ClassWork 04 Experiments with Blocking Factors

Exercise 1 [M]

Three different washing solutions are being compared to study their effectiveness in retarding
bacteria growth in 5-gallon milk containers. The analysis is done in laboratory, and only three trials
can be run on a day. The experimenter suspects that days could represent a potential source of
variability. Observations are taken for four days, and the data are shown here.

Day
Solution 1 2 3 4
1 13 24 18 39
2 16 24 17 44
3 2 4 1 22

a) Does the washing solution affect the retarding bacteria growth? (α=0.05)
b) Which washing solutions are different with respect to the retarding bacteria growth? Use the
method of Tukey.

Solution

a) Does the washing solution affect the retarding bacteria growth? (α=0.05)
Main Effects Plot for Growth Interaction Plot for Growth
Data Means Data Means
Days Solution 50 Days
1
35
2
40 3
4
30

30
25
Mean
Mean

20 20

15 10

10
0
1 2 3
1 2 3 4 1 2 3 Solution

The factor washing solution seems to influence the response variable; also the block factor seems to
be relevant. Instead, the interaction lines of factor and block appear parallel; this allows us to not
suspect any interaction. Remember that this kind of design assumes that the interaction between the
factor and the block in absent. Vice versa, the estimated variance will be greater and the experiment
could become inefficient to discover the significance of the factor.

General Linear Model: Growth versus Days; Solution

Factor Information
Factor Type Levels Values
Days Fixed 4 1; 2; 3; 4
Solution Fixed 3 1; 2; 3

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Days 3 1147,33 382,444 65,25 0,000
Solution 2 788,17 394,083 67,24 0,000
Error 6 35,17 5,861
Total 11 1970,67

1
Model Summary
S R-sq R-sq(adj) R-sq(pred)
2,42097 98,22% 96,73% 92,86%

Before drawing the conclusion, the residual assumptions have to be checked.

Scatterplot of SRES1 vs FITS1 ; Days; Solution Probability Plot of SRES1
Normal
FITS1 Days
2 99
Mean -2,0261 6E-1 5
StDev 1 ,044
1
95 N 12
AD 0,354
90
0 P-Value 0,400
80
-1 70

Percent
60
SRES1

0 10 20 30 40 1 2 3 4 50
40
Solution
2 30
20
1
10
5
0

-1 1
-3 -2 -1 0 1 2 3
1 ,0 1 ,5 2,0 2,5 3,0 SRES1

All the residual values belong to the interval (-3;+3), thus no outliers appear. The hypothesis of
normality cannot be refused. From the analysis of the scatterplot, we can assume that the hypothesis
of homogenous variance is satisfied.
In conclusion, the factor washing solution affects the bacteria growth and we rightly add the block
factor to the analysis.

b) Which washing solutions are different with respect to the retarding bacteria growth? Use
the method of Tukey.

The exercise text suggests to use the Tkey method:

Grouping Information Using the Tukey Method and 95% Confidence
Solution N Mean Grouping
2 4 25,25 A
1 4 23,50 A
3 4 7,25 B
Means that do not share a letter are significantly different.
The washing solutions (1) and (2) are not statistically different.
Just to practice, the Tukey’s constant computation follows:
1 1 1 4.34
Tα = qα ( a, df E )= qα ( a, ( a − 1)( b − 1) )= q0.05 ( 3, 6 )= = 3.0688
2 2 2 2
MS E 5,86
is Tα 2
The critical value= 3.0688
= 2 5.2529
b 4
The matrix of the differences is:
Levels (1) (3)
(2) 1.75 18
(1) 16.25

(3) (1) (2)

2
Exercise 2 [M]

Consider the ratio control algorithm experiment described in the exercise 3.8(M). 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 follows:

Algorithm Observations
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. Does the choice of ratio control algorithm affect the
average cell voltage? (α=0.05)
b) Does the choice of ratio control algorithm affect the pot noise cell voltage? (α=0.05)
c) Which ratio control algorithm would you select if your objective is to reduce both the average
cell voltage and the pot noise?

