Stat 336-Design of Experiments - Dr. Eric Nyarko

DESIGN OF EXPERIMENTS
Instructor
Eric Nyarko (PhD)
Department of Statistics and Actuarial Science
University of Ghana
ericnyarko@ug.edu.gh / nyarkoeric5@gmail.com
Introduction to DOE
• Why is this trip necessary? Goals of the

course
• An abbreviated history of DOE
• Some basic principles and terminology
• The strategy of experimentation
• Guidelines for planning, conducting and
analyzing experiments
Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 2
Introduction to DOE
Chapter Learning Objectives
• Learn about the objectives of experimental design and the role it plays
in the knowledge discovery process
• Learn about different strategies of experimentation
• Understand the role that statistical methods play in designing and
analyzing experiments
• Understand the concepts of main effects of factors and interaction
between factors
• Know about factorial experiments
• Know the practical guidelines for designing and conducting
experiments

Introduction to DOE
• An experiment may be defined as a test or series of runs in which
purposeful changes are made to the input variables/factors of a process or
system so that we may observe and identify the reasons for changes that
may be observed in the output response.
• We may want to determine which input factors are responsible for the
observed changes in the response.
• Develop a model relating the response to the important input variables, and
use this model for process or system improvement or other decision
making
• “All experiments are designed experiments, some are poorly designed,

some are well-designed”.
Introduction to DOE
• A well-designed experiment is important because the results and
conclusions that can be drawn from the experiment depend to a large
extent on the manner in which the data were collected.
• Investigators perform experiments in virtually all fields

– Discover something about a process or system
– Confirm previous experience or theory
• Experiments are used widely in the engineering world

– Process characterization and optimization
– Evaluation of material properties
– Product design and development
– Component and system tolerance determination

Introduction to DOE
• The use of experimental design in product realization can result in
– Products that are easier to manufacture and that have enhanced field
performance and reliability
– Lower product cost
– Shorter product design and development time
• Designed experiments also have extensive applications
– Marketing
– Market research
– Transactional and service operations, and general business operations
• Some examples are presented in the following

Introduction to DOE

Introduction to DOE
• In general, experiments are used to study the

performance of processes and systems
• The process or system can be represented by
the model shown in this figure
• The process can be visualized as a combination
of operations, machines, methods, people, and
other resources that transforms some input
(often a material) into an output (one or more
observable response variables)
• Some of the process variables and material
properties 𝑥1 , 𝑥2 ,…, 𝑥𝑝 are controllable
• Other variables such as environmental factors
or some material properties 𝑧1 , 𝑧2 ,…, 𝑧𝑞 are
uncontrollable

Four Eras in the History of DOE
• The agricultural origins, 1908 – 1940s
– Field Trials at Rothamsted (1843 – 1846)
– W.S. Gossett and the t-test (1908)
– R. A. Fisher and his co-workers
– Profound impact on agricultural science
– Factorial designs, ANOVA
• The first industrial era, 1951 – late 1970s
– Box and Wilson, response surfaces
– Applications in the chemical and process industries
• The second industrial era, late 1970s – 1990
– Quality improvement initiatives in many companies
– Taguchi and robust parameter design, process robustness
• The modern era, beginning circa 1990
William Sealy Gosset (1876-1937)
Gosset's interest in barley cultivation
led him to speculate that design of
experiments should aim, not only at
improving the average yield, but also
at breeding varieties whose yield was
insensitive (robust) to variation in soil
and climate.
Gosset was a friend of both Karl

Pearson and R.A. Fisher, an
achievement, for each had a
monumental ego and a loathing for the
other.
Gosset was a modest man who cut

short an admirer with the comment that
“Fisher would have discovered it all
anyway.”
R. A. Fisher (1890 – 1962) George E. P. Box (1919 – 2013)
Experimental Design Terminology
• Factor is a controllable independent variable that is being
investigated to determine its effect on a response, and each
specific value of a factor is called a level
– Fixed factors can take on a finite number of levels and
these are the only level of interest
– Random factors can take on a wide range of values and
one wants to generalize from specific values to all possible
values.
• Treatments are the different levels of factors we want to

compare in an experiment. E.g., Different kinds or amount of
fertilizer, etc.

Experimental Design Terminology
• Experimental unit is the entity on which measurement or an
observation is made.
– Homogeneous experimental units are units that are as
uniform as possible on all the characteristics that could
affect the response.
• Experimental error is the random variation present in all

experimental results; different experimental units will give
different responses to same treatment.
• Confounding variable is a variable not accounted for in the
study but may cause variation in the results

The Basic Principles of DOE
• The process of planning the experiment so that appropriate
data will be collected and analyzed by statistical methods,
resulting in valid and objective conclusions is known as
statistical design of experiments
• The statistical approach to experimental design is necessary if

we wish to draw meaningful conclusions from the data
• There are two aspects to any experimental problem

