CH 04

Solutions from Montgomery, D. C.
(2008) Design and Analysis of Experiments, Wiley, NY
Chapter 4
Randomized Blocks, Latin Squares, and Related Designs
Solutions
4.1. The ANOVA from a randomized complete block experiment output is shown below.
Source DF SS MS F P
Treatment 4 1010.56 ? 29.84 ?
Block ? ? 64.765 ? ?
Error 20 169.33 ?
Total 29 1503.71
(a) Fill in the blanks. You may give bounds on the P-value.
Completed table is:
Source DF SS MS F P
Treatment 4 1010.56 252.640 29.84 < 0.00001
Block 5 323.82 64.765
Error 20 169.33 8.467
Total 29 1503.71
(b) How many blocks were used in this experiment?
Six blocks were used.
(c) What conclusions can you draw?
The treatment effect is significant; the means of the five treatments are not all equal.
4.2. Consider the single-factor completely randomized experiment shown in Problem 3.4. Suppose that
this experiment had been conducted in a randomized complete block design, and that the sum of squares for
blocks was 80.00. Modify the ANOVA for this experiment to show the correct analysis for the randomized
complete block experiment.
The modified ANOVA is shown below:
Source DF SS MS F P
Treatment 4 987.71 246.93 46.3583 < 0.00001
Block 5 80.00 16.00
Error 20 106.53 5.33
Total 29 1174.24
4-1
Solutions from Montgomery, D. C. (2008) Design and Analysis of Experiments, Wiley, NY
4.3. A chemist wishes to test the effect of four chemical agents on the strength of a particular type of
cloth. Because there might be variability from one bolt to another, the chemist decides to use a randomized
block design, with the bolts of cloth considered as blocks. She selects five bolts and applies all four
chemicals in random order to each bolt. The resulting tensile strengths follow. Analyze the data from this
experiment (use α = 0.05) and draw appropriate conclusions.
Bolt
Chemical 1 2 3 4 5
1 73 68 74 71 67
2 73 67 75 72 70
3 75 68 78 73 68
4 73 71 75 75 69
Design Expert Output

Response: Strength
ANOVA for Selected Factorial Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob > F
Block 157.00 4 39.25
Model 12.95 3 4.32 2.38 0.1211 not significant
A 12.95 3 4.32 2.38 0.1211
Residual 21.80 12 1.82
Cor Total 191.75 19
The "Model F-value" of 2.38 implies the model is not significant relative to the noise. There is a
12.11 % chance that a "Model F-value" this large could occur due to noise.
Std. Dev. 1.35 R-Squared 0.3727

Mean 71.75 Adj R-Squared 0.2158
C.V. 1.88 Pred R-Squared -0.7426
PRESS 60.56 Adeq Precision 10.558
Treatment Means (Adjusted, If Necessary)

Estimated Standard
Mean Error
1-1 70.60 0.60
2-2 71.40 0.60
3-3 72.40 0.60
4-4 72.60 0.60
Mean Standard t for H0

Treatment Difference DF Error Coeff=0 Prob > |t|
1 vs 2 -0.80 1 0.85 -0.94 0.3665
1 vs 3 -1.80 1 0.85 -2.11 0.0564
1 vs 4 -2.00 1 0.85 -2.35 0.0370
2 vs 3 -1.00 1 0.85 -1.17 0.2635
2 vs 4 -1.20 1 0.85 -1.41 0.1846
3 vs 4 -0.20 1 0.85 -0.23 0.8185
There is no difference among the chemical types at α = 0.05 level.
4-2
4.4. Three different washing solutions are being compared to study their effectiveness in retarding
bacteria growth in five-gallon milk containers. The analysis is done in a laboratory, and only three trials
can be run on any day. Because days could represent a potential source of variability, the experimenter
decides to use a randomized block design. Observations are taken for four days, and the data are shown
here. Analyze the data from this experiment (use α = 0.05) and draw conclusions.
Days
Solution 1 2 3 4
1 13 22 18 39
2 16 24 17 44
3 5 4 1 22

Response: Growth
Sum of Mean F
Block 1106.92 3 368.97
Model 703.50 2 351.75 40.72 0.0003 significant
A 703.50 2 351.75 40.72 0.0003
Residual 51.83 6 8.64
Cor Total 1862.25 11
The Model F-value of 40.72 implies the model is significant. There is only
a 0.03% chance that a "Model F-Value" this large could occur due to noise.

C.V. 15.68 Pred R-Squared 0.7255

Estimated Standard
Mean Error
1-1 23.00 1.47
2-2 25.25 1.47
3-3 8.00 1.47

1 vs 2 -2.25 1 2.08 -1.08 0.3206
1 vs 3 15.00 1 2.08 7.22 0.0004
2 vs 3 17.25 1 2.08 8.30 0.0002
There is a difference between the means of the three solutions. The Fisher LSD procedure indicates that
solution 3 is significantly different than the other two.
4-3
4.5. Plot the mean tensile strengths observed for each chemical type in Problem 4.3 and compare them to
a scaled t distribution. What conclusions would you draw from the display?
Scaled t Distribution
(1) (2) (3,4)
70.0 71.0 72.0 73.0

Mean Strength
MS E 1.82
S yi . = = = 0.603
b 5
There is no obvious difference between the means. This is the same conclusion given by the analysis of
variance.
4.6. Plot the average bacteria counts for each solution in Problem 4.4 and compare them to an
appropriately scaled t distribution. What conclusions can you draw?
Scaled t Distribution
(3) (1) (2)
5 10 15 20 25
Bacteria Growth
MS E 8.64
S yi . = = = 1.47
b 4
4-4
There is no difference in mean bacteria growth between solutions 1 and 2. However, solution 3 produces
significantly lower mean bacteria growth. This is the same conclusion reached from the Fisher LSD
procedure in Problem 4.4.
4.7. Consider the hardness testing experiment described in Section 4.1. Suppose that the experiment was
conducted as described and the following Rockwell C-scale data (coded by subtracting 40 units) obtained:
Coupon
Tip 1 2 3 4
1 9.3 9.4 9.6 10.0
2 9.4 9.3 9.8 9.9
3 9.2 9.4 9.5 9.7
4 9.7 9.6 10.0 10.2
(a) Analyize the data from this experiment.
There is a difference between the means of the four tips.

Response: Hardness
Analysis of variance table [Terms added sequentially (first to last)]
Sum of Mean F
Bock 0.82 3 0.27
A 0.38 3 0.13 14.44 0.0009
Residual 0.080 9 8.889E-003
Cor Total 1.29 15


Estimated Standard
Mean Error
1-1 9.57 0.047
2-2 9.60 0.047
3-3 9.45 0.047
4-4 9.88 0.047

1 vs 2 -0.025 1 0.067 -0.38 0.7163
1 vs 3 0.13 1 0.067 1.87 0.0935
1 vs 4 -0.30 1 0.067 -4.50 0.0015
2 vs 3 0.15 1 0.067 2.25 0.0510
2 vs 4 -0.27 1 0.067 -4.12 0.0026
3 vs 4 -0.43 1 0.067 -6.37 0.0001
(b) Use the Fisher LSD method to make comparisons among the four tips to determine specifically which
tips differ in mean hardness readings.
Based on the LSD bars in the Design Expert plot below, the mean of tip 4 differs from the means of tips 1,
2, and 3. The LSD method identifies a marginal difference between the means of tips 2 and 3.
4-5
One Factor Plot

10.2
9.95
Hardness
9.7
9.45
9.2
1 2 3 4
A: T ip
(c) Analyze the residuals from this experiment.
The residual plots below do not identify any violations to the assumptions.
Normal Plot of Residuals Residuals vs. Predicted

0.15
99
95
0.0875
90
Normal % Probability
80
Res iduals
70
50 0.025 2
30
20
10
-0.0375
5
1 2
-0.1
-0.1 -0.0375 0.025 0.0875 0.15 9.22 9.47 9.71 9.96 10.20
Residual Predicted
4-6
Residuals vs. Tip

0.15
0.0875
Res iduals
0.025
-0.0375
-0.1
1 2 3 4
Tip
4.8. A consumer products company relies on direct mail marketing pieces as a major component of its
advertising campaigns. The company has three different designs for a new brochure and want to evaluate
their effectiveness, as there are substantial differences in costs between the three designs. The company
decides to test the three designs by mailing 5,000 samples of each to potential customers in four different
regions of the country. Since there are known regional differences in the customer base, regions are
considered as blocks. The number of responses to each mailing is shown below.
Region
Design NE NW SE SW
1 250 350 219 375
2 400 525 390 580
3 275 340 200 310
(a) Analyze the data from this experiment.
The residuals of the analsysis below identify concerns with the normality and equality of variance
assumptions. As a result, a square root transformation was applied as shown in the second ANOVA table.
The residuals of both analysis are presented for comparison in part (c) of this problem. The analysis
concludes that there is a difference between the mean number of responses for the three designs.

Response: Number of responses
Sum of Mean F
Block 49035.67 3 16345.22
A 90755.17 2 45377.58 50.15 0.0002
Residual 5428.83 6 904.81
Cor Total 1.452E+005 11
4-7


Estimated Standard
Mean Error
1-1 298.50 15.04
2-2 473.75 15.04
3-3 281.25 15.04

1 vs 2 -175.25 1 21.27 -8.24 0.0002
1 vs 3 17.25 1 21.27 0.81 0.4483
2 vs 3 192.50 1 21.27 9.05 0.0001
Design Expert Output for Transformed Data

Response: Number of responses Transform: Square root Constant: 0
Sum of Mean F
Block 35.89 3 11.96
A 60.73 2 30.37 60.47 0.0001
Residual 3.01 6 0.50
Cor Total 99.64 11


Estimated Standard
Mean Error
1-1 17.17 0.35
2-2 21.69 0.35
3-3 16.69 0.35

1 vs 2 -4.52 1 0.50 -9.01 0.0001
1 vs 3 0.48 1 0.50 0.95 0.3769
2 vs 3 4.99 1 0.50 9.96 < 0.0001
(b) Use the Fisher LSD method to make comparisons among the three designs to determine specifically
which designs differ in mean response rate.
Based on the LSD bars in the Design Expert plot below, designs 1 and 3 do not differ; however, design 2 is
different than designs 1 and 3.
4-8
One Factor Plot

24.083
Sqrt(Number of responses)
21.598
19.113
16.627
14.142
1 2 3
A: Design
The first set of residual plots presented below represent the untransformed data. Concerns with normality
as well as inequality of variance are presented. The second set of residual plots represent transformed data
and do not identify significant violations of the assumptions. The residuals vs. design plot indicates a slight
inequality of variance; however, not a strong violation and an improvement over the non-transformed data.

