
changed? 17~19. Consider the data in Exercise 17-15. Set ..tp a u
Cd) 'iVhat is the actual faI:out experienced a....fter the chart for tlrls process. Compare it to the c chart in
shift ill the meal:? Exercise 17-15.
554 Chapter 17 Statistica: Quality Control and Reliability Engineering
17~20. Consider the oxide thickness d3.ia given in decision switch and only one unit is required for sub-
Exercise 17-7. Set up an EWMA control chart with system survival, detem.:i;ne tbe subsystem reliability.
;. == 0.20 and L;:;;;;: 2.962. Interpret t.':le chart. 17~28. One 111mdred unit;.; are p:aced on test and aged
11-21. Consider the oxide thickness data given in until all units have failed. The follo~ing results are
Exercise 17-7. Construct a CUSUM control ch..'ilrt with obtained, and a mean life oft:= 160 hours is calculated
k = 0.75 and h = 3.34 if the target L';leme" is 50. from the serial data.
Interpret the chart.
17-22. Consider the weigh:s provided in Exereise Time Interval Number off-allures
17·8. Set up an E\VMA eontro1 chart with .<=0.10
and L == 2.7. Interprl!t the chart. 0-100 50
17kl3. Consider t."Ie weights provided in Exercise 100-200 18
17~g. Set up a CUStr:M control chart with k;:;;;;: 0.50 and 200-300 :7
h:::::: 4,0. Interpret the chart. 300-400 8
17-24. A titue-to-failure distribution is giv::::n by a uni- 400-500 4
form distribution: After 500 hours 3
1
/(1)=--, Use the chi-square goodness-of-fit test to detennine
fJ-a whether you consider the exponential distribution to
=0 otheI'Vllse, represent a reasonable tirne~to--fai1ure model for thesl!
Ca) Determine the re:iability function. data.
(b; Show tha~
17-29. Fifty uaits are placed on a life test for 1000
hours. Eight units fail during the period. Estimate
(R(t)dt= (tf(t)dt.
R(lOOO) for these units. Determine a lower 95% con~
(c) Determine t.':le hazard function, fidence interval on R(lOOO).
(d) Show that 17-30. In Section 17-4.7 h was noted that for one-
partt.."nerer reliability functions, R(t;fi ),R(t;f1):= R(t;8).
e
where and Rare the maximum likelihood estima~
whereH is defmed as in eqcati.on 17~31. tors. Prove this statement for the case
17-25. TIrree units that operate and fail independently R(r,8) ~ e-<l'. t:?:O,
form a series configuration. as shown in the figure at
the bottom of this page. =0, othenv1.se.
The tirr.e-to-failure distribution for each UJI.it is expo- Hint: Ex.press the density function fin tenns of R.
nential with t.1e failure rates indicated,
(a) Find R(60) fur the systeIIL 17-31. For a nonreplacement test that is terminated after
200 hours of operation, it is noted that failures occur at
(b) ~'hat is the mean time-to-failurc Cv1TTF) for this
the following ti:nes: 9, 21, 40, 55, and 85 l:ours. The
sy&tem'f
units are a.<sl.llned to have an expOTh---ntial tittle-To-failure
17M26. Five identical units are arranged in an active distribution, and 100 uoits were on test initially,
redundancy to fonn a subsystelXl. Unit failure is inde-
(a) Estim.ate the mean time to failure.
pcadem, and a:: least t'No of the units must survive
1000 hours for the subsystem to perfonn its :nission. (b) Construct a 95% lower-confidence limit on t?-e
mean tir::te to failure.
(a) If the units have c:xpone!ltial time-to-failure dis-
tributions with failure rate 0,002. what is t.':le sub- 17~32. Use the statement in Ex.ercise 17-3!.
sys~ reliability? (2) Estimate R(300) and construct a 95% lower-
(0) Vihatis the reliability if only onc unit is reqci."'ed? confidence limit on R(300).
17~27. If the utrits described in the previous exercise (b) Estin-.ate the time for which the reliability will be
are operated in a standby redundancy with a perfect 0,9. and construct a 95% lower limit on te.g,
for t= O. 1r 2, ... ~ and every sequencej, i, it • •.. , it' This is equivalent to stating that the prob-
ability of an event at time t + 1 gi:ven only the outcome at time t is equal to the probability
of the event at time t + 1 given the entire state history of the system. In other words, the
probability of the event at t.,. 1 is not dependent on the state history prior to time t.
The conditional probabilities
P {Xt + 1 =j1X,. """ i} :;:::: Plj (18.2)
are called one..step transition probabilities, and they are said to be statior.ary if
for t= 0,1,2, .,., (18-3)
55S
556 Cbapter 18 Stochastic Processes and Queueing
so that the transition probabilities remain unchanged through time. These values may be
displayed in a matrix P :;;; {Pi)]' called the one-step transitio.o matrix. Tne matrix P has m +
1 rows and m + 1 columns, and
while
That is. each element of the P matrix is a probability, and each row of the matrix surns to one.
The existence of the one-step! stationary transition probabilities implies that
and
and they indicate that in passing from state i to state j in n steps the process will be in some
state, say I, after exactly v ,mps (v,,; n). Therefore pt' .
J'~-'J is the conditional probability
that given state i as the starting state, the process goes to state 1in v steps and from 1to j in
(n - v) steps. When summed over 1. the sUIIl of the products )'ie1dspiJ) .
By setting v lor V'=n -1. we obtain
i=O,1,2, ...• m,
j O.1.2.... ,m.
n=l,2, .. ..
It follows that the n-step transition probabilities, pt'" may be obtained from the one-step
probabilities, and
(18-6)
18-3 Classification of States and Chains 557
"
~
'" a~O) . p~;:)
I I) 1
j O,I,2" ..,m,
n:;;: 1,2, •...
i=C
Tnus, A(n) =A· p(n). Further, we note that the rule for matrix mcitiplication solves: the total
probability law of Theorem 1~8 so that A{n) A(n-I). p.
.lli~~l~i.~:r
In a eomputing system, the probabiHty of an error on each cyele depends on whether or not it was pre~
ceded by an error. We will define a as the error state and 1 as the nonerror sta:e. Suppose the proba-
bility of an error if preceded by an error is 0.75, the probability of a."'! error if p:eceded by a nonerror
is 0.50, :.he probability of a nonerror if pre~eded by an error is 0.25, a:r:d the probability of noneITOr
if preceded by nOl!error is 0.50. Thus.
p- r
O.75 0.25;
- L0.50 0.50' J
Two-step, threo-step. , .. , seven-s::.ep transition matrices arc sho'W'U below:
P""=
, rO. 688 0.312] p' =rO,672
0.328]
LO.625 0.375 ' LO.656 0.34 '
~~~!}
, jO.668 p-< = [0.667 0.333]
p =LO.664 0.666 0.334 '
p, = [0.667 0333] , _rO. 667 0.333J
0.667 0.333 ' P -,0.667 0.333
If we know that initially the system is in the nonerro: state, thc.'l ai~) I, d,,0) = O. and A(II) :::::; (ar] =
A· pro,. Thus, forcxample, A(1)= [0.667, 0.333J.
Thus, recursive computation from the one-step transition probabilities yields the probabil-
ity that the :first passage time is n for given i,j,
;:t>:an;~i~I§:Z:
Using the one-step transition probabilities presented in Exan::.ple 18~1, the distrlbutloo. of the passage
time index n for i = O,j = 1 is determined as
f~';1 =POI =: 0.250,
f~;' = (0.312) - (0.25)(0.5) = 0.187,
There are four such distribution, corresponding to i, j value: (0,0), (0, I), (I, 0), (I, 1).
If i and} are fixed, then I:::"tfi;} s 1. W'hen the sum is equal to one, the values for It).
n::;1, 2; ... , represent the probability distribution of first passage time for specific i,j. In
the case where a process in state i may never reach state j, L:tf~; < L
Where i = j and I::,/i;) :;:;: 1, the state i is termed a recurrent STate, since given that the
process is in state i it will always eventually return to i.
If Pa 1 for some State i, then that state is: called an absorbing state. and the process
will never leave it after it is entered.
The state i is called a transient state if
ifS tt
) <1,
n=!
since there is a positive probability that given the process is in state i, it will never retum to
this state. It is not always easy to classify a state as transient or recurrent, since it is some-
times difficult to calculate first passage time probabilities t;;i
for all n 1 as was the case in
Example 18-2. Nevertheless, the expected first passage time is
(18-9)
1) a simple conditioning argument shows that
There are no null recurrent states in a finite-state Markov chain. All of the states in such
chains are either positive recurrent or transient.
18~3 Classification of Sta:es and Chains 559
A state is called periodic with period 1'> 1 if a return is possible only in 'r, 2'r, 3-r, "'\ steps;
::
so p;~) 0 for all values of n that are not divisible by 1'> 1. and -ris the smallest integer hav~
ing this property.
A state j is termed accessible from state i ifp::) > 0 forsorne n= 1.2, .. , . In ol.!!'exam-
'J
pie of the computing system. each state. 0 and 1, is accessible from the other; since p;~') > 0
for all i, j and all n. If state j is accessible from i and state i is accessible from j. then the
states are said to communicate. This is the case in Example 18-1. We note that any state
conununicates with itself. If state i communicates withj,j also communicates with i, Also,
if i communicates with 1and I commuricates \vithJ, then i also communicates with).
If the state space is partitioned into disjoint sets (called equivalence classes) of states,
where COIIL.Yflunicating states belong to the same class, then the IvIarkov chain may consist
of one or more classes. If there is only one class so that at: states communicate. the Markov
chain is said to be irreducible. The chain represented by Example 18-1 is thus also irre-
ducible. For finite-state Markov chains, the states of a class are either aU positive recurrent
or all transient. In many applications, the states will all communicate. This is the case if
there is a value of n for which p;;) > 0 for all values of i andj,
If state i in a class is aperiodic (not periodic), and if the state is also positive recurrent.
then the state is said to be ergodic. An irreducible Markov chain is ergodic if all of its states
are ergodic. In the case of such Markov chains the distribution
A") = A· P"
converges as n -7 00 , and the limiting disuibution is independent of the initial probabilities,
A.In F.xamp1e 18-1, this was clearly observed to be the case, and after fiye steps (n > 5),
PIX, = OJ = 0.667 and PIX, 1) = 0.333 when three significant figures are used,
In general. for irreducible, ergodic 11arkov chains,
lim {r.) - lim {II)_
n--t"" Pi) - n~"''''' aj - Pj.
and, fur..hennore, these values Pi are independent of i. These "steady stare" probabilities, p/,
satisfy the following state equations:
(l8-Ha)
m
I,Pj =1, (18-Hb)
j=O
m
pj=LPi'Pij j == O.1,2,.".m. (l8-11e)
i=O
Since there are m + 2 equations in 18-11b and 18-11c, and since there are m + 1 un.!::nowns,
one of the equations is redundant. Therefore we will use 111 of the 111 + 1 equations in equa-
j
In the case of the computing system presented in E.xample 18-1. we have fro;:;. equario:1S IS-lIb and
1S-lle,
1 =Po+Pj'
Po=Po (0.75)+p,(0.50).
560 Chapter 18 Stochastic Processes and Queucing
or
Po =213 and p, = 113,
which agrees with l.1e emerging result as n:> 5 in Exampie 18~ 1.
The steady state probabilities and the mean recurrence time for irreducible~ ergodic
Markov chains have a reciprocal relationship,
1
j = O,1,2, .. "m, (18,12)
Pj
In Example 18,3 :lote that,lloo ~ 1ipo = 1,5 and)1" ~ lip, 3,
Tue mood of a corporate president is observed over a period of time by a psychologist in the opera~
tions research department. Being inclined toward mathematica:. modeling, the psychologist classifies
mood into three states as follows:
0: Good (cheerful)
1: Fair (so-so)
2: Poor (glum and depressed)
'The psychologist observes that mood changes OCClr only overnight: thus. the data allow estimation
of the transition probabilities
jO,6 0,2 0,2'
P~ l0.3 0,4 O.3J'
0,0 0,3 0,7
The equations
Po O.6pQ .... a.3pt + 0P2'
PI ;;;: O,2pJ + OAp; -'- O.3P2'
1 =Po+P,":"Pz
are solved simultaneously for the steady s+..ate probabilities
Pa=3/13.
p, =4113,
P2= 6/13.
Given that the president is in a bad mood. that is, state 2. the mean rime reql.l.ired to return to that state
is i-i;:;, where
1 [3
I-tn ,::::::-=-days,
p:! 6
As noted earlier ifpk}t= 1, state k is called an absorbing state. and the process remains
j
in state k once that state is reached, In this case, b" is called the absorption probability,
18~4 CQntinuous~Time Markoy Chains 561
which is the conditional probability of absorption into state k given state i. }'lathematically,
we have
bik.
'" ·bjl\.>
= LP(l l=0,1.2, ... ,m, (18-13)
i""O
where
and
for i recurrent. i *' k.
i 0,1,2, ... , m.
PiP)~P [X(t +s)=jlX(s) = ,1, j=O, 1, 2t .. " m,
s;;:: 0, t::?:O,
is the stationary transition probability function. It is noted that these probabilities are not
cependent on s but only on t for a specified i.j pair of states. Furthermore, at time t = 0, the
function is continuous with
There is a direct correspondence between the discrete ..time and continuous-time mod-
els, The Chapman-Kolmogorov equations become
Pij(t) '"
2::Pe(v), p/j(t-v) (18-14)
1=0
for 0:::;; v S; t, and for the specified state pair i,j and time t. If there are times t1 and '2 such
that Plj(t,) " 0 and Pj,{r,) > 0, then states i andj are said to communicate. Once agab states
that communicate form an equivalence claSs. and where the chaL, is irreducible (all states
form a single class)
p,P) > 0, for t > 0,
for eacb state pair iJ
We also have the property that
562 Chapter 18 Stochastic Processes and Queueing
where Pi exists and is independent of the initial state probability vector A, The values Pj a..--e
again called the steady state probabilities and they satisfy
Pj >0, j = 0.1.2....,m,
m
PI = '\' v·· p .. (t).
£..,.<1 1) ,
j = O,1.2,.,.,m, t;;' O.
f",C
(18·15)
where the limit exists and is finite, Likewise; the intensity of passage from state ito statej,
given that the system is in state i, is
(18·16)
again where the limit exists and is finite. The interpretation of the intensities is that they rep-
resent an instantaneous rate of transition from state i to j. For a small ill, Pij(ill) :::: u(p.t +
0(/;1), where 0(/;1)1/;1 ... 0 as /;1 ... 0, so that .ii is a proportionality constant by which
Pij(6t) is proportional to ill as ill ~ 0. The transition intensities also satisfy the balance
equations
These equations indicate that in steady state, the rate of transition out of state j is equal to
the rate of transition into j.
An electronic control mechanism for a chemical process is constructed with two identical modU!es,
operating as a parallel. active cedur:.dant pair, The function of at least one module is necessary for the
mechanism to operate. The m3i:ntenance shop has two identical repair stations for these modules and,
furtilerrnore, when a module fails and enters the shop, on,er work is moved aside and repair work is
i.:m..rLediately initiated_ The "system" here consists of tbe mechanism and repair facility and the states
are as fellows:
0: Both modules operating
1: One unit operating and one unit itt repair
2: 1\vo units in repair (mechanism down)
The random variable representing t.inle to faillL>-e for a module has an exponential density~ say
t,.(tl = k"", 1<: O.
=0, ,<0,
and the random variable describing repair time at a repair station also has an exponential density, say
r-;<t) ~ j1e-J.U, I ~ 0,
=0, 1<0.
18-4 Continuous-Tlffie Markov Chains 563
Inteti'ailo.re and interrepait times are independent ax.d {X(t) t can be shown to be a continuous-param-
eter, irred:J.cible Markov cb.ain with transitions only from a state to its neighbor states: 0 -) 1. 1 --t 0,
1 ~ 2, 2 -;. 1. Of course, there may be nO state cbange.
The transition htensities are
'" =2Jl.
L'sing eqaation 18-17.
2};J, = !1P,.
(A...,... ll) P J ::= 2}{Jo + 2p.p2'
fJ,'
Po=-(;-)-,'
A+Jl,
The system a.vailability (probability that the mechanism is up) in the s:eady state condition is thus
.. bill'ty= 1- - ';t'
Avaca' -,.
(},+fJ,t
p~[Pij(ru)J
!-uOru "olru ".
"0/'" ", uO,"At
u1O .6t 1-u, ru " . Ul/lt " . ulmM (18-18)
= ... , ..
UiO At uil At uljAt u:mAt
where
p,{l) = P (X(t) = jJ.
From tbejtb equation in tbe m + 1 equations of equation 18-19,
p/t+ ru) = Po(t). uo;ill+ ". + p,{t)· ui;ill + ... + p/t)(1- u;ill] + ... + Pm(t) , uc, ru,
564 Chapter 18 Stochastic Processes am! Queueing
d I
-p;,t)=
at J'
.
lim
6t t'
[PJ(t+b.J)- p,(t)] =-u.·p.(t)+
1\ ))
.L u.. ·p.(t).
IJ I
(18-20)
--)" ul I'*j
which may be solved when m is finite, given initial conditions (probabilities) A. and using
lhe result that 2:;.0
p,(t) = L The solution
JJ..'t ::: Service rate given that n customers are in the system
The birth-death process can be used to describe how X(I) cbanges through time. It will
be assumed bere that when X(t) lhe probability distribution of the time to the next birth
(arrival) is exponential \Vith parameter Aj'! j :: O. 1. 2; .... Furthermore, given X(r) the
remaining rime to the next service completion is taken to be exponential with parameter j1.i'
j 1, 2, .... Poisson-type postulates are assumed to hold, so tha, the probability of mo{~
than one birth or death at me same .instant is zero.
18-5 The Birth-Deaili Process in Queuei..'lg 565
System
~------------------!
: Service !
! facility 1
o I
I
Queue o I
lnpLi Arriva.ls o I
o 0 /) , •• 0 r------1- Departl.res
process I
I
I 0 I
I I
I I
------------------~
Figure 18-1 A simple queucing system,
po C 0 0 0
).,' Jt.:
0 0 0 i~-()'j ..,u:})& '"
0 0 0 J1 }+l Ar
0 0 0 C
\Ve note that Plj(&) ;;; 0 for j < i 1 or j > i + 1. Furthermore, the transition intensities and
intensities of passage shown in equation 18-17 are
The fact that the transition intensities and intensities of passage are constant with time
is important in the development of this model The nature of ttansiti9D car; be viewed to be
specified by assumption, or it may be considered as a result of the prior assumption about.
the distribution of time between occurrences (births and deat.'ls),
1
~
i"O
In the steady state (I -7 ~). we have p;(t) = 0, so the steady state equations are obtained
from equations 18-23 and 18-24:
11,p1 = A"P"
A"Po + p..,y, = ('-I + 11,) . p"
'-:PI + iLJP3 = (A., + J.iJl. p"
(18-25)
If we let
18~6 Considerations in Que:ucing Models 567
C '_;C!/.)-Z "'An
(18·26)
j- J1jJ1j-l'''fl: •
then
and since
we obtain
1
PC=-';"--'
1+ LC;
J.1
These steady state results assume that the A,h J1J values are such that a steady state can
°
be reached. This will be true if Ai = for j > k, so iliat there are a finite number of states, It
is also true if p = NS).t < 1, where A. and).t are constant and,$ denotes the number of serverS.
The steady state will not be reached if Ll",
I G.,;;; QQ,
L,=AW,
568 Chapter 18 Stochastic Processes and Queue!ng
(These results"are special cases of what is bown as Little's law.) If the ;"J are not equal, I
replaces A., where
~
The system utilization coefficient p = /Js(.1 is the fraction of time that the servers are busy.
