POLYTECHNIC UNIVERSITY OF THE PHILIPPINES

COLLEGE OF ENGINEERING
DEPARTMENT OF INDUSTRIAL ENGINEERING

MODULE 3
SINGLE FACTOR EXPERIMENTS

Experiments are a natural part of the engineering and scientific

decision-making process. Suppose that an industrial engineer is
investigating the effect of different manufacturing methods on the
mean quality characteristic of one of its products. The experiment
would consist of making up several product samples using each of
the manufacturing methods under study and measuring the average
quality characteristic of each group of samples.

If there are only two manufacturing methods of interest, this

experiment could be designed and analyzed using the statistical
hypothesis methods for two samples, i.e. the experimenter has a
single factor of interest---manufacturing method---and there are
only two levels of the factor.

Sometimes, we want to investigate more than two levels. To do

this, we use the Analysis of Variance (ANOVA).

Statistically-based experimental design techniques are particularly

useful in engineering as they help improve the performance of a
manufacturing process. By using designed experiments, engineers
can determine which among the different factors has the greatest
influence on process performance. The results of such an
experiment can lead to:

1. improved process yield

2. reduced variability in the process and closer conformance
to nominal or target requirements
3. reduced design and development time
4. reduced cost of operation

Page 1 of 23
I. The Completely Randomized Single-Factor Experiment

EXAMPLE: A manufacturer of paper used for making grocery

bags is interested in improving the tensile strength of the product.
Product engineering thinks that tensile strength is a function of the
hardwood concentration in the pulp and that the range of hardwood
concentrations of practical interest is between 5 and 20%. A team
of engineers responsible for the study decides to investigate four
levels of hardwood concentration: 5%, 10%, 15%, and 20%. They
decide to make up six test specimens at each concentration level,
using a pilot plant. All 24 specimens are tested on a laboratory
tensile tester, in random order. The data from this experiment are
shown in the table below:

Replicate, j
factor

Factor levels, i yij

This is an example of a completely randomized single-factor

experiment with four levels of the factor. The levels of the factors
are sometimes called treatments. In the table above, each treatment
has six observations or replicates.

Page 2 of 23
The Analysis of Variance

Suppose we have a different levels of a single factor that we wish

to compare. The response for each of the a treatments is a random
variable. An entry in the table above, say y ij , represents the jth
observation taken under treatment i.

Let us describe the observations in the table above by the linear

statistical model:

i 1, 2,, a
yij i ij
j 1, 2,, n
where y ij is a random variable denoting the (ij)th observation, is
the overall mean, i is treatment effect of the ith level, and ij is
a random error component. We assume that the random errors are
NID 0, 2 ; with this assumption, each treatment can be thought of
as a normal population with mean i and variance 2 .

The factor levels in the experiment can be chosen in two different

ways:

1. The experimenter can specifically choose the a treatments. This

is called the FIXED EFFECTS MODEL. In this model, we test
the hypotheses about the treatment means. Furthermore, we cannot
extend the conclusion made from the study to treatments that were
not considered.

2. The experimenter can choose the a treatments randomly from

a larger population of treatments. This is called the RANDOM
EFFECTS MODEL. For this situation, we test the hypotheses
about the variability of each treatment, and to try to estimate such

Page 3 of 23
variability. Conclusions made from this type of study can be
extended to all treatments in the population.

Fixed-Effects Model

In the fixed-effects model, the treatment effects, i , are usually

defined as deviations from the overall mean , so that

i 1
i 0

For the fixed-effect model, we will be testing the equality of the

different factor level means, that is:

H0 : 1 2 a 0
H1 : i 0 for at least one i

This means that if we fail to reject the null hypothesis, changing

the levels of the factor has no effect on the mean response. If we
reject the null hypothesis, our conclusion will be that there is at
least one level that has an effect on the mean response; this
means that changing the factor level to that particular level will
bring about a significant change on the mean response.

The table below shows how the ANOVA table is constructed:

Page 4 of 23
 a n
y..2 a
yi2. y..2
SS T y
2
ij SS Factor
i 1 j 1 an i 1 n an

SS E SST SS Factor
Consider the paper tensile strength described earlier. Let us use
ANOVA to test the hypothesis that different hardwood
concentrations do not affect the mean tensile strength of the paper.
Use a 90% confidence level.

H 0 :1 2 3 4 0
Step 1: H 1 : i 0 for at least one i

Step 2: 0.10

Step 3: (declare the appropriate ANOVA model) Completely

Randomized Design, Fixed-Effects Model

Step 4: Construct your ANOVA table:

Page 5 of 23
Step 5: Determine the critical region, i.e. f 0 f 0.10, a 1, a n 1
*Based on the statistical tables, f0.10,3, 20 2.38 , therefore the
critical region will be f 0 2.38 .

Step 6: Declare the result of your test: since the computed

value is within the critical region, i.e. greater than 4.94, then we
reject the null hypothesis.

Step 7: Declare your conclusion: Hardwood concentration in

the pulp significantly affects the mean strength of the paper.