36.5833
99
95
17
90
80
70
Residuals
50 -2.58333
30
20
10
-22.1667
5
-41.75
-41.75 -22.1667 -2.58333 17 36.5833 199.75 285.88 372.00 458.13 544.25
Residual Predicted
4-9
Residuals vs. Design

36.5833
17
Residuals
-2.58333
-22.1667
-41.75
1 2 3
Design
The following are the square root transformed data residual plots.

0.942069
99
95
0.476292
90
80
Res iduals
70
50 0.0105142
30
20
10
-0.455263
5
-0.921041
-0.921041 -0.455263 0.0105142 0.476292 0.942069 14.41 16.68 18.96 21.24 23.52
Residual Predicted
4-10
Residuals vs. Design

0.942069
0.476292
Res iduals
0.0105142
-0.455263
-0.921041
1 2 3
Des ign
4.9. The effect of three different lubricating oils on fuel economy in diesel truck engines is being studied.
Fuel economy is measured using brake-specific fuel consumption after the engine has been running for 15
minutes. Five different truck engines are available for the study, and the experimenters conduct the
following randomized complete block design.
Truck
Oil 1 2 3 4 5
1 0.500 0.634 0.487 0.329 0.512
2 0.535 0.675 0.520 0.435 0.540
3 0.513 0.595 0.488 0.400 0.510
(a) Analyize the data from this experiment.
From the analysis below, there is a significant difference between lubricating oils with regards to fuel
economy.

Response: Fuel consumption
Sum of Mean F
Block 0.092 4 0.023
Model 6.706E-003 2 3.353E-003 6.35 0.0223 significant
A 6.706E-003 2 3.353E-003 6.35 0.0223
Residual 4.222E-003 8 5.278E-004
Cor Total 0.10 14

4-11

Estimated Standard
Mean Error
1-1 0.49 0.010
2-2 0.54 0.010
3-3 0.50 0.010

1 vs 2 -0.049 1 0.015 -3.34 0.0102
1 vs 3 -8.800E-003 1 0.015 -0.61 0.5615
2 vs 3 0.040 1 0.015 2.74 0.0255
(b) Use the Fisher LSD method to make comparisons among the three lubricating oils to determine
specifically which oils differ in break-specific fuel consumption.
Based on the LSD bars in the Design Expert plot below, the means for break-specific fuel consumption for
oils 1 and 3 do not differ; however, oil 2 is different than oils 1 and 3.
One Factor Plot

0.675
0.5885
Fuel consumption
0.502
0.4155
0.329
1 2 3
A: Oil
The residual plots below do not identify any violations to the assumptions.
4-12

0.0223333
99
95
0.00678333
90
80
Res iduals
70
50 -0.00876667
30
20
10
-0.0243167
5
-0.0398667
-0.0398667 -0.0243167 -0.00876667 0.00678333 0.0223333 0.37 0.44 0.52 0.59 0.66
Residual Predicted
Residuals vs. Oil

0.0223333
0.00678333
Res iduals
-0.00876667
-0.0243167
-0.0398667
1 2 3
Oil
4.10. An article in the Fire Safety Journal (“The Effect of Nozzle Design on the Stability and Performance
of Turbulent Water Jets,” Vol. 4, August 1981) describes an experiment in which a shape factor was
determined for several different nozzle designs at six levels of jet efflux velocity. Interest focused on
potential differences between nozzle designs, with velocity considered as a nuisance variable. The data are
shown below:
Jet Efflux Velocity (m/s)

Nozzle
Design 11.73 14.37 16.59 20.43 23.46 28.74
1 0.78 0.80 0.81 0.75 0.77 0.78
2 0.85 0.85 0.92 0.86 0.81 0.83
3 0.93 0.92 0.95 0.89 0.89 0.83
4 1.14 0.97 0.98 0.88 0.86 0.83
5 0.97 0.86 0.78 0.76 0.76 0.75
4-13
(a) Does nozzle design affect the shape factor? Compare nozzles with a scatter plot and with an analysis
of variance, using α = 0.05.

Response: Shape
Sum of Mean F
Block 0.063 5 0.013
A 0.10 4 0.026 8.92 0.0003
Residual 0.057 20 2.865E-003
Cor Total 0.22 29


Estimated Standard
Mean Error
1-1 0.78 0.022
2-2 0.85 0.022
3-3 0.90 0.022
4-4 0.94 0.022
5-5 0.81 0.022

1 vs 2 -0.072 1 0.031 -2.32 0.0311
1 vs 3 -0.12 1 0.031 -3.88 0.0009
1 vs 4 -0.16 1 0.031 -5.23 < 0.0001
1 vs 5 -0.032 1 0.031 -1.02 0.3177
2 vs 3 -0.048 1 0.031 -1.56 0.1335
2 vs 4 -0.090 1 0.031 -2.91 0.0086
2 vs 5 0.040 1 0.031 1.29 0.2103
3 vs 4 -0.042 1 0.031 -1.35 0.1926
3 vs 5 0.088 1 0.031 2.86 0.0097
4 vs 5 0.13 1 0.031 4.21 0.0004
Nozzle design has a significant effect on shape factor.
4-14
One Factor Plot

1.14
1.04236
Shape
0.944718
0.847076 2
2
2
0.749435
1 2 3 4 5
Nozzle Design
(b) Analyze the residual from this experiment.
The plots shown below do not give any indication of serious problems. Thre is some indication of a mild
outlier on the normal probability plot and on the plot of residuals versus the predicted velocity.
Normal plot of residuals Residuals vs. Predicted

0.121333
99
95
0.0713333
Norm al % probability
90
80
Res iduals
70
50 0.0213333
30
20
10
-0.0286667
5
-0.0786667
-0.0786667 -0.0286667 0.0213333 0.0713333 0.121333 0.73 0.80 0.87 0.95 1.02
Res idual Predicted
4-15
Residuals vs. Nozzle Design

0.121333
0.0713333
Residuals
2
0.0213333
-0.0286667
-0.0786667
1 2 3 4 5
Nozzle Design
(c) Which nozzle designs are different with respect to shape factor? Draw a graph of average shape factor
for each nozzle type and compare this to a scaled t distribution. Compare the conclusions that you
draw from this plot to those from Duncan’s multiple range test.
MS E 0.002865
S yi . = = = 0.021852
b 6
R2= r 0.05 (2,20) S y =

i.
(2.95)(0.021852)= 0.06446
R3= r 0.05 (3,20) S y =
i.
(3.10)(0.021852)= 0.06774
R4= r 0.05 (4,20) S y =
i.
(3.18)(0.021852)= 0.06949
R5= r 0.05 (5,20) S y =
i.
(3.25)(0.021852)= 0.07102
Mean Difference R
1 vs 4 0.16167 > 0.07102 different
1 vs 3 0.12000 > 0.06949 different
1 vs 2 0.07167 > 0.06774 different
1 vs 5 0.03167 < 0.06446
5 vs 4 0.13000 > 0.06949 different
5 vs 3 0.08833 > 0.06774 different
5 vs 2 0.04000 < 0.06446
2 vs 4 0.09000 > 0.06774 different
2 vs 3 0.04833 < 0.06446
3 vs 4 0.04167 < 0.06446
4-16
Scaled t
(1) (5) (2 (3 (4
0.7 0.8 0.8 0.9 0.9

Shape
4.11. An article in Communications of the ACM (Vol. 30, No. 5, 1987) studied different algorithms for
estimating software development costs. Six algorithms were applied to several different software
development projects and the percent error in estimating the development cost was observed. Some of the
data from this experiment is show in the table below.
Project
Algorithm 1 2 3 4 5 6
1 (SLIM) 1244 21 82 2221 905 839
2 (COCOMO-A) 281 129 396 1306 336 910
3 (COCOMO-R) 220 84 458 543 300 794
4 (COCOMO-C) 225 83 425 552 291 826
5 (FUNCTION POINTS) 19 11 -34 121 15 103
6 (ESTIMALS) -20 35 -53 170 104 199
(a) Do the algorithms differ in their mean cost estimation accuracy?
The ANOVA below identifies the algorithms are significantly different in their mean cost estimation
error.
Response Cost Error

ANOVA for selected factorial model
Analysis of variance table [Classical sum of squares - Type II]
Sum of Mean F p-value
Source Squares df Square Value Prob > F
Block 2.287E+006 5 4.575E+005
Model 2.989E+006 5 5.978E+005 5.38 0.0017 significant
A-Algorithm 2.989E+006 5 5.978E+005 5.38 0.0017
Residual 2.780E+006 25 1.112E+005
Cor Total 8.056E+006 35

C.V. % 84.89 Pred R-Squared 0.0009
PRESS 5.764E+006 Adeq Precision 8.705
4-17
Estimated Standard
Mean Error
1-SLM 885.33 136.13
2-COCOMO-A 559.67 136.13
3-COCOMO-R 399.83 136.13
4-COCOMO-C 400.33 136.13
5-FUNCTION POINTS 39.17 136.13
6-ESTIMALS 72.50 136.13