Solution

a) Analyze the average cell voltage data. Does the choice of ratio control algorithm affect the
average cell voltage? (α=0.05)

Main Effects Plot for Average Interaction Plot for Average

Data Means Data Means
Algorithm Period 4,95 Algorithm
4,88 1
2
3
4
4,86 4,90
Mean

4,84 4,85
Mean

4,82
4,80

4,80
4,75

1 2 3 4 5 6
1 2 3 4 1 2 3 4 5 6 Period

The nuisance factor period might be significant. The interaction between the 2 factors seems relevant
but it cannot be evaluated in the model.

General Linear Model: Average versus Algorithm; Period

Factor Information
Factor Type Levels Values
Algorithm Fixed 4 1; 2; 3; 4
Period Fixed 6 1; 2; 3; 4; 5; 6

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Algorithm 3 0,002746 0,000915 0,19 0,901
Period 5 0,017437 0,003487 0,72 0,615
Error 15 0,072179 0,004812
Total 23 0,092363

3
Model Summary
S R-sq R-sq(adj) R-sq(pred)
0,0693682 21,85% 0,00% 0,00%

Before drawing the conclusions, let us check the residual assumptions.

Scatterplot of SRES1 vs FITS1 ; Algorithm; Period Probability Plot of SRES1
Normal
FITS1 Algorithm
2 99
Mean -2,631 23E-1 4
1 StDev 1 ,022
95 N 24
0 AD 0,31 8
90
P-Value 0,51 5
-1 80
70
-2

Percent
60
SRES1

4,80 4,85 4,90 1 2 3 4 50

Period 40
2 30
20
1
10
0
5

-1
1
-2 -3 -2 -1 0 1 2 3
2 4 6 SRES1

No outliers appear in the scatterplots, the hypothesis of normality cannot be refused. Observing the
scatterplots, the variance appears homogeneous. Thus, the residual assumptions are checked and the
ratio control algorithm does not affect the mean cell voltage (The interaction may have concealed the
significance of the factors).

b) Does the choice of ratio control algorithm affect the pot noise cell voltage? (α=0.05)

Main Effects Plot for std_dev Interaction Plot for std_dev

Data Means Data Means
Algorithm Period 0,16 Algorithm
0,12 1
2
0,14 3
0,11 4

0,10 0,12

0,09
0,10
Mean

0,08
Mean

0,08
0,07

0,06 0,06

0,05 0,04

0,04
0,02
0,03 1 2 3 4 5 6
1 2 3 4 1 2 3 4 5 6 Period

Both factors seem to influence the response and their interaction seems moderate.

General Linear Model: Dev.st versus Algorithm; Period

Factor Information
Factor Type Levels Values
Algorithm Fixed 4 1; 2; 3; 4
Period Fixed 6 1; 2; 3; 4; 5; 6

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Algorithm 3 0,026013 0,008671 50,76 0,000
Period 5 0,002721 0,000544 3,19 0,037
Error 15 0,002563 0,000171
Total 23 0,031296

Model Summary

4
 S R-sq R-sq(adj) R-sq(pred)
0,0130703 91,81% 87,45% 79,04%
Let us check the residual assumptions.
Scatterplot of SRES2 vs FITS2; Algorithm; Period Probability Plot of SRES2
Normal
FITS2 Algorithm
2 99
Mean -9,43690E-1 6
1 StDev 1 ,022
95 N 24
0 AD 0,454
90
P-Value 0,247

-1 80
70

Percent
-2 60
SRES2

0,00 0,05 0,1 0 0,1 5 1 2 3 4 50

40
Period
2 30
20
1
10
0 5

-1
1
-3 -2 -1 0 1 2 3
-2
2 4 6 SRES2

All the standardized residuals belong to the interval [-3;+3], thus there are not any outliers. Analyzing
the scatterplot, problems of variance inhomogeneity do not appear. The hypothesis of normality
cannot be refuse. In conclusion the ratio control algorithm does affect the pot noise.

c) Which ratio control algorithm would you select if your objective is to reduce both the average
cell voltage and the pot noise?