In the case where the mean service time is 11(.1 for allj ~ 1,
I
W )¥,,+-. (18.30)
I'
The birth-death process rates,~, 1 1, ••• , Aj; ••• and J.ljl J.L;" ••• , J.l;~ ••. may be assigned
j
any positive values as long as the assignment leads to a steady state solution. This allows
considerable flexibility in using the results given in equation 18-27. The specific models
subsequently presented will differ in the manner in which Ay and I'} vary as a function ofj.
L= Lj·(1-p)pJ
j=O
(18,34)
=2-
I-p
1&-7 Basic Single-Server Model \f.-ith Constant Rates 569
Lq=I,U-I),pj
j=l
=L-(I-po)
Using equations 18-28 and 18-34, we find that the expected waiting time ill the system is
(18-36)
;.' 1
W, = , (18-37)
A !1-(!1-- 1 )1 !1-(!1--).)'
These results could have been developed directly from the distributions of time in the sys-
tem and time in the queue. respectively. Since the exponential distribution reflects a mem-
oryless process, an arrival finding j units in the system will wait through j + 1 services,
including its own, and thus its waiting time ~. + I is the sum ofj + 1 independent, exponen-
tially distributed random variables, This random variable was shown in Chapter 6 to have a
gamma distribution. This is a conditional density given that the arrival fi:lds j units in the
system. Thus, if S represents time in the system.
... /+1
I, (l_p)pi . r~~e-I'Idl
i"O w r (J+l)
n-
w
p}!.le-I'II (P':,Y dt
1",,0 ),
(18-38)
= e-,u(I-P)"', w;;'O,
W<O,
which is seen to be the complement of the distribution function for an exponential random
variahle with parameter !1-(1 - pl. The mean value W = IIJl(1 - p) = 11(j.t - )_) follows
directly.
If we let S, represent time in the queue, excluding service time, then
P(S,=O)=p,= I-p.
570 Chapter 18 Stochastic Processes and Queueing
If we take T; as the sum of} service times. ~ will again have a gamma distribution. Then,
as in the previous manipulations.
p(Sq:>Wq ) 2>j·P(1j:>Wq )
1=1
~
-u(I-p)"
=pe' q,
=0,
and we find the distribution of time in the queue g(w<i); for \IV q > 0, to be
Wq >0,
Vlhen /" 2: J.i, the summation of the terms PI in equation 18-32 diverges, In this case.
there is no steady state solution since the steady state is never reached. That is, the queue
would grow without bound.
j~N,
(18-42)
=0, j>N.
18-8 Single Server with Limited Qt:.eue I..engt.l:! 571
j = O,I,2,,,,,N,
so that
N
pol>j =1
)"'0
and
I-p
1- pN+l' (18-43)
(18-45)
m
(18·..
(18·48)
It follows trom equation 18-27 that the state equations are developed as
p=!bi po
j ji
j >$
(18-51)
(18-52)
Then
L
W =-'L (18-53)
q I.
lS~12 Exercises 573
and
(18-54)
so that
(18-55)
In the case where service times are constant at 1Ij.L, the foregoing relationships yield the
measures of system performance by taking the variance cJl::::::; O.
18·11 SUlY1MARY
This chapter introduced the notion of discrete~state space stochastic processes for discrete~
time and continuous"time orientations. The Markov process was developed along with the
presentation of state properties and characteristics. TIllS was followed by a presentation of
the birth-death process and several important applications to queueing models for the
description of waiting~time phenomena.
18·12 EXERCISES
1s..1. A shoe repair shop in a suburban .!.'Call has one (b) Find the mean ll1!IDber of pairs in :.'I)e shcp and the
shoesmith. S=.toos are brought in for repair and arrive n::.ean nu::nber of pairs waiting for ser.'ice.
according ro a Poisson process with a constant arrival (c) Bnd the mea>. tumaround time for a pair of shoes
rate of two pairs per hour. The repair time distribution (time in the shop waiting plus repair, but exclud~
is exponential 'kith ame:an of20 minutes. and there is ing time waiting to be picked up),
independence betwee~ the reprur and arrival 1s..2. Weather data are analyzed for a particular local~
processes. Consider a pair of s:Qoes to be the unit to be
it)', and a Markov chain is employed as a model for
served, and do the following: weather change as follows. The conditional probahil~
(a) In the steady slate, find the probability tlJat the i:y of change from rain to clear weather in one day is
number of pairs of shoes in the system e~ceeds 5. 0.3. Likewise, the co;.:;.ditional probabili~ of transition
574 Chapter 18 Stochastie Processes and Qaeueing
from elearto rain in one day is 0.1. The model is to be If t Xn} can. be modeled as a Markov chain with one-
a discrete~time model, with transitions occurring only step tranSition matrix as
between days.
(a) Determine the matrix P of one-step transition o
probabilities, 1/6
(b) Find the steady state probabilities. 2/3
(c) If today is clear, find the probability that it will be o
dear exactly 3 days hence.
(d) Find the probability that the first passage from a do !he following:
clear day to a rainy day OCC'Ui.'S in exactly 2 days,
(a) Show that states 0 and 1 are absorbing states.
given a clear day is the initial state.
(b) If the initial state is state 1. compute the steady
(e) \Vhat is the mean reCW(ence time for the rainy
state probability Clat the system is in state O.
day state?
(c) If !he ~-utial probabilities are (0,112.112,0), com-
19..3~ A communication link: transmits binary chamc~ pute the steady state probability Pc'
tets, (O~ 1), There is a probability P t...'1at a trnnsmitted
character will be received corredy by a receiver,
(d) Repeat (c) "i!hA =(lJ4,1I4,1I4,1I4),
w:-.dch then transmits to another link, etc. If Xu is the 18-6, A gambler bets $1 on each hand of blackjack.
initial character and Xl is the character received after The probability of winning on any hand is p. and the
the first transmission, X2 after the second. etc., then probability of losing is 1 - P = q, Tlle gambler will
\\ith independence {X~} is a ~iarkov chain. End the continue to play until either $Yhas been accumulated.
one-step and steady state transitinn matrices. or he has no money left. Let XI denote the acct;.mu-
18-4. Consider a two--component active redundancy lated 'Winnings on hand t • .Kate that X,+ I =Xt + 1, with
where the components arc identical and the time-to- probability P. that Xl,. 1 :;:;;; Xr - 1. with probability q,
failure distributions are ex.ponential. When both units andXH I ::z:X. if XI = 0 or X,.=Y. The stochastic process
are operating. each cames load U1 and each has fail~ X, is a Markov chain.
ure rate .I•. However. when one unit fails. the load car- (a) Find the one-step transition matrix P.
ried by the other component is L, and its failure ::ate (b) For Y =4 and p 0.3, find the absorption proba-
under this load is (1.5)A. There is only one repair bilities: blO • b 14, bJo• and by..
facility available, and. repair time is exponentially dis~
tributed Wtth mean lip,. The system is considered 18-7. Au ohject moves between fot:! points on a circle,
failed when both components are in the failed state. which are labeled 1. 2, 3, and 4, 11:~ probability of
Both components are initially operat:ir:.g. Assume that moving one unit to the right is p, and the probability
J1:> (l.S))", Let the states be as follows: of mO'.'ing one unit to the left is 1-P =q. Assume that
the object starts at 1. and let Xn denote the location on
0: No components are failed. t.'1c circle after n steps.
1: O:J.e component is failed and is in repair. (a) Find the one....step transition matrix p.
2: Two components are failed, one is in repair,
(b) FilYJ an expression for the steady stilZe probabili~
one is waiting. and the system is in the failed
condition. ties Pi'
(c) Evaluate the probabilities Pj for p ::::: 0 ..5 and
(a; Detem:ine the !:latrix P of transition probabilities
p 0.8,
associated v.ith interval ill.
(b) Determine the steady state probabilities, 18..8. For the singie-set'/er queueing model presented
(c) Write the system of differential equations that in Section 18·7, sketch the graphs of the following
present the transicct or time-dependent relation- quantities as a function of p:;:;;; Alp.., for 0 < P < L
ships for transitioI!, (a) Probability of no units in the system.
(b) :Mean time in the system,
18 S, A communication satellite is launched via a
R
(b) \\lh31 is the average queue length? (b) Hov.' much time does it take, on average. for a
(c) The telephone company will ir.stall a second professor to get his or her jobs completed?
booth whe4 an arrival would expect to have to (c) If an. economy drive reduced the secretarial force
wait 3 minutes or more for the phone. By how to 1:'\¥o secretaries. what wi1: be the new answers
much must the rate of amvals be increased in to (a) and (b)?
order to justify a second booth?
18~ U. The mean frequency of arrivals at an airpor. is
(d) What is the probability that an arrival VIill ha.ve to 18 planes per hour, and the mean time that a runway is
wait more than 10 minutes for the phone? tied up with an arrival is 2 minutes. How many run~
(e) What is the p:obability that it will: take a person ways VoiD have to be provided so that the probability
more than 10 ;:nir~utes altogether, for the phone of a plane having to wait is 0.201 Ignore finite popu-
and to complete the call? lation effects and make the assumption of exponential
(0 Estimate the :fraction of <:I day that the phone will mJera.'1ival and service times.
be in use.
1&-13. A hotel reservations facility uses inward WATS
18-10. Automobiles arrive at a serviee station in a ran" lines to service customer requests. The mea.r, number
dam manner at a mean rate of 15 per hour. This station of ealls that arrive per hour is 50, and the mean serv~
has o:uy one service position. with a mean serviciI:g ice tim.e for a call is 3 minutes. Assume that interar~
rate of2i customers per hour. Service times are expo~ rival and service times are exponentially distributee..
nentially distributed. There is space fur only the auto~ Calls that arrive when all lines are busy obtaln a busy
mobile being sen'cd and two waiting. If all three signal and a..~ lost from the syste:n.
spaces are filled., an miving automobile will go on to (a) Flnd the steady state equations for dlls system,
another station. (b) How many WATS Jines must be provided to
(a) "''hat is the average ll'J.mber of units in the ensure that the probability of a customer obtain-
station? ing a busy signal is 0.05?
(b) \¥bat fraction of customers VIill be lost? (c) Mat fraetion of the time are all WATS lines
(e) \VbyisL,>'L-l? busy?
(d) Suppose that during the evening hours cal:
18-11. An engineering school has three secretaries in arrivals occur at a mean rate of 10 per hour, How
its general office.. Professors wim jobs for the secre- does this affect the WATS line utilization?
taries arrive at random. at an average rate of 20 per (e) Suppose the estimated mean serv:ce time (3 mill-
8-hour day. The amount of time that a secretary utes) is in enor, and the true mean service rin::e is
spe::1ds on a job has an exponential distribution v."ith a really.5 minutes. Wha: effect INill this have on the
mean of 40 minutes. probability of a customer finding all lines busy if
(a) What fraction of the time a."'C the secretaries busy? the number of lines in (b) are used?
Chapter 19
Computer Simulation
One of the ",ast widespread application. of probability and stati.tics lies in the use of com-
puter simulation methods. A simulation is simply an imitation of the operation of a real-
world system for purposes of evaluating that system. Over the past 20 years t computer
simulation has enjoyed a great deal of popularity in the manufacturing, production, logis-
tics. service, and financial industries) to name just a few areas of application. Simulations
are often used to analyze systems that are too complicated to attack via analytic methods
such as queueing theory, We are primarHy interested in simulations that are:
ice times. The subject of Section 19-213 the development of techniques to generate random
variables. Some of these techniques have already been alluded to in previous chapters, but
we will give a more complete and self-contained presentation here. After a simulation run
is completed, one must conduct a rigorous analysis of the resulting output, a task made dif-
ficult because simulation output, for example customer waiting times, is almost never inde-
pendent or identically distributed. The problem of output analysis is studied in Section 19-3.
A particularly attractive feature of computer simulation is its ability to allow the experi-
menter to analyze and compare certain scenarios quickly and efficiently, Section 19-4 dis-
cusses methods for reducing the variance of estimators arising from a single scenario, thus
resulting in more-precise statements about system performance, at no additional cost in
simulation run time. \Ve also extend this work by mentioning methods for selecting the best
of a number of competing scenarios. Vle point out here that excellent general references for
the topic of stochastic simulation are Banks. Carson, Nelson, and Nicol (2001) and Law and
Kelton (2000).
576
19-1 Motivational Examples 577
:EXam "'19";1
, p",:,
Coin Flipping
We are interested in simulating independent flips of a fair coin. Of course, :his is a trivial sequence of
Bernoulli trials with success probability p =. 112. but this example serves to show how OnC ea., use sim~
ulation to analyze such a system. First of all we need to generate realizations of heads (H) and tails
('T). each with probability 112. Assuming that the simulation can somehow produce a sequence of
independent uniform (0,1) random numbers. Vi' V2• ••• , we will a.,"bitrarily designate f'llp i as H if we
observe Ui < 0.5, and a flip as T if we observe Vi 2: 0.5. HQ'IN one generates independent uniforms is
the subject of Section 19~2. In any case, suppose tha: the following uniforr.:.s are observed:
032 0.41 0,06 0.93 0,82 0.L9 0,21 0,77 0,71 0,08.
This sequence of uuifonns corresponds to the outcomes HHHTIHHTTH. The reader is asked to
study this exan:ple in various ways in Exercise 19--1. This type of "static" simulation. in which we
simply repeat the same type of trials over and over, has come to be known as Monte Carlo simt;lation,
in honor of the European city-state, where gambling is a populru: recreational activity.
Estimate ::r
In this example, we 'Will estimate 1r using Mon~e Carlo simulation in conjunction with a simple geo~
metric relation. Referring to Fig. 19~1. consider a unit square with an L"lscribed eL."de, both centered
at {l/2.1/2). If one were to tlu:ow darts randomly at the square. the probability that a particular dart
willla..d in the circle is ref4, t."le ratio of the circle's area to that of the square. Hew can we use this
simple fact to estimat:e :r? We shall use Monte Carlo sirrl'..:lation to throw many darts at the square.
Specifically, genera:e indepehdent pairs of .independent '..Liiform (0,1) random variables. {U;!. Ud.
(U'2"' U'1.2)' .... Tnese pairs will fall randomlY on the square. If, for pair i. itbappens 'j}at
(19-1)
t."len that pair will also fall Vlithin the circle. Suppose we run the experiment for II. pairs {darts). Let
X, I if pair i satisfies ineqmlity 19-1. t.ltat is, if the itb dart falls in the c;rcle; otherwise,let Xi "'" O.
Now count up the number of d~'"tS X "'" L~JX{ falling in the circ:e, Clearly, X has the binoro.ial
dis;;ribution with parameters n and p "'" 1r/4. Then the proportionp .= Xlr. is the maximum likelihood
estimate for p = ;cf4, and so the rn.aximum likelihood estimator for ;cis justj"", 4ft. If, for instance,
(0.1)
........ ....... ..
r."
, _. ..•••••
~- .. ~
..... .
...
<t.#
~
:;
~
.
'[-.........
J.
•
................
. -,... -
.. ........ "
.. - . ':''':.... .
. . . , . "'. •
••
'.:~
\
~
I·
:. It
• ,. I: t , •• _:c -::
~..
.-:.
.·c·.... · ... '
!. . . ._ .. c"
• • • I
(O,Oj: • • • • •.,... ···:.·.·:(1,0)
Figure 19~1 Throv.ing darts ~o estimate i!..
578 Chapter 19 Computer Simulation
we conducted n = 1000 trials and obscrvedX =:: 753 dart5 in the circle, OUI estimate wocld be,r::::::: 3.12.
We will encounter this estimation technique again in Exercise 19-2.
(19-2)
As described in Fig. 19-2, we shall estimate the value of this integral by su:mmiog up n rectangles.
each ohvidth lin cen~ randorr.ly at point U; on [O,lJ. and of hcightj(a + (b -a)U), Then an esti-
mate for I is
(19-3;
One can show (see Exercise 19-3) that ~ is an unbiased estimator for I. that is, EV~J : : :. 1 for all n. Tris
makes Z, an in!:'citive and attractive estimator.
To illustrate. suppose that we wa:.t to est:ir:1ate the integra:
wbicb. is close to the actual answer of 1. See Exercise 19-4 for additi.o~a: Monte Carlo integration
examples,
A Single.Server Queue
Now the goal is to simulate the behavior of a single-server queueing system. Suppose that s:x cus-
tomers arrive at a bank at the following times, which have been generated from some approp;::ia:c
probability distribution:
3 4 6 10 15 20.
UpOn arrival, customers q~eue up in front of a single teller and are processed sequentially, in a fust-
come~:first~served manner. The &erVice times corresponding to the a..'1iving customers are
7 6 4 6 2.
For this example, we assume t."at the bank opens at time 0 and closes its doors at t~me 20 (just after
custorr.er 6 arrives), serving any remaicing customers.
Table 19~1 and Fig. 19-3 trace the evolution of the system as time prOg.Tesses. The table keeps
track of ~e times at which customers arrive, begin semce, <!Ild leave, Figure 19-3 g.Taphs the status
of the queue as a function of time; in particular, it graphsL(t), the number of customers:""l the system
(queue + service) at time t.
Note that cusromer i can begin service only at tinte ma~ (A" Df-.:)' that: is, the maximum of his
arrival time and the previous customer's depanu:re time. The table und figure are quite easy to inter·
pret. For insta.4.ce, the system is empty until time 3, when customer 1 arrives. At time 4, customer '2
a..'1ives, but must wait in line until Cus~mer 1 finishes service at time lO. We see from llc figure :hat
between times 2.0 and 26, eustomer 4 is in scrv::ee, while customers 5 and 6 wait in me queue. From
the table, tbe average waiting time for :he six customers is l:.~l W,16;;;;;; 4416. Further, the avcro.ge num-
ber of customers it:. the system is 1;9 L(t)d!129 "'" 70129, where we have computed the integral by
adding up :he rectangles in Fig. 19-3. Exercise 19-5 looks at extensions of :be single-server queue,
Many simulation software packages provide Simple ways to model und analyze more-eomplicated
queueing networks.
r 5
Queu e i
Customerl
I
i
3 4 4 5 6
i
-~
.-
2 2 3 3 4 5 6
i
1
i
'1 , I
I
I"~~:1"E> I ; 1 i
2
121
3
I
4 5 6
~ 15 16
3 4 6 10 20 2627 29
Figure 19~3 Number of customers L(t) in single~server queueing system.
. EXamJilei?,~'
(s, S) Inventory Policy
Customer orders for a particular good arrive at a store every day. DurIDg a certain one-week period.
the quantities ordered are
10 6 11 3 20 6 8.
The store starts the week off with an initiai stock of 20. If the stock falls to 5 or below, the owner
orders enough from a central warehouse to replenish the stock to 20. Such replenishment orders are
p:laced only at the end of the day and are rece~ved before the store opens the next day. There arc r.o
cusmmer back orders, so any custo:r.:ter orders that are DOt filled immediatelY a..--e lost. This is called
an (s, S) inventory sYStem, where the inventory is replenisbed to S -;: ;:; 20 whenever it hits level s =: 5.
The fonowing is a history for this system:
We see that at the end of days 2 and 5, replenishment orders were made, In .particular. on day 5, the
store ran out of stock and lost 14 orders as a result. See Exercise 19-6.
ice times, and order quantities. This section discusses methods to generate such random
variables automatically. The generation of uniform (0,1) random variables is a good pI",.
to start, especially since it turns out that uniform (a,l) generation forms the basis for the
generation of all other random variables.