Treatment Difference df Error Coeff=0 Prob > |t|
1 vs 2 325.67 1 192.51 1.69 0.1031
1 vs 3 485.50 1 192.51 2.52 0.0184
1 vs 4 485.00 1 192.51 2.52 0.0185
1 vs 5 846.17 1 192.51 4.40 0.0002
1 vs 6 812.83 1 192.51 4.22 0.0003
2 vs 3 159.83 1 192.51 0.83 0.4143
2 vs 4 159.33 1 192.51 0.83 0.4157
2 vs 5 520.50 1 192.51 2.70 0.0122
2 vs 6 487.17 1 192.51 2.53 0.0181
3 vs 4 -0.50 1 192.51 -2.597E-003 0.9979
3 vs 5 360.67 1 192.51 1.87 0.0727
3 vs 6 327.33 1 192.51 1.70 0.1015
4 vs 5 361.17 1 192.51 1.88 0.0724
4 vs 6 327.83 1 192.51 1.70 0.1010
5 vs 6 -33.33 1 192.51 -0.17 0.8639
(b) Analyze the residuals from this experiment.
The residual plots below identify a single outlier that should be investigated.
N
orma
lPlo
tofR
esid
uals
99
95
90
80
70
50
30
20
10
5
-1000 -500 0 500 1000
R
esid
uals
4-18
R
esid
ualsv
s.Pre
dicte
d
1000
500
Residuals
-500
-1000
-500.00 0.00 500.00 1000.00 1500.00
P
red
icte
d
Pre
dicte
dvs.Actu
al
2500.00
2000.00
1500.00
Predicted
1000.00
500.00
0.00
-500.00
-500.00 0.00 500.00 1000.00 1500.00 2000.00 2500.00
A
ctu
al
4-19
R
esid
ualsv
s.Alg
orith
m
1000
500
Residuals
-500
-1000
1.00 2.00 3.00 4.00 5.00 6.00
A
:Alg
orith
m
(c) Which algorithm would you recommend for use in practice?
The FUNCTIONAL POINTS algorithm has the losest cost estimation error.
4.12. An article in Nature Genetics (2003, Vol. 34, pp. 85-90) “Treatment-Specific Changes in Gene
Expression Discriminate in vivo Drug Response in Human Leukemia Cells” studied gene expressionas a
function of different treatments for leukemia. Three treatment groups are: mercaptopurine (MP) only;
low-dose methotrexate (LDMTX) and MP; and high-dose methotrexate (HDMTX) and MP. Each group
contained ten subjects. The responses from a specific gene are shown in the table below:
Project
MP ONLY 334.5 31.6 701 41.2 61.2 69.6 67.5 66.6 120.7 881.9
MP + HDMTX 919.4 404.2 1024.8 54.1 62.8 671.6 882.1 354.2 321.9 91.1
MP + LDMTX 108.4 26.1 240.8 191.1 69.7 242.8 62.7 396.9 23.6 290.4
(a) Is there evidence to support the claim that the treatment means differ?
The ANOVA below identifies the treatment means are significantly different.
Response Gene Expression

Block 9.206E+005 9 1.023E+005
Model 5.384E+005 2 2.692E+005 3.68 0.0457 significant
A-Treatment 5.384E+005 2 2.692E+005 3.68 0.0457
Residual 1.316E+006 18 73130.15
Cor Total 2.775E+006 29

C.V. % 92.04 Pred R-Squared -0.9714
PRESS 3.657E+006 Adeq Precision 5.288
4-20

Estimated Standard
Mean Error
1-MP Only 237.58 85.52
2-MP + HDMTX 478.62 85.52
3-MP + LDMTX 165.25 85.52

1 vs 2 -241.04 1 120.94 -1.99 0.0616
1 vs 3 72.33 1 120.94 0.60 0.5572
2 vs 3 313.37 1 120.94 2.59 0.0184
(b) Chec the normality assumption. Can we assume these samples are from normal populations?
The normal plot of residuals below identifies a slightly non-normal distribution.
N
orma
lPlo
tofR
esid
uals
99
95
90
80
70
50
30
20
10
5
-600 -400 -200 0 200 400 600
R
esid
uals
(c) Take the logarithm of the raw data. Is there evidence to support the claim that the treatment means
differ for the transformed data?
The ANOVA for the natural log transformed data identifies the treatment means as only moderately
different with an F value of 0.07
Response Gene Expression

Transform:Natural Log Constant: 0
Block 14.75 9 1.64
Model 6.30 2 3.15 3.09 0.0700
A-Treatment 6.30 2 3.15 3.09 0.0700
Residual 18.32 18 1.02
Cor Total 39.37 29
The Model F-value of 3.09 implies there is a 7.00% chance that a "Model F-Value"
this large could occur due to noise.
4-21

C.V. % 19.83 Pred R-Squared -1.0672

Estimated Standard
Mean Error
1-MP Only 4.79 0.32
2-MP + HDMTX 5.73 0.32
3-MP + LDMTX 4.74 0.32

1 vs 2 -0.95 1 0.45 -2.10 0.0505
1 vs 3 0.050 1 0.45 0.11 0.9122
2 vs 3 1.00 1 0.45 2.21 0.0405
(d) Analyze the residuals from the transformed data and comment on model adequacy.
The residual plots below identify no concerns with the model adequacy.
N
orma
lPlo
tofR
esid
uals
99
95
90
80
70
50
30
20
10
5
-2 -1 0 1 2
R
esid
uals
4-22
R
esid
ualsv
s.Pre
dicte
d
2
1
Residuals
-1
-2
3.00 4.00 5.00 6.00 7.00
P
red
icte
d
Pre
dicte
dvs.Actu
al
7.00
6.00
Predicted
5.00
4.00
3.00
3.00 4.00 5.00 6.00 7.00
A
ctu
al
4-23
R
esid
ualsv
s.T
rea
tme
nt
2
1
Residuals
-1
-2
1.00 1.50 2.00 2.50 3.00
A
:Tre
atm
ent
4.13. Consider the ratio control algorithm experiment described in Section 3.8. The experiment was
actually conducted as a randomized block design, where six time periods were selected as the blocks, and
all four ratio control algorithms were tested in each time period. The average cell voltage and the standard
deviation of voltage (shown in parentheses) for each cell are as follows:
Ratio Control Time Period

Algorithms 1 2 3 4 5 6
1 4.93 (0.05) 4.86 (0.04) 4.75 (0.05) 4.95 (0.06) 4.79 (0.03) 4.88 (0.05)
2 4.85 (0.04) 4.91 (0.02) 4.79 (0.03) 4.85 (0.05) 4.75 (0.03) 4.85 (0.02)
3 4.83 (0.09) 4.88 (0.13) 4.90 (0.11) 4.75 (0.15) 4.82 (0.08) 4.90 (0.12)
4 4.89 (0.03) 4.77 (0.04) 4.94 (0.05) 4.86 (0.05) 4.79 (0.03) 4.76 (0.02)
(a) Analyze the average cell voltage data. (Use α = 0.05.) Does the choice of ratio control algorithm
affect the cell voltage?

Response: Average
Sum of Mean F
Block 0.017 5 3.487E-003
Model 2.746E-003 3 9.153E-004 0.19 0.9014 not significant
A 2.746E-003 3 9.153E-004 0.19 0.9014
Residual 0.072 15 4.812E-003
Cor Total 0.092 23

Mean 4.84 Adj R-Squared -0.1560

Estimated Standard
Mean Error
4-24
1-1 4.86 0.028

2-2 4.83 0.028
3-3 4.85 0.028
4-4 4.84 0.028

1 vs 2 0.027 1 0.040 0.67 0.5156
1 vs 3 0.013 1 0.040 0.33 0.7438
1 vs 4 0.025 1 0.040 0.62 0.5419
2 vs 3 -0.013 1 0.040 -0.33 0.7438
2 vs 4 -1.667E-003 1 0.040 -0.042 0.9674
3 vs 4 0.012 1 0.040 0.29 0.7748
The ratio control algorithm does not affect the mean cell voltage.
(b) Perform an appropriate analysis of the standard deviation of voltage. (Recall that this is called “pot
noise.”) Does the choice of ratio control algorithm affect the pot noise?

Response: StDev Transform: Natural log Constant: 0.000
Sum of Mean F
Block 0.94 5 0.19
Model 6.17 3 2.06 33.26 < 0.0001 significant
A 6.17 3 2.06 33.26 < 0.0001
Residual 0.93 15 0.062
Cor Total 8.04 23

Mean -3.04 Adj R-Squared 0.8432
C.V. -8.18 Pred R-Squared 0.6654

Estimated Standard
Mean Error
1-1 -3.09 0.10
2-2 -3.51 0.10
3-3 -2.20 0.10
4-4 -3.36 0.10

1 vs 2 0.42 1 0.14 2.93 0.0103
1 vs 3 -0.89 1 0.14 -6.19 < 0.0001
1 vs 4 0.27 1 0.14 1.87 0.0813
2 vs 3 -1.31 1 0.14 -9.12 < 0.0001
2 vs 4 -0.15 1 0.14 -1.06 0.3042
3 vs 4 1.16 1 0.14 8.06 < 0.0001
A natural log transformation was applied to the pot noise data. The ratio control algorithm does affect the
pot noise.
4-25
(c) Conduct any residual analyses that seem appropriate.
Normal plot of residuals Residuals vs. Predicted

0.288958
99
95
0.126945
90
80
Res iduals
70
50 -0.0350673
30
20
10
-0.19708
5
-0.359093
-0.359093 -0.19708 -0.0350673 0.126945 0.288958 -3.73 -3.26 -2.78 -2.31 -1.84
Res idual Predicted
Residuals vs. Algorithm

0.288958
0.126945
Res iduals
-0.0350673
-0.19708
-0.359093
1 2 3 4
Algorithm
The normal probability plot shows slight deviations from normality; however, still acceptable.
(d) Which ratio control algorithm would you select if your objective is to reduce both the average cell
voltage and the pot noise?
Since the ratio control algorithm has little effect on average cell voltage, select the algorithm that
minimizes pot noise, that is algorithm #2.
4-26
4.14. An aluminum master alloy manufacturer produces grain refiners in ingot form. The company
produces the product in four furnaces. Each furnace is known to have its own unique operating
characteristics, so any experiment run in the foundry that involves more than one furnace will consider
furnaces as a nuisance variable. The process engineers suspect that stirring rate impacts the grain size of
the product. Each furnace can be run at four different stirring rates. A randomized block design is run for a
particular refiner and the resulting grain size data is as follows.
Furnace
Stirring Rate 1 2 3 4
5 8 4 5 6
10 14 5 6 9
15 14 6 9 2
20 17 9 3 6
(a) Is there any evidence that stirring rate impacts grain size?