The ratio control algorithm does not affect the average cell voltage, thus any algorithm can be
selected. Concerning the pot noise, we are interested in the control algorithm:
a ( a − 1) 4 ( 4 − 1)
Bα t α
= ( df E=
) rA = = 6= df E 15
2 rA 2 2
=Bα t0.05/12
= (15) 3.03624

1 1 4.08
Tα
= qα ( a, df
= E) q0.05 ( 4,15
= ) = 2.88
2 2 2

Sα =( a − 1) Fα ( a − 1, df E ) =3F0.05 ( 3,15 ) =3*3.28738 =3.1404

MS E 0.000171
The critical value
= is: Tα 2 2.88
= 2 0.02174
b 6
Using MINITAB we get:
Grouping Information Using the Tukey Method and 95% Confidence
Algorithm N Mean Grouping
3 6 0,113333 A
1 6 0,046667 B
4 6 0,036667 B
2 6 0,031667 B
Means that do not share a letter are significantly different.

In order to reduce the pot noise, there is not difference among the algorithms 1, 4 and 2. We would suggest a
future experiment focused on these algorithms.

5
Exercise 3 [M]

An aluminum 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 a
nuisance variable. The process engineers suspect that stirring rate affects the grain size of the
product. Each furnace can be run at four different stirring rates. A RCDB is run for a particular
refiner, and the resulting grain size data is as follow.

Furnace
Stirring rate (rpm) 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 affects grain size? (use α=0.05)
b) What should the process engineers recommend concerning the choice of stirring rate and
furnace for this particular grain refine if small grain size is desirable?
c) Suppose that the observation for stirring rate 4 in furnace 3 is missing.

Solution

a) Is there any evidence that stirring rate affects grain size? (use α=0.05)
Main Effects Plot for Size Interaction Plot for Size
Data Means Data Means
Rate Furnace 18 Rate
14 5
16 10
13 15
14 20
12
12
11
10
Mean

10
Mean

8
9
6
8
4
7
2
6
0
5 1 2 3 4
5 10 15 20 1 2 3 4 Furnace

The stirring rate does not seem to be relevant; instead the furnace (factor block) seems to influence
the response. The interaction seems to be fairly relevant but we have to assume that it is absent.

General Linear Model: Size versus Rate; Furnace

Factor Information
Factor Type Levels Values
Rate Fixed 4 5; 10; 15; 20
Furnace Fixed 4 1; 2; 3; 4

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Rate 3 22,19 7,396 0,85 0,500
Furnace 3 165,19 55,063 6,35 0,013
Error 9 78,06 8,674
Total 15 265,44

6
Model Summary
S R-sq R-sq(adj) R-sq(pred)
2,94510 70,59% 50,99% 7,05%

Let us check the residual assumptions.

Scatterplot of SRES1 vs FITS1 ; Rate; Furnace Probability Plot of SRES1
FITS1 Rate
Normal
99
1 Mean 1 ,040834E-1 6
StDev 1 ,033
95 N 16
0 AD 0,374
90
P-Value 0,374
-1 80
70

Percent
-2 60
SRES1

5,0 7,5 1 0,0 1 2,5 1 5,0 5 10 15 20 50

Furnace 40
30
1 20

10
0
5

-1
1
-3 -2 -1 0 1 2 3
-2
1 2 3 4 SRES1

From the analysis of the scatterplot, no outliers appear and, moreover, the hypothesis of equal
variance cannot be refused. The hypothesis of normality cannot be rejected, in fact the p-value is
lower than 0.05.
In conclusion, the factor stirring rate does not affect the grain size, which instead is influenced by the
furnace (block factor).

b) What should the process engineers recommend concerning the choice of stirring rate and
furnace for this particular grain refine if small grain size is desirable?

The choice of the stirring rate is irrelevant for the grain size. We suggest using the stirring rate that
allows the manufacturer to decrease the production costs or the production time.
From an optimization point of view, we can compare the furnaces. The furnace (1) negatively
influences the grain size. We can suggest to analysis this furnace and to understand why it is not
working well as the other furnaces.