– Design of the experiment
– Statistical analysis of the data

• The three basic principles (Randomization, Replication and
Blocking) of experimental design.
• Randomization
– By randomization we mean that both the allocation of the
experimental material and the order in which the
individual runs of the experiment are to be performed are
randomly determined
– Running the trials in an experiment in random order
– Notion of balancing out effects of “lurking” variables
• Replication
– By replication we mean an independent repeat run of each
factor combination
– Sample size (improving precision of effect estimation,
estimation of error or background noise)
• Blocking
– Blocking is a design technique used to improve the
precision with which comparisons among the factors of
interest are made.
– Blocking is used to reduce or eliminate the variability
transmitted from nuisance factors
– Nuisance factors are factors that may influence the

experimental response but in which we are not directly
interested

Strategy of Experimentation
• Usually, an objective of the experimenter is to determine the influence
that some factors have on the output response of the system
• The general approach to planning and conducting the experiment is
called the strategy of experimentation
• “Best-guess” experiments
– Used a lot
– More successful than you might suspect, but there are
disadvantages…
• One-factor-at-a-time (OFAT) experiments
– Sometimes associated with the “scientific” or “engineering”
method
– Devastated by interaction, also very inefficient
• Statistically designed experiments
– Based on Fisher’s factorial concept

Factorial Designs
• In a factorial experiment, all
possible combinations of
factor levels are tested
• The golf experiment:
– Type of driver
– Type of ball
– Walking vs. riding
– Type of beverage
– Time of round
– Weather
– Type of golf spike
– Etc, etc, etc…

Factorial Design

Factorial Design

Factorial Designs with Several Factors


A Fractional Factorial

Planning, Conducting and Analyzing an
Experiment
1. Recognition of and statement of problem
2. Choice of factors, levels, and ranges
Guidelines
3. Selection of the response variable(s) for designing
4. Choice of design experiment
5. Conducting the experiment
6. Statistical analysis
7. Drawing conclusions, recommendations
Assignment: Suppose that you want to design an experiment to study the

proportion of unpopped kernels of popcorn. Complete steps 1-3 of the
guidelines for designing experiments. Are there any major sources of variation
that would be difficult to control?
Planning, Conducting and Analyzing an
Experiment
• The proper use of statistical techniques in

experimentation:
– Get statistical thinking involved early

– Your non-statistical knowledge is crucial to
success
– Pre-experimental planning (steps 1-3) vital
– Think and experiment sequentially

This master guide
can be used to help
plan and design an
experiment.
It serves as a
checklist to improve
experimentation and
ensures that results
are not corrupted for
lack of careful
planning.

Important Issues and Concerns about
Experimental Design
• If something can go wrong in conducting an experiment, it will.
• The probability of successfully completing an experiment is inversely
proportional to the number of runs.
• Never let one person design and conduct an experiment alone, particularly if that
person is a subject-matter expert in the field of study.
• All experiments are designed experiments; some of them are designed well, and
some of them are designed really badly. The badly designed ones often tell you
nothing.
• About 80 percent of your success in conducting a designed experiment results
directly from how well you do the pre-experimental planning (steps 1-3 in the
7-step procedure).
• It is impossible to overestimate the logistical complexities associated with
running an experiment in a “complex” setting, such as a factory or plant.
• Finally, without good experimental design, we often end up doing Planning
After the Research is Complete (PARC) analysis.

Experiments with a Single Factor
• Understand how to set up and run a Completely Randomized Design (CRD).

• Understand how to perform a single-factor analysis of variance for a CRD.
• Know the assumptions underlying the ANOVA and how to check for
departures from these assumptions.
• Know how to apply methods for post-ANOVA comparisons for individual
differences between means.
• Know how to interpret computer output from some standard statistics packages.
• Understand several approaches for determining appropriate sample sizes in
designed experiments.

• The t-test does not directly apply when there are more than
two levels of a factor
• There are lots of practical situations where there are either
more than two levels of interest, or there are several factors
of simultaneous interest
• The analysis of variance (ANOVA) is the appropriate
analysis “engine” for these types of experiments
• The ANOVA was developed by Fisher in the early 1920s,
and initially applied to agricultural experiments
• Used extensively today for industrial experiments
(The Analysis of Variance)
• In general, there will be a levels of the factor, or a treatments, and n

replicates of the experiment, run in random order…a completely
randomized design (CRD)
• N = an total runs
• We consider the fixed effects case…the random effects case will be
discussed later
• Objective is to test hypotheses about the equality of the a treatment
means

The Analysis of Variance
• The name “analysis of variance” stems from a
partitioning of the total variability in the response
variable into components that are consistent with a
model for the experiment
• The basic single-factor ANOVA model is
 i = 1, 2,..., a
yij =  +  i +  ij , 
 j = 1, 2,..., n
 = an overall mean,  i = ith treatment effect,

 ij = experimental error, NID(0,  2 )
Models for the Data
There are several ways to write a model for the

data:
yij =  +  i +  ij is called the effects model

Let i =  +  i , then
yij = i +  ij is called the means model
Regression models can also be employed