E~~iiii?-6
Consider the "'toy" generator Xi = (SXf-1 + 1) mod (8), v.-ith seed Xo ;: O. This produces the .:nteger
sequence X: 1,X2= 6, X:) ::.:: 7, X 4 =4,Xs =5, X6 = 2., X, 3, Xl! O. whereupon things start repeat-
ing, or "cycling." The PRl'i's corresponding to the sequence sta."ting with seedXo;..:;;; 0 are therefore VI ;;;;;
li8, Uz = 618, UJ =7/8, U4 ;:4/8, U~ =5/8, U6 := 2/8, U, "'" 3fS. Us:::; O. Since an;y seed ev·enJ:ual.:y pr0-
duces all integers 0, 1, •.. , 7. we say that this is ajt.tU-cycle ~orfull period) generator.
~Pl.el<j:t
Not all generators are full period. Consider another ''toy'' generator Xi = (3Xi _ 1;.. 1) mod (7), 'With seed
Xc.;; O. This ptoeuces the integer sequence X) =: 1,Xz =4, X3 6, X", =5,Xj :::; 2,X6 =0, whereupon
cycling ensues. Further, notice that for this generator, a seed of Xc =.3 produces the sequence Xl := 3 =
Xl =X3"'" "', not very :random looking!
~-----------------------
The cycle length oft."e generator from ExampJe 19·7 obviously depends on the seed
chosen,. which is a disadvantage, Full-period generators; such as that studied in Example
19-6 obviously avoid this problem, A full~period generator with a long cycle length is given
7
.~~".T;iI'I~J?;~
The generator XI = 16807 X,_l mod (2 31 -1) is full period. Since c;;= 0, this generator is termed a mul-
tipiicative LeG and must be \!sed with a seed XO ~ O. This generator is used in many rea1~world a.ppli~
cations and passes most s~atistical !ests for uniformity and randomness, Lrt order to avoid integer
overllow and real-arithmetic round~off problems. Bratley, Fox, and Schrage (1987) oner the follow-
ing Fortran implementation scheme for this algorithm.
F.mCTION ltrJIF(IX)
Kl '" IX/l27773
IX = 16807*(IX - K1~127773) - K1~2836
:p (IX.LT.O)IX ~ IX + 2147483647
UNIF ~ IX ~ 4,6566l2S75E-lO
=uRN
END
L'1 the above prognu:n, we input a.."I. integer seed IX and receive a PRN UNIF. The seed IX is auto-
rr.atical2.y updated for the next calL ?-;'ote t.1.at b Fortran, integer division results in trun.cation., for
example 15/4 3; thus I<J. is an wtege:.
Theorem 19·1
If X is a random variable with cumulative distribution function (CDF) F(x), then the random
variable Y = FCX') has the unifOffil (0,1) distribution.
Proof
For ease of exposition. suppose thatX is a continuous random variable. Then the CDF of fis
G(y)=P(Y"y)
P(F(X')" y)
= P(X" F'; (y)) (the inverse exists since FCx) is continuous)
= F(P-I(y))
=y.
Since G(y) = y is the CDF of the unifonn (0,1) distribution, we are done,
With Theorem 19-1 in hand. it is easy to generate certain random variables. All one has
to do is the following:
1_ Find the CDF of X, say F(x).
19-2 Generation of Random Variables 583
l dislributions.
I Here we generate an exponential random variable with rate A, fol1owi!:g the :ecipe outlined above.
1. The COP is F(x) = l-e''-'.
2. Se: F(X) = 1 - e-'X ~ U.
Now we try to generate a standa."'d normal random variable, ca!l itZ. Using the special notation <D(.) for
the standard nor=! (0,1) COP, we set <1>(2) U. so that Z = <I>-I(U). t:::.fortunate1y, :he ,>verse COP
does not exist in closed fonn, so one must reso~ to the use of standard Doana! tables (or other approxi-
mations). For instan~ if we have U"" 0.72, then Table II (Appendix) yields Z"".p-l(O,72) ""0583.
'~~I~1?i'(i'"
\Ve can ex:end the previous example to generate any normal random variable, that is. one with a...'""hi-
crary !!lean and variance. 'This fQilows easily. since ifZ is standard normal, then X "")1+ aZ is llormcl
with mean )1 and variance (5'2. For inst3nce, suppose we are interested in generating a r.orroal ,'Briate
X 'Rith mean)1 = 3 and variance (J(.::;::: 4. Then if, as in the previous example, U= 0.72, we obtainZ""
0.533, and, as.a consequence, X o:::! 3 + 2(0.583) =4.166.
~:E;~~p,I:e'Q,I~
We C3.f~ ilio use the ideas:from Theorem 19# 1 to generate realizations from discrete random variables,
SUppose that the discrete random variable X has probability function
rO.3 ifx=-l,
-' \-to.6
P\b/-
ifx=2.3.
0.1 ifx=?
o otherwise.
To generate variates frore this distribution. we set up:he following table, where F(x) is the associated
CDF and U denotes the set ofunifonn (0,1) PR.,'\'"s cor;esponding to each x~value:
x F(x) U
-1 0.3 D.3 [0,0.3)
2.3 0.6 0.9 [0.3,0.9)
7 0.1 1.0 (0.9.1.0)
To generate a realization of X. we first generate a PRN U and then read the corresponding x-value
from the table. For instance. if U::;::: 0.43, then X =: 2.3.
l
584 Ch.apter 19 Computer Simulation
Box-M.illier M.ethod
The Box-Muller (1958) method is an exact technique for generating independent and iden-
tically distributed (lID) standard normal (0,1) random variables. The appropriate theorem.
stated without proof. is
Thwrem 19-2
Suppose that U, and U, are lID uniform (0,1) random variables. Then
Z[ = ,j-2ln{U[) cos(2n;U,)
and
Suppose that Vt ::::: 0.35 and U. . ;;; 0.65 are two Ill) PRNs. Using the Box-Milllermethod to generate
two normal (0.1) random variates. we obtain
Z?[U, - E[l:;"U,]
c··· "
,IV ar( l:Z" U, I
" ,
:=:
I:7-1Ui -
r-'
Ii-lE[Ud
"l:~.,var(U,)
= 17",1UI -(n/2)
-In/12
=:-:1(0,1).
- ,
19·2 Generation of Random Variables 585
In particular, the choice n = 12 (which turns out to be "large enough") yields the conven-
ient approximation
12
2..:Ui - 6 ~ N (0,1),
j""l
Convolution
Another popular trick involves the generation of random variables via convolution, indi-
cating that some sort of sum is involved.
Suppose that X~, Xl' ... , X" are lID exponential random variables with rate A.. Then Y = L:IX1 is said
to have an Erlang distribution with para."1leters n and A.. It turn.s out that this distribution has proba-
bility density function
(19-4)
whid:, readers may recognize as a special ca.qe of the gamma distribution (see Exercise 19-16),
This distribution's CDF is too difficult to invert directly. One way that comes to mi.'1d to gener-
a:e a ~tion from the BrIang is simply to generate and then add up n TID cxponcntia1().,) random
variables. The following scheme is an efficient way to do precisely that. Suppose that Vj • V z• .. "' Ur
are TID PR..~s< From Example 19~9, wc know thatXr=-xln(l- UI)' i = 1, 2 • ... ,n. are lID expone:r:.~
tial()..) rand.om variables. 'I1?erefore, we can ",,'rite
This iu1plcmematiou is quite efficient., sir.ce it requires only one execution of a natural log operation,
In fact:., we can even do slightly better from an efficiency point of vicw-si.mp:y note that both [I" a:..d
(1- U) a;e uniform (0.1). Then
To illustrate, suppose that we have three lID PRNs at our disposal. VI = 0.23, U;. =0.97, and U3 =
0.48. To generate an Erlang realization V.1th parameters n::.;:: 3 and ,1.= 2.we simply take
Acceptance-Rejection
One of the most popular classes of random variate generation procedures proceeds by sam-
pling PRl.'is until some appropriate "acceptance)1 criterion is met.
',J':ii~tn~t~!9,;i~~
An easy example of the acceptance-rejection technique involves the generation of a geometric :taIl-
dom variable with snccess probability p. To this end, consider a sequence of PRNs UI • U'), • .... Our
aim is to genet3.te a geometric realizatior. X, t.hat is,. one that has probability function
'(
p(xi=1)-P )"-''p lfx =1.2.....
, l O. othenvise.
In words. X represents the number of Bernoulli trials nntil the first success is obse.....'ed, 'This English
characterizatior. immcdiately suggests an elemenTary acceptance-rejection algorithm.
1. Initialize i (- O.
2. Leti t:-i ..... 1.
3. Take a Bemoulli(p) observation,
{I
11 = 0,
ifUi <po
otherwise.
4. If Y,::::: I, then we have our first snccess Md We StOp. in which case We accept X=::> i, Other~
wise, if Y/=O, then we reject and go back to step 2.
To illustrate, let US generate a geometric variate having success probabl.!ity p ;::: 0.3. Suppose we have
ore at our disposal the fonowing PRNs:
Since U~ =0,38 cp, we have Y1-=0, and so we reject X = 1. Since U2 ::;; 0,67 cp. we have 1:2 ::;;0, and
so we reject X =2. Since [13::;; 0,24 <p, werurve Y3;;:::: 1, and so we acceptX=3,
The problem is that simulations almost never produce convenient raw output that is ffi)
normal data, For example. consecutive customer waiting times from a queueing system
are not independent-typically, they are serially correlated; if one customer at the
post office waits in line a long time, then the next customer is also likely to wait a
longtime.
• are not identically distributed; customers showing up early in the morning might
have a much shorter walt than those who show up just before closing time.
are not normally distributed-they are usually skewed to the right (and are certainly
never less than zero).
The point is that it is difficult to apply "classical'" statistical techniques to the analysis
of simulation Output Our purpose here is to give methods to perform statistical analysis of
output from discrete-event computer simulations. To facilitate the presentation, we identify
two types of simulations with respect to output analysis: Terminati.'1g and steady state
simulations.
1. 'terminating (or transient) simulations, Here, the nature of the problem explicitly
defines the length oftbe simulation run. For instance, we might be interested in sim~
ulating a bank that closes at a specific time each day.
2. Nollterminating (steady stare) simulations. Here, the long-run behavior of the sys-
tem is studied. Presumedly this "steady state>~ behavior is independent of the simula-
tion's initial conditions, l\.n example is that of a continuously running production line
for which the experimenter is interested in some long~run performance measure.
Techniques to analyze output from tenninating simulations are based on the method of
independenlreplications, discussed in Section 1Y,3. I. Additional prob1ems arise for steady
state simulations. For instance, we must now worry about the problem .of starting the
simulation-how should it be initialized at time zero, and how long must it be run before
data representative of steady state can be collected? Initialization problems are considered
in Section 19,3.2. Finally, Section 19-3.3 deals with point and confidence interval estima-
tion for steady state simulation performance parameters.
observations m can be a constant or a random variable. For example} the experimentcr can
specify the number m of customer waiting times Yl • Yt.? .. " Ym to be taken from a queueing
simulation, Or m could denote the random number of customers observed during a speci-
fied time period [0, TJ.
Alternatively, we might observe continuous simulation output {Y(t)IO ~ t ~ T} over a
specified interval [0, TJ. For instance, if we are interested in estimating the time-averaged
number of customers waiting in a queue during [0, 1], the quantity Y(t) would be the num-
ber of customers in the queue at time I.
The easiest goal is to estimate the expected value of the sam.ple mean of the
observations,
e" E[YmJ,
where the sample mean in the discrete case is
_ 1 m
Ym=-I;r,
m "'.1
588 Chapter 19 Computer Sirculation
(with a similar expression for the continuous case). For instance. we might be interested in
estimating the expected average waiting time of all customers at a shopping cemer during
the period 10 a.ill, to 2 p,m.
e,
Although Ym is an unbiased est.imator for a proper statistical analysis requires that we
also pro\'ide an estimate of Var{Ym). Since the Y" are not necessarily IID random variables,
it may be that Var(Ym) :;i: Var(Y{)/m for any i a case not covered in elementary statistics
j
courses,
For this reason. the familiar sample variance
2. 1 m. _ 2
S =-~(y,-y)
m- 1~'
i=l
t m
is likely to be highly biased as an estimator of mVar(i'm)' Thus, one sbould not use [{'1m to
estimate Var{i'm)'
The way around the problem is via the method of independent replications (IR). IR
estimates Var(i'm) by conducting b independem simulation runs (replications) of the system
under study, where each replication consists of m observations.l! is eru;y to make the repli-
cations independent-simply reinltialize each replication with a different pseudorandom
number seed.
To proceed, denote the sample mean from replication i by
, 1 b _ 2
VR "-2:;(Z, -Z,) ,
b-l ;",1
Notice how dosely the forms of lin and Slim resemble each other, But si:l.ce the replicate
sample means are IID, VR is usually much less biased for Var(Ym) than is S2/m.
In light of the above discussion, we see that V"Ib is a reasonable estimator for Var(Z,,).
Further. if the number of observations per replication, m, is large enough, the Central Limit
Theorem tells us that the replicate sample means are approximately lID normal.
Then basic statistics (Chapter 10) yields an approximate 100(1- a)% two-sided con-
fidence interval (eI) for e,
(19-5)
where '''''''-' is the 1 - cd2 percentage point of the t distribution with b -1 degrees of freedom.
19-3 OutpUt Analysis 589
Suppose we want to estimate the expected avc:age waiting time for the first 5000 custOmerS i::i a ccr~
tam queueing system, We wiil make five indepe:.:dent replications of th.: system, v.itl:. each run ini-
tialized e:npty and idle and consisting of 5000 waiting times. The resulting replicate means are
2 3 4 5
Z, 32 4.3 5.1 4.2 4.6
Then Zs = 4,28 andV:~ "'" 0.487, For level a::= 0.05. we have tom1,4:= 2.72. and equation 19~5 gives
[3.41.5.153 as a 95% CI for the expec'".ed average waiting time for the first 5000 customers.
Independent replications can be used to calculate variance estimates for statistics other
t.i.an sample means. Then the method can be used to obtain CIs for quantities oilier than
E[Ym ], for example quantiles, See any of the standard simulation teXts for additional uses
of independent replications.
Since initialization bias raises important concerns, how do we detect and &a1 with it?
We first list methods to detect it.
2. Conduct statistica.l testsJor initialization bias, Kelton and Law (1983) give anintu-
itively appealing sequential procedure to detect bias. Various omer tests check to see
whether the initial portion of the simulation output contains mOre variation than lat~
ter portions,
If initialization bias is detected, one may want to do something about it. There are tv/o
simple methods for dealing with bias. One is to truncate the output by allowing the simu~
!ation to '''warm up" before data are retained for analysis. The experimenter would then
hope that the remaining data are representative of the steady state system. Output truncation
is probably the most popular method for dealing with initialization bias, and all of the
major simulation languages have built-in truncation functions. But how can one find a good
truncation point? If the output is truncated "too early," significant bias might still exist in the
remaining data. If it is truncated "too late," then good observations might be wasted. Unfor-
tunately, all simple rules to determine tr..mcation pomts do not perfort:l well in general, A
common practice lS to average observations across several replications and then visually
choose a truncation point hased on the averaged run. See Welch (1983) for a good
visual/graphical approach,
TIle second method is to rr.ake a very long mn to overwhelm the effects of initializa-
tion bias. This method of bias control is conceptually simple to carry out and may yield
point estimators having lower mean squared errors than the analogous estimators from
truncated data (see, <,g.• Fisbman 1978). However, a problem with :his approach is that it
can be wasteful with observations; for some systems, an excessive run length might be
required before the initialization effects are rendered negligible.
BatchMcam
The method of batch means is often used to estimate 11lr(Yn) or calculate as for the steady
state process mean JL The idea is to divide one long simulation ron into a number of con-
tiguous batches, and then to appeal to a Central Limit Theorem to assume that the resulting
batch sample means are approximately IID normal.
In particular, suppose that We partition Yt , Y:;:> •. , Y" into b nonoverlapping, contiguous
j
batches, eacb consisting of m observations (assume that n =bm). Thus, the itb hatcb, i;::; 1,
2, ... ! b, consists of the randoI<l variables
The ith batch mean is siIDply the sample mean of the mobservations from batch i, i= 1,2, ... , b,
1 m
Z, =- L~H)m+r
m 1",,1
19-4 Comparison of Systems 591
wbere
_ _ 1 b
y" = Zb = -;- I,Z,
/) ;.>1
is the grand sample mean. If m is large. then the batcb means are approxixnately IID nor-
mal, and (as for IR) we obrain an approximate 100(1 - a)% Clfar f1,
Independent Replications
Of the difficulties encountered when using batch means, the possibility of correlation
among the batch means might be the most troublesome. This problem is explicitly avoided
by the method of IR, described in the context of terminating simulations in Section 19-3.1-
In fact, the replicate means are independent by their construction. Unfortunately, since
each of the b replications has to be started properly. initialization bias presents more trou-
ble when using IR than when using batch means, The usual recommendation. in the context
of steady state analysis., is to use batch means over IR because of the possible initialization
bias in each of the replications.
and
where the approximate degrees of freedom v (a function of the sample variances) is given
in Chapter 10,
Suppose (as in airline example) that small cost is good. Jf the interval lies entirely to
the left [rigbt] of zero, then system I [2] is better; if the interval contains zero, then the two
systems must be regarded, in a statistical sense, as about the same,
An alternative classical strategy is to use a C1 that is analogous to a paired Hest. Here
we take b replications from both strategies and set the differences Dj =Z',J - z,J forj = I,
2, ... ) b. Then we calculate the sample mean and variance of the differences:
_ 1b ? 1 b _2
Db=-LD, and S;;=-L(Dj-Db) '
b i"",l • b-l J",l
These paired tintervals are very efficient if Corr(Z;,J' 2zJ) > O,j= I, 2, '''' b (where we
still assume that 2"" Z,," ' .. , Z"b are lID and z,,l' z"" "" z"" are lID), In that case, it rums
out that
could be very large, in which case our declaration might lack conviction. H we could reduce
V(9A -fiB)' then We could be much more confident about our declaration.
eRN sometimes induces a high positive correlation between the point estimators ~A
and 88' Then we have
V(~A 1I,l = V(~A) + V(ils) - 2eav(llA' e.)
< V(~A) - V(~D)'
and we obtain a savings in variance.
19·5 SUMMARY
This chapter bega!1 vrith SOme simple motivational examples illustrating ,,'arious simulation
concepts. After this) the discussion t'.lrned to the generation of pseudorandom numbers, that
is. numbers that appear to be lID unifO!D1 (0,1). PRNs are important because they drive the
generation of a number of other important random variables) for example normal. expo~
nential, and Erlang. We also spent a great deal of discussion On simulation output analy-
sis--simulation output is almost never IID, so special care must be taken if we are to make
594 Chapter 19 Computer Si..llulation
19·6 EXERCISES
19~1. Extension of Example 19.. 1. (c) Vlhat is the maximum number of customers in the
(a) F'.up a coin 100 times. How n-..:my heads to do you system? When is this maximum achieved?
observe? (d) What is the average number of customers in line
(b) How many times do you observe two heads in a during the first 30 minutes?
row? T1m::e in a row? Four? Five? (e) Now repeat parts (a)-{d) assuming t.tlat the serv~
(c) Find 10 friends and repeat (a) and (b) based on a ices are performed last-in-mst-out.
rotal of 1000 flips. 19-6_ Repeat Example 19-5, whichdea], with an (s, S)
(d) Now simulate eoin flips via a spreadsheet pro- im'entory pOlicy, except now use order level s.:c::: 6.
gram. Flip the simulated coin 10,000 times and
19-7. Consider the pseudorandom number generator
answer (a) and (h).
X, = (5X,-, + I) mOO (16), with seed A;, = O.
19-2. Extension ofRxampie 19,.2. Throw 11 dat"sran-
(a) Calculate Xl andX1, along with the eorresponding
domly at a unit square cootai.ning an inscribed circle.