Grouping Information Using the Tukey Method and 95% Confidence

furnace N Mean Grouping
1 4 13,25 A
2 4 6,00 B
4 4 5,75 B
3 4 5,75 B
Means that do not share a letter are significantly different.

c) Suppose that the observation for stirring rate 4 in furnace 3 is missing.

If a data is missing, we have to add a * in the corresponding cell in Minitab and then we can analyze
the data as usual.

General Linear Model: Size versus Rate; Furnace

Factor Information
Factor Type Levels Values
Rate Fixed 4 5; 10; 15; 20
Furnace Fixed 4 1; 2; 3; 4

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Rate 3 37,94 12,648 1,94 0,202

7
 Furnace 3 146,94 48,981 7,50 0,010
Error 8 52,22 6,528
Total 14 242,00

Model Summary
S R-sq R-sq(adj) R-sq(pred)
2,55495 78,42% 62,24% 25,74%

The conclusions do not change. The Stirring rate is still insignificant and the factor furnace influences
the grain size.

8
Exercise 4

yi2 y2
Consider a p x p Latin square without replications. Prove that=
SS A ∑i p p2
− and obtain least

squares estimates of the model parameters.

Solution
The model is:
 i = 1, 2.. p

yijk = µ + τ i + β j + α k + ε ijk  j = 1, 2.. p
k = 1, 2.. p

∑(y − µ − τ i − β j − α k ) by solving the following system of linear
2
We must minimize ∑=
ε ijk2 ijk
i , j ,k i , j ,k

equations:
∂ ∑ ( yijk − µ − τ i − β j − α k )
2

i , j ,k
=0
∂µ
∂ ∑ ( yijk − µ − τ i − β j − α k )
2 ∑τ i
i =0
i , j ,k

∂τ i
= 0 considering the additional constraints ∑β j
j =0

...
∑α k =0
∂ ∑ ( yijk − µ − τ i − β j − α k )
2 k

i , j ,k
=0
∂α k

The normal equations are:

µˆ : y =µˆ p 2 + ∑τˆi p + ∑ βˆ j p + ∑ αˆ k p
i j k

τˆi : yi  = µˆ p + τˆi p + ∑ βˆ j + ∑ αˆ k
j k

βˆ j : y j  =µˆ p + ∑τˆi + βˆ j p + ∑ αˆ k
i k

αˆ k : y k =µˆ p + ∑τˆi + ∑ βˆ j + αˆ k p

i j

The solutions are:

y
=µˆ = y
p2
τˆi yi  − y
=
β
=ˆ y −y
j  j 

ˆ k y k − y
α=

The total sum of squares is:

9
yijk = µˆ + τˆi + βˆ j + αˆ k + eijk

∑( y )
2
ijk − µˆ= ∑δ τˆ + ∑ δ ijk βˆ j2 + ∑ δ ijkαˆ k2 + ∑ δ ijk eijk2
ijk i
2

ijk ijk ijk ijk ijk

where δ ijk = 1 if the treatment k appears in the blocks i and j.

and
 yi2 2   yi2 yi  y y2 
∑ δ ijkτˆ=i2 p∑ ( yi  − y )= p∑  ∑i  p 2 p p 2 + p 4=
2
SS=
A 2
− 2 y y
 i  + y =
  p − 2
ijk i i p   
2 2 2 2 2 2
y y yi  y yi  y y y y y
=∑ − 2 p 2
i 
∑ + p∑ =∑
4
− 2 p 2  + p 2 4 = ∑ i  − 2
i p p i p i p i p p p p i p p

10
 Exercise 5 [M]

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 this source of variability,
the engineer uses the Latin square design shown below.

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

a) Analyze the data from this experiment (use α=0.05) and calculate manually the residual of the
data (C,1,1);
b) Verify whether it is advantageous to conduct an experiment with two block factors.

Solution

a) Analyze the data from this experiment (use α=0.05);

Notice that the factor Operator should be considered as a Random Factor, if the operators were
randomly selected from a larger number. Here, we consider Operator as a fixed factor.