• Total variability is measured by the total sum of
squares:
a n
SST =  ( yij − y.. ) 2
i =1 j =1
• The basic ANOVA partitioning is:

a n a n
 ij ..  i. .. ij i.
( y − y
i =1 j =1
) = [( y −
2
y ) + ( y − y
i =1 j =1
)]2
a a n
= n ( yi. − y.. ) 2 +  ( yij − yi. ) 2
i =1 i =1 j =1
SST = SSTreatments + SS E
SST = SSTreatments + SS E
• A large value of SSTreatments reflects large differences in
treatment means
• A small value of SSTreatments likely indicates no differences in
treatment means
• Formal statistical hypotheses are:
H 0 : 1 =  2 = = a
H1 : At least one mean is different

• While sums of squares cannot be directly compared to test
the hypothesis of equal means, mean squares can be
compared.
• A mean square is a sum of squares divided by its degrees
of freedom:
dfTotal = dfTreatments + df Error

an − 1 = a − 1 + a (n − 1)
SSTreatments SS E
MSTreatments = , MS E =
a −1 a(n − 1)
• If the treatment means are equal, the treatment and error
mean squares will be (theoretically) equal.
• If treatment means differ, the treatment mean square will
be larger than the error mean square.
The Analysis of Variance Table
• The reference distribution for F0 is the Fa-1, a(n-1) distribution

• Reject the null hypothesis (equal treatment means) if
F0  F , a −1, a ( n −1)


The Analysis of Variance (Example 3.1)
• We will use the analysis of

variance to test
𝐻0 : 𝜇1 = 𝜇2 = 𝜇3 = 𝜇4
against
𝐻1 : Some means are different

ANOVA Table

Estimation of the Model Parameters
• Estimators for the parameters in the single-factor model
are given by

Confidence Intervals on the Treatment
Means
• We would base the Confidence Interval (CI) on the 𝑡 distribution.
• Therefore, a 100 1 − 𝛼 % CI on the 𝑖𝑡ℎ treatment mean 𝜇𝑖 would be

Confidence Intervals on the Treatment
Means
• Differences in treatments are frequently of great practical interest.
• A 100 1 − 𝛼 % CI on the difference in any two treatment means, say

𝜇𝑖 − 𝜇𝑗 , would be

Unbalanced Data
• Unbalanced design is when in some single-factor experiments,
the number of observations taken within each treatment is
different.
• The analysis of variance described previously may still be used,
but slight modifications must be made in the sum of squares
formulas.
• Let 𝑛𝑖 observations be taken under treatment 𝑖 𝑖 = 1, 2, … 𝑎
and σ𝑎𝑖=1 𝑛𝑖 .
• The 𝑆𝑆𝑇 and 𝑆𝑆𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡𝑠 become
• Note that no other changes are required in the analysis of

variance.
Model Adequacy Checking in the ANOVA
• Checking assumptions is important

• Normality
• Constant variance
• Independence
• Have we fit the right model?
• What to do if some of these assumptions are violated?

• Examination of residuals
eij = yij − yˆ ij
= yij − yi .
• Computer software
generates the residuals
• Residual plots are very
useful
• Normal probability plot
of residuals
• Examination of the residuals
should be an automatic part of
any analysis of variance.
• If the model is adequate, the
The general impression from
residuals should be examining this display is that the
structureless; that is, they error distribution is approximately
normal.
should contain no obvious
patterns.
• A very common defect that often shows up on normal
probability plots is one residual that is very much larger
than any of the others
• Such a residual is often called an outlier (may result from
a mistake in calculations or a data coding or copying error)
• The presence of one or more outliers can seriously distort
the analysis of variance
• So when a potential outlier is located, careful investigation
is called for
• A rough check for outliers may be made by examining the
standardized residuals

• Standardized residuals
• If the standardized residuals 𝑑𝑖𝑗 are approximately normal with mean zero
and unit variance, then about
– 68% of the 𝑑𝑖𝑗 should fall within the limits ±1, about 95% of them
should fall within ±2, and virtually all of them should fall within ±3.
– A residual bigger than 3 or 4 standard deviations from zero is a
potential outlier.
• The normal probability plot (Example 3.1) gives no indication of outliers.

• Furthermore, the largest standardized residual (which should cause no
concern) is

Other Important Residual Plots
• Plotting the residuals in time order of data collection is helpful in
detecting strong correlation between the residuals.
• A tendency to have runs of positive and negative residuals indicates
positive correlation.
• This would imply that the independence assumption on the errors has
been violated.
• The plots below do not reveal any obvious pattern. Hence there is no
reason to suspect any violation of the independence or constant
variance assumptions.