Response: Grain Size
Sum of Mean F
Block 165.19 3 55.06
A 22.19 3 7.40 0.85 0.4995
Residual 78.06 9 8.67
Cor Total 265.44 15

Mean 7.69 Adj R-Squared -0.0382

Estimated Standard
Mean Error
1-5 5.75 1.47
2-10 8.50 1.47
3-15 7.75 1.47
4-20 8.75 1.47

1 vs 2 -2.75 1 2.08 -1.32 0.2193
1 vs 3 -2.00 1 2.08 -0.96 0.3620
1 vs 4 -3.00 1 2.08 -1.44 0.1836
2 vs 3 0.75 1 2.08 0.36 0.7270
2 vs 4 -0.25 1 2.08 -0.12 0.9071
3 vs 4 -1.00 1 2.08 -0.48 0.6425
The analysis of variance shown above indicates that there is no difference in mean grain size due to the
different stirring rates.
4-27
(b) Graph the residuals from this experiment on a normal probability plot. Interpret this plot.
Normal plot of residuals
99
95
90
80
70
50
30
20
10
5
-3.8125 -2.0625 -0.3125 1.4375 3.1875
Res idual
The plot indicates that normality assumption is valid.
(c) Plot the residuals versus furnace and stirring rate. Does this plot convey any useful information?
Residuals vs. Stirring Rate Residuals vs. Furnace

3.1875 3.1875
1.4375 1.4375
Res iduals
Residuals
2
-0.3125 -0.3125
-2.0625 -2.0625
-3.8125 -3.8125
1 2 3 4 1 2 3 4
Stirring Rate Furnace
The variance is consistent at different stirring rates. Not only does this validate the assumption of uniform
variance, it also identifies that the different stirring rates do not affect variance.
(d) What should the process engineers recommend concerning the choice of stirring rate and furnace for
this particular grain refiner if small grain size is desirable?
There really is no effect due to the stirring rate.
4-28
4.15. Analyze the data in Problem 4.4 using the general regression significance test.
µ : 12µˆ +4τˆ1 +4τˆ2 +4τˆ3 +3βˆ1 +3βˆ2 +3βˆ3 +3βˆ4 =

225
τ1 : 4 µˆ +4τˆ1 + βˆ 1 + βˆ 2 + βˆ 3 + βˆ4 =
92
τ2 : 4 µˆ +4τˆ2 + βˆ1 + βˆ2 + βˆ3 + βˆ4 =
101
τ3 : 4 µˆ +4τˆ3 + βˆ 1 + βˆ 2 + βˆ 3 + βˆ4 =
32
β1 : 3µˆ +τˆ1 +τˆ2 +τˆ3 +3βˆ1 =
34
β1 : 3µˆ +τˆ1 +τˆ2 +τˆ3 +3βˆ2 =
50
β1 : 3µˆ +τˆ1 +τˆ2 +τˆ3 +3βˆ3 =
36
β : 3µˆ +τˆ1 +τˆ2 +τˆ3 +3βˆ4 =
105
Applying the constraints ∑τˆ = ∑ βˆ

i j = 0 , we obtain:
225 51 78 −129 −89 −25 −81 195

µ̂ = , τ̂ 1 = , τ̂ 2 = , τ̂ 3 = , β̂1 = , β̂ 2 = , β̂ 3 = , β̂ 4 =
12 12 12 12 12 12 12 12
       −   −   − 
R(µ ,τ , β ) =  (225) +  (92) +  (101) +  (32) +  (34) +  (50) +
225 51 78 129 89 25
 12   12   12   12   12   12 
 − 81   195 
 (36 ) +  (105) = 6029.17
 12   12 
∑∑ y 2
ij = 6081 , SS E = ∑∑ y 2
ij − R(µ ,τ , β ) = 6081 − 6029.17 = 51.83
Model Restricted to τ i = 0 :
µ : 12µˆ +3βˆ1 +3βˆ2 +3βˆ3 +3βˆ4 =

225
β1 : 3µˆ +3βˆ1 =
34
β2 : 3µˆ +3βˆ2 =
50
β3 : 3µˆ +3βˆ3 =
36
β4 : 3µˆ +3βˆ4 =
105
Applying the constraint ∑ β̂ j = 0 , we obtain:
225 ˆ −25 −81 195

µ̂ = , β1 = −89 / 12 , β̂ 2 = , β̂ 3 = , β̂ 4 = . Now:
12 12 12 12
 225   − 89   − 25   − 81   195 
R(µ , β ) =  (225) +  (34 ) +  (50 ) +  (36 ) +  (105) = 5325.67
 12   12   12   12   12 
R (τ µ , β ) = R(µ ,τ , β ) − R(µ , β ) = 6029.17 − 5325.67 = 703.50 = SS Treatments
Model Restricted to β j = 0 :
4-29
µ : 12µˆ +4τˆ1 +4τˆ2 +4τˆ3 =

225
τ 1 : 4µˆ +4τˆ1 =92
τ 2 : 4µˆ +4τˆ2 =
101
τ 3 : 4µˆ +4τˆ3 =32
Applying the constraint ∑ τ̂ i = 0 , we obtain:
225 51 78 −129
µ̂ = , τ̂ 1 = , τ̂ 2 = , τ̂ 3 =
12 12 12 12
 225   51   78   − 129 
R(µ ,τ ) =  (225) +  (92 ) +  (101) +  (32 ) = 4922.25
 12   12   12   12 
( )
R β µ ,τ = R(µ ,τ , β ) − R(µ ,τ ) = 6029.17 − 4922.25 = 1106.92 = SS Blocks
4.16. Assuming that chemical types and bolts are fixed, estimate the model parameters τi and βj in
Problem 4.3.
Using Equations 4.18, applying the constraints, we obtain:
35 −23 −7 13 17 35 −65 75 20 −65

µ̂ = , τ̂ 1 = , τ̂ 2 = , τ̂ 3 = , τ̂ 4 = , β̂ = , β̂ 2 = , β̂ 3 = , β̂ 4 = , β̂ 5 =
20 20 20 20 20 1 20 20 20 20 20
4.17. Draw an operating characteristic curve for the design in Problem 4.4. Does this test seem to be
sensitive to small differences in treatment effects?
Assuming that solution type is a fixed factor, we use the OC curve in appendix V. Calculate
b∑τ i2 4∑τ i2
=
Φ2 =
aσ 2 3 ( 8.64 )
using MS E to estimate σ2. We have:
υ1 = a − 1 = 2 υ 2 = (a − 1)(b − 1) = (2)(3) = 6 .
If ∑ τˆ 2
i
= σ 2 = MS E , then:
4
Φ= = 1.15 and β ≅ 0.70
3(1)
If ∑ τˆ i = 2σ 2 = 2 MS E , then:
4 ( 2)
=Φ = 1.63 and β ≅ 0.55 , etc.
3 (1)
This test is not very sensitive to small differences.
4-30
4.18. Suppose that the observation for chemical type 2 and bolt 3 is missing in Problem 4.3. Analyze the
problem by estimating the missing value. Perform the exact analysis and compare the results.
ay2.' + by.3' − y..' 4 ( 282 ) + 5 ( 227 ) − 1360

y 23 is=
missing. yˆ 23 = = 75.25
( a − 1)( b − 1) ( 3)( 4 )
Therefore, y 2. =357.25, y .3 =302.25, and y .. =1435.25
Source SS DF MS F0
Chemicals 12.7844 3 4.2615 2.154
Bolts 158.8875 4
Error 21.7625 11 1.9784
Total 193.4344 18
F 0.05,3,11 =3.59, Chemicals are not significant. This is the same result as found in Problem 4.3.
4.19. Consider the hardness testing experiment in Problem 4.7. Suppose that the observation for tip 2 in
coupon 3 is missing. Analyze the problem by estimating the missing value.
ay2.' + by.3' − y..' 4 ( 28.6 ) + 4 ( 29.1) − 144.2

y 23 is=
missing. yˆ 23 = = 9.62
( a − 1)( b − 1) ( 3)( 3)
Therefore, y 2. =38.22, y .3 =38.72, and y .. =153.82
Source SS DF MS F0
Tip 0.40 3 0.133333 19.29
Coupon 0.80 3
Error 0.0622 9 0.006914
Total 1.2622 15
F 0.05,3,9 =3.86, Tips are significant. This is the same result as found in Problem 4.7.
4.20. Two missing values in a randomized block. Suppose that in Problem 4.3 the observations for
chemical type 2 and bolt 3 and chemical type 4 and bolt 4 are missing.
(a) Analyze the design by iteratively estimating the missing values as described in Section 4.1.3.
4 y'2. + 5 y'.3 − y'.. 4 y' + 5 y'.4 − y'..

ŷ23 = and ŷ44 = 4.
12 12
0
Data is coded y-70. As an initial guess, set y23 equal to the average of the observations available for
2
0
chemical 2. Thus, y 23 = = 0.5 . Then ,
4
4(8) + 5(6 ) − 25.5
0
ŷ 44 = = 3.04
12
4(2) + 5(17 ) − 28.04
ŷ 123 = = 5.41
12
4-31
4(8) + 5(6 ) − 30.41

ŷ 144 = = 2.63
12
4(2 ) + 5(17 ) − 27.63
2
ŷ 44 = = 5.44
12
4(8) + 5(6 ) − 30.44
2
ŷ 44 = = 2.63
12
∴ ŷ23 = 5.44 ŷ 44 = 2.63

Sum of Mean F
Block 156.83 4 39.21
A 9.59 3 3.20 2.08 0.1560
Residual 18.41 12 1.53
Cor Total 184.83 19
(b) Differentiate SSE with respect to the two missing values, equate the results to zero, and solve for
estimates of the missing values. Analyze the design using these two estimates of the missing values.
SS E = ∑∑ y 2
ij − 15 ∑y 2
i. − 14 ∑y 2
.j + 20
1
∑y 2
..
SS E = 0.6 y23
2
+ 0.6 y44
2
− 6.8 y23 − 3.7 y44 + 0.1y23 y44 + R
∂SS E ∂SS E
From = = 0 , we obtain:
∂y23 ∂y44
1.2 ŷ23 + 0.1 ŷ44 = 6.8

⇒ ŷ23 = 5.45 , ŷ 44 = 2.63
0.1 ŷ23 + 1.2 ŷ44 = 3.7
These quantities are almost identical to those found in part (a). The analysis of variance using these new
data does not differ substantially from part (a).
(c) Derive general formulas for estimating two missing values when the observations are in different
blocks.
( y′ + y ) + ( y′ ) − ( y′ ) + ( y′ ) + ( y′ )
2 2 2 2 2
i. iu k. + ykv .u + yiu .v + ykv .. + yiu + ykv
SS E =
y +y −
2
iu
2
kv
b a ab
∂SS E ∂SS E
From = = 0 , we obtain:
∂y23 ∂y44
 (a − 1)(b − 1)  ay' i . +by' . j − y' .. ŷ kv

ŷ iu  = −
 ab  ab ab
 ( a − 1 )( b − 1 )  ay'k . +by'.v − y'.. ŷiu
ŷkv  = −
 ab  ab ab
4-32
whose simultaneous solution is:
y 'i. a 1 − ( a − 1) ( b − 1) − ab  + y '.u b 1 − ( a − 1) ( b − 1) − ab  − y '.. 1 − ab ( a − 1) ( b − 1) 

