5.

Analysis of Variance
Outline
5.1 Introduction to Experimental Design
5.2 One-Way Analysis of Variance
5.3 Two-Way Analysis of Variance
5.4 Two-Way Analysis of Variance with Interaction

Objectives
At the end of this chapter, the students should be able to
1. Understand different experimental designs.
2. Apply One-way ANOVA technique to determine whether there is a
significant difference among 3 or more means.
3. Apply Two-way ANOVA technique to determine whether there is a
significant difference in the main effects or whether a possible
interaction effect exists.

5.1 Introduction to Experimental Design

Many studies involve data collection from experiment. Before data collection can be
done, it is wise to organize and design the experiment properly so that the right type of data
is collected for the desired statistical analysis. This process is called experimental design.
Different experimental designs enable us to study different effects of the response in the
study.

Three types of the experimental designs are presented here, that are Completely
Randomized Design, Randomized Block Design and Completely Randomized Factorial
Design.

(a) Completely Randomized Design

Completely Randomized Design is the simplest experimental designs, in which

random samples are selected independently from each of k populations. This design
involves only one factor. The following Case 1 is an example of Completely Randomized
Design.

Case 1:

For example, a researcher designs three different diets for athletes, which are Diet I,
Diet II and Diet III. The researcher is interested to know whether there is any significant
effect on the weight gain of athletes following these 3 different diets. In another word, the
researcher would like to see whether there is any difference in the weight gains of athletes
who follow different diets. Fourteen athletes are randomly assigned to three groups and
follow the assigned diet for 1 month.

The randomized assignment in this case can be done by assigning a number to each
of these 14 athletes, say number 1 to 14. Using a randomized device, random numbers can
be generated and the athletes with the respective random numbers are placed in three
groups. For example, when first 5 random numbers are generated, the athletes who are
associated with these 5 numbers are assigned to Diet I. Then, another 5 random numbers
are generated and the athletes who are associated with these numbers are assigned to Diet
II. The rest of the athletes are assigned to Diet III.

1
 Note that the athletes are the experimental units. The weight gains of athletes (in kg)
after following the diets for one month are recorded in the following table.

Diet
Diet I Diet II Diet III
7 3 1
8 6 2
12 3 1
6 2 3
8 5

In this case, the researcher is interested to know whether there is any effect on the
weight gain of athletes following different diets. Thus, there is only 1 factor that affects the
weight gain in this study, which is diet. The factor (diet) is also called independent variable
while the weight gain is the dependent variable.

The factor in this case consists of 3 levels, or, 3 different treatments. Treatment is
something the researcher administers to the experimental units. In this case, the
experimental units (athletes) are ‘treated’ with 3 different diets to see the effect on their
weight gains.

The hypotheses for this study would be as follows:

Ho: μI = μII = μIII

(There is no difference in mean weight gain for different diets)
H1: At least one mean is different from the others
(There is a difference in mean weight gain for different diets)

Alternatively, we may also write:

Ho: No effect on mean weight gain due to different diets

H1: There is an effect on mean weight gain due to different diets

Analysis of Variance (ANOVA) is a procedure used to test the null hypothesis that the
means of three or more populations are equal. In this case, only 1 factor (that is 1
independent variable) involves, thus One-way ANOVA is used to solve this hypothesis test.

2
(b) Randomized Block Design

Sometimes, undesired sources of variation exist in an experiment and affect the

dependent variable measurement. The undesired sources of variation come from the
nuisance variable. Randomized Block Design is a type of experimental design that is used to
isolate the effect of nuisance variable by blocking procedure. Blocking procedure is done by
grouping the homogeneous experimental units into same block. After grouping into blocks,
the members of each block are then randomly assigned to different treatment groups. The
following Case 2 is an example of Randomized Block Design.

Case 2:

For example, the objective of a study is to investigate whether there is a significant

difference in the English language test results by using 4 different teaching methods.

However, the background of the students (the undergraduate course that they are
pursuing) might affect the English language test results. Thus, the background of the
students is a nuisance variable. In order to isolate the effect of this nuisance variable,
blocking procedure is done by grouping the students with same background into same
block. Then, the students in each block are randomly assigned to different teaching
methods on learning English language. After three months, these students sit for an English
language test. The test results are presented in the following table.