Other Important Residual Plots

Dealing with Nonconstant Variance
• The usual approach to dealing with nonconstant variance when it occurs is to
apply a variance-stabilizing transformation and then to run the analysis of
variance on the transformed data.
• In this approach, one should note that the conclusions of the analysis of variance
apply to the transformed populations.
• Information about the theoretical distribution of the observations may be utilized
in choosing a transformation.
• For example, if the observations follow the
– Poisson distribution, we would utilize the square root transformation
– Lognormal distribution, we would utilize the logarithmic transformation
– For binomial data expressed as fractions, the arcsin transformation is

appropriate

Post-ANOVA Comparison of Means
• The analysis of variance tests the hypothesis of equal
treatment means
• Assume that residual analysis is satisfactory
• If that hypothesis is rejected, we don’t know which specific
means are different
• Determining which specific means differ following an
ANOVA is called the multiple comparisons problem
• We will use pairwise t-tests on means…sometimes called
Fisher’s Least Significant Difference (or Fisher’s LSD)
Method

Post-ANOVA Comparison of Means
• To use the Fisher LSD procedure, we simply compare the observed difference
between each pair of averages to the following
If the design is unbalanced
If the design is balanced
• If we conclude that the population means 𝜇𝑖 and 𝜇𝑗 are

significantly different

Design-Expert Output

The Random Effects Model
• There are a large number of possible levels for the factor

(theoretically an infinite number)
• The experimenter chooses a of these levels at random
• Inference will be to the entire population of levels

• The linear statistical random effects model
where both the treatment effects 𝜏𝑖 and 𝜀𝑖𝑗 are random variables.
• We will assume that 𝜏𝑖 and 𝜀𝑖𝑗 and independent. Because 𝜏𝑖 is
independent of 𝜀𝑖𝑗 , the variance of any observation is
• The variances 𝜎𝜏2 and 𝜎 2 are called variance components
• We can express the covariance structure of the observations in the single-

factor random effects model through the covariance matrix of the
observations.

• To illustrate, suppose that we have a = 3 treatments and n = 2
replicates. There are N = 6 observations, which we can write as a
vector
and the 6 × 6 covariance matrix of these observations is
• The main diagonals of this matrix are the variances of each individual
observation and every off-diagonal element is the covariance of a pair
of observations.
Analysis of Variance for the Random
Model
• is still valid
The basic ANOVA sum of squares identity
𝑆𝑆𝑇 = 𝑆𝑆𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡𝑠 + 𝑆𝑆𝐸
• Testing hypotheses about individual treatment effects is not very meaningful
because they were selected randomly
• We are more interested in the population of treatments, so we test hypotheses
about the variance component 𝜎𝜏2 .
• Under the null hypothesis 𝜎𝜏2 = 0, the following ratio is distributed as F with
a − 1 and N − a degrees of freedom

Estimating the Model Parameters for the
Random Model
• We are usually interested in estimating 𝜎𝜏2 and 𝜎 2 in the model.
• The analysis of variance method can be used to obtain these
estimates. A different method based on maximum likelihood can also
be used to estimate 𝜎𝜏2 and 𝜎 2 .
• The estimators of 𝜎𝜏2 and 𝜎 2 are
• For unequal sample sizes, replace the corresponding n by

Estimating the Model Parameters for the
Random Model
• The ANOVA variance component estimators are moment estimators
• Normality not required
• They are unbiased estimators
• Finding confidence intervals on the variance components is “clumsy”
• Negative estimates can occur – this is “embarrassing”, as variances are
always non-negative

Confidence Interval for the Error Variance
• A 100 1 − 𝛼 % CI for 𝜎 2 is
2 2
• Since 𝑀𝑆𝐸 = 1.90, 𝑁 = 16, 𝑎 = 4, 𝜒0.025,12 = 23.3364 and 𝜒0.975,12 =
4.4038, the 95% CI on 𝜎 2 is 0.9770 ≤ 𝜎 2 ≤ 5.1775
• Notice that an exact confidence interval for 𝜎𝜏2 cannot be constructed.

However, there are approximate procedures.

Randomized Blocks, Latin Squares, and
Related Designs
• Know how the blocking principle can be effective in reducing the variability
arising from controllable nuisance factors.
• Learn about the randomized complete block design.
• Understand how the analysis of variance can be extended to the randomized
complete block design.
• Know how to do model adequacy checking for the randomized complete block
design.
• Understand how a Latin square design can be used to control two sources of
nuisance variability in an experiment.

The Blocking Principle
• Blocking and nuisance factors

• The randomized complete block design or the RCBD
• Extension of the ANOVA to the RCBD
• Other blocking scenarios…Latin square designs

• Blocking is a technique for dealing with nuisance factors
• A nuisance factor is a factor that probably has some effect on the
response, but it’s of no interest to the experimenter…however, the
variability it transmits to the response needs to be minimized
• Typical nuisance factors include batches of raw material, operators,
pieces of test equipment, time (shifts, days, etc.), different
experimental units
• Many industrial experiments involve blocking (or should)
• Failure to block is a common flaw in designing an experiment
(consequences?)

