Professional Documents
Culture Documents
Chapter 4 Experimental Design
Chapter 4 Experimental Design
Chapter 4 Experimental Design
Chapter 4 Experimental design 4.1 Overview A designed experiment is a test or series of tests in which purposeful changes are made to the input variables of a process so that the resulted changes in the output variables (responses) may be observed and identified. Many of the earlier applications of experimental design were in the agriculture and biological sciences. It has found broad applications in many disciplines now. Consider the tension bond strength of cement mortar which is an important characteristic of ready-mixed cement. An engineer is interested in comparing the strength of a modified formulation in which polymer latex emulsions (referred to as a factor) have been added during the mixing to the strength of unmodified formulation. The engineer will collect observations on tested items selected these two formulations. The two different formulations are referred to as treatments. A two-sample t-test can be used to test whether the modified one on the average has larger tension bond strength. If we have more than two formulations to compare the statistical techniques ANOVA we discussed earlier may be used to test the difference in the bond strength of different formulations. Formally, this involves several steps in the experimental process. 1. Statement of the problem objectives 2. Factors choose the factors to be varied in the experiment, and the specific levels at which runs will be made. Need practical experience and theoretical understanding. 3. Selection of response variable provides useful information about the process under study. 4. Choice of experimental design sample size, replications, run order, blocking, randomization. Keep design and analysis as simple as possible. Should consider costs and other resources. 5. Performing the experiment 6. Data analysis appropriate statistical methods, results and conclusions should be objective rather than judgmental in nature. 7. Conclusions and recommendations practical conclusions about results and recommendations for action.
Experiment design
4.2 Factorial experiments Factorial experiments are employed to study simultaneously the effects of two or more factors, each covers a pre-determined number of levels. By a factorial design we mean that in each complete trial or replicate of the experiment all possible combinations of the levels of the factors are investigated. Thus if there are a levels of factor A and b levels of factor B, then each replicate contains all ab treatment combinations. We shall have a twoway or higher-way classification experiment. The effect of a factor is defined as the change in response produced by a change in the level of the factor. This is called a main effect because it refers to the primary factor under study. In some experiments, the difference in response between the levels of one factor is not the same at all levels of the other factors. When this occurs there is an interaction effect between the factors. Factorial designs are more efficient than one-factor-at-a-time experiments. A factorial design is necessary when interactions are present. 4.2.1 Two-factor factorial design Let these factors be A and B. Each observation will be denoted by xijt . Formally, we assume
Experiment design
a factor A treatment sums = X i.. , (nA = nb observations each) b factor B treatment sums = X . j . , (nB = na observations each)
2 X i2 X ... .. = SSA N i =1 n A a
Treatment B SS = na ( x. j . x... ) 2 =
j =1
n
j =1 a
X .2j .
B b
Total SS
( xijt x... )2 =
i =1 j =1 t =1
x
i =1 j =1 t =1
2 X ... N
When factors A and B interact the interaction variation is: 2 a b X2 a b X ... ij . 2 - SSA - SSB AB interaction SS = n ( xij . xi.. x. j . + x... ) = N i =1 j =1 n i =1 j =1
The total SS is decomposed as Total SS = Treatment A SS + Treatment B SS + AB Interaction SS + Error SS Error SS is obtained by subtraction. As before, Error SS =
2 2 2 ( xijt xi.. )2 ~ 2 n21 = 2 ab ( n 1) = N ab i =1 j =1 t =1 i =1 j =1 a b n a b
Here we do not require H0. The error SS always has a 2 - distribution whether or not H0 is true. Under H0, Treatment A SS = nb ( xi.. x... ) 2 ~ 2 a 1 Treatment B SS = na ( x. j . x... ) 2 ~ 2 b 1
