CSCI 688

Homework 7a

Megan Rose Bryant

Department of Mathematics
William and Mary

November 19, 2014

8.1 Suppose that in the chemical process development experiment described in Problem 6.7, it was only
possible to run a one-half fraction of the 24 design. Construct the design and perform the statistical analysis,
using the data from replicate 1.
We will use the following 24−1 design with the defining relation I = ABCD.
Fractional Factorial Design
Factors: 4 Base Design: 4, 8 Resolution: IV
Runs: 8 Replicates: 1 Fraction: 1/2
Blocks: 1 Center pts (total): 0

Design Generators: D = ABC

Defining Relation: I = ABCD

Alias Structure
I + ABCD
A + BCD
B + ACD
C + ABD
D + ABC
AB + CD
AC + BD
AD + BC

Design Table
Run A B C D
1 - - - -
2 + - - +
3 - + - +
4 + + - -
5 - - + +
6 + - + -
7 - + + -
8 + + + +
We see from our initial analysis shown below that factors A, AB, and AD seem to be the most significant.
We chose these factors because A is aliased with a three factor interaction, which we assume to be more
negligible. Furthermore, AB and AD are considered to be significant over their aliases since they include
factor A, which is believed to be the most significant.
Analysis of Variance
Source DF Seq SS Contribution Adj SS Adj MS F-Value P-Value
Model 7 448.000 100.00% 448.000 64.000 * *
Linear 4 324.000 72.32% 324.000 81.000 * *
A 1 288.000 64.29% 288.000 288.000 * *
B 1 2.000 0.45% 2.000 2.000 * *
C 1 32.000 7.14% 32.000 32.000 * *
D 1 2.000 0.45% 2.000 2.000 * *
2-Way Interactions 3 124.000 27.68% 124.000 41.333 * *
A*B 1 72.000 16.07% 72.000 72.000 * *
A*C 1 2.000 0.45% 2.000 2.000 * *
A*D 1 50.000 11.16% 50.000 50.000 * *
Error 0 * * * *
Total 7 448.000 100.00%
Therefore, we will rerun the analysis focusing on these factors and including factors B and D for heirarchy.

Analysis of Variance

Source DF Seq SS Contribution Adj SS Adj MS F-Value P-Value

Model 5 414.000 92.41% 414.000 82.800 4.87 0.179
Linear 3 292.000 65.18% 292.000 97.333 5.73 0.152
A 1 288.000 64.29% 288.000 288.000 16.94 0.054
B 1 2.000 0.45% 2.000 2.000 0.12 0.764
D 1 2.000 0.45% 2.000 2.000 0.12 0.764
2-Way Interactions 2 122.000 27.23% 122.000 61.000 3.59 0.218
A*B 1 72.000 16.07% 72.000 72.000 4.24 0.176
A*D 1 50.000 11.16% 50.000 50.000 2.94 0.228
Error 2 34.000 7.59% 34.000 17.000
Total 7 448.000 100.00%

Model Summary
S R-sq R-sq(adj) PRESS R-sq(pred)
4.12311 92.41% 73.44% 544 0.00%

Coded Coefficients
Term Effect Coef SE Coef 95% CI T-Value P-Value VIF
Constant 85.00 1.46 ( 78.73, 91.27) 58.31 0.000
A -12.00 -6.00 1.46 (-12.27, 0.27) -4.12 0.054 1.00
B -1.00 -0.50 1.46 ( -6.77, 5.77) -0.34 0.764 1.00
D -1.00 -0.50 1.46 ( -6.77, 5.77) -0.34 0.764 1.00
A*B 6.00 3.00 1.46 ( -3.27, 9.27) 2.06 0.176 1.00
A*D -5.00 -2.50 1.46 ( -8.77, 3.77) -1.71 0.228 1.00

Regression Equation in Uncoded Units

Yield = 85.00 - 6.00 A - 0.50 B - 0.50 D + 3.00 A*B - 2.50 A*D

We see from our new analysis of variance that our revised model has anR2 of 92.41%, meaning that we have
accounted for a fair amount of variance in the model, but there is still room for improvement. This is further
evidenced by the fact that none of our factors are significant at the 5% significance level. We see this in the
normal probability plot below.