• If the nuisance variable is known and controllable, we use blocking
• If the nuisance factor is known and uncontrollable, sometimes we can

use the analysis of covariance to remove the effect of the nuisance
factor from the analysis
• If the nuisance factor is unknown and uncontrollable (a “lurking”

variable), we hope that randomization balances out its impact across
the experiment
• Sometimes several sources of variability are combined in a block, so

the block becomes an aggregate variable

The Hardness Testing Example
• We wish to determine whether 4 different tips produce different
(mean) hardness reading on a Rockwell hardness tester
• Gauge and measurement systems capability studies are frequent areas

for applying design of experiments
• Assignment of the tips to an experimental unit; that is, a test coupon
• Structure of a completely randomized experiment
• The test coupons are a source of nuisance variability
• Alternatively, the experimenter may want to test the tips across

coupons of various hardness levels
• The need for blocking

• To conduct the hardness testing example as a RCBD, assign all 4 tips
to each coupon
• Each coupon is called a “block”; that is, it’s a more homogenous

experimental unit on which to test the tips
• Variability between blocks can be large, variability within a block

should be relatively small
• In general, a block is a specific level of the nuisance factor
• A complete replicate of the basic experiment is conducted in each

block
• A block represents a restriction on randomization
• All runs within a block are randomized

• Suppose that we use b = 4 blocks:
• Notice the two-way structure of the experiment

• Once again, we are interested in testing the equality of treatment
means, but now we have to remove the variability associated with the
nuisance factor (the blocks)

Extension of the ANOVA to the RCBD
• Suppose that there are a treatments (factor levels) and b blocks
• A statistical model (effects model) for the RCBD is
 i = 1, 2,..., a
yij =  +  i +  j +  ij 
 j = 1, 2,..., b
• The relevant (fixed effects) hypotheses (for testing the equality of treatment
means)
where the 𝑖th treatment mean
• An equivalent way to write the above hypotheses in terms of the treatment

effects

• Let 𝑦𝑖. be the total of all observations taken under treatment 𝑖, 𝑦.𝑗 be
the total of all observations in block 𝑗, 𝑦.. be the grand total of all
observations, and N = ab be the total number of observations.
• Expressed mathematically,
• Similarly, 𝑦ത𝑖. = 𝑦𝑖.Τ𝑏 is the average of the observations taken under

𝑦
treatment 𝑖, 𝑦ത.𝑗 = .𝑗Τ𝑎 is the average of the observations in block 𝑗,
and 𝑦ത.. = 𝑦..Τ𝑁 is the grand average of all observations.

• ANOVA partitioning of total variability:
a b a b
(y
i =1 j =1
ij − y.. ) =   [( yi. − y.. ) + ( y. j − y.. )
2
i =1 j =1
+ ( yij − yi . − y. j + y.. )]2

a b
= b ( yi . − y.. ) + a  ( y. j − y.. ) 2
2
i =1 j =1
a b
+   ( yij − yi . − y. j + y.. ) 2
i =1 j =1
SST = SSTreatments + SS Blocks + SS E

• The degrees of freedom for the sums of squares in
SST = SSTreatments + SS Blocks + SS E

are as follows:
ab − 1 = a − 1 + b − 1 + ( a − 1)(b − 1)
• Therefore
– The ratios of sums of squares to their degrees of freedom result in
mean squares
– The ratio of the mean square for treatments to the error mean
square is an F statistic that can be used to test the hypothesis of
equal treatment means

ANOVA Display for the RCBD
• To test the equality of treatment means, we

would reject 𝐻0 : 𝜏𝑖 = 0 if 𝐹0 > 𝐹𝛼,𝑎−1,(𝑎−1)(𝑏−1)
• To test the equality of block means, we would
reject 𝐻0 : 𝛽𝑗 = 0 if 𝐹0 > 𝐹𝛼,𝑏−1,(𝑎−1)(𝑏−1)

Vascular Graft Example

• To conduct this experiment as a RCBD, assign all 4 pressures to
each of the 6 batches of resin
• Each batch of resin is called a “block”; that is, it’s a more

homogenous experimental unit on which to test the extrusion
pressures

The ANOVA is shown in the table below.

Using 𝛼 = 0.05, the critical value of 𝐹 is
𝐹0.05,3,15 = 3.29. Because 8.11>3.29, we
conclude that extrusion pressure affects the
mean yield.

(Design-Expert Output)

Vascular Graft Example (JMP Output)

Residual Analysis for the


• Basic residual plots indicate that normality, constant
variance assumptions are satisfied
• No obvious problems with randomization
• No patterns in the residuals vs. block
• Can also plot residuals versus the pressure (residuals by
factor)
• These plots provide more information about the constant
variance assumption, possible outliers

Multiple Comparisons for the Vascular Graft
Example – Which Pressure is Different?
To compute LSD for RCBD, replace n and 𝑑𝑓𝐸 = 𝑁 − 𝑎 in the LSD formula
for CRD by b and 𝑑𝑓𝐸 = (𝑎 − 1)(𝑏 − 1), respectively.
Random Blocks and/or Treatments
• Suppose that there are a treatments (fixed) and b blocks (random)
• A statistical model for the RCBD is
 i = 1, 2,..., a
yij =  +  i +  j +  ij 
 j = 1, 2,..., b
• This is a special case of a mixed model (because it contains both fixed and
random factors).
• In this situation, it can be shown that