Student background
(Blocking factor)
Engineering Science Pharmacy
Teaching method Method A 13 12 13
(Treatment factor) Method B 10 15 11
Method C 14 11 12
Method D 11 14 14

The factor, teaching method, is known as treatment factor while the student
background is blocking factor. In this study, blocking procedure is done in order to remove
the block-to-block variability that might hide the effect of the treatment factor. Note that the
factors are also called independent variables while the English language test result is the
dependent variable.

The hypotheses for this study would be as follows:

For teaching method,

Ho: μA = μB = μC = μD
(There is no difference in mean result using different teaching methods)
H1: At least one mean is different from the others
(There is a difference in mean result using different teaching methods)

For student background,

Ho: μE = μS = μP
(There is no difference in mean result for different student background)
H1: At least one mean is different from the others
(There is a difference in mean result for different student background)

3
 Alternatively, we may write the hypotheses as follows:

For teaching method,

Ho: No effect on mean result due to different teaching methods

H1: There is an effect on mean result due to different teaching methods

For student background,

Ho: No effect on mean result due to different student background

H1: There is an effect on mean result due to different student background

As mentioned previously, ANOVA is a procedure that we use to test the null

hypothesis that the means of three or more populations are equal. When only 1 factor
involves in the study, One-way ANOVA is used to solve the hypothesis test. In this case,
there are 2 factors, thus Two-way ANOVA is used to solve the hypothesis tests.

4
(c) Completely Randomized Factorial Design

In the case when the researcher is interested to investigate the effects of two factors
at different levels and to test the interaction effect between the factors, Completely
Randomized Factorial Design is the right experimental design to be carried out. The
following Case 3 is an example of Completely Randomized Factorial Design.

Case 3:

For example, a researcher wishes to test the effects of 2 different diets and 3
different exercise programs on the amount of weight loss. The researcher is also interested
to detect whether interaction exists between these two factors. To conduct this experiment, 6
treatment combinations have been formed. Two adults are randomly assigned to each
treatment combination for two months. The amount of weight loss for each of the adults (in
kg) is recorded after two months and it is shown in the following table.

Exercise program

A B C

Diet I 9, 8 5, 3 2, 2

II 6, 7 2, 2 1, 1

In this study, there are 2 factors, which are diet and exercise program. The diet
consists of 2 levels while the exercise program consists of 3 levels. The diet and the
exercise program are also called the independent variables while the weight loss is the
dependent variable.

The hypotheses for this study would be as follows:

For interaction,

Ho: There is no interaction effect between diet and exercise program

H1: There is an interaction effect between diet and exercise program

For diet,

Ho: μI = μII
(There is no difference in mean weight loss for different diets)
H1: At least one mean is different from the others
(There is a difference in mean weight loss for different diets)

For exercise program,

Ho: μA = μB = μC
(There is no difference in mean weight loss for different exercise programs)
H1: At least one mean is different from the others
(There is a difference in mean weight loss for different exercise programs)

5
 Alternatively, we may write the hypotheses as follows:

For interaction,

Ho: There is no interaction effect between diet and exercise program

H1: There is an interaction effect between diet and exercise program

For diet,

Ho: No effect on mean weight loss due to different diets

H1: There is an effect on mean weight loss due to different diets

For exercise program,

Ho: No effect on mean weight loss due to different exercise programs

H1: There is an effect on mean weight loss due to different exercise programs

Note that the first set of hypotheses concerns about the interaction effect while the
second and third sets of hypotheses concern about main effects. In this case, the procedure
of Two-way ANOVA with Interaction is used to solve the hypothesis tests.

It is noticed that hypothesis tests on main effects consider whether the null
hypothesis Ho: μI = μII or Ho: μA = μB = μC should be rejected or not. In other words,
hypothesis test on main effect is concerning whether there is any difference among several
mean levels of each independent variable.

In the case when interaction effect exists between two independent variables
(factors), the effect of one independent variable on the dependent variable depends on the
level of another independent variable.

6
5.2 One-Way Analysis of Variance

ANOVA (Analysis of Variance) is a procedure used to test the null hypothesis that the
means of three or more populations are equal.

Assumptions of one-way ANOVA:

1. The populations from which the samples are drawn are approximately normally
distributed.
2. The variances of the populations from which the samples are drawn are equal.
3. The samples drawn from different populations are random and independent.