Therefore, we shall run the model once more, this time considering only factor A.

2
Analysis of Variance
Source DF Seq SS Contribution Adj SS Adj MS F-Value P-Value
Model 1 288.0 64.29% 288.0 288.00 10.80 0.017
Linear 1 288.0 64.29% 288.0 288.00 10.80 0.017
A 1 288.0 64.29% 288.0 288.00 10.80 0.017
Error 6 160.0 35.71% 160.0 26.67
Total 7 448.0 100.00%

Model Summary
S R-sq R-sq(adj) PRESS R-sq(pred)
5.16398 64.29% 58.33% 284.444 36.51%

Coded Coefficients
Term Effect Coef SE Coef 95% CI T-Value P-Value VIF
Constant 85.00 1.83 ( 80.53, 89.47) 46.56 0.000
A -12.00 -6.00 1.83 (-10.47, -1.53) -3.29 0.017 1.00

Regression Equation in Uncoded Units

Yield = 85.00 - 6.00 A
This indicates that factor A is significant in determining the yield output. This is supported by the normal
probability plot below.

We see that while A has become significant in this model at the 5% level, our R2 has decreased dramatically.
Therefore, we would recommend that the confirming experiment of the non-primal fraction be run so that
we may further improve our understanding.
An analysis of the 4 − in − 1 residual plots does not reveal any reasons to question our normality and equal
variance assumptions.

3
8.3 Consider the plasma etch experiment described in Example 6.1. Suppose that only a one-half fraction
of the design could be run. Set up the design and analyze the data.
We will use the followin 23−1 ,2-replicate design. Note tha D = ABC.
Defining Relation: I = ABC

Aliases

I + ABCD
A + BCD
B + ACD
C + ABD
D + ABC
AB + CD
AC + BD
AD + BC

Design Table
Run A B C D
1 - - - -
2 + + - -
3 + - + -
4 - + + -
5 + - - +
6 - + - +
7 - - + +
We see from the initial normal probability plot of effects included below that A, D, and AD are the significant
factors.

Analysis of Variance
Source DF Seq SS Contribution Adj SS Adj MS F-Value P-Value
Model 3 279051 99.36% 279051 93017 207.05 0.000
Linear 2 201039 71.58% 201039 100519 223.75 0.000
A 1 32258 11.49% 32258 32258 71.80 0.001
D 1 168781 60.10% 168781 168781 375.69 0.000
2-Way Interactions 1 78012 27.78% 78012 78012 173.65 0.000
A*D 1 78012 27.78% 78012 78012 173.65 0.000
Error 4 1797 0.64% 1797 449
Total 7 280848 100.00%
Model Summary

4
 S R-sq R-sq(adj) PRESS R-sq(pred)
21.1955 99.36% 98.88% 7188 97.44%

We see in the above analysis of variance that factors A, D, and AD are significant. Now we must analyze
the residual plots.

The versus fits give us some cause for concern as the left-most points hint at a possible inequality of variance.
The normal probablity plot of effects is satisfactory and doesn’t give us any reason to question our normality
assumptions.

8.6 R.D. (”Experimenting with a Large Number of Variables” in Experiments in Industry: Design, Analysis
and Interpretation of Results, by R.D. Snee, L.B. Hare, and J.B. Trout, Editors ASQC, 1985) describes an
experiment in which a 25−1 design with I = ABCDE was used to investigate the effects of five factors on the
color of a chemical product The factors are A = solvent/reactant, B = catalyst/reactant, C = temperature,
D = reactant purity, and E = reactant pH. The results obtained were as follows.
e = −0.63 d = 6.79
a = 2.51 ade = 5.47
b = −2.68 bde = 3.45
abe = 1.66 abd = 5.68
c = 2.06 cde = 5.22
ace = 1.22 acd = 4.38
bce = −2.09 bcd = 4.30
abc = 1.93 abcde = 4.05

a) Prepare a normal probability plot of the effects. Which effects seem active?
We see from our intial analysis that factors A, B, D, AB, and AD appear to be active (significant),
however factor A is the only factor to register as such in this model.

5
 If we re-analyze the model focusing on only these factors, the following normal probablity plot of effects
is generated.