• The expected mean squares from the usual ANOVA partitioning of the total
sum of squares are
• The appropriate statistic for testing the null hypothesis of no treatment

effects (all 𝜏𝑖 = 0)
• An ANOVA-type estimator of the variance component for blocks is
• For the vascular graft experiment (Example 4.1), the estimate of 𝜎𝛽2 is


Estimating Model Parameters and the
General Regression Significance Test
• Suppose that there are a treatments (fixed) and b blocks (fixed)
• A statistical model for the RCBD is
 i = 1, 2,..., a
yij =  +  i +  j +  ij 
 j = 1, 2,..., b
• We may estimate the parameters in the RCBD model by least squares

• The normal equations may be simplified as
• The estimated or fitted values of 𝑦𝑖𝑗 is

Estimating Missing Observations in RCBD
• When using the RCBD, sometimes an observation in one of the blocks

is missing.
• This may happen because of carelessness or error or for reasons
beyond the experimenter’s control, such as unavoidable damage to an
experimental unit.
• The missing value causes the design to be unbalanced.
• A missing observation introduces a new problem into the analysis.
• Thus, treatments are no longer orthogonal to blocks (i.e., every
treatment does not occur in every block)

• There are two general approaches to the missing value problem
– Approximate procedure
– Exact procedure (utilizes the general regression significance test)
• The approximate procedure consists of estimating the missing value so

that the error mean square is minimized.
• The missing observation is estimated and the usual analysis of variance

is performed just as if the estimated observations were real data, with
the error degrees of freedom reduced by 1.

• Suppose the observation 𝑦𝑖𝑗 for treatment i in block j is missing.
Denote the missing observation by x.
• As an illustration, suppose that in the vascular graft experiment of
Example 4.1 there was a problem with the extrusion machine when the
8700 psi run was conducted in the fourth batch of material, and the
observation 𝑦24 could not be obtained as shown in the table below.

• In general, let y′𝑖𝑗 represent the grand total with one missing
observation, y′𝑖. represent the total for the treatment with one missing
observation, and y′.𝑗 be the total for the block with one missing
observation.
• Suppose we wish to estimate the missing observation x so that x will
have a minimum contribution to the error sum of squares.
• The estimate of the missing observation is given by
• From the vascular graft experiment with one missing value, we find
y′2. = 455.4, y′.4 = 267.5, and y′.. = 2060.4.
• Therefore,

• The usual analysis of variance may now be performed using 𝑦24 =
91.08 and reducing the error degrees of freedom by 1.
• The analysis of variance is shown below.
NOTE:
• We may use the above “missing observation equation” iteratively to estimate several missing
values.
• To illustrate the iterative approach, suppose that two values are missing. Arbitrarily estimate the
first missing value, and then use this value along with the real data and the “missing
observation equation” to estimate the second. Now the “missing observation equation” can be
used to reestimate the first missing value, and following this, the second can be reestimated.
This process is continued until convergence is obtained. In any missing value problem, the error
degrees of freedom are reduced by one for each missing observation.
The Latin Square Design
• These designs are used to simultaneously control (or eliminate) two
sources of nuisance variability
• A significant assumption is that the three factors (treatments, nuisance
factors) do not interact
• If this assumption is violated, the Latin square design will not produce
valid results
• Latin squares are not used as much as the RCBD in industrial
experimentation
• However, it can be useful in situations where the rows and columns
represent factors the experimenter actually wishes to study and where
there are no randomization restrictions.
• Thus, three factors (rows, columns, and letters), each at p levels, can be
investigated in only 𝑝2 runs.

• This is a 5 × 5 Latin square design

• Notice that the design is a square arrangement and that the five
formulations (or treatments) are denoted by the Latin letters A, B, C, D,
and E; hence the name Latin square.
• Both batches of raw material (rows) and operators (columns) are
orthogonal to treatments (each letter occurs once and only once in each
row and column)
• Some examples of Latin squares are
A Latin square in which the first row and column consists of the letters
written in alphabetical order is called a standard Latin square

Statistical Analysis of the
Latin Square Design
• The statistical (effects) model is
 i = 1, 2,..., p

yijk =  +  i +  j +  k +  ijk  j = 1, 2,..., p
k = 1, 2,..., p

where 𝑦𝑖𝑗𝑘 is the observation in the ith row and kth column for the jth
treatment, 𝜇 is the overall mean, 𝛼𝑖 is the ith row effect, 𝜏𝑗 is the jth
treatment effect, 𝛽𝑘 is the kth column effect, and 𝜀𝑖𝑗𝑘 is the random error.
• Note that this model is completely additive; that is, there is no

interaction between rows, columns, and treatments
• The statistical analysis (ANOVA) is much like the analysis for the
RCBD.