One-Way ANOVA

Case 1 presented in section 5.1 is an example of Completely Randomized Design. It

involves 1 factor, thus One-way ANOVA is used to solve the hypothesis test. The procedure
of One-way ANOVA is presented here.

Diet
Diet I Diet II Diet III
7 3 1
8 6 2
12 3 1
6 2 3
8 5
Can we conclude that there are differences in mean weight gain among 3 different diets?
Use   0.01 .

Solution:
Step 1 Hypothesis:

Ho: μI = μII = μIII

(There is no difference in mean weight gain for different diets)
H1: At least one mean is different from the others
(There is a difference in mean weight gain for different diets)

Alternatively, we may write the hypotheses as follows:

Ho: No effect on mean weight gain due to different diets

H1: There is an effect on mean weight gain due to different diets

Step 2 Test statistic:

Diet
Diet I Diet II Diet III
7 3 1
8 6 2
12 3 1
6 2 3
8 5
ni n1 = 5 n2 = 5 n3 = 4 N = 14
 xi 41 19 7  xi = 67
x i
2
357 83 15  x i
2
= 455

7
 Total sample size = N = 14

  x   67 
2 2

CF    320.643
i

N 14

SST   x i2  CF  455  320.643  134.357

(41)2 (19)2 (7)2

SSB     CF
n1 n2 n3
(41)2 (19)2 (7)2
    320.643
5 5 4
 100.007

SSE  SST  SSB  34.35

ANOVA Table
Source df Sum of Squares Mean Square Fo
(SS) (MS)
Diet k–1 SSB = 100.007 SSB MSB
(Between =3–1 MSB   50.004 F0   16.01
k 1 MSE
groups) =2
Error N– k SSE = 34.35 SSE
(Within = 14 – 3 MSE   3.123
Nk
groups) = 11
Total N–1 SST = 134.357
= 14 – 1
= 13

Step 3 Rejection area:

Reject H0 if Fo  F0.01,2,11  7.21

Step 4 Decision:

Since Fo  16.01  7.21, reject Ho.

Step 5 Conclusion:

There is enough evidence to conclude that there are differences in mean weight gain
among 3 different diets.

8
5.3 Two-Way Analysis of Variance

Case 2 presented in section 5.1 is an example of Randomized Block Design. It

involves 2 factors, which are 1 treatment factor and 1 blocking factor. Hence, Two-way
ANOVA is used to solve the hypothesis tests. The procedure of Two-way ANOVA is
presented here.

Student background
(Blocking factor)
Engineering Science Pharmacy
Teaching method Method A 13 12 13
(Treatment factor) Method B 10 15 11
Method C 14 11 12
Method D 11 14 14

i) Do the data provide enough evidence to indicate that a difference exists in the
average result using different teaching methods? Use   0.05 .
ii) Is there enough evidence to indicate that mean result differs for different
student background? Test at 5% level of significance.

Solution:
Step 1 Hypothesis:

For teaching method,

Ho: μA = μB = μC = μD
(There is no difference in mean result using different teaching methods)
H1: At least one mean is different from the others
(There is a difference in mean result using different teaching methods)

For student background,

Ho: μE = μS = μP
(There is no difference in mean result for different student background)
H1: At least one mean is different from the others
(There is a difference in mean result for different student background)

Alternatively, we may write the hypotheses as follows:

For teaching method,

Ho: No effect on mean result due to different teaching methods

H1: There is an effect on mean result due to different teaching methods

For student background,

Ho: No effect on mean result due to different student background

H1: There is an effect on mean result due to different student background

9
Step 2 Test statistic:
MSR
F1  for teaching method (Row)
MSE

MSC
F2  for student background (Column)
MSE

Student background
(Blocking factor)
Engineering Science Pharmacy Total
Teaching Method A 13 12 13 38
method Method B 10 15 11 36
(Treatment Method C 14 11 12 37
factor) Method D 11 14 14 39
Total 48 52 50 150

 x = 13 + 12 +…+ 14 + 14 = 150

 x 2
= 132 + 122 +…+ 142 + 142 = 1902

Total sample size = rc = (4)(3)=12

( x)2
(150)2
CF    1875
rc 12

SST =  x 2  CF  1902  1875  27

For teaching method (Row):