We see now that factors A, B, D, AB, and AD all register as significant in the reduced model. This is
supported by the following analysis of variance performed using the reduced model.

Analysis of Variance

Source DF Seq SS Contribution Adj SS Adj MS F-Value P-Value

Model 5 106.039 92.46% 106.039 21.2078 24.53 0.000
Linear 3 92.192 80.39% 92.192 30.7308 35.54 0.000
A 1 6.864 5.99% 6.864 6.8644 7.94 0.018
B 1 7.182 6.26% 7.182 7.1824 8.31 0.016
D 1 78.146 68.14% 78.146 78.1456 90.37 0.000
2-Way Interactions 2 13.847 12.07% 13.847 6.9233 8.01 0.008
A*B 1 6.503 5.67% 6.503 6.5025 7.52 0.021
A*D 1 7.344 6.40% 7.344 7.3441 8.49 0.015
Error 10 8.647 7.54% 8.647 0.8647
Total 15 114.686 100.00%

b) Calculate the residuals. Construct a normal probability plot of the residuals and plot the residuals
versus the fitted values. Comment on the plots. The following residuals and fits were calculated using
Minitab.

6
 Treatment Color FITS RESI
Combination
E -0.63 0.4725 -1.1025
A 2.51 1.8625 0.6475
B -2.68 -2.1425 -0.5375
ABE 1.66 1.7975 -0.1375
C 2.06 0.4725 1.5875
ACE 1.22 1.8625 -0.6425
BCE -2.09 -2.1425 0.0525
ABC 1.93 1.7975 0.1325
D 6.79 6.2475 0.5425
ADE 5.47 4.9275 0.5425
BDE 3.45 3.6325 -0.1825
ABD 5.68 4.8625 0.8175
CDE 5.22 6.2475 -1.0275
ACD 4.38 4.9275 -0.5475
BCD 4.30 3.6325 0.6675
ABCD 4.05 4.8625 -0.8125

These residuals were used to construct the following graphs.

We see that the residual plots are satisfactory and do not give us any reason to question our normality
and equal variance assumptions.
c) If any factors are negligible, collapse the 25−1 design into a full factorial in the active factors. Comment
on the resulting design, and interpret the results.
Since we found factors A,B,D, AB, and AD significant, we can collapse the 25−1 into a full factorial 23 .
The results of this collapse are shown below. It is possible to relabel D as C, however we have chosen
to keep the original labeling format for consistentcy.

7
 Analysis of Variance

Source DF Seq SS Contribution Adj SS Adj MS F-Value P-Value

Model 7 106.510 92.87% 106.510 15.2156 14.89 0.001
Linear 3 92.192 80.39% 92.192 30.7308 30.07 0.000
A 1 6.864 5.99% 6.864 6.8644 6.72 0.032
B 1 7.182 6.26% 7.182 7.1824 7.03 0.029
D 1 78.146 68.14% 78.146 78.1456 76.46 0.000
2-Way Interactions 3 14.087 12.28% 14.087 4.6956 4.59 0.038
A*B 1 6.503 5.67% 6.503 6.5025 6.36 0.036
A*D 1 7.344 6.40% 7.344 7.3441 7.19 0.028
B*D 1 0.240 0.21% 0.240 0.2401 0.23 0.641
3-Way Interactions 1 0.230 0.20% 0.230 0.2304 0.23 0.648
A*B*D 1 0.230 0.20% 0.230 0.2304 0.23 0.648
Error 8 8.177 7.13% 8.177 1.0221
Total 15 114.686 100.00%

Model Summary
S R-sq R-sq(adj) PRESS R-sq(pred)
1.01099 92.87% 86.63% 32.7072 71.48%

Coded Coefficients
Term Effect Coef SE Coef 95% CI T-Value P-Value VIF
Constant 2.707 0.253 ( 2.125, 3.290) 10.71 0.000
A 1.310 0.655 0.253 ( 0.072, 1.238) 2.59 0.032 1.00
B -1.340 -0.670 0.253 (-1.253, -0.087) -2.65 0.029 1.00
D 4.420 2.210 0.253 ( 1.627, 2.793) 8.74 0.000 1.00
A*B 1.275 0.638 0.253 ( 0.055, 1.220) 2.52 0.036 1.00
A*D -1.355 -0.677 0.253 (-1.260, -0.095) -2.68 0.028 1.00
B*D 0.245 0.123 0.253 (-0.460, 0.705) 0.48 0.641 1.00
A*B*D -0.240 -0.120 0.253 (-0.703, 0.463) -0.47 0.648 1.00