PR..1I.fs U; and Uz'
Use the results of your tosses to estimate 1C Let r. =:: 2'<
(b) Is this afullwperiod generator?
for k ::.:: 1. 2, ...• 15. and graph your estimates as a
function of k. (c) What is XJ5I)?
19~3. Extension of Example 19~3~ Show thd t. 19~8. Consider the "recoIU."Tlended" pseudorandom
defined in equation 19-3, is ur.biased for the integral I, n'J:JLber geoerator XI = 16807 Xi-I mod (231 - 1\ with
defined in equation 19~2. seed Xc = 1234567.
19-4. Other c-x1.ensions ofRxample 19>-3. (a) Calculate X; andXz, along with the corresponding
(a) Use Monte Carlo :.ntegration with n=10 observa- PR-'1s Vi and V,.
tions to estimate ~e-tl/2.dx. Now use r. =
r2 (h; What.isXtoo.ooo?
J() 2n 19-9~ Show how to use the inverse transform method
1000. Compare to the answer '±tat you c.a.."1 obtain
to generate an exponential ralldom va:i.able with rate
via nonna] tables.
A;;;;;: 2. Demonsnte your technique using :he PRN
(b) Vt'hat would you do if you had to estimate V=0.75.
r1Q J..... e-;;.1/2dx? 19-10. Consider the inverse transformrnethod to gen-
Jo 21r erate a standard normal (0, 1) random variable.
(c) Use Monte Carlo integration withn;;;;;: 10 observa~
(<1) DemOllSt.n1te your technique using the PRN
tions to estimate I~ COS(21CC) dx. Now use n =: V=O.25.
1000. Compare to the actual answer,
(b) Using your answer in (a). generate a:;N 0,9) ran-
19-5. Extension of Example 194. Suppose that 10 dom variable.
customers arrive at a post office at the following
times: 19~11. SU:pPose thatXhas probability density funetion
j(x) = !xl41, -2 <x< 2.
3 4 6 7 13 14 20 25· 28 30
(a) Develop an inverse transfonn technique to gener-
Upon arrival, eustomers queue up in £::ont of a single ate a realization of X
clerk and are processed in a first-comewfirst-served
(b) Demo:J.strate yQW'tech:uque usiDg U:: 0.6.
JJ1a!mer. The service times corresponding to the arriv-
ing customers are as follows:: (c) Sketch out/ex) and see if you can come up with
another method to geoerate X.
6.0 5,5 4.0 1.0 2.5 2.0 2.0 2.5 4.0 2.5
19-12. Suppose that:he discrete random variable X
Assume that the post office opens at time 0, and closes
has probability function
its doors at time 30 Gust after customer 10 arrives),
SIDing any remaining customers. (0.35 if x =-2.5.
(a) wnen does the last customer fillally leave the ifx =: 1.0,
system?
p(x)= 10.25
0. 40 ifx= 10.5.
(b) 'What is the average waitiIlg time for :he 10
C'Jstomers ? 1 O. otherwise.
19-6 E.--.::ercises 595
As ia Exa.-r.ple 19· ~2, set up a table 'X) generate real- 19~22. The yearly unemployment rates for Andorra
izations from this distribution. Illustrate yO'GC tech- daring the past 15 years are as follows:
nique with the PRN U = 0,86, 6,9 8,3 8.8 11A 11.8 12.1 10,6 11.0
19-13. The Weibull (a, (3) distributiou. popular inreli· 9.9 9.2 12.3 13.9 9.2 8.2 8.9
ability theory and o:her applied statistics disciplines, Use the method of batch means on the above data to
has CDP obtain a two-sided 95% confidence interval for the
mean unemployment. Use five batches, each consist~
ifx> 0,
ing of tbJ:ee years' data
otherwise.
19,,23. Suppose that we are interested in steady state
(a) Show how to us.e the inverse transform method confidence intervals for the mean of simulation ou~ut
to generate a realization from the Weibull Xl' X2• X1COOO" (You can pre:end :hat t.'iese are wail-
"'j
iii, Would you expect that Cov(X;, 11) < 07 where the Ui a:e IID uniform (0.1). 'Vv'hat ki!1d of dis~
Answer. Yes, tribution does it look like?
(b) LetI'" and f,; denote the sample means of the-X, 19~28. Another miscellaneous computer exercise.
aed Y" res;;ectively, each based on n observations. Let us see if the Cent:ral Limit Theorem works. In
'Ni::hout actua:ly calculating Cov(Xi' Yi), state how Exercise 19~27> you generated 20,000 exponcntial(l)
V(O!" ~ ))12) compares [0 V(X",). k other words. observations, Now form :000 averages 0[20 observa~
should we do two negativelY correlated runs, each tiollS each fro:.r.. theoriginal20,QOO. More precisely, let
consisting of n observations, or just one run con-
sisting of 2r. observations? I '"
' " X,,(,
Y, : -20..i..J . '.
-'" ,-.J"'J
(e) What if you tried to use tills trick when using .1",1
Monte Carlo simulation to estimate J~ sin(n:x) d:x?
Make a histogram of the Yr DQ they look approxi M
Suppose we SOITh':how know the expected value of cise. Let us generate some Donnal observations via
some other RV Y. and we also know that Cov(X, 1) > the Box-Mtl1cr method. To do so, fust generate 1000
0, where X is the sample me<l.."l. Obviously, X is the pairs of liD uniform (0,1) flndoID nu...rnbers, (U J ,!'
"usual" estimator for jL Let us look at another est:bla M Ul. l )' (0;,::, U2;2)' .". (UU:lJo, U2 ,JOIX;). Set
tor for j.!., namely. the control-variate estimator, r·-·-, .
XI ~ r21n(Ulj)cos(2rrU"iJ
c~X - kiY -E(l']),
where k is some consmnt. and
(a) Show that C is unbiased for jl,
(b} Find an expression for V{C). Commems'?
(c) Minimize Vee) with respect to k.
for i;; 1,2, ,." 1000, Make a hisrogram of the result-
19~.27.A miscellaneous computer exercise. Make a ingX" [The Xl'S areN(O,l),) Now gtaphX; vs. Y,> Any
histogram of XI ~ -In(U), for i ~ I, 2, ... , 20,000, commcn~s?
Appendix
597
598 Appendix
c:= iJ.
x 0,70 0,80 0,90 LOO LlO 1.20 1.30 lAO
0 0.496 0,449 0.406 0,367 0.332 0.301 0.272 0.246
Co:::: Ar
x 1.50 1,60 1.70 1.80 1.90 2.00 2,10 2,20
C=N
x 230 2.40 2.50 2,60 2.70 2,80 2.90 3,00
c::;;: j"t
x 3.50 4.00 4.50 5.00 5.50 6.00 6.50 7,00
{I,030
°
1 0,135
0.018
0,091
0.011
0,061
0,006
0,040
0.004
0.026
0.002
0,017
0,001
0.011
0.000
0,007
2 0.320 0,238 0.173 Q,!24 0,088 0.061 0.043 0,029
3 0536 0,433 0,342 0,265 0,201 0.151 0,111 0.081
4 0,725 0.628 0.532 0,440 0.357 0.285 0.223 0.172
5 0,857 0,785 0,702 0.615 0.528 0,445 0,369 0.300
6 0,934 0,889 0,831 0,762 0,686 0,606 0526 0.449
7 0,973 0,948 0,913 0,866 0,809 0<743 0,672 0.598
8 0,990 0,978 0,959 0,931 0,894 0,847 0,791 0.729
9 0.996 0,991 0,982 0.%8 0,946 0,916 0,877 0,830
10 0,998 0,997 0,993 0.986 0,97L 0,957 0.933 0,901
11 0.999 0,999 0,997 0,994 0,989 0.979 0,966 0.946
12 0,999 0,999 0,999 0,997 0,995 0<991 0,983 0,973
13 0,999 0,999 0,999 0.999 0.998 0,996 0.992 0,987
:4 0,999 0,999 0,999 0.999 0,998 0,997 0.994
15 0,999 0,999 0,999 0,999 0,998 0,997
16 0,999 0,999 0,999 0,999 0,999
17 0.999 0,999 0,999 0,999
18 0.999 0.999 0.999
,9 0,999 0.999
20 0.999
(continues)
600 Appendix
<I>(z) = J~ 1
-,=~e
_uu .,
"du
- ,21t
.'_. ~ 0.995 0.990 0.975 0.950 0.900 0.500 0.100 0.050 0.G25 0.010 0.005
1 0.00+ 0.00+ 0.00+ 0.00+ 0.02 0.45 2.71 3.84 5.02 6.63 7.88
2 om 0.02 0.05 0.10 0.21 1.39 4.61 5.99 7.38 9.21 10.60
3 0,07 0.11 0.22 0.35 0.58 2.37 6.25 7.81 9.35 11.34 12.84
4 0.21 0.30 0,48 0.71 1.06 3.36 7.78 9,49 11.14 13.28 14.86
5 0041 0.55 0.83 Ll5 1.61 4.35 9.24 11.07 12.83 15.09 16.75
6 0.68 0.87 :.24 1.64 2.20 5.35 10.65 12.59 14.45 16.81 18.55
7 0.99 124 ~_69 2.17 2.83 6.35 12.02 14.07 16.01 18.48 20.28
,,34 1.65 2.18 2.73 13.36 15.51 17.53 20.09 21.96
9\ 1.73 2.09 2.70 3.33 14.68 16.92 19.02 21.67 23.59 \,:
10 2.16 2.56 325 3.94 15.99 r...J.U! 20,48 23.21 25.19
11 2.60 3.05 3.82 4.57 5.58 10.34 17.28 19.68 21.92 24.72 26.76
r-,''',
'~::'
3.07 3.57 4.40 5.23 6.30 . 11.34. 18.55 21.03 23.34 26.22 28.30
13 3.57 4.11 5,()1 5.89 7.04 12:34' 19.81 22.36 24.74 27.69 29.82
14 4.07 4.66 5.63 657 7.79 13.34 21.06 23.68 26.12 29.14 31.32
.15 4.60 5.23 627 7.26 8.55 14.34 22.31 25.00 27,49 30.58 32.80
16 5.14 5.81 6.91 7.96 9.31 15.34 23.54 26.30 28.85 32.00 34.27
17 5.70 6.41 7.56 8.67 10.09 16.34 24.77 2759 30.19 33.41 35.72
18 6.26 7.01 8.23 9.39 10.87 17.34 25.99 28.87 31.53 34.81 37,16
19 6.84 7.63 8.91 10.12 11.65 18.34 27.20 30.14 32.85 36.19 38.58
20 7.43 8.26 9.59 10.85 12."4 19.34 28.41 31.41 34.17 3757 40.00
21 g.03 8.90 10.28 11.59 1324 2034 29.62 32.67 35.48 38.93 41.40
22 8.64 9.54 10.98 12.34 14.04 21.34 30.81 33.92 36.78 40.29 42.80
23 9.26 10.20 11.69 13.09 14.85 22.3~ 32.01 35.17 38.08 41.64 44.18
24 9.89 10.86 12.40 13.85 15.66 23.34 33.20 36.42 39.36 42.98 45.56
25 JO.52 11.52 13.12 14.61 16.47· 24.3~ .34.28 37.65 4>3.65 44.31 46.93
26 ILl6 12.20 13.84 15.38 17.29 25.34 35.56 38.89 41.92 45.64 48.29
27 11.81 12.88 14.57 16.15 18.11 26.34 36.74 40.11 43.19 46.96 49.65
28 12.46 1357 15.31 16.93 18.94 27.34 37.92 41.34 44.46 48.28 50.99
29 13.12 14.26 16.05 >7.71 19.77 2834 39.09 42.56 45.72 49.59 52.34
30 13.79 14.95 16.79 18.49 20.60 29.34 40.26 43.77 46.98 50.89 53.67
4J 20.71 22,16 24.43 26.51 29.05 39.34 51.8155.76 59.34 63.69 66.77
50 27.99 29.71 32.36 34.76 37.69 49.33 63,17 67.50 7!.42 76.)5 79.49
60 35.53 37.48 40.48 43.19 46.46 59.33 74.40 79.08 83.30 88.38 91.95
70 43.28 45.44 48.76 5).74 55.33 69.33 85.53 90.53 95.02 100.42 104.22
80' 51.17 5354 57.15 60.39 64.28 79.33 96.58 10l.88 106.63 112.33 116.32
90 59.20 61.75 65.65 69.13 73.29 89.33 107.57 113.:4 118.14 124.12 128.30
100 67.33 70.06 7~.22 77.93 82.36 99.33 118.50 12><.34 129.56 135.81 14{i.17
'v:;:: cegreos off:ecdor:n.
604 Appendix
F 025, V!,>'l
v, Degrees of fr~edom for the numerator (VI)
v, 2 4 6 7 8 9 to 12 15 20 24 30 40 60 120
5.83 7.50 8.20 8.58 8.82 8.98 9.lD 9.19 9.26 9.32 9,41 9.49 9.58 9.63 9.67 9.71 9.76 9.80 9.85
2 2.57 3.00 3.15 3.23 3.28 3.31 3.34 3.35 3.37 3.38 3.39 3041 3,43 3,43 3.44 3,45 3,46 3.47 3,48
3 2.02 2.28 2.36 2.39 2,41 2042 2,43 2.44 2.44 2.44 2,45 2046 2,46 2,46 2,47 2,47 2.47 2047 2.47
4 1.81 2.00 2.05 2.06 2.(17 2.08 2.08 2.08 2.0& 2.08 2.0& 2.08 2.08 2.08 2.08 2.08 2.08 2.08 2.08
5 1.69 LaS 1.88 1.&9 1.&9 1.&9 1.89 1.89 1.89 1.89 1.89 1.89 \.88 1.88 1.88 1.88 1.87 1.87 1.87
6 1.62 1.76 1.78 1.79 1.79 1.78 1.78 1.78 1.77 1.77 1.77 1.76 1.76 1.75 1.75 1.75 1.74 1.74 1.74
LS7 1.70 1.72 1.72 1.71 1.71 1.70 1.70 1.70 1.69 1.68 1.68 1.67 1.67 1.66 1.66 1.65 1.65 1.65
1.54 1.66 1.67 1.66 1.66 1.65. I.M 1.64 1.63 1.63 1.62 1.62 1.61 1.60 1.6() 1.59 1.59 1.58 1.5&
9 LSI 1.62 1.63 1.63 1.62 1.61 1.6() 1.60 LS9 1.59 1.58 1.57 U6 LS6 LS5 LS4 1.54 1.53 1.53
ill 1.49 1.60 1.60 1.59 1.59 1.5& 1.57 LS6 LS6 LS5 1.54 1.53 152 LS2 1.51 LSI LSO 1.49 1.48
11 1,47 1.58 LS8 157 LS6 155 1.54 1.53 LS3 1.52 LSI LSO 1.49 1.49 1.48 1,47 1,47 1.46 1.45
"£ 12 1.46 1.56 1.56 1.55 154 1.53 1.52 LSI LSI 1.50 1.49 1,48 1,47 1.46 1,45 1,45 1.44 1.43 1.42
~ 8 1.45 LS5 LS5 LS3 1.52 LSI 1.50 1.49 1.49 1048 1,47 1.46 lAS 1.44 1.43 1.42 1.42 1,41 1,40
i 19
20
21
1,41
1,40
1,40
1,49
1.49
1,48
1,49
1,48
1,48
1,47
1,47
1.46
1.46
1,45
1,44
1.44
1.44
1.43
1.43
1.43
1,42
1,42
1.42
1.41
1.41
1.41
1,40
1,41
1,40
1.39
1,40
1.39
1.38
1.3&
l.37
1.37
1.37
1.36
1.35
1.36
1.35
1.34
1.35
1.34
1.33
1.34
1.33
1.32
1.33
1.32
1.31
1.32
1.31
1.30
1.30
1.29
1.28
'" n 1,40 1,48 1,47 1,45 1.44 1.42 1.41 lAO 1.39 1.39 1.37 LJ6 1.34 1.33 1.32 1.31 1.30 1.29 1.28
£ D
24
e
1.39
1.39
1.39
1,47
1,47
1,47
1,47
1.46
1.46
1,45
1.44
1.44
1.43
1,43
1,42
1.42
1.41
1.41
1.41
1,40
1,40
lAO
1.39
1.39
1.39
1.38
1.38
1.38
1.38
1.37
1.37
'1.36
1.36
1.35
1.35
1.34
1.34
1.33
1.33
1.33
1.32
1.32
1.32
1.31
1.31
1.31
1.30
1.29
1.30
1.29
1.28
1.28
1.28
1.27
1.27
1.26
1.25
U 1.38 1.46 1,45 1.44 1.42 1,41 1.39 1.3& 1.37 1.37 1.35 1.34 1.32 1.31 1.30 1.29 1.28 1.26 1.25
D 1.38 1.46 1,45 1,43 1,42 lAO 1.39 1.38 1.37 1.36 1.35 1.33 1.32 1.31 1.30 1.28 1.27 1.26 1.24
U 1.38 1.46 1,45 1.43 1,41 lAO 1.39 1.38 1.37 1.36 1.34 1.33 1.31 1.30 1.29 1.28 1.27 1.25 1.24
B 1.38 1,45 1,45 1.43 1.41 lAO 1.38 1.37 1)6 1.35 1.34 1.32 1.31 1.30 1.29 1.27 1.26 1.25 1.23
1.28 1.27
30
~
1.38
1.36
1,45
1.44
1.44
1,42
1.42
1.40
1.41
1.39
1.39
[.37
1.38
1.36
1.37
1.35
1.36
1.34
1.35
1.33
1.34
1.31
1.32
1.30
1.30
1.28
1.29
1.26 1.25 1.24
1.26
1.22
1.24
1.21
1.23
Ll9 ;g
60 1.35 1.42 1.41 1.38 1.37 US 1.33 1.32 1.31 1.30 1.29 1.27 1.25 1.24 1.22 1.21 1.19 1.17 LIS g
120 1.34 1,40 1.39 1.37 1.35 1.33 1.31 1.30 1.29 1.28 1.26 1.24 1.22 1.2\ 1.19 LI8 Ll6
Ll2
L13
1.08
1.10
~
1.32 1.39 1.37 1.35 1.33 1.31 1.29 1.2& 1.27 1.25 1.24 1.22 1.19 1.18 1.16 1.14 1.00
Source: Adapted with permhiion from Blome/rikn Tllbfes illr Swtisticillns, Vol. 1, 3rd edition, by E. S. Pear-sOD and H. O. HlITtI<lY, Cambridge Univenity Press, Cambridge, 1966, (ContiHlles)
~
Table V Percentage Points of !he F Distrihution (cmltinlJed)
§l
FO$I.",,\.~,_ _ _-:-_ _ _ _ _ _ _ _ _ _ _ _ _ _ __
D.!grees of fr~om for the numerator (VI} .6'
Y,
v,
2 3 4 6 7 8 9 10 U 15 20 24 30 40 60 120 g
31,l.86 49JO 53.59 55.!B 57.24 58.20 58.91 59.44 5!tS6 60.l9 60.71 61.22 61.74 62.00 62,26 6253 61,79 63JJ6 6133 If
2
3
it~3
5.$4
,.'5A6" 9.16
5.39
9.24
5.34
9.29
5.3]
9.33
5.28
!U5
5,27
9.37
525
9.3&
5,24
9.39
'5.23
9.41
5.22
9.42
5.20
'.44
5.IR
9,45
5.18
9A1i
5,17
9.47
5.16
9.47
'5.15
9,48
5.14
!JA9
5.13
4 4.54 4.'\2 4.19 4,11 4.{l5 4.Ot 198 3.95 3.94 3.92 3.9{) 3.87 3,84 ),Sl 1.S2 ).BO 3.79 1.78 3.76
5 4.06 3.78 3.62 352 't45 'AD 3.17 1.34 3.32 ,UO ]'27 ),24 1.21 3.19 3.17 3.16 ).14 3.12 3.10
6 3.78 1.46 3.29
1.1l7
'1.18 3.11
2.&8
3.05 3.01 2.98 2.96
2.72
2.94 2.90 '.In
2.63
'.84 ,." 2.80
2.56
2.78
254
2.76 2:'4
2.'19
,."