Let us do the analysis by hand.

y = 164
y A 30
= = yB  37
= yC  53
= yD  44
y1 39
= = y 2  44
= y3 35
= y 4  46
y1 32
= = y 2 52
= y3 41
= y 4 39
1 y2 1 y2
SS Methods
= ∑
p i
yi2 − =
p2
72.5 SS =
Operator ∑
p k
y 2
 k − =
p2
51.5

1 y2 y2
SSOrder
= ∑
p j
y2j  − 2= 18.5
p
SSTOT= ∑y
i , j ,k
2
ijk −
p2
= 153.0

The ANOVA table is:

Source SS df MS F0 p
Method 72.5 3 24.2 13.8 0.004
Operator 51.5 3 17.2 9.8 0.01
Order 18.5 3 6.2 3.5 0.089
Error 10.5 6 1.75
Total 153.0 15

11
The residual of the data (C,1,1) is:
eC11 = yC11 − yˆC11 where yˆC11 = µˆ + τˆC + βˆ1 + αˆ1
y 164
µˆ =2 = =
10.25 τˆC = 3 βˆ1 =
yC  − y = −0.5 αˆ1 =
y1 − y = −2.25
y1 − y =
p 16
yˆC11 =10.25 + 3 − 0.5 − 2.25 =10.5 → eC11 =yC11 − yˆC11 =−0.5

Let’s do the analysis with MINITAB

Main Effects Plot for Time Interaction Plot for Time
Data Means Data Means
Order Operator Methods 1 2 3 4 1 2 3 4
14
Order
15 1
13 2
Order 3
10
4
12
5

Operator
11
Mean

15 1
2
3
10 Operator 10
4

5
9

8
Methods
7
1 2 3 4 1 2 3 4 1 2 3 4

Method seems to be relevant, the interactions appear moderate but they can not be considered in the
model.

General Linear Model: Time versus Order; Operator; Methods

Factor Information
Factor Type Levels Values
Order Fixed 4 1; 2; 3; 4
Operator Fixed 4 1; 2; 3; 4
Methods Fixed 4 1; 2; 3; 4

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Order 3 18,50 6,167 3,52 0,089
Operator 3 51,50 17,167 9,81 0,010
Methods 3 72,50 24,167 13,81 0,004
Error 6 10,50 1,750
Total 15 153,00

Model Summary
S R-sq R-sq(adj) R-sq(pred)
1,32288 93,14% 82,84% 51,20%

Before drawing the conclusions, let us check the residual assumptions.

Scatterplot of SRES1 vs FITS1 ; Order; Operator; Methods Probability Plot of SRES1
Normal
FITS1 Order
2 99
Mean -7,78897E-1 6
1 StDev 1 ,033
95 N 16
AD 0,456
0 90
P-Value 0,232

-1 80
70
Percent

-2 60
SRES1

5 10 15 1 2 3 4 50
Operator Methods 40
2 30
20
1
10
0 5

-1
1
-3 -2 -1 0 1 2 3
-2
1 2 3 4 1 2 3 4 SRES1

12
All the standardized residuals belong to the interval [-3;+3], thus no outliers appear. Moreover, the
hypothesis of homogeneous variance is assumed from the analysis of the scatterplot. The hypothesis
of normality cannot be refused. Finally the assembly method does affect the assembly time. Moreover
the assembly order and the operator are nuisance factors that influence the assembly time, thus we
can suggest a deep study to better understand the differences among the operators and the assembly
orders.

b) Verify whether it is advantageous to conduct an experiment with two block factors.

In order to understand the advantageous, we can do the analysis without the block factors.

General Linear Model: Time versus Methods

Factor Information
Factor Type Levels Values
Methods Fixed 4 1; 2; 3; 4

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Methods 3 72,50 24,167 3,60 0,046
Error 12 80,50 6,708
Total 15 153,00

Model Summary
S R-sq R-sq(adj) R-sq(pred)
2,59005 47,39% 34,23% 6,46%

Without using the block factors, we can conclude with some difficulty (the p-value is almost 5%) that
the factor method is significant.