T2
SSR=  i  CF
c
2
38 362 372 392
     CF
3 3 3 3
5630
  1875  1.667
3

For student background (Column):

T2
SSC =  i  CF
r
482 522 502
    CF
4 4 4
7508
  1875  2
4

SSE = SST – SSR – SSC = 27 – 1.667 – 2 = 23.333

10
 ANOVA Table
Sources of df Sum of Mean square F
variation squares
Teaching r-1=3 SSR=1.667 SSR MSR
MSR   0.556 F1   0.143
method r 1 MSE
(row)
Student c-1=2 SSC=2 SSC MSC
MSC  1 F2   0.257
background c 1 MSE
(column)
Error (r-1)(c-1) SSE=23.333 SSE
MSE   3.889
=6 r  1 c  1
Total rc-1 =11 SST=27

Step 3 Rejection area:

For teaching method: Reject Ho if F1  F,r 1,r 1c 1  F0.05,3,6  4.76
For student background: Reject Ho if F2  F,c1,r 1c1  F0.05,2,6  5.14

Step 4 Decision:

For teaching method: Since F1  4.76 , do not reject Ho

For student background: Since F2  5.14 , do not reject Ho

Step 5 Conclusion:

There is not enough evidence to indicate that a difference exists in the average result
using different teaching methods.

There is not enough evidence to indicate that mean result differs for different student
background.

11
5.4 Two-Way Analysis of Variance with Interaction

Case 3 presented in section 5.1 is an example of Completely Randomized Factorial

Design. The researcher is interested to investigate the effects of two factors at different
levels and to test the interaction effect between the factors. Thus, Two-way ANOVA with
Interaction would be used to solve the hypothesis tests. The procedure of Two-way ANOVA
with Interaction is presented here.

Exercise program
A B C
I 9, 8 5, 3 2, 2
Diet
II 6, 7 2, 2 1, 1

i) Is there any interaction effect between diet and exercise program? Use   0.05 .

ii) Do the data provide sufficient evidence to indicate that there is a significant effect
on mean weight loss due to different diets? Test at 5% level of significance.

iii) Do the data provide enough evidence to indicate that there is a significant effect
on mean weight loss due to different exercise programs? Use   0.05 .

Solution:

Step 1 Hypothesis:
For interaction,
Ho: There is no interaction effect between diet and exercise program
H1: There is an interaction effect between diet and exercise program

For diet,
Ho: μI = μII
(There is no difference in mean weight loss for different diets)
H1: At least one mean is different from the others
(There is a difference in mean weight loss for different diets)

For exercise program,

Ho: μA = μB = μC
(There is no difference in mean weight loss for different exercise programs)
H1: At least one mean is different from the others
(There is a difference in mean weight loss for different exercise programs)

Alternatively, we may write the hypotheses as follows:

For interaction,
Ho: There is no interaction effect between diet and exercise program
H1: There is an interaction effect between diet and exercise program

For diet,
Ho: No effect on mean weight loss due to different diets
H1: There is an effect on mean weight loss due to different diets

For exercise program,

Ho: No effect on mean weight loss due to different exercise programs
H1: There is an effect on mean weight loss due to different exercise programs

12
 MSR
Step 2 Test statistic: F1  for diet (Row),
MSE
MSC
F2  for exercise program (Column),
MSE
MSRC
F3  for interaction
MSE

Exercise program

A B C Total

Diet I 9, 8 5, 3 2, 2 29
(17) (8) (4)
II 6, 7 2, 2 1, 1 19
(13) (4) (2)
Total 30 12 6 48

 x = 9 + 8 + … + 1 + 1 = 48
 x 2
= 92 + 82 +…+ 12 + 12 = 282

Total sample size = rcn = (2)(3)(2)=12

( x)2
482
CF =   192
rcn 2(3)(2)

SST =  x 2  CF = 282 - 192 = 90

For diet (Row):

T2 292 192
SSR    CF    CF  8.33
cn 3(2) 3(2)

For exercise program (Column):

T2 302 122 62
SSC    CF     CF  78
rn 2(2) 2(2) 2(2)

For interaction:

(172  82  42  132  42  22 )
SSRC   SSR  SSC  CF
n
558
  SSR  SSC  CF  0.67
2

SSE = SST – SSR – SSC – SSRC = 3

13
 ANOVA Table
Source of df Sum of square Mean Square F
Variation
Diet (Row) r-1=1 SSR = 8.33 SSR MSR
MSR  = 8.33 F1  =16.66
r 1 MSE
Exercise c-1=2 SSC = 78 SSC MSC
program MSC  = 39 F2  =78
c 1 MSE
(Column)
Interaction (r-1)(c-1)=2 SSRC = 0.67 SSRC MSRC
MSRC  F3  =0.67
(r  1)(c  1) MSE
= 0.335
Error rc(n-1)=6 SSE = 3 SSE
MSE  = 0.5
rc(n  1)
Total rcn-1=11 SST = 90

Step 3 Rejection area:

For diet: Reject H0 if F1  F,r 1,rcn1 = F0.05,1,6 = 5.99
For exercise program: Reject H0 if F2  F,c 1,rcn1 = F0.05,2,6 = 5.14
For interaction: Reject H0 if F3  F,(r 1)(c 1),rc(n1) = F0.05,2,6 = 5.14

Step 4 Decision:
For diet: Since F1  16.66  5.99 , reject H0.
For exercise program: Since F2  78  5.14 , reject H0.
For interaction: Since F3  0.67  5.14 , do not reject H0.

Step 5 Conclusion:
There is not enough evidence to indicate that there is an interaction effect between
diet and exercise program.

There is enough evidence to indicate that there is a significant effect on mean weight
loss due to different diets.

There is enough evidence to indicate that there is a significant effect on mean weight
loss due to different exercise programs.

Note: Interesting results

The means for each cell (or group) are shown in the following table. The graph of the means is given
by the following figure. In this case, there is no significant interaction effect between diet and exercise
program. The lines in the graph are approximately parallel. The main effects can be interpreted
independently of each other.

10
Exercise 8.5
program 8
A B C 6.5
6
Diet Mean 4 2
I = 8.5 4 4
Diet Mean 2 1
2 Diet I
II = 6.5 2 2
1
0 Diet II
A B C
14
 Chapter 5 Analysis of Variance

Exercise 5.1 - Introduction to Experimental Design

1. Fill in the type of experimental design and the related ANOVA.

Experimental design ANOVA

a) Completely Randomized Design One-way ANOVA

b) Randomized Block Design ( )
c) ( ) Two-way ANOVA with interaction

2. State the assumptions of one-way ANOVA.

Exercise 5.2 - One-way ANOVA

1. From time to time, unknown to its employees, the research department at a company
observes various employees for their work productivity. Recently this company wanted
to check whether the four planners at a brand of the company, on average, the number
of orders are planned. The following table gives the number of orders planned by four
planners.

Planner A Planner B Planner C Planner D

19 14 11 24
21 16 14 19
26 14 21 21
24 13 13 26
18 17 16 30
13 18

At the 5% significance level, test the null hypothesis that the mean number of
customers served per hour by each of these four planners is the same. Assume that all
the assumptions required to apply the one-way ANOVA procedure hold true.

[ Answer: F = 10.57, reject Ho ]

2. Suppose that an automobile manufacturer designed a new engine and wants to

recommend which grade of gasoline to use. The four unleaded grades of gasoline are
regular, below regular, premium and super premium. The test car made 3 trial runs in
the test track using each of the 4 grades of gasoline. The data shown is the distance
travel.
Grades of gasoline
Below Regular Premium Super
Regular Premium
39.31 36.69 38.99 40.04
39.89 40 40.02 39.89
39.87 40.01 39.99 39.93

a) Is there a difference in the mean distance using 4 grades of gasoline? Use   0.05 .
b) What is the type of experimental design used in this study?

[ Answer: F = 0.6, do not reject Ho ]

15
3. A laboratory has many voltmeters that are used interchangeably to make voltage
measurements. A sample of three voltmeters is chosen at random. Each voltmeter
was used to measure a constant emf of 100 volt. The following voltage readings were
recorded.
Voltmeter
I II III
48 55 84
73 85 68
51 70 95
65 69 74
87 90 67
Is there a difference in the mean voltage among these 3 voltmeters? Use   0.05 .