2.47
7 3.59
1.46
1.26
3.11 2.92
2!J6
2.81 '.73
2.33
,."
2.78
2.62
2.75
2.59 2.56
2.70
2.54
2.67
2JO 2.46
'39
2.42
258
2040 2.38
2.25
2:,6
2.21
'"
234
2.21
2.32 2.29
2.16
9 3.36 3.01 2.8\ '~9 2.61 255 B1 2A7 2.44 2.42 2.38 2.34 2.30 2.28 2.18
10 3.29 ,.'" 273 2~1 252 2A6 2A1 2.38 2.35 2.12 2.Ut 2.24 220 Vi! 2.16 2.l3 2.11 2.0& 2.06
~
II ,."
3.lB
2.86
281
2.66
2~1
254
2048
2A5
239
'39
2.33
234
2.26
2.10
2.24
2.27
2.21
2.25
2,19
2.21
2.15
2.17
2.10
2.12
2J16
2.10
,{);
2.0Il
2.01
2,05
1.99
2.03
1.96
'.00
1.93
1.97
1.00
§ I'
13 3.14 2."/6 256 2,43 2.35 2.28 2.23 2.20 :Uti 2,14 2.10 2,05 2.01 1.98 1.96 1.93 1.00 La& 1.85
.~ 14 3.10 2,73 2.52 2.39 2.31 2.24 2.19 2.15 2.12 2.tO 2.05 2.01 1.96 1.94 1.91 Ul9 1.86 l.83 L80
~ IS 3.07 2.70 249 2.% 22"/ 1.21 2.16 2,12 2.09 2.06 2.02 1.97 1.92 1.00 1.87 1.35 l.S2 1.79 Uti
16 3.05 2.67 2,46 233 224 2.18 2.13 2.09 2.M 2.03 1.99 1.94 1.89 L87 Lll4 l.81 1,78 1.75 1.72
<l 17 3.03 '.64 2.44 211 222 2.15 2.10 2.06 2.03 2.00 1.96 1.91 1.86 I.Il4 L31 1.78 U5 1.72 1.69
.ll
" 3.01 1.62 2.42 2,29
= 2.13 2.0B 2.04 2.00 1.98 1.93 Ul9 U4 1.81 Ug 1.75 1.72 1.69 ,Jj6
] 19
20
21
2.99
2.97
,.%
2.61
2.59
2.57
2M)
238
1.36
227
2.21
2.13
2.18
2.16
2.14
2.ll
'J)9
2.08
2.06
2.04
2.02
2.02
2JJO
1.98
1.98
1.96
1.95
1.91i
1.94
1.92
1.91
1.89
l.B7
1.86
I.Il4
1.83
L81
1.79
US
1.79
1.7'1
L75
1.76
1.74
1.72
1.73
1.71
1.69
1.70
1.6&
1.66
1.67
1.64
1.62
If~
1.6J
1.";9
'a
22 2.1,l5 2.56 2.35 2.13 2.06 2.01 1.97 L93 1.90 L86 LSI 1.76 1.73 1.70 1.67 1.<>1 1.60 1,57
I
2.22
23 '.94 255 2:>1 2.21 2.11 ,.OS 1.99 1.95 1.92 1.89 1.84 UlU 1.74 1.72 1.69 1.66 1.62 159 1.55
24 293 254 2.33 2.19 2.10 2M 1.9& 1.94 1.91 1.88 U3 l:/8 1.73 1.70 1.67 1.64 1.61 1.57 1.53
25 2.92 2.53 232 2.18 2J)9 2J)'l 1.97 un L89 1.87 1.82 1.71 1.72 1.69 1.66 1.63 LSI) 156 1.52
U 2.91 252 '31 2.17 2.08 2.01 1.96 1.92 U& Uti 1.81 1.76 1.71 1.68 1.65 1.61 1.58 1.54 1.50
27 '.90 2.51 2JO 2,17 '1JfJ '.00 1.95 1.91 1.87 I.SS 1.80 1.75 1.70 1.67 1.<>1 U50 1.57 153 1.49
28 2.89 2.50 2.29 2.16 2.06 :2.00 1.94 }.90 1.87 U!4 1.79 1.74 1.69 1.66 1.63 1.59 1.56 1..52 1048
21' 1.99 1.91 1.89 L86 un 1.78 1.68 1.6'\ US 1.55 1.51 IA7
"
30
40
2.&9
2.B8
'.ll4
250
2,49
244
2.28
22'
2,23
2.l4
,.'"
'.06
:un
2.'"
1.98
1.93
1.91
L87
UIS
1.83
U15
L79
1.82
1.76
1.77
l.71
1.71
1.72
1.66
1.67
U;1
1.64
LS7
1.62
1.1;1
1.54
LS7
151
1.54
1.47
l50
1042
lA6
US
60 '79 219 1.18 '.M ).95 1.81 U2 1.77 1.74 1.71 1.66 1.60 1.54 1.51 lAS 1.44 lAO US 1.29
1'0 2.75 235 2.13 1.99 1.00 1.82 1.77 , L72 1.68 1.65 1.60 155 lAg iA5 \.41 137 1.32 L26 U9
2.71 2.30 2.03 1.94 1.85 1.77 1.72 1..67 1,63 1.60 U5 1,49 JA2 U8 1.34 1.30 1.24 Ll7 1.00
Table V Pen.:entage Points of the FDistrlbution (comitmed)
FilM,v"",
Vt Degrees of freedom for the numerll.tOr (VI)
\ '3 4 5 6 7 8 9 10 12 15 20 24 30 40 6(J 120
iol.4 199.5 l1S.1 224.6 230.2 234.0 236.8 238.9 240.5 2'11.9 243.9 245.9 248.0 249.1 250.1 2')Ll 252.2 2~3.3 254,3
2 ]8.5l 19.00 19.16 19.25 19.30 19.11 1935 19.37 19.33 19.40 J9.41 19.43 19.45 19.45 19.46 19.41 t9.48 19.49 19.50
3
4
5
]0.13
,~
UI
~
9.55
= ~
~
9.23 9.12
~
~~
9.01
~
'M
8.-94
U6
om
3.R9
_
_
lUiS
~
_
= _
ItSl
~ ~
8.19 8.74
~
_
8.70
~
~
8.66
~
~
8.64
~
~D
= =
8.62
~ _
8.59 8.51
5,69
4,41
8.SS
5.66
4AD
8.53
S.63
4.36
6
7
8
5.99
~
~
5,14
~
-
4.76
~
~
4.53
41'
~
4.19
~
~
4.28
1m
=.
4.21
=
~
4.15
D'
~
4.10
~
~
4.M
~
,~
=
4.00
= =
3.94
~
3,87
~
~
184
~
~
=
3.&1
~
1.71
~
~
>.74
3.)0
3.01
3,70
311
2.97
1.67
3.23
2.93
9
IO
5.12
~
_ =
4.26
410
~l.8(j 1.61
~
1,48
~ =
_ = =
3.37
= = 3.29
1M
1.21 :'\.18 3.l4
I.
3,07
UI
3.01
'8
.n
2,94
~J
2.90
~ =
2.&6 :un
1M
2.79
2.62
2.?S
258
2.11
2..54
~.
11
12
~
W ~
~
~
~
Uri
~
ill ~ = ~ ~
~
~
,n
D9
~ ~
~
I~
W
UI
UJ
~J
'"
,~
2.49
2Jf!
2.45
2.34
2.40
2.30
.j 13 4.61 :UH 3.41 3,18 1.03 2.92 2.83 2.77 2.7l 2.67 2.60 2.53 2.46 2.42 2.38 2.:34 2.30 2.25 2.21
j
.4
15
16
4.60
_
~
3.14
~
~
334
~
~
3.11
.m
~
2.%
~
~
2.85
~
~
2.]6
_
~I
2.70
~
~
2iiS
~
~
2.60
~
~
253
~
~
2.46
~
~
2.39
=
~
2.35
~
~
= =
2.3\
~~
2.2/
~B
2.22
2.16
2.11
2.18
2.11
2,06
2.13
2.07
2,01
-!l lJ 4.45 3.59 3.20 2.% Hi 2.70 2.61 2:5') 2,49 2.4S 2.3g 2.31 2.21 2.19 2.15 2,H) 2.1l6 2.01 L%
.!i 18 4.41 3.55 3.16 2.93 2.77 2.66 253 7..51 2.46 2.41 234 2~27 2,19 2.15 2.ll 2.06 '.ill un 1.92
Il
]
19
20
21
4.38
W
4.32
352
~
3.'17
1.J 3
~IO
3.07
2.90
10
2.84
2:J4
DI
2.68
2,61
I~
2.~7
254
=
2.49
2AS
~
2,42
2.42
~
2,37
=
2.38
2:12
2.31
~
2.25
2.23
~
2J3
•..
2.16
2.10
2.l1
~
2.05
2.111
~
2.01
.•
2.00
1.96
1.9&
1,95
1.92
193
L9{)
1.87
!JIg
UN
1.81
'~" "
.. 23
4.30
~
3,44
~
3.05
~
2.8;>
~
2,66
1M =
2.55 2.46
~ = =
;>AO 2.34 2,30
In
2.23
110
2. t5
,., =
2.07 2.03
WI
1.98
1M
1.94
1m
1.89
UII'i
1.M
1.81
L7&
1.76
~
,."
24 4.26
4.24
lAO
'l39
3"01
2.99
2.78
2.76
2,62
2,60
251
2.49
2,42
2.40
2.36
2,]4
2,30
2.211
2.25
2.24
2,18
2.16
f.ll
2.09
2,03
2.01
1.98
1.96
1.9~
J.92
1.89
1.87
1.84
1.82
L79
1.77
1.73
J.7l
27
"
'"
30
4.23
4.21
~
~.,
4.17
3.37
= =
3.35
~
:t32
2.98
2.%
~
2.92
2,74
2.71
v.
,~
2.69
:2-59
2.57
=
~
2.53
2.47
2.46
~
~
2,42
2.39
= =
2.37
~
2.33
2,32
2.11
ill
2.27
2.27
2,25
~
~
2.2'1
2.22
2.20
U9
~
2.16
2.15
2.13
1.2
2JO
2.09
2JJ7
2.06
=
~
7 O:l
loY')
L97
'M
I~
1.91
1.95
1.93
~
.~
U9
1.9(1
U:Hs
.m
1M
1.84
I.
1.85
U!4
•••
L79
LSO
L79
1.17
US
1.74
us
1.73
1.71
1.10
t68
1,69
1.67
LG~
1.41
1.62
40 4,08 3.71 2.84 2.61 2,45 2.'14 2.25 7.18 2.12 2.0S: 2.00 1.92 1.8.4 1.19 1.74 1.69 1.64 1.58 1.51 £-
'0
60 4.00 3.15 2.76 2.51 2.37 2.2.') 2.17 2.10 UM 1.99 L92 1.84 1.75 1.70 1.6'5 1.59 .>} IA7 1.19 g
120 192 ::to., 2.68 2,45 2.29 2.17 2J)9 2.02 1.96 1.91 1..83 1.75 }'66 1.6l 1.5) 1.55 1.43 US 1.25 ~
3.84 ::tOo 2.60 2,1'1 2.2l 2.10 2.01 1.94 LSS 1.113 1.75 1.67 1.57 1.52 1.46 1.39 1.32 1.22 LOa
(cO'njimlfS)
13
Table V Percenfage POints of lhe FDistribulion (colltr1H1ed) g
FOf)'.!.5.Y,,~!
, - vIr"
~
~ I
1
647.& 799.5
2 3
864.7.
4
399.6
5
921.8
6
Sm.l 943.2
7
lJ.!gTC;:S of fre.edom for lb(." mlmt13lOr (Y,)
956,7
&
963,3
9
963.6
10 12
976.7
15
9&4.9
20
993.1
24
997.2 100[
30
1006
40
lOW
60 120
lOJ4 HHB
f!'l'
2 I 3851 39.00 39.17 39.25 3930 19,33 19.36 'l,9.31 39.39 39.40 3M) 39.41 39.4'> :¥.lAo 39.46 39,47 19.4.& 39.49 3950
3 17,44 16'(H l5A4 15.l0 14,8& 14,73 14,62 14,~4 14.41 14.42 J4.34 \4.15 \4.17 i4J2 14.0& 1<1,/)4 13.99 13.95 13.90
• "D _ 9. _ ~ ~ _ ~ ~ w m ~ ~ "1 W UI U. g" ~
5
•
HUll
,..
&.43
~
7.76
_
7.39
~
.n
7.l5
m
6.9&
m
.n
=
6.&5
_
6.76
,~
_
=
6.6& 662
~
6.52
~ =
6.43 6.33
.n
6.2&
'J2 =
6.13 6.1&
,m
612
~
6.07
_
6.02
~
, = _
7 ~ ~ ~
~ ~
~
~ ~ ~ ~
~
W
~
~
W
~
.~
un
~
~
W
~
W
I.
W
~
W
~
~
.n
.~
9
10
7,2-1
6.94
6.72
5.11
~
~
SAil!
~
~
4.72
W
~
4.48
~
.M
4.32
~
~
4.20
=
_
4,f(}
1n
oM
4.03
~
~
3.96
D,
~
=
3,87
W
3,71
=
~
3.67
~
w
=
3.61
.n
356
UI
~n
3.51
~
~
3.45
~
~
3.39
'"
,~
3.33
m
~
.'£ "
12 6.55 ~.10 4A7 4.12 3.89 3.13 3.61 3.51 3.44 3.37 3,28 3.1& 3J.!7 3.0-1 2,% 2,91 2.85 2,79 2.72
11 6Al 4.97 435 4.00 3.77 3.60 JAS ].39 3.31 125
=
115 3.05 2.95 2.89
=
2.84 2.1& 27'J; 2.66 2.60
1"
"
I"
15
6.30
6.2Q
6.12
~
4,77
4.69
~
4.15
4,08
~
3.80
3,73
~
358
350
~
141
3.3'1
g
3.29
3.22
~
3.20
3.12
U'
3.12
3.05
10
:t06
=
2.99
2.96
2.89
~
2.86
=
2J9
~
2.16
2.68
~
110
2.63
1.64
= = =
2.57
~
259
2.~1
W
252
2A5
,~
2.46
2:3&
~
2AO
=
2.32
-5
"q
17 6.;)4
5.98
~
.~
~
~
~ ~
.~ =U, 116
1m = =
~ ~
~ = ,~
=
=
~
W
~
~
~ ~
= =
=
~ 'D
~ ~
t '"1
UI
\9 5.92 ~ ~ ~ .~ 1U ~ 2% = = UI ~ ,~ 2~ 2JI
10 5.8:7 4.46 3.116 151 3.29 '3.13 3.01 2.91 2.84 2.17 HiS 257 2.'16 2.4J 2.35 2.29 2.22 2.16 2.09
5.8"3 -1,42 3.&2 3A8 3.?-5 3.M 2.97 2.&7 2.80 /.73 2.64 2.53 2,42 2,37 2.31 2.15 2.18 1.1 t 2.04
i" ""
22 5.79 4.3& 17& 1.44 :1.22 3.05 2m 2.&4 2.76 210 2-60 :1.50 139 133 2.21 2.21 2,14 2.0& 2.00
5,15 4.1~ 115 3Al 3,18 3.02 2.90 2,81 2.13 2.61 2,51 2A7 236 1.30 2.24 2.JR ;UJ 204 1.97
5.72 -1.32 172 ),38 3.1'1 2.91:1 2jn 2_78 2.70 2.64 254 2.44 2.3':1 2,?7 2.21 2.15 2.08 2.01 1.94
II ~69 4,29 \69 ':U5 3.13 2.97 /..85 2.75 2.68 2.61 2.51 '.:AI 2.30 :1-2<1 2.18 2.12 2.05 1.98 L9J
26
""
5,66
5.613
5.61
4.27
~
4,22
3.61
W
3,63
3.33
UI
3.29
3.}0
~
106
=
2.44
2.90
2,&2
m
/.,?8
2,73
'"
2.69
2.M
~
2.61
=
2.59
:t55
2,49
W
2.45
2.39
~
2,34
= .,,
228
2.23
2.22
;1.17
2.16
'n
2.11
2.09
~
2.05
;.'! OJ
,.
1.98
1.95
I.
1.91
1.88
~
l.B3
29
30
I
I 559
5.57
4.20
'LIS
3.6]
359
3.17
3.2)
3.{}4
3.03
2.3S
2.B7
2.76
2.75
2.67
2.65
2.59
257
2.53
2.51
2.43
2,41
232
2.:H
2.21
2.20
2,15
2.14
2,09
2.07
Hj:l
2.0t
1.96
L94
1.89
un
LSI
1:19
I~
5042 4,05 3.46 3:13 2.W 2,74 2,62 2.53 2A5 2,39 2.29 U8 2.07 2JH 1.94 til£! L~O 1.72 }.6<1
5,19 3.93 3.:)4 3.01 2.79 2.63 2.SI 2.41 2.33 2.27 2.17 2.06 1.94 1.8& 1.82 1.74 un 1.58 1.48:
5.15 3,go 3.23 2.89 2.67 l.52 239 2.30 2.22 2,16 2.05 1.94 1.&2 L76 1.69 L61 1.53 lA3 1.31
)
5.02 3.69 3.12 2.79 2.57 2.41 229 2.19 2,11 2.05 1.94 U3 L71 1.64 157 1.'18 1.39 1.27 1.00
---
--~-.-,--
Table V Percentage Points: of the FDislrlbutlon (continued)
F(lm,~",,:>
v, '. , 3 4 6 7
Degrel!is of fr~~dorn fer the namenter (v,1
8 9
._---------
10 12 15 20 24 30 40 60 120
4052 4m,5 5403 5625 5764 5859 "23 WR2 6022 6056 6W6 6j<;7 6.209 613<; 6261 6287 6313 6339 6366
2 98.50 9'),00 99.17 99,25 99.30 99.33 99,36 99.37 99.39 99.40 99,42 99.43 99,45 99,46 99,47 99A7 99.4S 99,49 99.50
3 34.12 30.32 29A6 2lt71 2S,,24 27.91 27.G7 ZiA9 27.35 27.23 nos 26.87 :2(i}:;9 26.00 26.50 26.41 26.32 26.22 26.13
4 21.20 18,00 16.69 l't98 15.52 15.21 14.98 14.&0 14.66 14.55 14.3·' 14.20 14.02 13.93 13.84 11.7<; 11.M 11;56 13.46
5 1616 1121 12.06 11.39 10,91 10,67 10A6 10.29 10.16 10.05 9.89 9.72 9.55 9A7 9.38 9.29 9.20 9.11 9.m
6 D.75 10.92 9.78 9.l5 K75 8.47 ~" IUO 7.93 7.&7 7.n 7.56 7.'10 7.31 7,23 7.14 7.06 6.9'} 6.8g
7 1.2,25 9.55 8,45 U5 7.46 7.19 M9 6.1<4 6.72 6.62 6,47 6.11 6.16 6.07 '.99 5.91 5.8:2 5.74 5.65
g 1L26 8.65 73' 7,01 6,63 6.37 6.[8 6,03 5.91 5.81 5.61 <;,";2 <;.16 5.28 5.20 5.12 S.UJ 4.95 4.86
9 lQS6 g,{)2 6.99 6.42 6.00 5,80' 5.61 5A7 5.35 5.26 5.11 4.96 4.81 473 4.65 4/17 4,48 4.40 4.31
10 1O.Q.I. 7.56 6.55 5.99 5.64 5.39 5.20 5.06 4.94 4.85 4.71 4.56 4..11 4.33 4.23 4.17 4,OR 4.00 3.9l
11 9.65 7.21 6.22 5.67 5,32 5.07 4.89 4.74 4.63 4.54 4.40 4.25 4.10 4.02 3." 3.86 3.78 3.69 HiO
£ 6.93 'AI 5.06 4.64 354
J2 9.33 5.95
5.21
4.82
4.62
4.50
4,30
4.39
4,19
4.30
4.10
4.16
3.%
4.01 3.86
'"
'},59
3:10
3~1
3.62 3.45 3.36
I
13 9.07 6.70 5.74 4.86 4.44 3.82 1.66 JA3 3.31 3.25 3.17
14 8,86 6.51 5,";6 5.04 4.69 4,46 428 4.14 4.03 3.94 3.W 3.66 151 3,43 135 3.27 3.111: 3.00 3.00
lS 8.68 6.36 5,42 4,89 4,36 4.32 4.14 4.00 3.89 3.W 3.67 3.52 3.37 3.29 3.21 .1,13 1,0-; 2.96 2.117
]
..• 16
17
8.53
8,40
6.'n
6.11
5.29
5.1&
4.'f'1
4.67
4,44
4.34
4.20
4.10
4.01
1.93
3.89
1.79
3.78
3.68
3.69
3.59
1.55
1.46
3.4}
3.31
3.26
3.16
3.1&
3.OS
3,10
3.00
3.02
2.92
2,93
2JiJ
2.84
2:15
2J5
2.65
"]e l'
8.29 6.01 5.09 458 4.25 4.01 3.84 l71 3,60 3.51 3.37 3.23 1,08 3.00 2.92 2-'14 2.75 2.66 257
19 8.18 5m 'Di 450 4.17 3S14 3.77 3.63 3.52 J.43 3.30 3.15 3.00 2.92 2.1<4 2.76 2.67 2.58 2.'"