2 2 2 2 2 2
yîu      +
( )( )  ( ) ( ) 
− −  − − 
2
−
2
a 1 b 1 1 a 1 b 1
ab [ ay 'k . + by '.v − y '.. ]
1 − ( a − 1)2 ( b − 1)2 
 
ay 'i. + by '.u − y '.. − ( b − 1)( a − 1) [ ay 'k . + by '.v − y '.. ]
yˆ kv =
1 − ( a − 1)2 ( b − 1)2 
 
(d) Derive general formulas for estimating two missing values when the observations are in the same
block. Suppose that two observations y ij and y kj are missing, i≠k (same block j).
SS E = yij2 + ykj2 −
(y ′ i. + yij ) + (y ′
2
k. + ykj ) − (y ′
2
.j + yij + ykj ) + (y ′
2
.. + yij + ykj )
2
b a ab
∂ SS ∂ SS E
From =E
= 0 , we obtain
∂ yij ∂ ykj
ayi′. + by.′j − y..′

ŷij = + ŷkj (a − 1)(b − 1)2
(a − 1)(b − 1)
ayk′ . + by.′j − y..′
ŷkj = + ŷij (a − 1)(b − 1)2
(a − 1)(b − 1)
whose simultaneous solution is:
( b − 1)  ayk′ . + by.′j − y..′ + ( a − 1)( b − 1) ( ayi′. + by.′j − y..′ )

2
ayi′. + by.′j − y..′
=yîj +
( a − 1)( b − 1) 1 − ( a − 1)2 ( b − 1)4 
 
ayk′ . + by.′j − y..′ − ( b − 1) ( a − 1)  ayi′. + by.′j − y..′ 
2
yˆ kj =
( a − 1)( b − 1) 1 − ( a − 1) ( b − 1) 
2 4
4.21. An industrial engineer is conducting an experiment on eye focus time. He is interested in the effect
of the distance of the object from the eye on the focus time. Four different distances are of interest. He has
five subjects available for the experiment. Because there may be differences among individuals, he decides
to conduct the experiment in a randomized block design. The data obtained follow. Analyze the data from
this experiment (use α = 0.05) and draw appropriate conclusions.
Subject
Distance (ft) 1 2 3 4 5
4 10 6 6 6 6
6 7 6 6 1 6
8 5 3 3 2 5
10 6 4 4 2 3
4-33

Response: Focus Time
Sum of Mean F
Block 36.30 4 9.07
A 32.95 3 10.98 8.61 0.0025
Residual 15.30 12 1.27
Cor Total 84.55 19


Estimated Standard
Mean Error
1-4 6.80 0.50
2-6 5.20 0.50
3-8 3.60 0.50
4-10 3.80 0.50

1 vs 2 1.60 1 0.71 2.24 0.0448
1 vs 3 3.20 1 0.71 4.48 0.0008
1 vs 4 3.00 1 0.71 4.20 0.0012
2 vs 3 1.60 1 0.71 2.24 0.0448
2 vs 4 1.40 1 0.71 1.96 0.0736
3 vs 4 -0.20 1 0.71 -0.28 0.7842
Distance has a statistically significant effect on mean focus time.
4.22. The effect of five different ingredients (A, B, C, D, E) on reaction time of a chemical process is
being studied. Each batch of new material is only large enough to permit five runs to be made.
Furthermore, each run requires approximately 1 1/2 hours, so only five runs can be made in one day. The
experimenter decides to run the experiment as a Latin square so that day and batch effects can be
systematically controlled. She obtains the data that follow. Analyze the data from this experiment (use α =
0.05) and draw conclusions.
Day
Batch 1 2 3 4 5
1 A=8 B=7 D=1 C=7 E=3
2 C=11 E=2 A=7 D=3 B=8
3 B=4 A=9 C=10 E=1 D=5
4 D=6 C=8 E=6 B=6 A=10
5 E=4 D=2 B=3 A=8 C=8
The Minitab output below identifies the ingredients as having a significant effect on reaction time.
4-34
Minitab Output
General Linear Model
Factor Type Levels Values

Batch random 5 1 2 3 4 5
Day random 5 1 2 3 4 5
Catalyst fixed 5 A B C D E
Analysis of Variance for Time, using Adjusted SS for Tests
Source DF Seq SS Adj SS Adj MS F P

Catalyst 4 141.440 141.440 35.360 11.31 0.000
Batch 4 15.440 15.440 3.860 1.23 0.348
Day 4 12.240 12.240 3.060 0.98 0.455
Error 12 37.520 37.520 3.127
Total 24 206.640
4.23. An industrial engineer is investigating the effect of four assembly methods (A, B, C, D) on the
assembly time for a color television component. Four operators are selected for the study. Furthermore,
the engineer knows that each assembly method produces such fatigue that the time required for the last
assembly may be greater than the time required for the first, regardless of the method. That is, a trend
develops in the required assembly time. To account for this source of variability, the engineer uses the
Latin square design shown below. Analyze the data from this experiment (α = 0.05) draw appropriate
conclusions.
Order of Operator
Assembly 1 2 3 4
1 C=10 D=14 A=7 B=8
2 B=7 C=18 D=11 A=8
3 A=5 B=10 C=11 D=9
4 D=10 A=10 B=12 C=14
The Minitab output below identifies assembly method as having a significant effect on assembly time.
Minitab Output

Order random 4 1 2 3 4
Operator random 4 1 2 3 4
Method fixed 4 A B C D

Method 3 72.500 72.500 24.167 13.81 0.004
Order 3 18.500 18.500 6.167 3.52 0.089
Operator 3 51.500 51.500 17.167 9.81 0.010
Error 6 10.500 10.500 1.750
Total 15 153.000
4-35
4.24. Consider the randomized complete block design in Problem 4.4. Assume that the days are random.
Estimate the block variance component.
The block variance component is:
=σˆ β2
[ MS
= Blocks − MS E ] [=
368.97 − 8.64]
120.11
a 3
4.25. Consider the randomized complete block design in Problem 4.7. Assume that the coupons are
random. Estimate the block variance component.
=σˆ β2
[=
MS Blocks − MS E ] [ 0.27 − 0.008889]
= 0.06528
a 4
4.26. Consider the randomized complete block design in Problem 4.9. Assume that the trucks are random.
Estimate the block variance component.
=σˆ β2
[=
MS Blocks − MS E ] [ 0.023 − 0.0005278]
= 0.007491
a 3
4.27. Consider the randomized complete block design in Problem 4.11. Assume that the software projects
that were used as blocks are random. Estimate the block variance component.
=σˆ β2
[=
MS Blocks − MS E ] [ 457500 − 111200]
= 57716.67
a 6
4.28. Consider the gene expression experiment in Problem 4.12. Assume that the subjects used in this
experiment are random. Estimate the block variance component
=σˆ β2
[=
MS Blocks − MS E ] [102300 − 73130.15]
= 9723.28
a 3
4.29. Suppose that in Problem 4.22 the observation from batch 3 on day 4 is missing. Estimate the
missing value and perform the analysis using this value.
y 354 is missing. ŷ 354 =

[ ]
p y i′.. + y .′j . + y ..′ k − 2 y ...
′
=
5[28 + 15 + 24] − 2(146 )
= 3.58
( p − 2)( p − 1) (3)(4)
4-36
Minitab Output

Day random 5 1 2 3 4 5

Catalyst 4 128.676 128.676 32.169 11.25 0.000
Batch 4 16.092 16.092 4.023 1.41 0.290
Day 4 8.764 8.764 2.191 0.77 0.567
Error 12 34.317 34.317 2.860
Total 24 187.849
4.30. Consider a p x p Latin square with rows (αi), columns (βk), and treatments (τj) fixed. Obtain least
squares estimates of the model parameters αi, βk, τj.
p p p
µ : p µˆ + p
2
∑
i =1
αˆ i + p ∑
j =1
τˆ j + p ∑ βˆ
k =1
k = y...
p p
α i : pµˆ + pαˆ i + p ∑j =1
τˆ j + p ∑ βˆ
k =1
k = yi .. , i = 1,2 ,..., p
p p
τ j : pµˆ + p ∑
i =1
αˆ i + pτˆ j + p ∑ βˆ
k =1
k = y. j . , j = 1,2 ,..., p
p p
β k : pµˆ + p ∑
i =1
αˆ i + p ∑τˆj =1
j + pβˆ k = y..k , k = 1,2 ,..., p
There are 3p+1 equations in 3p+1 unknowns. The rank of the system is 3p-2. Three side conditions are
p p p
necessary. The usual conditions imposed are: ∑ i =1
αˆ i = ∑ j =1
τˆ j = ∑ βˆ
k =1
k = 0 . The solution is then:
y...
µˆ
= = y...
p2
αˆ i =yi.. − y... , i =1, 2,..., p
τˆ j =y. j . − y... , j =1, 2,..., p
βˆk =yi.. − y... , k =
1, 2,..., p
4.31. Derive the missing value formula (Equation 4.27) for the Latin square design.
yi2.. y.2j . y..2k  y2 

SS E = ∑∑∑ 2
yijk − ∑ p
− ∑ p
− ∑ p
+ 2 ...2 
p 
 
4-37
Let y ijk be missing. Then
( y′ + y ) − ( y′ ) − ( y′ ) ( )
2 2 2 2
i .. ijk . j. + yijk ..k + yijk 2 y...′ + yijk
SS E =
y − 2
ijk + +R
p p p p2
∂SS E
where R is all terms without y ijk .. From = 0 , we obtain:
∂y ijk
( p − 1)( p − 2) = ( )
p y' i .. + y' . j . + y' ..k − 2 y' ... ( )
p y' i .. + y' . j . + y' ..k − 2 y' ...
y ijk , or y ijk =
p 2
p 2 ( p − 1)( p − 2)
4.32. Designs involving several Latin squares. [See Cochran and Cox (1957), John (1971).] The p x p
Latin square contains only p observations for each treatment. To obtain more replications the experimenter
may use several squares, say n. It is immaterial whether the squares used are the same are different. The
appropriate model is
 i = 1,2,..., p
 j = 1,2,..., p

y ijkh = µ + ρ h + α i( h ) + τ j + β k ( h ) + ( τρ ) jh + ε ijkh 
k = 1,2,..., p
 h = 1,2,..., n
where yijkh is the observation on treatment j in row i and column k of the hth square. Note that α i ( h ) and
β k ( h) are row and column effects in the hth square, and ρh is the effect of the hth square, and ( τρ) jh is the
interaction between treatments and squares.
(a) Set up the normal equations for this model, and solve for estimates of the model parameters. Assume
that appropriate side conditions on the parameters are ρ̂ h = 0 , α̂ i (h ) = 0 , and∑β̂ k (h ) = 0
h ∑ i ∑ k
for each h, ∑ τ̂ j = 0, ∑ (τˆρ ) jh = 0 for each h, and ∑ (τˆρ ) jh = 0 for each j.