[ Answer: F = 1.093, do not reject Ho ]

4. Eight different machines are being considered for use in manufacturing bottle caps.
The tensile strength of the bottle cap produced by each machine is of interest. A
random sample of 4 bottle caps produced by each machine was randomly selected
and their tensile strength recorded. The ANOVA table below summarizes the results
obtained.

Source df Sum of square Mean Square Fcalc

Machine 0.305
Error 3.45
Total

Complete the above table. Test if there is a difference in the mean tensile strength
between the bottle caps produced by the eight machines. Use   0.025 .

[ Answer: F = 0.305, do not reject Ho ]

5. The effect of three different paints on the corrosion of pipes was studied. One of the
three paints was randomly assigned so that each paint would be used on four
segments of pipe. The segments were painted and were observed for a period of six
months. The accompanying corrosion readings were then obtained. Is there sufficient
evidence to indicate a difference in the mean levels of corrosion for three paints? Use
ANOVA with   0.01 .

Paint A Paint B Paint C

10.1 13.4 12.7
11.4 12.9 11.9
12.1 13.3 12.5
10.8 13.1 12.3

[ Answer: F = 14.75, reject Ho ]

16
Exercise 5.3 - Two-way ANOVA

1. A study is conducted on the output (number of units produced) in a factory. The

purpose is to investigate whether the output is affected by 2 different locations. For
each location, the output is recorded for 3 shifts.

Shift
Morning Afternoon Night
Location A 700 702 621
B 869 823 700

a) At the 0.01 level of significance, do the data provide sufficient evidence to indicate
that there is a significant effect on output due to location?
b) At the 0.01 level of significance, do the data provide sufficient evidence to indicate
that there is a significant effect on output due to shift?
c) What is the type of experimental design used in this study? Identify the treatment
and blocking factor in this study.
d) Give a reason why blocking is carried out in an experiment.

[ Answer: a) F = 22.38, do not reject Ho, b) F = 8.634, do not reject Ho ]

2. A group of engineering workers wants to determine if humidity conditions have an

effect on the force required to pull apart pieces of glued plastic. Three types of plastics
(M, N, P) were tested at four different levels of humidity (10%, 30%, 50%, 70%). The
following results, in kilograms, were recorded. In this study, plastic type is assigned as
blocking factor.

Humidity
Plastic type 10% 30% 50% 70% Total
M 39 33 33.6 32 137.6
N 36.9 27.2 29.7 28.5 122.3
P 29.6 30.2 27.7 32.5 120.0
Total 105.5 90.4 91 93 379.9

At the level 1% level of significance, do the data provide sufficient evidence to indicate
there is an effect of
i) Types of plastic (Blocking factor)
ii) Humidity (Treatment factor)
on force required to pull apart glued plastic?

[ Answer: i) F = 2.905, do not reject Ho, ii) F = 2.136, do not reject Ho ]

17
Exercise 5.4 - Two-way ANOVA with Interaction

1. The number of rejected products produced by 3 workers operating of 2 different

machines are shown in the following table. In each case, the first number represents
the number of rejected products produced on Monday and the second number
represents the number of rejected products produced on the next day, that is Tuesday.

Worker
Worker 1 Worker 2 Worker 3
Machine A 700, 660 682, 742 627, 700
B 860, 689 813, 789 704, 711

a) What is the type of experimental design used in this study? How many treatment
combinations are there?

b) At the 0.01 level of significance, do the data provide sufficient evidence to indicate that
there is
i) a significant effect on number of rejected products due to worker?
ii) a significant effect on number of rejected products due to machine?
iii) a significant effect on number of rejected products due to interaction of worker and
machine?

[ Answer: i) F = 5.125, do not reject Ho, ii) F = 1.513, do not reject Ho, iii) F = 0.228, do not
reject Ho ]

2. The response time (in milliseconds) in an electric circuit was the subject of a recent
experiment. Three different circuits and five different testers were chosen at random.
Each tester measured the response time in each circuit three times. The results are
summarized in the following ANOVA table.

Source of df Sum of Mean Fcalc

Variation Squares, SS Squares, MS
Circuit 12.34
Testers 6.35
Interaction
Error 1.24
Total 76.12

i) Complete the given ANOVA table

ii) Test all appropriate hypothesis at the 5% level of significance

[ Answer: F1 = 4.976, reject Ho, F2 = 1.2802, do not reject Ho, F3 = 2.0393, do not reject Ho ]

18