20 8.10 5.S5 4.94 4.43 4,10 ~87 3,70 3.56 lA6 3.37 3.13 3m 2,94 2,1:16 :2.18 2.69 '2.61 2.52 2.42
21 8.02 5.13 4.87 4.37 4.04 3.81 364 3.5J lAO 33J 3.17 3m 2JP:l :2.80 2.72 2.64 2:'iS lAG 2.36
." 22 7.95 5.n 4.82 4,31 3.99 3.76 3.59 3,4<; 335 3.26 3.12 2.98 2.33 2." 2.67 2,58 250 lAO 2.31
J "" 24
7.88
7.82
7.77
5.66
5.61
5,51
4.76
4.n
4,6&
4,26
4,22
4,l8
3,91
3.90
3,8':5
3,71
3.67
3.63
354
3.50
JAG
3,41
3.36
3.32
330
3.26
3.22
3.21
3.17
3.13
3.07
3.03
2.'"
2.91
2.89
2.85
2)3
214
:Uo
2.70
2.66
2.62
2.62
2,58
2.54
2~1
2.'19
2,45
2.45
2A0
2.36
2,)5
2,31
2.27
2.26
2.21
2.17
26 7.n 5.53 4.64 4.14 3J!2 3." 3.42 3.29 :1.18 3,09 2.% 2.81 vS6 2.58 2.:50 2.42 'U'3 2.23 2.11
27 7.£8 5A9 4.60 4.11 3.78 356 3.39 3a6 3.15 3.06 2.93 2.78 2.63 23' 2,47 2.38 2.29 220 2,10
7.64 5.45 3.21 3.12 2.90 2.44 2.26 2.06
"2. 7.60 5.42
4.57
4.>1
4.07
4.01
3.75
3.73
3.53
350
3.36
3.33 'UO 3.(l9
3.01
3.00 2.87
2.75
2,73
Via
2.57
2~2
2,49 2.41
2.35
2.33 2.23
2.\'1
2.14 2,03
30 7.56 5.39 4.51 4.02 3.70 3.17 3,1/) 3J7 '-07 2.98 2.84 2,70 '55 2.47 2,39 2.30 2,21 2.11 2.01
3,12 2)19 2.66 2,29 2.1 1 LSD .i;'
40
60
7.31
7.OS
5.13
4.93
4.31
4.13
3.1\3
3.65
3.51
3.>4
3.29
3.12 2,95
2.99
2,82 2,72
lJlO
2.63 2.50
2.51
:,",35
2:17
2,20 2,12
2.20
2.03 1.94
2.02
1.84
1.92
1.73 1.60 g
120 6,35 4.79 3.')"; 3AS 3.n 2.% 2.79 2.66 2.56 2.47 2.34 :,",19 2.03 1.95 LS6 L76 1,66 1..53 1.38
~
6i;) 4.til 118 3.31 3.01 2.30 2.64 2.51 2.41 2-32 2.18 2.04 L88 1:/9 1.70 1.59 1.17 1.32 1.00
§
610 Appendix
.~J
i
r-
2 3 4 5
d
(0) OC curves for different values of n for the tvfo~sided normal test for a leve~ of significance
,,~O.05.
I
I
0.00 --- _., .. ,.. l-
I
£
-~
~
'" 0.60 L._
~
U
d
i"•
D
O.4C,~-
6:
0.20
(b; OC curves fot cii.fferent values of n for the two-sided nonnal rest for a level of significance
"~O.OL
SO!4rce: Cia.'t VIa. e,/: k, r.; md q are reproduced Vlifr. pe.'1D:issior. from "'Operating Cha.-a~eristiC$ for the
Common St.J.istical Tests of Significance;" by C. L. Ferris. F, E. Grebbs. and C. L. W¢aver,.A1w11s of
14athematicai Statistics, June 1946.
Cha..'1:s Vlb. c, d, g, k, i,j, I, n, 0, p. and r are reproduceC ...1m pemissio!1 from Er.gineering Srar~tics. 2.'1.d edi~
non, by A H. Bowker and G. 1. Lieber:na:n, Pre!J.oce-Hal~, Englewood Cliffs, NJ, 1972.
Appendix 611
0.80 f---+---+
:F
~
Co 0.60 r-----r--··-+-\\Ii\\\-~\\,'Ii\_'c-\-·_\f·-·-_'<+·-·-·-·--I-----·I·-·
~
~
:c 0.40
•
~
"
0.20
o~-~----~--~~~~~~~~~~~~~~
-1.00 -0.50 0.0 0.50 1.00 1.50 2.00 2.50 3.00
d
(c) OC curves for different values of n for the one-sided normal test for a level of significance
a= 0.05.
1.00
I
,
,
....
I
0.80
:F
rn
~ 0.60
•8
•
":E£ 0.40
•
~
e
" ,
,
- __I.
0.20
0
-1.00 -0.50 0.0 0.50 1.00 1.50 2.00 2.50 3.00
d
(d) OC curves for different values of n for the one-sided normal test for a level of significance
a=O.OL
(continues)
612 Appendix
0.8
:!?
~ 0,8
1l
•
':5
~
0.4 '
~
~
e
~
02
(e) OC curves for di:ff~"ent values of n for the t\'l.to-sided r test for a level of significance ct= 0.05.
(f) OC curves for different values of n for the "No-sided t test for a level of significance 0:;: 0.01,
Appentli< 613
0.80
t' 0.70
~
<
~
~
0
O.fie
~ 0.5C
':5
£? 0.4{}
~
D
2 0.30
'-
(g) OC curves for different values of II for the oae-sided t test for a level of significance Ct= 0,05,
1.";;--;-~"i"'''''
0.90
::r~ j-~
0.30
~
-f-+- fi-\I\ \--H+
;.~-- t - t-
0,20 r"
C.10 I
(h) OC ctn"\'e$ for differeat values cf n for the one-sided t test for a level of significaacc a = 0.01.
614 A?pendix
"£ 0:70
g>
O.'0UiJjJjJJJ/j[:r;j~~~~
CAO 2.00 2.4C 2,SO 320 3.60 4,CO
1
(0 OC C'J..'"VCS for different values of r. for the t\Vo~sided cru·square test for a level of s~gnificance a= 0.05.
1.00 ;--.,--:--=
ia 0.70,-
j.60 - II
, ,
I
c
II
'0 0.50 I
'
~
'§ OAO i
£
01 OC c'U..""Vcs for different values of r. for the twoMsided chi~square test for a level of signillcance a,: 0.01.
Appendix 615
0.80
:i!:
~
c
=n 0,60 . -- - - - --- --
8u
•
'0
if
D 0.40
•
D
e
Q.
0.20
0; i.O
,
2.0 3,0 A,O
(k) OC curves for d.iffL'!ent values of r. for me one~sided (upper tail) chl-squa.-e test fo:: a leve; of
significance a= 0.05.
1.00
C.80
:1'
0
.§
n 0,60
•8
•
"15l? 0.40
D•
e
"-
75
50
0.2:0
'20"
"5
A
(m) ac curves for different values of n for the one-sided (lower taU) chi-square test for a level of siguificaucc
<:t= 0.05.
:f
~ I
•8 '.60
~
I,
• II
":5z. i
•
~
GAOl
E
a.
'.2<l
>
(1'1) oc C4cves for different v~ucs of n for the one-sided (lower taU) chi~squa:re test for a level of sign.if.cance
a=O.01.
Appendix 617
0.80
:!!:
§ O.BO
~
E
0
"
"5
~
:5 O.LD
i
<L
020
(0) OC curves for di!feren: va2ues of n for the two-sided F-test for a level of significance cc= 0.05.
~ .DC
0.80
:E
~
:§
~ 0.60
§
d
"5
~
0.40
~
~
.<L
020
0
0
(p) OC curves for different values of n for the two-sided F-test for a level of sig::.ificance a::::: 0,01
(cor.rinues)
618 Appendix
(q) OC curves for different values of n for tb.e one-sided F-test for a level of significance a:;:;:; 0.05.
1,00
0.80
:if
:E 0,6Q
~
~
15
:s'"•
n
0.40
:t
o~~~~~~~~~~~==~~
o 1.00 .2.00 4.00 5,00 8.00 10.00 12.00 14.00 16,00
'-
(r) OC CUlVes for different values of n for the one-sided F-test for a level of significance a = 0.01.
Appcndix 619
Chart VII Operating Characteristic Curves for the Fixed-Effects Model Analysis of Variance
1.00,---~----c--,--------~-------,--------,--------,--------,----,
¢(!ora "'0.Q1)-2 3 4 5
1.00,----------,----------,-----------,----------,-----------------,
0.40
0.30
~
5
0
'"
•
0.20
---T
~
J
§'
0 0.10
~ 0.08
0.Q7
--- - - - - -
":B'" 0.06
0.05
E
~
e 0.04
0.03
0.02 ._--
0,01
1 2 3 - ¢(fora: ",0.05)
<fJ(fora .. 0.Q1) -~1 2 3 4 5
ChartVll Opemtiug Characteristic Curves for the FIxed-Effects Model Analysis of Variance
(continued)
030
1
'i
I
~
0,20 i
g 0.10
~ 0.05
M7
"?i 0,," t"
t
0,05 I ~ •
i {L(J.t
" i \l,;)S -, ···,t
oror
,
(},01
tp(rore",i).C~)-1 2 3 4
1,00
o.en
0.70
0,60
0.50 i
.*
"
~
~
~
02C:
! :),10 I"
~ 0,06", -
e,C7:
'il
0,00 C·
'"
~ ;::,05 !'"
1 0,,,1
o,osl
Appendix 621
Chart vn Operating Characteristic Curves for the Fixerl~Effects Model Analysis of Variance
(continued)
,-~--~----,-----~----~----~~
II
',00"- - , - - , - - - - - - - - - , - - - - - , - - - - - - - , - - - - ,
0,';0'
0,70
0,50
0.50
OAO ~,
0.0
'~
O.03C"0_0'~' 0"1""\11,\"\
O.011!-------!:-L.l..Ll-'-.Ll~_:'::::::':_=:_::__;:_;S:_..L..LL'-.L.L-'-l-'-"'--.J
2 4
(continues)
622 Appendix
Chart vn Ope(ating Characteristic Curves for the Fixed-Effects Model A!talysis of Variance
(continued)
1.00 ~----:-'--------'--T'-'-'~--"--'--'--,-----'-'~'---'
O.llIJ~~--"'--'-;:=--=::~:~~~=:=:j
(),10
I
I
0.01 ",-----'-;;'-'...\..-'-'-l...\..'-;;3~~=·'-.·{;1~":·a-.·-;:o.~0S):;-'-'--''--''-...\..-'--'--'--'--'---~
¢:!cra .. O,O'l-_1 2 4
Appendix 623
Chart VIII Operating Olaractcristic Cur¥eS ror the Random-Effects Model Analysis ofVa..iance
'.00
O.1l<l
O.1C
0.00
4!• 0.5C
OA:
I• 0.30
020
"
g>
.~
"- 0.10
S
u Q,08
•
15
0.Q7
0.06 ,.
O.OS
g 0.04
jj om
D
e
"- M2
0.00
0]0
0.00 --+_ ........ _ ... _........ .
O.SO
0.40
0.20
"
"i 020
'•"
,s
g>
'K 0.10
•
~0.08
'00.C7
C 0.06 f---+--+..-
:a• J.CS
E
U.
3.04- !----I---+--+-
Mai--"""-+-
oo,L....--~--~__-1~--L---JL--~--~~~~~~~~~~~~~~~~~.
;3 5 1 9 11 i3 15 17 19 21 23 25""O:-.l.(!orcu.<.l.OSj
5 7 ;; 11 '3 15 17 19 21 23 25
So~rce: Reproduced with pennission from Ettgineer.n8 Statistics. 2nd edition, by A. H, Bowker and G,], Lieberman.
P~ntice"HalJ., Englewood Cliffs, NJ, 1972.
(continues)
624 Appendix
Chart VDI Operating Characteristic CtU-YCS for the Random-Effects Model Analysis of
Va..>iance (continued)
0,03
002
I
I,
1
I
. 1t
00'
t 3
,
4 5 6 7 ., 0 10 11 -}.{for;;t ",0.05)
5
.l,(for'X"'O.Ol} _1 3 4 6 7 8 9 10
" 12 13
1
2 4 5
• '0
Appendix 625
i!' i
'3. 0.10:
i 0.00: .
i ~:: f-~~~t_.
£Jl C.04 - --- :~- ..
0.03
C.02
0,0.1 "L-:2;--";---;--'.!~5~""~'L'-j-:;~;\;;;:;;';;~l"~~-".~~"-':'~"---_J
:t(fpra "'v.Ol}----1 , , 1
2 3 4
;:.(bra .. C.05)-1
5
,
6
,
(con.tin.ues)
626 Appendix
, 7 ,
tOO ~ -, -~:Y----r~--.,. ::--~'--;'--'T---,-',-~'----r---~-~-'- --)
0.83
0.70
CuD i-
0.50 i
GAO
OOSO i
j I
~ 020 I r
:l - I
~ 000-
;2:" 0:07\
OAiS .
-
~ 0.C5
il i
" ::\
0,02 '
1':: nl 2 3 4 5 6 7 8 9 10 II 12 13 14 15
,~--
5 15
6 10 16 23
7 10 17 24 32
8 II 17 25 34 43
9 6 11 18 26 35 45 56
10 6 12 19 27 37 47 58 71
11 6 12 20 28 38 49 61 74 87
12 7 13 21 30 40 51 63 76 90 106
13 7 14 22 31 41 53 65 79 93 109 125
14 7 14 22 32 43 54 67 81 96 112 129 147
15 8 15 23 33 44 56 70 84 99 115 133 151 171
16 8 15 24 34 46 58 72 86 102 119 137 155
17 8 16 25 36 47 60 74 89 105 122 140
18 8 16 26 37 49 62 76 92 108 125
19 3 9 17 27 38 50 64 78 94 111
20 3 9 18 28 39 52 66 81 97
21 3 9 18 29 40 53 68 83
22 3 10 19 29 42 55 70
23 3 10 19 30 43 57
24 3 10 20 31 44
25 3 il 20 32
26 3 11 21
27 4 11
28 4
Appendix 629
"For n:> 40, R is approximately normally distributed. with mean nI2 and variance nl4.
630 Appendix
EI 0,10 0.05
~-~--,----.-~--
0.02 O,O[
" "I
28
0,[0 0.05 0,02
.~-~--.~-~---.~-----~"~~
0,01
9[
4
5 I
I 0 29
130
140
116
126
101
110 100
6 2 0 30 151 137 120 109
7 3 2 0 31 163 147 130 118
8 5 3 0 32 175 159 140 128
9 8 5 3 1 33 187 170 151 138
10 10 8 5 3 34 200 182 162 148
11 13 10 7 5 35 213 195 173 159
12 17 13 9 7 36 227 208 185 171
13 21 17 12 9 37 241 221 198 182
14 25 21 15 12 38 256 235 211 194
15 30 25 19 15 39 271 249 224 207
16 35 29 23 19 40 286 264 238 220
17 41 34 27 23 41 302 279 252 233
18 47 40 32 27 42 319 294 266 247
19 53 46 37 32 43 336 310 281 261
20 60 52 43 37 44 353 327 296 276
21 67 58 49 42 45 371 343 312 291
22 75 65 ,5 48 46 389 361 328 307
23 83 73 62 54 47 407 378 345 322
24 91 81 69 61 48 426 396 362 339
25 100 89 76 68 49 446 415 379 355
26 110 98 84 75 50 466 434 3'17 373
27 119 107 92 83
--~
Source: Adapted with peo:nission from "Extended Table.'> of the Wilcoxon Matched Pair Signed Rank Statistic"
by Robert L. McComnck. Joumw. of the American. Statistical Assoc".atiOI'l, Vol. 60, September, 1965.
"If n > 50, R is approximately normally distributed, with mean n(n 1" 1)/4 and varianoe rt{r. + lX2Jt 1" 1)124.