Regression Equation in Uncoded Units

Color = 2.707 + 0.655 A - 0.670 B + 2.210 D + 0.638 A*B - 0.677 A*D + 0.123 B*D - 0.120 A*B*D

We see that the same factors are active in the collapsed model as were in the original model, though the
SSE has decreased slightly to accomodate the addition of B, BD, andABD.
Furthermore, we note that in the residual plots included below there is no significant cause for concern.
However, the normal probability plot is potentially exhibiting a pattern which could indicate curvature that
was unaccounted for. This should be looked into further to determine if the model cannot be made to be
more accurate.

8
8.7 An article written by J.J. Pignatiello Jr. and J.S. Ramberg in the Journal of Quality Technology (Vol.
17, 1985, pp. 198-206) describes the use of a replicated fractional factorial to investigate the effect of five
factors on the free height of leaf springs used in an automotive application. The factors are A = furnace
temperature, B = heating time, C = transfer time, D = hold down time, and E = quench oil temperature.
The data are shown in the table below.
A B C D E Free Height
- - - - - 7.78 7.78 7.81
+ - - + - 8.15 8.18 7.88
- + - + - 7.50 7.56 7.50
+ + - - - 7.59 7.56 7.75
- - + + - 7.54 8.00 7.88
+ - + - - 7.69 8.09 8.06
- + + - - 7.56 7.52 7.44
+ + + + - 7.56 7.81 7.69
- - - - + 7.50 7.25 7.12
+ - - + + 7.88 7.88 7.44
- + - + + 7.50 7.56 7.50
+ + - - + 7.63 7.75 7.56
- - + + + 7.32 7.44 7.44
+ - + - + 7.56 7.69 7.62
- + + - + 7.18 7.18 7.25
+ + + + + 7.81 7.50 7.59

a) Write out the alias structure for this design. What is the resolution of this design?
The defining relation of this design is I = ABCD which means that is Resolution IV. The alias

9
 structure is as follows.

I+ABCD
A+BCD
B+ACD
C+ABD
D+ABC
E+ABCDE
AB+CD
AC+BD
AD+BC
AE+BCDE
BD+ACDE
CE+ABDE
DE+ABCE

b) Analyze the data. What factors influence the mean free height?
We see from the below intial analysis of variance with two-factor effects that the factors A, B, D, E,
and BE are significant influences on the mean free height.

Analysis of Variance

Source DF Seq SS Contribution Adj SS Adj MS F-Value P-Value

Model 12 2.21646 76.67% 2.21646 0.184705 9.59 0.000
Linear 5 1.83846 63.60% 1.83846 0.367692 19.08 0.000
A 1 0.70325 24.33% 0.70325 0.703252 36.50 0.000
B 1 0.32177 11.13% 0.32177 0.321769 16.70 0.000
C 1 0.02950 1.02% 0.02950 0.029502 1.53 0.224
D 1 0.09992 3.46% 0.09992 0.099919 5.19 0.029
E 1 0.68402 23.66% 0.68402 0.684019 35.50 0.000
2-Way Interactions 7 0.37800 13.08% 0.37800 0.054000 2.80 0.020
A*B 1 0.01050 0.36% 0.01050 0.010502 0.55 0.465
A*C 1 0.00002 0.00% 0.00002 0.000019 0.00 0.975
A*D 1 0.00630 0.22% 0.00630 0.006302 0.33 0.571
A*E 1 0.04877 1.69% 0.04877 0.048769 2.53 0.121
B*E 1 0.28060 9.71% 0.28060 0.280602 14.56 0.001
C*E 1 0.01300 0.45% 0.01300 0.013002 0.67 0.417
D*E 1 0.01880 0.65% 0.01880 0.018802 0.98 0.330
Error 35 0.67432 23.33% 0.67432 0.019266
Lack-of-Fit 3 0.04726 1.63% 0.04726 0.015752 0.80 0.501
Pure Error 32 0.62707 21.69% 0.62707 0.019596
Total 47 2.89078 100.00%

Therefore, we shall reanalyze the data to focus on significant factors and reduce noise.