j j h
µˆ = y....
ρˆ h y...h − y....
=
τˆ j y. j .. − y....
=
αˆ i=
(h) yi..h − y...h
βˆk=
(h) y..kh − y...h
 ^ 
τρ  = y. j .h − y. j .. − y...h + y....
  jh
4-38
(b) Write down the analysis of variance table for this design.
Source SS DF
y.2j .. 2
∑ np − np
y....
Treatments 2
p-1
y...2 h 2
∑p
y....
Squares 2
− n-1
np 2
y.2j .h 2
∑
y....
Treatment x Squares − − SSTreatments − SS Squares (p-1)(n-1)
p np 2
yi2..h y...2 h
Rows ∑ p
−
np
n(p-1)
y..2kh y...2 h
Columns ∑ p − np n(p-1)
Error subtraction n(p-1)(p-2)

2
∑∑∑∑ y
y....
Total 2
ijkh − 2
np2-1
np
4.33. Discuss how the operating characteristics curves in the Appendix may be used with the Latin square
design.
For the fixed effects model use:
Φ2 =
∑ pτ = ∑ τ
2
j
2
j
, υ1 = p − 1 υ 2 = ( p − 2)( p − 1)
pσ2
σ 2
For the random effects model use:
pσ τ2
λ = 1+ , υ1 = p − 1 υ 2 = ( p − 2)( p − 1)
σ2
4.34. Suppose that in Problem 4.22 the data taken on day 5 were incorrectly analyzed and had to be
discarded. Develop an appropriate analysis for the remaining data.
Two methods of analysis exist: (1) Use the general regression significance test, or (2) recognize that the
design is a Youden square. The data can be analyzed as a balanced incomplete block design with a = b =
5, r = k = 4 and λ = 3. Using either approach will yield the same analysis of variance.
Minitab Output

Day random 4 1 2 3 4
4-39

Catalyst 4 119.800 120.167 30.042 7.48 0.008
Batch 4 11.667 11.667 2.917 0.73 0.598
Day 3 6.950 6.950 2.317 0.58 0.646
Error 8 32.133 32.133 4.017
Total 19 170.550
4.35. The yield of a chemical process was measured using five batches of raw material, five acid
concentrations, five standing times, (A, B, C, D, E) and five catalyst concentrations (α, β, γ, δ, ε). The
Graeco-Latin square that follows was used. Analyze the data from this experiment (use α = 0.05) and draw
conclusions.
Acid Concentration
Batch 1 2 3 4 5
1 Aα=26 Bβ=16 Cγ=19 Dδ=16 Eε=13
2 Bγ=18 Cδ=21 Dε=18 Eα=11 Aβ=21
3 Cε=20 Dα=12 Eβ=16 Aγ=25 Bδ=13
4 Dβ=15 Eγ=15 Aδ=22 Bε=14 Cα=17
5 Eδ=10 Aε=24 Bα=17 Cβ=17 Dγ=14
The Minitab output below identifies standing time as having a significant effect on yield.
Minitab Output

Time fixed 5 A B C D E
Catalyst random 5 a b c d e
Acid random 5 1 2 3 4 5
Analysis of Variance for Yield, using Adjusted SS for Tests
Time 4 342.800 342.800 85.700 14.65 0.001
Catalyst 4 12.000 12.000 3.000 0.51 0.729
Batch 4 10.000 10.000 2.500 0.43 0.785
Acid 4 24.400 24.400 6.100 1.04 0.443
Error 8 46.800 46.800 5.850
Total 24 436.000
4.36. Suppose that in Problem 4.23 the engineer suspects that the workplaces used by the four operators
may represent an additional source of variation. A fourth factor, workplace (α, β, γ, δ) may be introduced
and another experiment conducted, yielding the Graeco-Latin square that follows. Analyze the data from
this experiment (use α = 0.05) and draw conclusions.
Order of Operator
Assembly 1 2 3 4
1 Cβ=11 Bγ=10 Dδ=14 Aα=8
2 Bα=8 Cδ=12 Aγ=10 Dβ=12
3 Aδ=9 Dα=11 Bβ=7 Cγ=15
4 Dγ=9 Aβ=8 Cα=18 Bδ=6
4-40
Minitab Output

Method fixed 4 A B C D
Order random 4 1 2 3 4
Operator random 4 1 2 3 4
Workplac random 4 a b c d
Method 3 95.500 95.500 31.833 3.47 0.167
Order 3 0.500 0.500 0.167 0.02 0.996
Operator 3 19.000 19.000 6.333 0.69 0.616
Workplac 3 7.500 7.500 2.500 0.27 0.843
Error 3 27.500 27.500 9.167
Total 15 150.000
Method and workplace do not have a significant effect on assembly time. However, there are only three
degrees of freedom for error, so the test is not very sensitive.
4.37. Construct a 5 x 5 hypersquare for studying the effects of five factors. Exhibit the analysis of
variance table for this design.
Three 5 x 5 orthogonal Latin Squares are:
ABCDE αβγδε 12345

BCDEA γδεαβ 45123
CDEAB εαβγδ 23451
DEABC βγδεα 51234
EABCD δεαβγ 34512
Let rows = factor 1, columns = factor 2, Latin letters = factor 3, Greek letters = factor 4 and numbers =
factor 5. The analysis of variance table is:
Source SS DF
5 2
1 y
Rows ∑ yi2.... − 25
5 i =1
..... 4
1 5 2 y2
Columns ∑
5 m =1
y....m − .....
25
4
2
1 5 2 y.....
Latin Letters ∑
5 j =1
y. j ... −
25
4
1 5 2 y2
Greek Letters ∑
5 k =1
y..k .. − .....
25
4
2
1 5 2 y.....
Numbers ∑
5 l =1
y...l . −
25
4
Error SS E by subtraction 4
5 5 5 5 5
y2
Total ∑∑∑∑∑ y
=i 1 =j 1 =
k 1 =l 1 =
m 1
2
ijklm − .....
25
24
4-41
4.38. Consider the data in Problems 4.23 and 4.36. Suppressing the Greek letters in 4.36, analyze the data
using the method developed in Problem 4.32.
Square 1 - Operator
Batch 1 2 3 4 Row Total
1 C=10 D=14 A=7 B=8 (39)
2 B=7 C=18 D=11 A=8 (44)
3 A=5 B=10 C=11 D=9 (35)
4 D=10 A=10 B=12 C=14 (46)
(32) (52) (41) (36) 164=y …1
Square 2 - Operator
Batch 1 2 3 4 Row Total
1 C=11 B=10 D=14 A=8 (43)
2 B=8 C=12 A=10 D=12 (42)
3 A=9 D=11 B=7 C=15 (42)
4 D=9 A=8 C=18 B=6 (41)
(37) (41) (49) (41) 168=y …2
Assembly Methods Totals

A y .1.. =65
B y .2.. =68
C y .3.. =109
D y .4.. =90
Source SS DF MS F0
Assembly Methods 159.25 3 53.08 14.00*
Squares 0.50 1 0.50
AxS 8.75 3 2.92 0.77
Assembly Order (Rows) 19.00 6 3.17
Operators (columns) 70.50 6 11.75
Error 45.50 12 3.79
Total 303.50 31
Significant at 1%.
4.39. Consider the randomized block design with one missing value in Problem 4.19. Analyze this data
by using the exact analysis of the missing value problem discussed in Section 4.1.4. Compare your results
to the approximate analysis of these data given from Problem 4.19.
To simplify the calculations, the data in Problems 4.19 was transformed by multiplying by 10 and
substracting 95.
4-42
µ : 15µˆ +4τˆ1 +3τˆ2 +4τˆ3 +4τˆ4 +4 βˆ1 +4 βˆ2 +3βˆ3 +4 βˆ4 =

17
τ1 : 4 µˆ +4τˆ1 + βˆ 1 + βˆ2 + βˆ3 + βˆ 4 =
3
τ2 : 3µˆ +3τˆ2 + βˆ1 + βˆ2 + βˆ4 =
1
τ3 : 4 µˆ +4τˆ3 + βˆ1 + βˆ2 + βˆ3 + βˆ 4 =
−2
τ4 : 4 µˆ +4τˆ4 + βˆ1 + βˆ2 + βˆ3 + βˆ4 =
15
β1 : 4µˆ +τˆ1 +τˆ2 +τˆ3 +τˆ4 +4 βˆ1 =
−4
β2 : 4 µˆ +τˆ1 +τˆ2 +τˆ3 +τˆ4 +4 βˆ2 =
−3
β3 : 3µˆ +τˆ1 +τˆ3 +τˆ4 +3βˆ3 =
6
β4 : 4 µˆ +τˆ1 +τˆ2 +τˆ3 +τˆ4 +4 βˆ4 =
18
Applying the constraints ∑ τˆ = ∑ βî j = 0 , we obtain:
41 −14 −24 −59 94 −77 −68 24 121

µ̂ = , τ̂ 1 = , τ̂ 2 = , τ̂ 3 = , τ̂ 4 = , β̂1 = , β̂ 2 = , β̂ 3 = , β̂ 4 =
36 36 36 36 36 36 36 36 36
4 4
R(µ ,τ , β ) = µˆ y.. + ∑ τˆ y + ∑ βˆ y
i =1
i i.
j =1
j .j = 138.78
With 7 degrees of freedom.
∑∑ y 2
ij = 145.00 , SS E = ∑∑ y 2
ij − R(µ ,τ , β ) = 145.00 − 138.78 = 6.22
which is identical to SS E obtained in the approximate analysis. In general, the SS E in the exact and
approximate analyses will be the same.
To test H o : τ i = 0 the reduced model is yij = µ + β j + ε ij . The normal equations used are:
µ : 15µˆ +4 βˆ1 +4 βˆ2 +3βˆ3 +4 βˆ4 =