j =1 a
i =1 b
Experiment design
Mean squared SS
Sum of squares
2 X i2 X ... .. N i =1 n A
Degree of freedom
Mean square
F-ratio
2 sA s2 2 sA s2 2 s AB s2
a-1
b-1
(a-1)(b-1) By subtraction = ab(n-1) = N- ab
2 A
j =1
a
X .2j . nB
b
2 X ... N
2 sB
i =1 j =1
2 X ij .
2 s AB
s2
By subtraction
x
i =1 j =1 t =1
2 ijt
2 X ... N
N-1
Note: When the interaction effect is not significant or when there are no replications (i.e. n = 1), we may omit the row of interaction. In the former situation, the random SS will be the original random SS plus the AB interaction SS and the d.f. will accordingly be increased. In the latter, no interaction SS can be calculated due to no replications.
The randomized block design discussed earlier may be regarded as a two-factor (one factor is plot and the other is fertilizer) experiment without replication. The interaction effect cannot be separated from the error SS.
Experiment design
Example
A department store conducted an experiment to study the sales of a product for different package designs and different selling prices. There were three different package designs at two different selling prices. (This is equivalent to a randomized block design with two replications. Package is the block and price is the treatment.)
Package design 1 2 3
They want to test: 1. The hypothesis that there are no interactions for any combination of package design and a selling price. 2. The hypothesis that the mean population amount of sales is the same for three package designs. 3. The hypothesis that the mean population sales amount is the same for two selling prices. 4. Locate where the difference are if the hypothesis is significant. Totals: the values inside table are interaction totals X ij . the row and column totals are factor totals X i.. , X . j . the overall total is grand treatment total X ... Package Selling price design (Higher) 1 26 2 28 3 78 132 X i.. a = 2, b = 3, n =2, N = 12, Selling price (Lower) 30 20 76 126
X . j.
56 48 154 258
Total SS = 142 + 122 + 172 + 132 + 372 + 412 + 352 + 412 + 122 + 162 + 112 + 92 - 2582/12 = 1809 Package SS = 562/4 + 1542/4 + 482/4 - 2582/12 = 1742 Price SS = 1322/6 + 1262/6 + 482/4 - 2582/12 = 3 Interaction SS = 262/2 + 302/2 + 782/2 + 762/2 + 282/2 + 202/2 - 2582/12 - 1742 3 = 18
Experiment design
ANOVA Table
Source of variation Package Price Package x Price Error Total Sum of squares 1742 3 18 46 1809 Degree of freedom 2 1 2 6 11 Mean square 871 3 9 7.67 F-ratio 113.6 0.391 1.173 F-critical at 5% 5.143 5.987 5.143
Conclusion At 5% level, the difference in package design effect on the mean sales is significant, whereas the difference in price effect and interaction effect of price and package design are not significant.
Experiment design
Total 4 56 14 4.666667
2 78 39 8
2 76 38 18
4 154 38.5 9
2 28 14 8
2 20 10 2
4 48 12 8.666667
Count Sum Average Variance ANOVA Source of Variation Sample Columns Interaction Within Total
6 132 22 177.2
6 126 21 184
SS 1742 3 18 46 1809
df 2 1 2 6 11
MS 871 3 9 7.666667
Experiment design
Material B
100 122 110 332 330 640 500 196 170 386 365 810 725
220 108 85 276 310 612 500 136 130 333 330 779 670
715 66 50 248 275 612 610 108 75 472 350 893 890
870 80 60 276 310 696 610 136 75 499 390 1820 890
Material B
Speed total
Experiment design
Speed x Feed