Analysis of Variance
Source DF Seq SS Contribution Adj SS Adj MS F-Value P-Value

10
 Model 5 2.08956 72.28% 2.08956 0.41791 21.91 0.000
Linear 4 1.80896 62.58% 1.80896 0.45224 23.71 0.000
A 1 0.70325 24.33% 0.70325 0.70325 36.86 0.000
B 1 0.32177 11.13% 0.32177 0.32177 16.87 0.000
D 1 0.09992 3.46% 0.09992 0.09992 5.24 0.027
E 1 0.68402 23.66% 0.68402 0.68402 35.86 0.000
2-Way Interactions 1 0.28060 9.71% 0.28060 0.28060 14.71 0.000
B*E 1 0.28060 9.71% 0.28060 0.28060 14.71 0.000
Error 42 0.80122 27.72% 0.80122 0.01908
Lack-of-Fit 10 0.17415 6.02% 0.17415 0.01742 0.89 0.554
Pure Error 32 0.62707 21.69% 0.62707 0.01960
Total 47 2.89078 100.00%

Model Summary
S R-sq R-sq(adj) PRESS R-sq(pred)
0.138118 72.28% 68.98% 1.04649 63.80%

It is clear that the factors A, B, D, E, and BE are all significant at the 5% level. Our normal probability
plot of the data supports this conclusion.

Now, we must analyze the residuals for signs of abnormality.

There is nothing particularly concerning regarding the residual plots and we have no reason to question
our assumptions of normality and equal variances.
c) Calculate the range and standard deviation of the free height for each run. Is there any indication that
any of these factors affects variability in the free height.

11
 Term Effect Coef SE Coef 95% CI T-Value P-Value VIF
Constant 7.6256 0.0200 ( 7.5850, 7.6663) 380.62 0.000
A 0.2421 0.1210 0.0200 ( 0.0804, 0.1617) 6.04 0.000 1.00
B -0.1638 -0.0819 0.0200 (-0.1225, -0.0412) -4.09 0.000 1.00
C -0.0496 -0.0248 0.0200 (-0.0655, 0.0159) -1.24 0.224 1.00
D 0.0913 0.0456 0.0200 ( 0.0050, 0.0863) 2.28 0.029 1.00
E -0.2387 -0.1194 0.0200 (-0.1600, -0.0787) -5.96 0.000 1.00
A*B -0.0296 -0.0148 0.0200 (-0.0555, 0.0259) -0.74 0.465 1.00
A*C 0.0013 0.0006 0.0200 (-0.0400, 0.0413) 0.03 0.975 1.00
A*D -0.0229 -0.0115 0.0200 (-0.0521, 0.0292) -0.57 0.571 1.00
A*E 0.0637 0.0319 0.0200 (-0.0088, 0.0725) 1.59 0.121 1.00
B*E 0.1529 0.0765 0.0200 ( 0.0358, 0.1171) 3.82 0.001 1.00
C*E -0.0329 -0.0165 0.0200 (-0.0571, 0.0242) -0.82 0.417 1.00
D*E 0.0396 0.0198 0.0200 (-0.0209, 0.0605) 0.99 0.330 1.00

d) Analyze the residuals from this experiment, and comment on your findings. The following residual
plots do not show anything of particular concern. They appear to be satisifactory.

e) Is this the best possible design for five factors in 16 runs? Specifically, can you find a fractional design
for five factors in 16 runs with a higher resolution than this one?
No. This is not the best possible design. A design with 5 factors (i.e. 25 factorial) can have a maximum
resolution of V. This can be acheived by setting the generator to I = ABCDE, a five-letter word.

8.9 Consider the leaf spring experiment in Problem 8.7. Suppose that factor E (quench oil temperature)
is very difficult to control during manufacturing. Where would you set factors A, B, C, and D to reduce
variability in the free height as much as possible regardless of the quench oil temperature?

12
Wee see from the above interaction of effects graph that we should run the process with A high, B and C
low, and D at whichever level is more economical. This will reduce the variability in the free height as much
as possible regardless of where E is set.

13