17
β1 : 4 µˆ +4 βˆ1 =
−4
β2 : 4 µˆ +4 βˆ2 =
−3
β3 : 3µˆ +3βˆ3 =
6
β4 : 4 µˆ +4 βˆ4 =
18
4
−35 −31
. Now R(µ , β ) = µˆ y .. + ∑ βˆ
19 13 53
µ̂ = , β̂1 = , β̂ 2 = , β̂ 3 = , β̂ 4 = j y . j = 99.25
16 16 16 16 16 j =1
with 4 degrees of freedom.
R (τ µ , β ) = R(µ ,τ , β ) − R(µ , β ) = 138.78 − 99.25 = 39.53 = SS Treatments
4-43
( )
with 7-4=3 degrees of freedom. R τ µ , β is used to test H o : τ i = 0 .
The sum of squares for blocks is found from the reduced model y ij = µ + τ i + ε ij . The normal equations
used are:
µ : 15µˆ +4τˆ1 +3τˆ2 +4τˆ3 +4τˆ4 =
17
τ 1 : 4µˆ +4τˆ1 =3
τ 2 : 3µˆ +3τˆ2 =1
τ 3 : 4µˆ +4τˆ3 =
−2
τ 4 : 4µˆ +4τˆ4 =
15
Applying the constraint ∑τˆ = 0 , we obtain:

i
13 −4 −9 −19 32
µˆ = , τˆ1 = , τˆ2 = , τˆ3 = , τˆ4 =
12 12 12 12 12
4
R(µ ,τ ) = µˆ y.. + ∑ τˆ y
i =1
i i. = 59.83
R (β µ ,τ ) = R(µ ,τ , β ) − R(µ ,τ ) = 138.78 − 59.83 = 78.95 = SS Blocks
with 7-4=3 degrees of freedom.
Source DF SS(exact) SS(approximate)

Tips 3 39.53 39.98
Blocks 3 78.95 79.53
Error 8 6.22 6.22
Total 14 125.74 125.73
Note that for the exact analysis, SST ≠ SSTips + SS Blocks + SS E .
4.40. An engineer is studying the mileage performance characteristics of five types of gasoline additives.
In the road test he wishes to use cars as blocks; however, because of a time constraint, he must use an
incomplete block design. He runs the balanced design with the five blocks that follow. Analyze the data
from this experiment (use α = 0.05) and draw conclusions.
Car
Additive 1 2 3 4 5
1 17 14 13 12
2 14 14 13 10
3 12 13 12 9
4 13 11 11 12
5 11 12 10 8
4-44
There are several computer software packages that can analyze the incomplete block designs discussed in
this chapter. The Minitab General Linear Model procedure is a widely available package with this
capability. The output from this routine follows. The adjusted sums of squares are the appropriate sums of
squares to use for testing the difference between the means of the gasoline additives. The gasoline
additives have a significant effect on the mileage.
Minitab Output

Additive fixed 5 1 2 3 4 5
Car random 5 1 2 3 4 5
Analysis of Variance for Mileage, using Adjusted SS for Tests

Additive 4 31.7000 35.7333 8.9333 9.81 0.001
Car 4 35.2333 35.2333 8.8083 9.67 0.001
Error 11 10.0167 10.0167 0.9106
Total 19 76.9500
4.41. Construct a set of orthogonal contrasts for the data in Problem 4.40. Compute the sum of squares for
each contrast.
One possible set of orthogonal contrasts is:
H 0 : µ 4 + µ5 = µ1 + µ 2 (1)
H 0 : µ1 = µ 2 (2)
H 0 : µ 4 = µ5 (3)
H 0 : 4 µ3 = µ 4 + µ5 + µ1 + µ 2 (4)
The sums of squares and F-tests are:
Brand -> 1 2 3 4 5
Qi 33/4 11/4 -3/4 -14/4 -27/4 ∑ ci Qi SS F0
(1) -1 -1 0 1 1 -85/4 30.10 33.06
(2) 1 -1 0 0 0 22/4 4.03 4.426
(3) 0 0 0 -1 1 -13/4 1.41 1.55
(4) -1 -1 4 -1 -1 -15/4 0.19 0.21
Contrasts (1) and (2) are significant at the 1% and 5% levels, respectively.
4-45
4.42. Seven different hardwood concentrations are being studied to determine their effect on the strength
of the paper produced. However the pilot plant can only produce three runs each day. As days may differ,
the analyst uses the balanced incomplete block design that follows. Analyze this experiment (use α = 0.05)
and draw conclusions.
Hardwood Days
Concentration (%) 1 2 3 4 5 6 7
2 114 120 117
4 126 120 119
6 137 117 134
8 141 129 149
10 145 150 143
12 120 118 123
14 136 130 127
There are several computer software packages that can analyze the incomplete block designs discussed in
this chapter. The Minitab General Linear Model procedure is a widely available package with this
capability. The output from this routine for Problem 4.35 follows. The adjusted sums of squares are the
appropriate sums of squares to use for testing the difference between the means of the hardwood
concentrations.
Minitab Output

Concentr fixed 7 2 4 6 8 10 12 14
Days random 7 1 2 3 4 5 6 7
Analysis of Variance for Strength, using Adjusted SS for Tests

Concentr 6 2037.62 1317.43 219.57 10.42 0.002
Days 6 394.10 394.10 65.68 3.12 0.070
Error 8 168.57 168.57 21.07
Total 20 2600.29
4.43. Analyze the data in Example 4.5 using the general regression significance test.
µ : 12µˆ +3τˆ1 +3τˆ2 +3τˆ3 +3τˆ4 +3βˆ1 +3βˆ2 +3βˆ3 +3βˆ4 =

870
τ1 : 3µˆ +3τˆ1 + βˆ1 + βˆ 2 + βˆ 4 =
218
τ2 : 3µˆ +3τˆ2 + βˆ2 + βˆ3 + βˆ4 =
214
τ3 : 3µˆ +3τˆ3 + βˆ1 + βˆ 2 + βˆ3 =
216
τ4 : 3µˆ +3τˆ4 + βˆ1 + βˆ3 + βˆ4 =
222
β1 : 3µˆ +τˆ1 +τˆ3 +τˆ4 +3βˆ1 =
221
β2 : 3µˆ +τˆ1 +τˆ2 +τˆ3 +3βˆ2 =
224
β3 : 3µˆ +τˆ2 +τˆ3 +τˆ4 +3βˆ3 =
207
β4 : 3µˆ +τˆ1 +τˆ2 +τˆ4 +3βˆ4 =
218
Applying the constraints ∑ τi = ∑ β j = 0 , we obtain:
µ = 870 / 12 , τ1 = −9 / 8 , τ 2 = −7 / 8 , τ 3 = −4 / 8 , τ 4 = 20 / 8 ,
4-46
β1 = 7 / 8 , βö2 = 24 / 8 , βö3 = −31/ 8 , β4 = 0 / 8

4 4
R (µ ,τ , β ) = µöy.. + ∑τöi yi. + ∑ βöj y. j = 63,152.75
i =1 j =1
∑ ∑ y ij2 = 63,156.00
SS E = ∑ ∑ y ij2 − R ( µ , τ , β ) = 63156.00 − 63152.75 = 3.25 .
To test H o : τ i = 0 the reduced model is yij = µ + β j + ε ij . The normal equations used are:
µ : 12µö +3βö1 +3βö2 +3βö3 +3βö4 = 870

β1 : 3µö +3βö1 = 221
β2 : 3µö +3βö2 = 224
β3 : 3µö +3βö3 = 207
β4 : 3µö +3βö4 = 218
870 ö 7 ö 13 ö −21 ö 1
µö = , β1 = , β 2 = , β3 = , β4 =
12 6 6 6 6
4
R(µ , β ) = µöy .. + ∑ βö
j =1
j y . j = 63,130.00
R (τ µ , β )= R(µ ,τ , β )− R(µ , β ) = 63152.75 − 63130.00 = 22.75 = SS Treatments
( )
with 7 – 4 = 3 degrees of freedom. R τ µ , β is used to test H o : τ i = 0 .
The sum of squares for blocks is found from the reduced model y ij = µ + τ i + ε ij . The normal equations
used are:
µ : 12µö +3τö1 +3τö2 +3τö3 +3τö4 = 870

τ 1 : 3µö +3τö1 = 218
τ 2 : 3µö +3τö2 = 214
τ 3 : 3µö +3τö3 = 216
τ 4 : 3µö +3τö4 = 222
The sum of squares for blocks is found as in Example 4.5. We may use the method shown above to find an
adjusted sum of squares for blocks from the reduced model, y ij = µ + τ i + ε ij .
4-47
∑
a
k Qi2
i =1
4.44. Prove that is the adjusted sum of squares for treatments in a BIBD.
(λa )
We may use the general regression significance test to derive the computational formula for the adjusted
treatment sum of squares. We will need the following:
∑n y
kQi
τöi = , kQ = kyi . −
(λa ) i i =1
ij . j
a b
R(µ ,τ , β ) = µöy .. + ∑
i =1
τöi y i . + ∑ βö
j =1
j y. j
and the sum of squares we need is:
a b b y .2j
R (τ µ ,β )= µöy .. + ∑ τöi y i . + ∑ βö j y . j − ∑
i =1 j =1 j =1
k
The normal equation for β is, from equation (4.35),
a
β : kµö + ∑ n τö + kβö
i =1
ij i j = y. j
and from this we have:
a
ky. j βö j = y.2j − ky. j µö − y. j ∑ n τö
i =1
ij i
therefore,
 a 
a b

 y . j kµöy . j
2
y . j n ij τ
öi 2 
y. j  ∑
R (τ µ ,β )= µöy .. + ∑τöi y i . + ∑  k − − i =1
−
i =1 j =1 
k k k 
 
 
a
 1 a  a  kQ  a
 Qi2 
R(τ µ , β ) = ∑τöi  yi. − ∑ nij y. j  = ∑ Qi  i  = k ∑   ≡ SSTreatments ( adjusted )
i =1  k i =1  i =1  λ a  i =1  λ a 
4-48
4.45. An experimenter wishes to compare four treatments in blocks of two runs. Find a BIBD for this
experiment with six blocks.
Treatment Block 1 Block 2 Block 3 Block 4 Block 5 Block 6