13
Exercise 6 [September 21st 2005]

A restaurateur was studying the performance of five different brands of oil. A tester, after testing
each brand, had to assign a score for each one. The tester could taste only three brands per day. An
experiment was conducted and the data are shown below:

Day
Brand 1 2 3 4 5 6 7 8 9 10
1 32.4 37.3 38.2 36.4 41.4 40.8
2 30.7 31.1 34.4 33.2 37.0 41.2
3 32.3 29.4 32.1 34.8 40.7 46.6
4 32.7 34.1 31.3 41.5 43.3 43.3
5 36.0 34.1 41.4 35.7 45.6 44.3

a) Which type of experiment was performed?

b) Analyze the data with α=5%.

Solution

a) Which type of experiment was performed?

The experiment was conducted using a Randomized Incomplete Block Design (RIBD). This design
is used when a nuisance factor exists and we may not be able to run all the treatment combinations in
each block. In our case, the design is not symmetric.

b) Analyze the data with α=5%.

We start by plotting the ME and the Interaction plot.

Main Effects Plot for performance Interaction Plot for performance
Data Means Data Means
brand day 48 brand
46 1
46 2
44 3
44
4
5
42 42

40
40
Mean
Mean

38
38
36
36
34

34 32

32 30

30 1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 1 2 3 4 5 6 7 8 9 10 day

The block factor seems to affect the response and to be more relevant than the factor brand. The
interaction does not appear significant.

General Linear Model: performance versus brand; day

Factor Information
Factor Type Levels Values
brand Fixed 5 1; 2; 3; 4; 5
day Fixed 10 1; 2; 3; 4; 5; 6; 7; 8; 9; 10

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
brand 4 69,10 17,275 3,35 0,036

14
 day 9 533,83 59,314 11,49 0,000
Error 16 82,56 5,160
Total 29 701,11

Model Summary
S R-sq R-sq(adj) R-sq(pred)
2,27163 88,22% 78,66% 58,60%

Before drawing the conclusions, let us check the residual assumptions.

Scatterplot of SRES1 vs FITS1 ; brand; day Probability Plot of SRES1
FITS1 brand
Normal
2 99
Mean 1 ,582068E-1 5
1 StDev 1 ,01 7
95 N 30
0 AD 0,390
90
P-Value 0,361

-1 80
70

Percent
-2 60
SRES1

30 35 40 45 1 2 3 4 5 50
day 40
2 30
20
1
10
0 5

-1
1
-3 -2 -1 0 1 2 3
-2
0,0 2,5 5,0 7,5 1 0,0 SRES1

No outliers appear in the scatterplots and the hypothesis of normality cannot be refuse.
We conclude that both factors are significant.

15
Exercise 7 [M]

The factors that influence the breaking strength of synthetic fiber are being studied. Four
production machines and three operators are chosen to run a factorial experiment. Two batches of
fiber are chosen to perform two replications. The results are as follows:

Machine
Operator 1 2 3 4
109 110 108 110
1
110 115 109 108
110 110 111 114
2
112 111 109 112
116 112 114 120
3
114 115 119 117
In blue the first batch.
Analyze the data (use α=0.05),

Solution

Analyze the data (use α=0.05)

Operator is a Random Factor (“three operators are chosen”) but “Mixed effect” problems (a mix of
random and fixed factors) are not tackled in this class. Thus, the factor operator is considered as fixed.

The batch is clearly a nuisance factor. Then, the analysis is done considering two factors of interest
(Operator and Machine) and one blocking factor (Batch). Let us notice that for each level of the
blocking factor, a complete plan is done. The model for the analysis will be complete for the factors
operator and machine and additive for the factor batch.
Main Effects Plot for Strength Interaction Plot for Strength
Data Means Data Means
Operator Machine Batch 1 2 3 4 1 2
120
116 Operator
1
115 2
115 3
Operator

110
114
120
Machine
Mean

1
113 2
115
Machine 3
4
112
110

111

110 Batch

1 2 3 1 2 3 4 1 2

The operator seems to be relevant. It affects the breaking strength more than the machine and batch
factors. The interaction between operator and machine seems to be fairly important. The interaction
between machine and batch is suspected but it cannot be verified with this kind of model.