Feed .004 .008 .014 Speed total 100 598 1413 2675 4686 220 459 1249 2561 4269 Speed 475 375 1191 2553 4119 715 299 1345 3005 4649 870 351 1475 4016 5842 Row total 2081 6673 14810 23565
Material x Feed
Material B V Feed total .004 849 1233 2081 .008 2900 3773 6673 Feed .014 5773 9037 14810 Material total 9522 14043 23565 = 5831591.25
Speed SS = 46862/12 + 42692/12 + 41192/12 + 46492/12 + 58422/12 - 235652 / 60 = 152453.17 Material SS = 95222/30 + 140432/30 - 235652 / 60 Feed SS = 20822/20 + 66732/20 + 148102/20 - 235652 / 60 SxM SS = 20342/6 + 18912/6 + + 27882/6 + 38102/6 - 235652/ 60 - 152453.17 - 340657.35 SxF SS = 5982/4 + 4592/4 + + 30052/4 + 40162/4 - 235652 / 60 - 152453.17 - 4154833.9 MxF SS = 8492/10 + 29002/10 + + 37732/10 + 90372/10 - 235652 / 60 - 340657.35 - 4154833.9 = 340657.35 = 4154833.9
= 88111.56
= 255471.43
= 237506.7
Experiment design
SxMxF SS = 2322/2 + 1932/2 + + 17832/2 + 27102/2 - 235652 / 60 - 152453.17 - 340657.35 -4154833.9 - 88111.56 - 255471.43 - 237506.7
= 113084.64
Higher Interaction SS = sum of interaction total squares/replication number - Total correction SS All main factor SS and lower interaction SS
* at 0.01
Since the three factor interaction effect is not significant, we may assume no three factor interaction so that its SS is combined with Error SS to form a new Error SS with 38 degrees of freedom. ANOVA Table without three factor interaction Source Speed Material Feed SxM SxF MxF Error Total SS 152453.17 340657.35 4154833.9 88111.56 255471.43 237506.7 602557.14 5831591.25 DF 4 1 2 4 8 2 38 59 MS 38113.3 340657.35 2077416.95 22027.89 31933.9 118753.35 15856.77 F 2.404 21.483 131.011 1.389 2.0139 7.489
The conclusions are not affected. The latter is a more powerful test because it has larger d.f. for the Error SS.
10
Experiment design
Main effect of Feed at different levels of Material Feed Material .004 .008 .014 B 84.9 290 577.3 V 123.3 377.3 903.7
11
Experiment design
S1 S2 S3 S4 S5
Response
Feed
Response
580
B
480 380 280 180 80 F1 F2 F3
Feed
12
Experiment design
Computer output from MINITAB General Linear Model: Thrust versus Speed, Feed, Material
Factor Speed Feed Material Type fixed fixed fixed Levels 5 3 2 Values 100, 220, 475, 715, 870 0.004, 0.008, 0.014 B, V
Analysis of Variance for Thrust, using Adjusted SS for Tests Source Speed Feed Material Speed*Feed Speed*Material Feed*Material Speed*Feed*Material Error Total DF 4 2 1 8 4 2 8 30 59 Seq SS 152453 4154834 340657 255471 88112 237507 113085 489473 5831591 Adj SS 152453 4154834 340657 255471 88112 237507 113085 489473 Adj MS 38113 2077417 340657 31934 22028 118753 14136 16316 F 2.34 127.33 20.88 1.96 1.35 7.28 0.87 P 0.078 0.000 0.000 0.088 0.275 0.003 0.555
S = 127.733
R-Sq = 91.61%
R-Sq(adj) = 83.49%
Unusual Observations for Thrust Obs 55 60 Thrust 1820.00 890.00 Fit 1355.00 1355.00 SE Fit 90.32 90.32 Residual 465.00 -465.00 St Resid 5.15 R -5.15 R
13
Experiment design
Mean of Thrust
200 100 800 600 400 200 B V 220 475 Material 715 870 0.004 0.008 0.014
Interaction Plot
0.004 1000 0.008 0.014 S peed 100 220 475 715 870 1000 F eed 0.004 0.008 0.014
500
Speed
Feed
500
1000
500
14
Experiment design
A B C D
D A B C
C D A B
A B C D E
B C D E A
E A B C D
A C E B D
B D A C E
D A C E B
E B D A C
Model
xijt = + Ri + C j + Tt + ijt , i, j, t = 1, 2, , n. Ri : ith row effect Cj : jth column effect Tt : tth treatment effect
Assumptions
R C
i
=0 =0 =0
T
t
15