Latin Square Design

Latin Square Design

Latin Square Design

Latin Square Design
• As in any design problem, the experimenter should investigate the
adequacy of the model by inspecting and plotting the residuals.
• For a Latin square, the residuals are given by
Assignment:
Find the residuals for the rocket propellant problem and construct
appropriate plots.

Estimating Missing Observations in
Latin Square Design
• Occasionally, one observation in a Latin square is missing.
• For a p × p Latin square, the missing value may be
estimated by
• where y′𝑖.. , y′.𝑗. and y′..𝑘 indicate totals for the row, column,
and treatment with the missing value, respectively, and y′...
is the grand total with the missing value.

Replication of Latin Squares
• A disadvantage of small Latin squares is that they provide a relatively

small number of error degrees of freedom.
• For example, a 3 × 3 Latin square has only two error degrees of

freedom, a 4 × 4 Latin square has only six error degrees of freedom,
and so forth.
• When small Latin squares are used, it is frequently desirable to

replicate them to increase the error degrees of freedom.
• A Latin square may be replicated in several ways.

• To illustrate, suppose that the 5 × 5 Latin square used in the rocket
propellant problem is replicated n times.
• This could have been done as follows:
1. Use the same batches and operators in each replicate.
2. Use the same batches but different operators in each replicate (or,
equivalently, use the same operators but different batches).
3. Use different batches and different operators.
• The analysis of variance depends on the method of replication.

• Consider case 1, where the same levels of the row and column
blocking factors are used in each replicate.
• Let 𝑦𝑖𝑗𝑘𝑙 be the observation in row i, treatment j, column k, and
replicate l. There are 𝑁 = 𝑛𝑝2 total observations.
• The ANOVA is summarized below:

Crossover Designs
• Occasionally, one encounters a problem in which time periods are a factor in
the experiment.
• In general, there are p treatments to be tested in p time periods using np
experimental units.
• For example, a human performance analyst is studying the effect of two
replacement fluids on dehydration in 20 subjects as shown below.
• In the first period, half of the subjects (chosen at random) are given fluid A and
the other half fluid B.
• At the end of the period, the response is measured and a period of time is
allowed to pass in which any physiological effect of the fluids is eliminated.
• Then the experimenter has the subjects who took fluid A take fluid B and those
who took fluid B take fluid A.
• This design is called a crossover design
The Graeco-Latin Square Design
• Consider a p × p Latin square, and superimpose on it a second p × p
Latin square in which the treatments are denoted by Greek letters.
• If the two squares when superimposed have the property that each
Greek letter appears once and only once with each Latin letter, the two
Latin squares are said to be orthogonal, and the design obtained is
called a Graeco-Latin square.
• An example of a 4 × 4 Graeco-Latin square is shown below:

The Graeco-Latin Square Design
• The Graeco-Latin square design can be used to control systematically
three sources of extraneous variability, that is, to block in three
directions.
• The design allows investigation of four factors (rows, columns, Latin
letters, and Greek letters), each at p levels in only 𝑝2 runs.
• Graeco-Latin squares exist for all p ≥ 3 except p = 6.
• The statistical model for the Graeco-Latin square design is
where 𝑦𝑖𝑗𝑘𝑙 is the observation in row i and column l for Latin letter j and Greek
letter k, 𝜃𝑖 is the effect of the ith row, 𝜏𝑗 is the effect of Latin letter treatment j, 𝜔𝑘
is the effect of Greek letter treatment k, Ψ𝑙 is the effect of column l, and 𝜖𝑖𝑗𝑘𝑙 is an
𝑁𝐼𝐷(0, 𝜎 2 ) random error component.
ANOVA for Graeco-Latin Square Design

Graeco-Latin Square Design
• The complete ANOVA is

summarized in the table below.
• Reject 𝐻0 if 𝐹0 > 𝐹𝛼,𝑝−1,(𝑝−3)(𝑝−1)
• Formulations are significantly

different at 1%.

Graeco-Latin Square Design

Balanced Incomplete Block Designs
• In certain experiments using randomized block designs, we may not be
able to run all the treatment combinations in each block.
• This may be as a result of shortages of experimental apparatus or

facilities or the physical size of the block.
• For example, in the vascular graft experiment (Example 4.1), suppose

that each batch of material is only large enough to accommodate
testing three extrusion pressures. Therefore, each pressure cannot be
tested in each batch.
• For this type of problem it is possible to use randomized block

designs in which every treatment is not present in every block.
• These designs are known as randomized incomplete block designs

Balanced Incomplete Block Designs
• A balanced incomplete block design (BIBD) is an incomplete block

design in which any two treatments appear together an equal number
of times.
• Suppose that there are a treatments and that each block can hold
exactly k (k < a) treatments.
𝑎
• A balanced incomplete block design may be constructed by taking 𝑘
blocks and assigning a different combination of treatments to each

block.

Statistical Analysis of the BIBD
• For the BIBD, we assume that there are a treatments and b blocks.
• In addition, we assume that each block contains k treatments, that each

treatment occurs r times in the design (or is replicated r times), and
that there are N = ar = bk total observations.
• Furthermore, the number of times each pair of treatments appears in

the same block is
• The design is said to be symmetric if a = b.
• Notice that the parameter λ must be an integer.