General Linear Model: Strength versus Operator; Machine; Batch

Factor Information
Factor Type Levels Values
Operator Fixed 3 1; 2; 3
Machine Fixed 4 1; 2; 3; 4
Batch Fixed 2 1; 2

Analysis of Variance

16
Source DF Adj SS Adj MS F-Value P-Value
Operator 2 160,333 80,167 20,29 0,000
Machine 3 12,458 4,153 1,05 0,409
Batch 1 2,042 2,042 0,52 0,487
Operator*Machine 6 44,667 7,444 1,88 0,172
Error 11 43,458 3,951
Total 23 262,958

Model Summary
S R-sq R-sq(adj) R-sq(pred)
1,98765 83,47% 65,44% 21,33%

Let’s check the residual assumptions.

Scatterplot of SRES1 vs FITS1 ; Operator; Machine; Batch Probability Plot of SRES1
Normal
FITS1 Operator
2 99
Mean -1 ,05656E-1 4
1 StDev 1 ,022
95 N 24
AD 0,492
0 90
P-Value 0,1 98

-1 80
Percent 70
-2 60
SRES1

1 1 0,0 1 1 2,5 1 1 5,0 1 1 7,5 1 20,01 ,0 1 ,5 2,0 2,5 3,0 50

40
Machine Batch
2 30
20
1
10
0 5

-1
1
-3 -2 -1 0 1 2 3
-2
1 2 3 4 1 ,00 1 ,25 1 ,50 1 ,75 2,00 SRES1

From the analysis of the scatterplot, no outliers appear and the hypothesis of homogeneous variance
cannot be refused. Also the hypothesis of normality cannot be rejected. Finally, the factor operator is
the only significant. Instead, the nuisance factor does not affect the breaking strength.

17
Exercise 8 [September 20th 2012]

The effect of five different ingredients (A, B, C, D, E) on the 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 hour and half, so only five runs can be made in one
day. The data of the experiment follows.

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)

a) Which kind of design has been used?

b) Analyze the data (α=5%);
c) What should the experiment recommend if we want to reduce the reaction time? (Use Tukey)

Solution

a) Which kind of design has been used?

The design used is the Latin Square Design that has been used to eliminate two nuisance sources of
variability (Batch and Day) and to focus the attention on the interest factor (Ingredient).

b) Analyze the data (α=5%)

Main Effects Plot for Time Interaction Plot for Time

Data Means Data Means
Batch Day Ingredient 1 2 3 4 5 1 2 3 4 5
9
10 Batch
1
2
8 3
Batch 5
4
5
7 0
10 Day
Mean

1
6 2
3
Day 5
4
5
5
0

Ingredient
3
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

The factor Ingredient seems to be relevant and affects the reaction time more than the nuisance
factors. The interactions appear moderate but they cannot be tested with this model.

General Linear Model: Time versus Batch; Day; Ingredient

Factor Information
Factor Type Levels Values
Batch Fixed 5 1; 2; 3; 4; 5
Day Fixed 5 1; 2; 3; 4; 5
Ingredient Fixed 5 1; 2; 3; 4; 5

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Batch 4 15,44 3,860 1,23 0,348
Day 4 12,24 3,060 0,98 0,455

18
 Ingredient 4 141,44 35,360 11,31 0,000
Error 12 37,52 3,127
Total 24 206,64

Model Summary
S R-sq R-sq(adj) R-sq(pred)
1,76824 81,84% 63,69% 21,19%

Before drawing the conclusions, we have to check the residual assumptions.