Experiment design
( x
i =1 j =1
ijt
( x
i =1 j =1
ijt
N = n2
Row total Column total Treatment total Grand total = X i.. = xijt
j =1 n n
= X . j . = xijt
i =1 n
= X ..t =
i or j =1 j n
xijt
= X ... = xijt
i =1 j =1
2 X ... = xijt N i =1 j =1
n n
X i2 .. i =1 n n X2 = . j. j =1 n
n
2 X ... N
2 X ... N
2 X ..2t X ... N t =1 n n
16
Experiment design
DF n1 n1 n1 (n 1)( n 2) N1
MS st2
2 sR
2 sC
s2
2 xijt i =1 j =1
2 X ... N
Example
An experimenter is studying the effect of 5 different formulations of an explosive mixture in the manufacture of dynamite on the observed explosive force. Each formulation is mixed from a batch of raw materials that is only enough for five formulations to be tested. Furthermore, the formulations are prepared by several operators. The appropriate design for this problem consists of testing each formulation exactly once in each batch of raw material and for each formulation to be prepared exactly once by each of five operators. A Latin square design is resulted for dynamite formulation Operator 3 C=19 D=30 E=26 A=26 B=20 121
A, B, C, D and E are the five formulations. Treatment totals are A = 143, B = 101, C = 112, D = 149, E = 130
17
Experiment design
= 1072/5 + 1432/5 + 1212/5 + 1302/5 + 1342/5 - 6352/5 = 1112/5 + 1342/5 + 1302/5 + 1282/5 + 1322/5 - 6352/5 = 1432/5 + 1012/5 + 1122/5 + 1492/5 + 1302/5 - 6352/5 = 242 + 172 + + 312 - 6352/5
ANOVA Table
Source Formulations Batches Operators Error Total SS 330 68 150 128 676 DF 4 4 4 12 24 MS 82.50 17.00 37.50 10.67 F 7.73
Conclusion: Different formulations have different explosive forces. Though we may test the batches and operator effects these usually are not of prime interest.
18
Experiment design
Practical Exercise
Question 1. (Training.xls) The personnel manager of a large insurance company wished to evaluate the effectiveness of four different sales-training programs designed for new employees. A group of 32 recently hired college graduates were randomly assigned to four programs so that there were eight subjects in each program. At the end of the month-long training period a standard exam was administered to the 32 subjects; the scores are given below: Programs A 66 74 82 75 73 97 87 78 B 72 51 59 62 74 64 78 63 C 61 60 57 60 81 55 70 71 D 63 61 76 84 58 65 69 80
Determine whether there is evidence of a difference in the four sales-training programs at the 0.05 level of significance. Question 2. (Cholesterol.xls) The following are the cholesterol contents, in milligrams per package, which four laboratories obtained for 6-ounce packages of three very similar diet foods: Diet food A 3.4 3.0 3.3 3.5 Diet food B 2.6 2.7 3.0 3.1 Diet food C 2.8 3.1 3.4 3.7
Laboratory 1 2 3 4
Test whether the cholesterol contents differ among the 3 diet foods at the 0.05 level of significance.
19
Experiment design
Question 3. (Effluent.xls) Ozonization as a secondary treatment for effluent, following absorption by ferrous chloride, was studied for three reaction times and three pH levels. The study yielded the following results for effluent decline. Reaction time (min) 20 pH level 7.0 23 21 22 20 22 19 21 20 19 pH level 9.0 16 18 15 14 13 12 13 12 12 pH level 10.5 14 13 16 12 11 10 11 13 12
40
60
a) Test for significant interaction. b) Test for significant differences among reaction times. c) Test for significant differences among pH levels.
20