Thh]e XII Percentage Point~ of the Studentized Range St'Jlisli(;'
QO_01(P.f)
----------~~~~~~~~
p
------------------------
---------~~~~~~~~
f 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
1 90~0 135 164 186 202 ll!> 227 237 246 253 260 266 272 272 282 286 290 294 298
2 14~0 19~0 22.3 24.7 26.6 28.2 29.5 30.7 31.7 32.6 33.4 31.4 34.8 35.4 3M 365 37.0 37.5 37.9
3 8.26 10.6 12.2 13.3 14.2 15.0 15.6 16.2 16.7 17.1 17.5 17.9 18.2 18.5 18.8 19.1 19.3 19.5 19.8
4 6.51 8.12 9.17 9.96 10.6 ILl ll.5 11.9 12.3 12.6 12.8 13.1 ll.3 13.5 13.7 J3.9 14.1 14.2 14.4
5 5.70 6.97 7.80 8.42 8.91 9.32 9.67 9.97 lO.24 WAS 10.70 10.89 11.08 11.2,1 11.10 11.55 lL68 lLKl lL93
6 5.24 6.33 7.03 7.56 7.97 8.32' 8.61 8.87 9.10 9.30 9,49 9.65 9.81 9.95 lO.08 10.21 JO.32 10.13 10.54
7 4.95 5.92 6.54 7.01 7.37 7.68 7.94 S.17 8.37 8.55 B.7l 8.86 9.00 9.12 9.24 9.35 9.16 9.55 9.65
8 4,74 5.63 6.20 6,63 6.96 7.24 7.47 7.68 7.87 8.03 8.18 8.31 8.44 8.55 8.66 8.76 8.85 8.94 9.03
9 4.60 5.43 5.96 6.35 6.66 6.91 7.13 7.32 7.49 7.65 7.78 7.91 8.03 8.13 8.23 8.32 8.41 8.49 857
10 4A8 5.27 5.77 6.14 6.43 6.67 6.87 7.05 7.21 7.36 7,48 7.60 7.71 7.81 7.91 7.99 8.07 8.15 8.22
11 4.39 5.14 5.62 5.97 6.25 6.48 6.67 6.84 6.99 7.13 7.25 7.36 7.46 7.56 7.65 7.73 7.81 7.R8 7.95
12 4.32 5.04 5.50 5.84 6.10 6.32 6.51 6.67 6.81 6})4 7.0(, 7.17 7.26 7.36 7.44 7.52 7.59 7.66 7.73
13 4.26 4.96 5.40 5.73 5.98 6.19 6.37 6.53 6.67 6.79 6.90 7.01 7.10 7.19 7.27 7.34 7.42 7.48 7.55
14 4.21 4.89 532 5.63 5.88 6.08 (,.26 6.41 6.54 6.66 6.77 6.87 6.96 7.05 7.12 7.20 7.27 7.33 7.39
15 4.17 4.83 5.25 5.56 5.80 5,99 6,16 6.31 6,44 6.55 6.66 6.76 6.84 6.93 7.0{) U)7 7.14 7.20 7.26
16 4.13 4.78 5.19 5.49 5.72 5.92 6.08 6.22 6.35 6.46 6.56 6.66 6.74 6.82 6.90 6.97 7.03 7.09 7.15
17 4.10 4.74 5.14 5.• 3 5.66 5.85 6.01 6.15 6.27 6.38 6.18 6.57 6.66 6.73 6.80 6.87 6.94 7.00 7.05
18 4.07 4.70 5.09 5.38 5.60 5.79 5.94 6.08 6.20 6.31 6.11 6.50 6.58 6.65 6.72 6.79 6.85 6.91 6.96
19 4.05 4.67 5.05 5.33 5.55 5.71 5.89 6.02 6.14 6.25 6.34 6.43 6.51 6.58 6.65 6.72 6.78 6.84 6.89
20 4.02 4.64 5.02 5.29 5.51 5.69 5.84 5.97 6.09 6.19 6.29 6.37 6,45 6.52 6.59 6.65 6.71 6.76 6.82
24 3.96 4.54 4.91 5.17 5.37 554 5.69 5.81 5.92 6.02 6.11 6.19 6.26 6.33 6.39 6.45 6.51 6.56 6.61
30 3.89 4.45 4.80 5.05 5,24 5.40 5.54 5.65 5.76 5.85 5.93 6.01 6.0S 6.14 6.20 6.26 6.31 6.36 6.41
40 3.82 4.37 4.70 4.93 5.1l 5.27 5.39 550 5.60 5.69 5.77 >.84 5.90 5.96 6.02 6.07 6.12 6.17 6.21
60 3.76 4.28 4.60 4.82 4.99 5.13 5.25 5.36 5,45 5.53 5.60 5.67 5.73 5.79 5.84 5.89 5.93 5.98 6.02
120 3.70
3.64
4.20
4.12
4.50
4.40
4.71
4.60
4.87
4.76
5.01
4.88
5.12
4.91)
5.21
5.08
5.30
5.16
5.38
5.23
5A4
5.29
551
5.35
5.56
5.40
5.61
5,45
5.66
5.49
5.71
5.54
5.75
5.57
5.79
5.61
5.83
5.65 i
~
f"' degress of freedom.
}l'rom J. M. May, ''li'{leuded and Corrected Tables of tbe Upper Pe~enl'lge Points of tbe S[lldel'l!ized Rang~," JJiomelrika. Vbl. 39, pp. 192-193. 1952. Reproduced by permission of the
trus{eM of Biowetrlka. ,..el
(contllIues)
Table ~'1I Percentage Points of the Studcnlized Range Statistica(cQII/ilwed)
!li
q",(p,f)
f 2 3 4 5 6 7 8 9 10
p
II 12 13 14 15 16 11 18 19 20
f!t-
I 18.1 26.7 32.8 37.2 40.5 43.1 45.4 47.3 49.1 50.6 51.9 53.2 54.3 55.4 56.3 57.2 58.0 58.8 59.6
2 6.09 8.28 9.80 10.89 11.73 12.43 13.03 13.54 13.99 14.39 14.75 15.08 15.38 15.65 15.91 16.14 16.36 16.57 16.77
3 4.50 5.88 6.83 7.51 8.04 8.47 8.85 9.18 9.46 9.72 9.95 10.16 10.35 10.52 JO.69 10.84 10.98 11.12 11.24
4 3.93 5.00 5.76 6.31 6.73 7.06 7.35 7.60 7.83 8.03 8.21 8.37 8.52 8.67 8.80 8.92 9.03 9.14 9.24
5 3.64 4.60 5.22 5.67 6.03 6.33 6.58 6.80 6.99 7.17 7.32 7.47 7.60 7.72 7.83 7.93 8.03 8.12 8.21
6 3.46 4.34 4.90 5.31 5.63 5.89 6.12 6.32 6.49 6.65 6.79 6.92 7.04 7.14 7.24 7.34 1.43 7.51 7.59
7 3.34 4.16 4.68 5.06 5.35 5.59 5.80 5.99 6.15 6.29 6.42 6.54 6.65 <\.75 6.84 6.93 7.01 7.08 7.16
8 3.26 4.04 4.53 4.89 5.17 5.4Q 5.60 5.71 5.92 6.05 6.18 6.29 6.39 6A8 6.51 6.65 6.73 6.80 6.87
9 3.20 3.95 4.42 4.76 5.02 5.24 5.43 5.60 5.74 5.87 5.98 6.09 6.19 6.28 6.36 6.44 6.51 6.58 6.65
10 3.15 3.88 4.33 4.66 4.91 5.12 5.30 5.46 5.60 5.72 5.83 5.93 6.03 6.12 6.20 6.27 6.34 6.41 6.47
11 3.11 3.82 4.26 4.58 4.82 5.03 5.20 5.35 5.49 5.61 5.71 5.81 5.90 5.98 6.06 6.14 6.20 6.27 6.33
12 3.08 3.77 4.20 4.51 4.75 4.95 5.12 5.27 SAO 5.51 5.61 5.71 5.80 5.88 5.95 6.02 6.09 6.15 6.21
13 3.06 3.73 4.15 4.46 4.69 4.88 5.05 5.19 5.32 5.43 5.53 5.63 5.71 5.79 5.86 5.93 6.00 6.06 6.11
14 3.03 3.70 4.11 4.41 4.64 4.83 4.99 5.13 5.25 5.36 5.46 5.56 5.64 5.72 5.79 5.86 5.92 5.98 6.03
15 3.01 3.67 4.08 4.37 4.59 4.78 4.94 5.08 5.20 5.31 5.40 5,49 5.57 5.65 5.72 5.79 5.85 5.91 5.96
16 3.00 3.65 4.05 4.34 4.56 4.74 4.90 5.03 5.15 5.26 5.35 5.44 5.52 5.59 5.66 5.73 5.79 5.84 5.90
17 2.98 3.62 4.02 4.31 4.52 4.70 4.86 4.99 5.11 5.21 5.31 5.39 5.47 5.55 5.61 5.68 5.74 5.79 5.84
18 2.97 3.61 4.00 4.28 4,49 4.67 4.83 4.96 5.07 5.17 5.27 5.35 5,43 5.50 5.57 5.63 5.69 5.74 5.79
19 2.96 3.59 3.98 4.26 4.47 4.64 4.79 4.n 5.04 5.14 5.23 5.32 5.39 5.46 5.53 5.59 5.65 5.70 5.75
20 2.95 3.58 3.96 4.24 4.45 4.62 4.77 4.90 5.01 5.11 5.20 5.28 5.36 5.43 5.50 5.56 5.61 5.66 5.71
24 2.92 3.53 3.90 4.17 4.37 4.54 4.68 4.81 4.92 5.01 5.10 5.18 5.25 5.32 5.38 5.44 5.50 5.55 5.59
30 2.89 3.48 3.84 4.11 4.30 4.46 4.60 4.72 4.B3 4.92 5.00 5.08 5.15 5.21 5.21 5.33 5.38 5.43 5.48
4Q 2.86 3.44 3.79 4.04 4.23 4.39 4.52 4.63 4.74 4.B2 4.90 4.98 5.05 5.11 5.17 5.22 5.27 5.32 5.36
60 2.83 3.40 3.74 3.98 4.16 4.31 4.44 4.55 4.65 4.73 4.81 4.88 4.94 5.00 5.06 5.1 I 5.15 5.20 5.24
120 2.80 3.36 3.69 3.92 4.10 4.24 4.36 4.47 4.56 4.64 4.71 4.78 4.84 4.90 4.95 5.00 5.04 5.09 5.13
2.77 3.32 3.63 3.86 4.03 4.11 4.29 4.39 4.47 4.55 4.62 4.68 4.74 4.80 4.84 4.98 4.93 4.97 5.01
Appendix 633
Percent
Coverage 0.90 0.95 0.99 0.90 0.95 0.99 0.90 0..95 0.99
2 15.978 18.800 24.167 32.019 37.674 48.430 160.193 188.49: 242.300
3 5.847 6.919 8.974 8.380 9.916 12.861 18.930 22.401 29.055
4 4.166 4.943 6.440 5.369 6.370 8.299 9.398 ILl50 14.527
5 3.949 4.152 5.423 4.275 5.079 6.634 6.612 7.855 10.260
6 3.131 3.723 4.870 3.712 4.414 5.775 5.337 6.345 8.301
7 2.902 3.452 4.521 3.369 4.007 5.248 4.613 5.488 7.187
8 2.743 3.264 4.278 3.136 3.732 4.891 4,147 4.936 6.488
9 2.626 3.125 4.098 2.967 3.532 4.631 3.822 4.550 5.966
10 2.535 3.018 3.959 2.839 3.379 4.433 3.582 4.265 5.594
11 2.463 2.933 3.849 2.737 3.259 4277 3.397 4.045 5.308
12 2.404 2.863 3.758 2.655 3.162 4.150 3.250 3.870 5.079
13 2.355 2.805 3.682 2.587 3.081 4.044 3.130 3.727 4.893
14 2.314 2.756 3.618 2.529 3.012 3.955 3.029 3.608 4.737
15 2.278 2.113 3.562 2.480 2.954 3.878 2.945 3.507 4.605
16 2.246 2.676 3.514 2.437 2.903 3.812 2,'072 3.421 4.492
17 2.219 2.643 3.471 2.400 2.858 3.154 2.808 3.345 4.393
18 2.194 2,614 3.433 2.366 2.819 3.702 2.753 3.279 4.307
19 2.112 2.588 3.399 2.331 2.7&4 3.656 2,703 3.221 4.230
20 2.152 2.564 3,368 2.310 2.752 3.615 2.659 3.168 4.161
21 2.135 2.543 3,340 2.286 2.723 3.577 2.620 3.121 4.100
22 2.118 2.524 3.315 2.264 2.697 3.543 2.584 3.078 4.044
23 2.103 2.506 3.292 2.244 2.673 3.512 2.551 3,040 3.993
24 2.089 2.489 3,270 2.225 2.651 3.483 2,522 3.004 3.947
25 2.071 2.474 3.251 2.208 2.631 3.457 2.494 2.972 3.904
30 2.025 2.413 3.170 2.140 2.529 3.350 2.385 2.841 3.733
40 1.959 2.334 3.066 2.052 2.445 3.213 2.247 2.677 3.518
50 1.916 2.284 3.001 1.996 2.379 3.126 2.162 2.576 3.385
60 1.887 2.248 2.955 1.958 2.333 3.066 2.103 2.506 3.293
70 :.865 2.222 2.920 1.929 2.299 3.021 2.060 2.454 3,225
80 1.848 2.202 2.894 1.907 2.272 2.986 2,026 2.414 3.173
90 1.834 2.185 2.872 l.889 2.251 2.958 1.999 2,382 3.130
100 1.822 2.172 2.854 1.874 2.233 2.934 1.977 2.355 3.096
636 Appendix
Agresti, A. and B. Coull (1998), "Approximate is Cook, R. D. (1977). "Detection of1Dlluential Ooser-
Better than "Exact' for Jr.terval Estimation of vations in Linear Regression," Technoff'..etrics, Vot
B:nomiai Proportions." The Aw..erican Statistician, 19, pp, 15-18,
52(2). Crowder, S, (1987), "A Simple Method for Studying
.t\nderson, V. L., ar.dR A. McLean (1974), Design Run~LeIlc,trth Distributions of Exponentially
of Experiments: A Realistic Approach, ~1arcel Weighted Moving Average Olarts," Technometn'cs,
Del:ker. New York. VoL 29. pp, 401-407,
Banks.]., J, S, Carsoo, B,L, Nelson., andD. M.Nicol Daniel. C" and F. S, Wood (1980), Fitting Equations to
(2001).Discrete~E)len.t System Simulation, 3rd edi- Daw.. 2nd edition, John Wiley & Sons, New York.
tion, Prentice~Hall. Upper Saddle River, NJ. Davenport, W. B" and W, 1" Root (1958), An Intro-
Bartlett, M. S. (1947), "The Use ofTransforma· duction to the Theory ofRarulom Signals and
tions," Biometrics, Vol. 3, pp. 39-52. Noise, McGraw~Hill. ~ewYork.
Bechhofer, R. E .• T. 1, Santner, and D. Goldsman Draper, N. R., and W. G. Hanter (1969), "Transfor-
(1995), Design. and Analysis of Experiments for mations: Some Examples Revisited,'" Technomet-
Statistical Selection" Screen.ing and Multiple Com~ rics, Vol. 11, pp, 23-4{),
parnaM, lohn Wiley and Sor.s, New York. Draper, N. R., and 'H. Smith (1998), Applied Regres·
BelsleY, D, A., E, Kuh, and R. E. Welsch (1980). sion Atudysts, 3rd edition., John Wiley & Sons,
Regression Diagnostics, John Wiley & Sons, New NewYotk.
York. D-.mcan, A, I, (1986), Quality Control and Industrial
Berrettoni, J. M. (1964), "Practical Applications of Statistics, 5th edition, Richard D. Irwin, Home-
the Weibull Distribution,"" Industrial Quality Con- wood, IL,
""I, Vol, 21,No, 2,pp. 71-79. Duncan, D, B. (1955). "Multiple Range and Multip!e
Box, G, E, P" and D, R. Cox (1964). "AllA:1alysis of FTests," Biometrics, Vol. III PP, 1-42,
Ttansfon:::lations." Journal of the Royal Statistical Efron, B. andR. TIbshi.'1mi(1993),Anln.rroti"wctionlO
Society, B, VoL 26, pp. 211-252, the Bootstrap, ~an ar.d Hall. ~ew lOCk.
Box, G, E, p" and M, F. Mdller (1958), "A Note on Elsayed, E. (1996), Reliability Engineering. Addison
the Generation of Normal RandgID Deviates," Wesley Longman, Reading. MA.
A.n1tals of Mathematical Statistics, Vol, 29, pp. Epstein, B, (1960). "'Estimation from Ufe Test Data,"
610-011, IRE Tran.sactions on Reliability, Vol. RQC"9.
Bradey, P,. B, L, Fox. and 1" E. Schrage (1987),A
Feller, W. (1968). An Introduction to Probability
Guide to Simulation, 2nd edition, Sp.ringer~Verlag,
Theory and Its Application.~, 3rd edition, John
New York.
Wiley & Sans, New York.
Cheng, R. C. (1977), "The Generation of Gamma
Fishman, G, S. (1978). PrinCiples of Discrete Event
Variables with NoninJ:cgra1 Shape Parameters;'
Simulation, John Wtiey &: Sons, ~C\\' York.
Applied Statistics, VoL 26, No, I. pp, 71-75,
Cochran, W, G, (1947), "Some Co!)Seq"e""es "''ben FumiYll!, G, :>1" and R. W, Wilson, Jr, (1974),
the .A.ssumptions for the Analysis of Variance Are "Regression by Leaps and Bounds," Tecrmomet-
Not Satisfied," Biometrics, Vol. ;., pp, 22-38. nus. Vol. 16, pp, 499-512,
CochrllJ1, W. G, (1977). Sampling Techniques, 3rd Hahn, G., and S. Shapiro (1967), Statistical Models
edition, Joha Wl1ey & Sons, New York, in Engineering, John Wliey &: Sons, New York.
Cochr:m, W. G,. and G, M, Cox (1957), Experimen- Hald,A. (1952), Statistical 'Dutory wiJI: Engineering
tal Designs. John Wiley & Sons, Ne\VYork. Application.s, John Wiley & Sons, New York.
Cook, R. D. (1979), "1Dlluential Obsel'V1loons in Lin- Hawkins, S. (1993), "Cumulative Sum Control
ear Regression," Jou.mal oftheAmencan: Sta;isti- Chartillg: All underntilized SPC Tool;' Quality
cal Association.. Vol. 169-174. Engineering. Vol, 5, pp, 463-477,
637
638 References
Hocking, R. R, (1976), "The Analysis and Selection Mood, A. M., F. A. Graybill, and D. C. Bees (1974),
of Variables in Linear Regression," Biometrics, IrtlroducMn to (roe Theory 0/ Statistics. 3rd edi~
Vo!' 32, pp. 1.-49. !ion, McGraw..Hill, New York.
Hocking, R. R .. F. M. Speed, and M. J. Lynn (1976). Neter.l. M. Kutner, C. Nachtsheim, and W. Wasser~
"A Class of Biased Estimators in Linear Regres~ man (1996), Applied Lir.ear Statistical Mrxiels.
SiOD," TecntU>metrics, Vol. 18, pp. 425-437. 4th edition, Irwin Press, Homewood.,.IL.
HOed, A. E., and R W. Kennard (1970a), "Ridge >lewman, D, (1939), 'The Distribution of the Range
Regression: Biased Estimation for Non-Orthogonal in Samples from a NoID13l Population Expressed
Problems,'" Technometdcs, Vol. 12, pp, 55-67. in Tern:s of an Iodependent Esti.JruIte of Standard
Hoed. A. E., andR. W. Kennard (197Gb), "Ridge Deviation," Biometrika, Vol. 31, p. 20.
Regression: Application to Non-Orthogonal Prob~ Odeh. R., and D. Owens (1980), Tables/or Nonr.al
leInS," TechtU>metrics, Vol. 12, pp. 69~82. Tolerance Limits, Sampling Plans, ami Screening,
Kelton, W. D .. and A. M. Law (1983), "A New Marcel Dekker, ~ew York.
Approach for Dealblg with the Startup Problem in Owen. D. B. (l962), Handbook of Statistical Tables,
Discrete Event Simulation:' Naval Research Addison·Wesley Publishing CompllIlY. Reading,
Logist;cs Q-,tarterly, \O\. 30, pp. 641-{558. Mass,
Kendall, M. G., and A. S:uart (1963), The Advanced Page, E. S. (1954). "Continuous Inspection
Theory of Statistics, Hafner Publishing Co:::npany, Schemes." Biometrika, Vol. 14. pp. 100-115.
>lew York, Roberts, S. (1959), "Control C'ha.ttTests Based on
Keuls, M. (1952). 'The Use of me S:Udentized Range Geometric Moving Averages," Technometrics. Vol
in Connection with aD Analysis of Variance," I, pp. 239-250.
Euphytics, Vol. 1, p. 112. Scbeff6, H. (1953), "A Memod for Judging All Con·
Law, A. M., and W. D. Kelton (2000). Simulation trasts in the Analysis of Variance," Biometrika,
Modeling andAnalysis, 3rd edition, McG:raw-Hill, VoL 4(], pp.87-104.
New York, Snee, R. D. (1977), "Validation of Regression Mod~
Lloyd.D. K, aDdM. Lipow (1972), Reliability: eIs; Methods and Examples," Technometrics, VoL i
Mana.gerr.ent, Methods, arul Mathenulrics. P1entice~
Hall, Englewood ClilIs, N.J.