Scatterplot of SRES1 vs FITS1 ; Batch; Day; Ingredient Probability Plot of SRES1
Normal
FITS1 Batch
2
99
Mean -6,21 725E-1 7
1 StDev 1 ,021
95 N 25
0 AD 0,347
90
P-Value 0,452

-1 80
70

Percent
-2 60
SRES1

2 4 6 8 10 1 2 3 4 5 50
Day Ingredient 40
2 30
20
1
10
0 5

-1
1
-3 -2 -1 0 1 2 3
-2
1 2 3 4 5 1 2 3 4 5 SRES1

From the analysis of the scatterplots, no outliers appear and the hypothesis of homogeneous variance
cannot be refused. The assumption of normality cannot be rejected with a p-value equal to 0.452. The
residual assumptions are verified and, in conclusion, the factor ingredient is significant whereas the
nuisance factors do not affect the reaction time.

c) What should the experiment recommend if we want to reduce the reaction time? (Use Tukey)

Grouping Information Using the Tukey Method and 95% Confidence

Ingredient N Mean Grouping
3 5 8,8 A
1 5 8,4 A
2 5 5,6 A B
4 5 3,4 B
5 5 3,2 B
Means that do not share a letter are significantly different.

In order to reduce the reaction time, there is not difference among the ingredients B, D, E.
A further experiment focused on these ingredients is suggested.

19
Exercise 9 [February 1st 2012]

The following experiment was described by Kuehl (1994) and run by J. Berry and A. Deutschman at
the University of Arizona to obtain specific information about the effect of pressure on percent
conversion of methyl glucoside to monovinyl isomers. The conversion is achieved by addition of
acetylene to methyl glucoside in the presence of a base under high pressure. Five pressures were
examined in the experiment, but only three could be examined at any one time under identical
experimental conditions. The data of the experiment follows.

Experimental condition
Pressure I II III IV V VI VII VIII IX X
250 16 19 20 13 21 24
325 18 26 19 13 10 24
400 39 21 33 34 30 31
475 32 46 35 47 31 37
550 45 61 55 48 52 50

a) Which kind of design has been used?

b) Analyze the data with α=5%.

Solution

a) Which kind of design has been used?

The experiment was conducted using a Randomized Incomplete Block Design (RIBD). This design
is used when a nuisance factor exists and we may not be able to run all the treatment combinations in
each block. In our case, the design is not symmetric.

b) Analyze the data with α=5%.

Main Effects Plot for Percent Interaction Plot for Percent

Data Means Data Means
Pressure Block Pressure
55 60 250
325
50 400
50 475
550
45

40
40
Mean
Mean

35 30

30
20
25

10
20
1 2 3 4 5 6 7 8 9 10
250 325 400 475 550 1 2 3 4 5 6 7 8 9 10 Block

The factor pressure seems to affect the percent of conversion. The interaction appears to be moderate
but it cannot be considered in this kind of model.

General Linear Model: Percent versus Pressure; Block

Method
Factor coding (-1; 0; +1)
Factor Information

20
 Factor Type Levels Values
Pressure Fixed 5 250; 325; 400; 475; 550
Block Fixed 10 1; 2; 3; 4; 5; 6; 7; 8; 9; 10
Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Pressure 4 3688,6 922,14 29,90 0,000
Block 9 346,9 38,55 1,25 0,334
Error 16 493,4 30,84
Total 29 5576,7
Model Summary
S R-sq R-sq(adj) R-sq(pred)
5,55328 91,15% 83,96% 68,89%

Scatterplot of SRES1 vs FITS1 ; Pressure; Block Probability Plot of SRES1

Normal
FITS1 Pressure
2 99
Mean -5,07002E-1 6
1 StDev 1 ,01 7
95 N 30
0 AD 0,469
90
P-Value 0,231

-1 80
70
Percent

-2 60
SRES1

20 30 40 50 60 300 400 500 50

Block 40
2 30
20
1
10
0 5

-1
1
-3 -2 -1 0 1 2 3
-2
0,0 2,5 5,0 7,5 1 0,0 SRES1

All the standardized residuals belong to the interval [-3;+3], thus no outliers appear. Observing the
scatterplot, no problems of inhomogeneous variance can be detected. The hypothesis of normality
cannot be refused. We conclude that the Blocking factor does not have an influence on the response,
while pressure has.

21