• The statistical model for the BIBD is
where 𝑦𝑖𝑗 is the ith observation in the jth block, 𝜇 is the overall mean, 𝜏𝑖 is the
effect of the ith treatment, 𝛽𝑗 is the effect of the jth block, and 𝜖𝑖𝑗 is the
𝑁𝐼𝐷(0, 𝜎 2 ) random error component.
• The total variability in the data is expressed by the total corrected sum of
squares:
• The block sum of squares (df=b-1) is
where 𝑦.𝑗 is the total in the jth block.

• The adjusted treatment sum of squares (df=a-1) is
where 𝑄𝑖 is the adjusted total for the ith treatment, which is computed as
with 𝑛𝑖𝑗 = 1 if treatment i appears in block j and 𝑛𝑖𝑗 = 0 otherwise.
• The error sum of squares (df=N − a − b + 1) is
• The statistic for testing the equality of the treatment effects is

• The ANOVA is summarized below:

• Consider the data below for the catalyst experiment


The analysis of variance is shown above.
Because the P-value is small, we conclude

that the catalyst employed has a significant
effect on the time of reaction.

Split-Plot Design
• Understand how a split-plot design handles easy-to-change and
hard-to-change factors in an experiment.
• Understand why there are two sources of variability in a split-plot
experiment.
• Know how to construct and analyze a split-plot design.

Split-Plot Design
• The split-plot design finds reasonably widespread
application in the industrial use of designed experiments.
• It also frequently involve one or more random factors
• It is a multifactor experiment where it is not possible to
completely randomize the order of the runs
• Example – paper manufacturing
– Three pulp preparation methods
– Four different temperatures
– Each replicate requires 12 runs
– The experimenters want to use three replicates
– How many batches of pulp are required?

Split-Plot Design
• Pulp preparation methods is a hard-to-change

factor
• Consider an alternate experimental design:
– In replicate 1, select a pulp preparation method,
prepare a batch
– Divide the batch into four sections or samples, and
assign one of the temperature levels to each
– Repeat for each pulp preparation method
– Conduct replicates 2 and 3 similarly

Split-Plot Design
• Each replicate (sometimes called blocks) has been
divided into three parts, called the whole plots
• Pulp preparation methods is the whole plot
treatment
• Each whole plot has been divided into four
subplots or split-plots
• Temperature is the subplot treatment
• Generally, the hard-to-change factor is assigned to
the whole plots
• This design requires only 9 batches of pulp
(assuming three replicates)
The Split-Plot Design
Model and Statistical Analysis
yijk =  +  i +  j + ( )ij +  k + ( )ik + (  ) jk
 i = 1, 2,..., r

+ ( )ijk +  ijk  j = 1, 2,..., a
 k = 1, 2,..., b

There are two error

structures; the
whole-plot error and
the subplot error

Split-Plot ANOVA
Calculations follow a three-factor ANOVA with one replicate

Note the two different error structures; whole plot and subplot

Stat 336-Design of Experiments - Dr. Eric Nyarko

Uploaded by

Copyright:

Available Formats

You might also like

Stat 336-Design of Experiments - Dr. Eric Nyarko

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Stat 336-Design of Experiments - Dr. Eric Nyarko

Uploaded by

Copyright:

Available Formats

DESIGN OF EXPERIMENTS

• Why is this trip necessary? Goals of the

Chapter Learning Objectives

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 3

• “All experiments are designed experiments, some are poorly designed,

• Investigators perform experiments in virtually all fields

• Experiments are used widely in the engineering world

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 5

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 6

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 7

• In general, experiments are used to study the

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 8

Gosset was a friend of both Karl

Gosset was a modest man who cut

• Treatments are the different levels of factors we want to

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 12

• Experimental error is the random variation present in all

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 13

• The statistical approach to experimental design is necessary if

• There are two aspects to any experimental problem

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 14

– Nuisance factors are factors that may influence the

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 16

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 17

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 18

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 19

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 20

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 21

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 22

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 23

Assignment: Suppose that you want to design an experiment to study the

• The proper use of statistical techniques in

– Get statistical thinking involved early

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 25

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 26

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 27

Chapter Learning Objectives

• Understand how to set up and run a Completely Randomized Design (CRD).

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 28

• In general, there will be a levels of the factor, or a treatments, and n

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 30

 = an overall mean,  i = ith treatment effect,

There are several ways to write a model for the

yij =  +  i +  ij is called the effects model

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 32

• The basic ANOVA partitioning is:

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 34

dfTotal = dfTreatments + df Error

• The reference distribution for F0 is the Fa-1, a(n-1) distribution

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 36

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 37

• We will use the analysis of

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 38

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 39

• Estimators for the parameters in the single-factor model

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 40

Dr. Eric Nyarko (UG) STAT 336: Design of Experiments 41

• Differences in treatments are frequently of great practical interest.