1 X X X
2 X X X
3 X X X
4 X X X
Note that the design is formed by taking all combinations of the 4 treatments 2 at a time. The parameters of
the design are λ = 1, a = 4, b = 6, k = 3, and r = 2
4.46. An experimenter wishes to compare eight treatments in blocks of four runs. Find a BIBD with 14
blocks and λ = 3.
The design has parameters a = 8, b = 14, λ = 3, r = 2 and k = 4. It may be generated from a 23 factorial
design confounded in two blocks of four observations each, with each main effect and interaction
successively confounded (7 replications) forming the 14 blocks. The design is discussed by John (1971,
pg. 222) and Cochran and Cox (1957, pg. 473). The design follows:
Blocks 1=(I) 2=a 3=b 4=ab 5=c 6=ac 7=bc 8=abc

1 X X X X
2 X X X X
3 X X X X
4 X X X X
5 X X X X
6 X X X X
7 X X X X
8 X X X X
9 X X X X
10 X X X X
11 X X X X
12 X X X X
13 X X X X
14 X X X X
4.47. Perform the interblock analysis for the design in Problem 4.40.
The interblock analysis for Problem 4.33 uses σö2 = 0.91 and σöβ2 = 2.63 . A summary of the interblock,
intrablock and combined estimates is:
Parameter Intrablock Interblock

τ1 2.20 -1.80
τ2 0.73 0.20
τ3 -0.20 -5.80
τ4 -0.93 9.20
τ5 -1.80 -1.80
4-49
4.48. Perform the interblock analysis for the design in Problem 4.42.
The interblock analysis for Pproblem 4.42 uses σö 2 = 21.07 and
 MS Blocks ( adj ) − MS E  (b − 1) [65.68 − 21.07 ](6 )

σ β2 =  = = 19.12 .
a (r − 1) 7 (2 )
A summary of the interblock, intrablock, and combined estimates is give below
Parameter Intrablock Interblock Combined

τ1 -12.43 -11.79 -12.38
τ2 -8.57 -4.29 -7.92
τ3 2.57 -8.79 1.76
τ4 10.71 9.21 10.61
τ5 13.71 21.21 14.67
τ6 -5.14 -22.29 -6.36
τ7 -0.86 10.71 -0.03
4.49. Verify that a BIBD with the parameters a = 8, r = 8, k = 4, and b = 16 does not exist.
r ( k − 1) 8(3) 24
These conditions imply that λ = = = , which is not an integer, so a balanced design with
a −1 7 7
these parameters cannot exist.
k ( (a − 1))σ 2
4.50. Show that the variance of the intra block estimators { τ
i } is
(λa ) 2
.
b b  b 
∑ ∑ ∑
kQi
nij y. j = (k − 1)yi . −  nij y. j − yi . 
1
Note that τˆ i = , and Qi = y i . − n ij y . j , and kQi = kyi . −
(λa ) k j =1 j =1

 j =1


y i. contains r observations, and the quantity in the parenthesis is the sum of r(k-1) observations, not
including treatment i. Therefore,
V (kQi ) = k 2V (Qi ) = r (k − 1)2σ 2 + r (k − 1)σ 2

or
[r (k − 1)σ {(k − 1)+ 1}]= r (k − 1)σ

2
V (Qi ) =
1 2
k2 k
To find V (τöi ), note that:
 k  r (k − 1) 2 kr (k − 1) 2
2 2
 k 
V (τöi ) =   V (Q )i =   σ = σ
λ
 a  λa  k (λa )2
However, since λ (a − 1) = r (k − 1), we have:
4-50
k (a − 1)
V (τöi ) = σ2
λa 2
(
Furthermore, the {τöi } are not independent, this is required to show that V τöi − τö j = ) 2k 2
λa
σ
4.51. Extended incomplete block designs. Occasionally the block size obeys the relationship a < k < 2a.
An extended incomplete block design consists of a single replicate or each treatment in each block along
with an incomplete block design with k* = k-a. In the balanced case, the incomplete block design will have
parameters k* = k-a, r* = r-b, and λ*. Write out the statistical analysis. (Hint: In the extended incomplete
block design, we have λ = 2r-b+λ*.)
As an example of an extended incomplete block design, suppose we have a=5 treatments, b=5 blocks and
k=9. A design could be found by running all five treatments in each block, plus a block from the balanced
incomplete block design with k* = k-a=9-5=4 and λ*=3. The design is:
Block Complete Treatment Incomplete Treatment

1 1,2,3,4,5 2,3,4,5
2 1,2,3,4,5 1,2,4,5
3 1,2,3,4,5 1,3,4,5
4 1,2,3,4,5 1,2,3,4
5 1,2,3,4,5 1,2,3,5
Note that r=9, since the augmenting incomplete block design has r*=4, and r= r* + b = 4+5=9, and λ = 2r-
b+λ*=18-5+3=16. Since some treatments are repeated in each block it is possible to compute an error sum
of squares between repeat observations. The difference between this and the residual sum of squares is due
to interaction. The analysis of variance table is shown below:
Source SS DF
Qi2
∑ aλ
Treatments
k a-1
(adjusted)
y .2j
y ..2
Blocks
∑ k N
− b-1
Interaction Subtraction (a-1)(b-1)

Error [SS between repeat observations] b(k-a)
y2
Total ∑∑ y ij2 − ..
N
N-1
4.52. Suppose that a single-factor experiment with five levels of the factor has been conducted. There are
three replicates and the experiment has been conducted as a complete randomized design. If the
experiment had been conducted in blocks, the pure error degrees of freedom would be reduced by (choose
the correct answer):
(c) 2
4-51
4.53. Physics graduate student Laura Van Ertia has conducted a complete randomized design with a single
factor, hoping to solve the mystery of the unified theory and complete her dissertation. The results of this
experiment are summarized in the following ANOVA display:
Source DF SS MS F
Factor - - 14.18 -
Error - 37.75 -
Total 23 108.63
The completed ANOVA is as follows:
Source DF SS MS F P
Factor 5 70.88 14.18 6.76 0.00104

Error 18 37.75 2.10
Total 23 108.63
Answer the following questions about this experiment.
(a) The sum of squares for the factor is 70.88.
(b) The number of degrees of freedom for the single factor in the experiment is 5.
(c) The number of degrees of freedom for the error is 18.
(d) The mean square for error is 2.10.
(e) The value of the test statistic is 6.67.
(f) If the significance level is 0.05, your conclusions are not to reject the null hypothesis. No.
(g) An upper bound on the P-value for the test statistic is 0.001.
(h) A lower bound on the P-value for the test statistic is 0.0001.
(i) Laura used 6 levels of the factor in this experiment.
(j) Laura replicated this experiment 4 times.
(k) Suppose that Laura had actually conducted this experiment as a random complete block design and the
sum of squares for the blocks was 12. Reconstruct the ANOVA display above to reflect this new
situation. How much has the blocking reduced the estimate of the experimental error?
Source DF SS MS F P
Block 3 12.00 4.00

Factor 5 70.88 14.18 9.91 0.00011
Error 18 25.75 1.43
Total 23 108.63
The blocking reduced the SS error by 12 and the MS error by 0.67 (32%).
4-52
4.54. Consider the direct mail marketing experiment in Problem 4.8. suppose that this experiment has
been run as a complete randomized design, ignoring potential regional differences, but that exactly the
same data was obtained. Reanalyze the experiment under this new assumption. What difference would
ignoring the blocking have on the results and conclusions?
The solution for Problem 4.8 used a square root transformation, so the solution below also includes this
same transformation. The results below are similar to Problem 4.8 in that the the difference in designs is
statistically significant; however, the F value changed from 60.46 to only 7.02. The corresponding P value
increased from 0.0001 to 0.0145.
Response: Number of responses Transform: Square root Constant: 0

Sum of Mean F
A-Design 60.73 2 30.37 7.02 0.0145
Pure Error 38.90 9 4.32
Cor Total 99.64 11

C.V. % 11.23 Pred R-Squared 0.3058

Estimated Standard
Mean Error
1-1 17.17 1.04
2-2 21.69 1.04
3-3 16.69 1.04

1 vs 2 -4.52 1 1.47 -3.07 0.0133
1 vs 3 0.48 1 1.47 0.33 0.7525
2 vs 3 4.99 1 1.47 3.40 0.0079
4-53

CH 04

Uploaded by

Copyright:

Available Formats

CH 04

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

CH 04

Uploaded by

Copyright:

Available Formats

Solutions from Montgomery, D. C.

(2008) Design and Analysis of Experiments, Wiley, NY

Completed table is:

(b) How many blocks were used in this experiment?

Six blocks were used.

(c) What conclusions can you draw?

The modified ANOVA is shown below:

Design Expert Output

Std. Dev. 1.35 R-Squared 0.3727

Treatment Means (Adjusted, If Necessary)

Mean Standard t for H0

There is no difference among the chemical types at α = 0.05 level.

Design Expert Output

Std. Dev. 2.94 R-Squared 0.9314

Treatment Means (Adjusted, If Necessary)

Mean Standard t for H0

(1) (2) (3,4)

70.0 71.0 72.0 73.0

(3) (1) (2)

(a) Analyize the data from this experiment.

There is a difference between the means of the four tips.

Design Expert Output

Std. Dev. 0.094 R-Squared 0.8280

Treatment Means (Adjusted, If Necessary)

Mean Standard t for H0

One Factor Plot

(c) Analyze the residuals from this experiment.

Normal Plot of Residuals Residuals vs. Predicted

Residuals vs. Tip

(a) Analyze the data from this experiment.

Design Expert Output

Std. Dev. 30.08 R-Squared 0.9436

Treatment Means (Adjusted, If Necessary)

Mean Standard t for H0

Design Expert Output for Transformed Data

Std. Dev. 0.71 R-Squared 0.9527

Treatment Means (Adjusted, If Necessary)

Mean Standard t for H0

One Factor Plot

(c) Analyze the residuals from this experiment.

Normal Plot of Residuals Residuals vs. Predicted

-41.75 -22.1667 -2.58333 17 36.5833 199.75 285.88 372.00 458.13 544.25

Residuals vs. Design

Normal Plot of Residuals Residuals vs. Predicted

Residuals vs. Design

(a) Analyize the data from this experiment.

Design Expert Output

Std. Dev. 0.023 R-Squared 0.6136

Treatment Means (Adjusted, If Necessary)

Mean Standard t for H0

One Factor Plot

(c) Analyze the residuals from this experiment.

Normal Plot of Residuals Residuals vs. Predicted

Residuals vs. Oil

Jet Efflux Velocity (m/s)

Design Expert Output

Std. Dev. 0.054 R-Squared 0.6407

Treatment Means (Adjusted, If Necessary)

Mean Standard t for H0