19, No.4, pp. 415-428.
Tucker, Ii. G. (1962), An Introduction to Probability
~
Lucas, 1., and M, Saccucd (1990), "Exponentizlly and Mathematical Statistics, Academic Press,
Weighted Moving Average Control Schemes: ~ewYork.
Properties and Enhancements." Technometrics, Tukey, J. W. (1953), 'The Problem of Multiple
Vol, 32, pp. 1-12. Comparisons," unpublished notes, Princeton
Marquardt, D. W.. and R. D. Snee (1975), "Ridge University.
Regression in Practice:' The American Tukey, J, W. (1977), Exploratory Data Analysis.
Statistician, Vol. 29, pp. 3-20. Addison-Wesley, Reading, MA.
Montgomery, D. C. (2001), Design and Analysis of United States Department of Defense (1957). Mili-
Experimeras, 5th edition, 10hn Wiley & Sons. tary Sttmdard Sampling Procedures and Tablesfor
New York. Inspection by Variables for Percent Defective
~fontgomery, D. C (2001),lntroduction to Statisti~ (MlL-STD-414), Government Printing Office,
cal Quality Control, 4th edition, John W:t1ey & Washing'.on, DC.
SOIlS, New York. Welch, P. D. (1983), 'The Statistical Analysis of
Montgomery, D. C, E. A. Peck, and O. G. vining Simulation Results." in The Corr.puter Perfor·
(2001), Introduction to Linear Regressicrn. mance M.odeling Handbook (ed. S. Lavenberg),
Analysis. 3rd edition, John Wiley & Sons, New . Academic Press, Orlando, FL.
York.
Montgomery, D.C., and G.C. Runger (2003),
Applied Statistics and Probability for Engineers,
3rd edtion, John Wiley & SODS, New York.
Answers to Selected Exercises
Chapter 1
1·1. (a) 0.75. (h) 0.18,
1·3. (a) it (' B= [5), (h) it c; B= {I, 3,4, 5, 6, 7, 8, 9.1O}, (e) A0!i = (2, 3, 4, 5).
(d) U=(l,2,3,4,5,6,7,8,9,lOj, (e) A('(B,-,C)= (1,2,5,6,7,8,9,IO).
1.5 9' = W, r,): t; ~ 0, I," 0),
't
A = (t t,):t, " 0, ,,0, ('l - :,)12 S; 0,15).
"
B = [('l' '2): " ~ 0,', ~ 0, max (t" t,)'; 0,15),
e = {(t t,l: 'l ,,0, t, ~ 0, It, - r,jt2 ,; 0,06),
1·7, "
9' = (NNNNN, NNN,ND, NNNDN, NNNDD. NND"IN, l'-CIDND, NNDD, ':'IDl'--:W, NDt-.'NIJ,
NDND, NDD, DNNNN, DNNND, DNND, miD, DD),
1~9. (a) N:::: not defective, D := defective.
9' = (NNN, ~'NIJ, NDN, NDD, DNN, DND, DDN, DDD) ,
(h) 9' = (NNNN, NNND, NNDN, NDi'.'N, DNNN),
1·11. 30 tomes. 1-13. 560.560 ways.
(3OOP "1(300(I- p'Yi
1·15. P(Accept Ip') =L \ ,
t li x IO-x
)I
,-0 (31~)
1-17. 28 comparisons, 1·19. (40)(39) = 1560 tests,
1-21. (a) (;)(;) = 25 ways, (h) @@= 100 ways,
1-23. R, =[1- (0,2)(0,1)(0.1)] [1 - (0,2)(0,1)] (0,9) =0,880,
1-25. S =Siberia. U = Ural, P(S) =0,6, P(U) =0.4, peElS) =P(FlS) =0,5,
P(F~U)=0,3, P(SIF) '( 0,6),0,))4-
/:~X~'\ 0.4 (0,3
0,714. )'
1 1 1 m-l m 2 -m-l
1-27, -_._+-,--= 2 1·29. P(womenI6,) '" 0,03226, 1·31. !l4,
m-1 m m m m (m-l)'
Chapter 2
639
640 Answers to Selected Questions
3:;'~S~l-Yes, (b) No, (eJ Yes. 2-7. a, b. 2-9. P(Xs 29) ~ 0.978.
~ 2.11 •.(.) k ~~, I (b) 11 ~ 2. ",' = l.
"-~/"'~ "> /"
x:::'4.
2-13. b 2. [14 - 2'12. 14 + 2£J.
2~15, (a) k=;. (b) .u::::~, cr"=i;,
(e) Fx(x) = 0" < l.
=-k.l S;x<2,
=r.f." 2 :5:x<3,
=1.x~3.
Chapter 3
3·1. (a) y p,fy) (b) E(l') = 14, Vel') = 564.
-~---
0 0.6
20 0.3
80 0,1
Qther.vise 0.0
3·3. (al 0.221. (b) SI55.80.
3~5. liz) :::: e-r., Z 2: 0,
= 0, other.vise.
3-7. 93.S c!gal.
3·9. {a} f,J;;) =:;; kfr 1f1 , trOfl)ill, y > D,
::::: 0, otherwise~
(b) Mv)= 2"'~, v > 0,
:::: 0, otherwise.
(c) f,)u)=e-('f"-u!,u>O,
::::; 0, other"Nise.
3·11. s = (513) x 10'.
3·13. (al !,(Y)=1<4-yr ",O';yS3,
"
"'" Q, otherwise.
(b) Jy(y):::' J. e:5: y.$ C,
:::;: 0. othGI"Nisc.
Answers [Q Selected Questions 641
6
3·15. Mx(t) = I,(t)e",
E(X)=M~(O)=t,
Chapter 4
4·1. Ca) x 0 I 2 3 4 5
pj.x) 27/50 II/50 6/50 3/50 'li50 1/50
Y 0 I 2 3 4
py(y) 20/50 15/50 10/50 4/50 1/50
(b) Y 0 I 2 3 4
Pno Cy ) llf27 8/27 4f27 3/27 In7
Ce) x 0 I 2 3 4 5
P.oCx) llnO 4/20 2120 1/20 Ino 1/20
4-3. Ca) k = 1.
,
Cb) Ix,Cx,) =100, 0 ';x,'; 100,
= 0, otherwise.
,
Ix,Cx,) = TI>, 0 ,; x,'; 10,
= 0, otherwise.
= 0," otherwise;
fr(Y)=~JI-y', O<y<I,
"
:::: 0, otherwise.
642 Answers to Selected Questions
E(Ylx)=t~l-x' .
2+3y
4-27. (a) E{Xly)= 3(h2Y)' O<y<l.
4-29. (a) k= (n - 1)(n ·2). (h) F(x. y) = I · (1 + x),". (1 + y)'" • (1 + x+ Y)"'.;P 0, y;> O.
4-31. (a) Independent. (b) Not independent (c) Not independent
4-33. (a) i (b) i
4-35. (a)F,(z)=Fx!(,.a)lb). (b) F,!.z)=I-Filh). (e) F,!.z) = Fx(i'). (d) F,!.z) = FJ)n z).
Chapter 5
5.1. P(X=x) =(~ p' (1- p)4.', x = 0, 1,2.3,4.
5·13. E(X) = ,\Ix (0) = 1., E(X') =M;(O) = ~,<1~ =E(X')-(E(X))' =3,.
p p' P
5·15. P(X= 36) '" 0.0083.
5·17. P(X = 4) =0.077. P(X <4) =0.896. 5·19. E(X) = 6.25, VeX) =1.5625.
5·21. p(3, 0, 0) + p(O, 3, 0) + P(O. 0, 3) = D.llS. 5-23. p(4, I, 3. 2) '" 0.005.
=
5.25. P(X" 2) 0.98. Binomial approx. P(X S 2) '" 0.97.
5·27, P(X" I) '" 0.95. Binomial'ppro.". "" n = 9.
P(XS;3)=e'''± 30',
..:..0 xl
P(X~5)=I-e'''± 3~ .
.r< X_I
Chapter 6
,,
6-1. ~,n·
Chapter 8
8-1. x = 131.30,? = 113.85, s = 10.67. 8-21. x = 74.002,? = 6.875 x 10", s = 0.0026.
8-23. (a) The sample average will be reduced by 63.
644 Answers to Selected Questions
(b) The sample mean and standard deviation will be 100 units larger. The sample vali.ance v.iJl be
10,000 units larger,
8·25. a = x. 8·29. (a) x= 120.22, s' 5.66. s =2.38. (b) x = 120. mode = 121.
8-31. For 8·29. cv = 0.0198; for 8·30. CY 9.72, 8·33. = x= 22.41, s' =208.25, x = 22.81, mode = 23.64.
Chapter 9 ,
9·1. j(x" x,. "., x,); (1I(2nO"»5!le·,n"' L (x,·· Il)'.
9·3. j(x" x" x" xJ = 1. 9·5. N(5.00,O.00125).
9·7. Use S I·r;;.
9·9.
- - :~-;;:z
ThestandatderrorofX,-X,is i'....L+:2.='/-·-·-+-=--·=0.47.
I(T 5)' (" 0)'
9·11. N(O,I).
- in! r-;. ,:25 30
Chapter 10
10-1. Both esti:nators are unbiased. ~ow,. V(X1):= (J'llln whereas V(X2) ::= a 211'1. Since V(X\) < V{X:J, XI is
a more efficient estimator ':han )(2'
10-3. $'l' because: it would have a smalier MSE.
10-7.
A fox, -
IX;£.. -=X. 10·9.
.
lit.
i=l n
2
10·25, , . r 6f
f(.uJx),xZ,,,,x,,)::::CJ, 2(2~rtl12Jexp~ __ J,l--11'--Z+-f
nX JJ.c ,\'i1 ]l,
I( L2. c\ a 0'0"
/j j
I·
n 1
wMre C=-.,+-,.
a" °0
10·21- The posterior density for p is a beta distribution with para.n:.eters a + n and b + L. Xi - n,
10·29, The posterior density fer ?.is gamma with parameters r =: m + Lx; + 1 and 0;:; n + (m +l)/l,:.
10·31. 0.967. 10·33. 0.3783.
j(x"e) 2x
f )=-':-(~)' =-'(2-2" (b)
10·35. (al j.elx, 112. 10-37. at:::;: a;::::: cd2 is shorter,
J x·, e , - xl,
10·39, (a) 74.03533 ~ 1">74.03666. (b) 74.0356" .u.
10..41, (al 3232.11 ;; i" ~ 3267.89. (b) 1004.80:$ J.l. 1043. 150 Or 151.
10-45. (a) 0.0723:$ J.l, -Ilz:$ 3267.89. (b) 0.049% Il: -Ilz:$ 0.33. (e) i", -Ilz;; 0.3076.
Answers to Selected Questions 645
1041. -3.68 S 1', - iLl S -2.12. 1049. 183.0 S I'S 256.6. 10·51. 13.
10·53. 94.282 S it ~ 111.518. 10-55. -D.839 ~ .u, - iLl ~ -D.679. 10·57. 0.355 $1', - iLl'; 0.455.
10-59. <a) 649.60'; a' $ 2853.69. (0) 714.56 s a'. (e) a' $ 2460.62. 10-61. 0.0039$ a' $ 0.0124.
10-63. 0.514" ,,',; 3.614. 10·65. O.ll S 0';1,,; ~ 0.86. 10-67. 0.088 $p S 0.152. 10·69. 165i7,
10-11. -D.0244 "p, -p,,,
0.0024. 10·13. -2038 $1', - iLl'; 3714.8.
10-75. -3.1529';!1, - iLl " 0.1529; -1.9015 "!1, - iLl S 0.9015; -D.1775 S!1, - iLl S 2.1775.
Chapter 11
11·1. <aJ:z;, -:.333, do not reject H,. (0) 0.05. U·3, (a) z;, = -12.65. reject lID. (b) 3.
11·5. :z;, = 2.50, reject II,. 11·7, <al:z;, = 1.349, do not r<;ject H" (b) 2. (e) L
11·9, Z, = 2.656, reject H~ n·11. z;, = -7.25, reject H,. 11·13. 10 = 1.842, do not reject Iio•
11-15. t o ;;d.47, do not rejectHo at a=: 0.05. 11-17. 3.
11·19. (a) '0 ~ 8.49, reject Ii", (b) '0 -2.35, do not reject II,. (e) 1. (d) 5.
11·21. Fe = 0.8832, do not reject He' 11·23. <a) Fo = 1.07. do not reject He' (b) 0.15. (e) 15.
11-25. to"" 0,56, do not reject Hey
11·27, (a) ~~43.75,rejectH,. (b) 0.3078 x 10"'. (e) 0.30. (d) 17.
11·29. (aj X; = 2.28. ,eject H,. (b) 0.58. 11·31. Fo = 30.69, rejeet H,; P'"' 0.65.
11..33. :;;;;: 2.4465, do not r<Bect Hr;> 1l<~5. to =: 5.21, reject HD, 11..37. Zo;;: 1.333, do not reject HQ•
1141. Zr;=-2.023,donotrejectHo. 11-47. ~=:2.9!5,donotrej:;ctHrr
1149. ~ = 4,724. do not reject Ho. 11",53. ~ = 0,0331, do not rejectHc~
11~55. ~ =: 2A65. do not reject Hrr 11s!. ~ =34.896. rejectHD• 11~59. ~ =: 22.06, reject He'
Chapter 12
12-1. (aj Fe =3.17. 12-3. (a) Fo =12.13. (bj );.fixing technique 4 is different from I, 2, and 3.
12-5. (a) Fe ~ 2.62. (b) jl = 21.70, '" 0.023. T, ~ -D. 166, i:, 0.029, "4 ~ 0.059,
12·7. (a) Fe=4.0L (b) Mean 3 differs from 2. (e) S5"=246.33. (d) 0.88.
12-9. (aJ Fe ~ 2.38. (b) None. 12-11. n = 3.
12-15. (aJ /l = 20.47, !; = 0.33, T, = 1.73. i:, ~ 2.07. (b) T, - i:, = -lAO.
Chapter 13
13-1. Source DF S5 MS J? P
CS 2 0.0317805 0.01.58903 1.5.94 0.000
DC 2 0.0271854 0.01.35927 13.64 0.000
CS'DC 4 0.0006873 0.0001.71.8 0.1.7 0.9S0
Error 1.8 0.017941.3 0.0009967
Total 26 0.0775945
13-9_ Source
- -.. ~~.~
DP SS MS F P
Cone 2 7_7639 3_8819 10_62 0.001
Freeness 2 19_3739 9.6869 26.50 0.000
Time 1 20.2500 20_2500 55.40 0.000
Conc'llFreeness 4 6.0911 1.5228 4.17 0.015
Conc*Time 2 2.0817 1. 0408 2_85 0.084
:E'reeness*Time 2 2.1950 1. 0975 3.00 0.075
Conc~Freeness~Tirne 4 1.9733 0.4933 1.35 0.290
Error 18 6.5800 0.3656
Total 35 66.3089
Concentration.lIDle, Freeness. and the interaction Tlllle·Freeness axe significant at 0.05.
13--15. Main effects A. B, D. E and the interaction AB are signi:ficant
13-11. Block 1: (1). abo ac, be. Block 2: a, b. c. abe.
13-19. Block 1: (ll, abo bed. !lCd, Block 2: a. b. cd, abed, Block 3: c, abc. bd. ad, Block 4: d, abd, be. ac.
13-21. A and C are significant. 13-25. (a) D; ABC. (b) A is significant.
13M27. 2>-\ wit.i two replicates. 13..29. 25-2 design.. Estimates for A, E, and AB are large.
Chapter 14
14-1. (a) y = 10.4391 - 0.00156.<. (b) F,; 2.052. (e) --0.0038'; fl, ,; 0.00068. (d) 7.316%.
14-3. (aJY;J 1.656 - 0.041x. (b) Fo ~ 57.639. (c) 81.59%. (d) (19.374.21.388).
14-7. (a) y= 93.3399 + 15.6485.. (b) Lack of fit not significant, regression significant.
(e) 7.997'; .8, s: 23.299. (d) 74.828 $ flo s 111.852. (e) (126.012.138.910).
14-9. (a) y; --6.3378 + 9.20836x. (b) Regression is significant. (e) to ~ -23.41. reject He.
(d) (525.58,529.91). (e) (521.22, 534.28).
14-11. (a) y; 77.7895 + 11.8634x. (b) Lack of fjt not signifi=~ regression is significru:;L
(e) 0.3933. (d) (4.5661. 19.1607).
14-13. Ca) y 3.96 + 0.00169x. (b) Rcgrcsiion is significant.
(e) (0.0015.0.0019). (d) 95.2%.
14-15. Ca) y~ 69.1044 + 0.4194.<. (b) 77.35%. (e) t, ~ 5.85, reject lit.
Cd) ZO~ 1.61. (e) (05513.0.8932).
Chapter 15
15-1. Ca) Y;7.30~0.0183xl-0.399x.,. (b) Fo~ 15.19.
15-3. 15-5. y; -1.808372 + 0.00359&x, + O.1939360x, - 0.0()4;l15x,.
(-0.8024.0.0044).
IS-7. Ca) ,;; -102.713 + 0.605x, + 8.924:<" + 1.437", + 0.014.<,.
(b) Fe; 5.106. (c) fJ,. Fa ~ 0.361; fJ,.Fo O.OOO<l.
15-9. Cal y = -13729 + 105.0""" - 0.18954>'.
15·13. (a) y ~ -4.459 + 1.384" + 1.467x'. (b) Significant lack of fit. (e) F, = 16.68.
15-15. to= 1.7898. 15-21. v'IF, ~VIF,; 1.4.
Chapter 16
1/i-l. R = 2. 16-5_ R = 2. 16-9. R ~ 88.5. 16-13. R, = 75.
16-15. Z, -2.117, 16-17. K = 4.835.
Chapter 17
17-1. (a) ~ ~ 34.32, R ~ 5.65. (b) PeR, ~ 1.228. (e) 0.205%.
17-5. DI2. 17·7. LCL; 34.55. CL; 49.85. UCL ~ 65.14. 17-9. Process is not in control.
A.nswers to Selected Questions 647
Chapter 18
18-1. (al ~ 0.088. (b) L = 2, L, = 1.33. (el W = 1 h.
18-3. l~PJ.
2
1/ 1-
1/2;
o
18-7. (al p= 0
[
I~ P I-p~
P 0 1- P
p
o
A=[lOOol·
, 9 ,
18·9, (a) ,0' (b)?ii' (e) 3.
(d) 0.03. (e) 0.10. (I) 10'
18-11. (a) 0.555. (b) 56.378 min. (e) 244.18 min.
""" 0, otherwise;
I
Po i (;.!!l)i .
j~ j!
(e) p, = 0.354. (d) From4L6 to 8.33%. (el ¢= 4.17. p, 0.377.
Chapter 19
t
=(b··o) f(a+{b-a)u)du
=I.
19-5. (a) 35. (b) 4.75. (e) 5 (at time 14).
19·7, Ca) X,I, V, 1116, Xi =6, V, =6/16. (b) Yes. (e) X"o 2.
19·9. (al X=-(II.1.) tn(l- U). (b) 0.693.
649
650 Index
Variar.ce of an estimator. 218 Weibull distribution, 137, 140, Wllcoxon signed rank test. 496
Va.-iance reduction JlleU:.ods in 540 critical values for, 630
simulation. 591, 593. 596 mean and variance of, 137
Venn diag:am... 4 \Veighted least squares, 435 X
\Vllcoxon ran..t:-sum test. 499 X chart, 510, 511
W critical values for, 627. 628
Waiting-line tb.coiy, see \V1lcoxon s.:gned rank test for y
queuing paired s3JLples, 498 . Yates' algorithm fot the2k, 385