Multiple Comparisons Following the ANOVA

When the null hypothesis is rejected in the ANOVA, we know that

some of the treatment levels significantly affect the output being
measured. However, ANOVA doesnt identify which level(s)
affect the output! There are several methods used to investigate
this, which are collectively called multiple comparison methods.
Let us try to learn one: the Fishers Least Significance Difference
(LSD) Method.

The Fisher LSD Method compares all pairs of means with the null
hypotheses H 0 : i j for all i j using the t-statistic

y i y j
t0
2 MS E
n

Assuming a two-sided alternative hypothesis, the pair of means

i and j would be declared significantly different if:

Page 6 of 23
 y i y j LSD

where LSD, the least significant difference, is given by the

equation:

2MS E
LSD t 2,a n 1
n

If the samples sizes are different in each treatment, the LSD

equation becomes:

1 1
LSD t 2,a n1 MS E
n n
i j

EXAMPLE: Let us apply the Fisher LSD method to the hardwood

concentration experiment. Recall from the table above that:

y 1 10.00 psi
y 2 15.67 psi
y 3 17.00 psi
y 4 21.17 psi

The value of LSD is:

2MS E
LSD t 2, a n 1
n
26.51
1.725
6
2.54

Page 7 of 23
Thus, any pair of treatments whose absolute value of the difference
of their means that differs by more than 2.54 implies that the
corresponding pair of treatment means is different, i.e. the effect of
one level is different from the effect to the one it is being
compared to.

The comparisons among the observed treatment averages are as

follows:

4vs.1 21.17 10.00 11.17 2.54

4vs.2 21.17 15.67 5.50 2.54
4vs.3 21.17 17.00 4.17 2.54
3vs.1 17.00 10.00 7.00 2.54
3vs.2 17.00 15.67 1.33 2.54
2vs.1 15.67 10.00 5.67 2.54

From the above computations, we can see that there are significant
differences between all pairs of means except 2 and 3. This implies
that 10% and 15% hardwood concentrations produce more or less
the same tensile strength and that all other concentration levels
tested produce different tensile strengths.

Determining the Sample Size

The choice of the sample size (or number of replicates) to use in

any experimental design is very important. The sample size helps
in determining the validity of your results through its power.

Operating Characteristic (OC) curves may be used to establish

the sample size.

RECALL: The OC curve is a plot of the probability of a type II

error for various sample sizes against a measure of the
difference in means that it is important to detect. Thus, if the

Page 8 of 23
experimenter knows the magnitude of the difference in means that
is of potential importance, the OC curves can be used to determine
how many replicates are required to obtain sufficient sensitivity.

The POWER (as mentioned earlier) of the ANOVA test is given

by the equation:

1 PReject H 0 H 0 is false

Operating Characteristic curves are used to evaluate . These

curves plot against a parameter , where

a
n i2
2 i 1
a 2

EXAMPLE: Suppose that five means are being compared in a

completely randomized experiment with 0.01. The
experimenter would like to know how many replicates to run if it is
important to reject H 0 with probability of at least 0.90 if
i1 i2 2 5.0 .
5

Given the situation above, our is:

a
n i2
2 i 1

n
5 n
a 2 5

We use the OC curve wherein 1 5 means or levels 1 4

degrees of freedom. To determine the appropriate sample size n,
we set up a table such as the one immediately proceeding below:

Page 9 of 23
 n 2 a(n 1) Power
4 4 2.00 15 0.38 0.62
5 5 2.24 20 0.18 0.82
6 6 2.45 25 0.06 0.94

As seen from the table above, a sample size of 6 is the most

applicable as our requirement is to reject H 0 with a probability of
at least 0.90.

Random-Effects Model

In many situations, the factor of interest has a large number of

possible levels. The analyst is interested in drawing conclusions
about the whole population of factor levels. If the experimenter
randomly selects a of those levels from the population of factor
levels, we can say that the factor is a random factor. Since the
factors used in the experiment were chosen randomly, the
conclusions made will be valid for the ENTIRE POPULATION
of factor levels. This is an important difference of the random-
effects model vis--vis with the fixed-effect model wherein the
conclusions apply only for the factor levels used in the experiment.

As mentioned above, the test of hypotheses to be done for the

random-effects model is:

H 0 : 2 0
H1 : 2 0

If 0 , all treatments are identical, but if 0 , there is

2 2

variability between treatments.

The ANOVA decomposition of total variability is still valid, i.e.:

Page 10 of 23
 SST SS treatments SS E

Thus, the relationship of the variances with the mean squares for
treatments and error are:

2 MS E
MS treatments MS E
2
n

EXAMPLE: A single-factor experiment involves the random-

effects model in which a textile manufacturing company weaves a
fabric on a large number of looms. The company is interested in
loom-to-loom variability in tensile strength. To investigate this
variability, a manufacturing engineer selects four looms at random
and makes four strength determinations on fabric samples chosen
at random from each loom. Data are presented in the table below:

Observations
Loom 1 2 3 4
1 98 97 99 96
2 91 90 93 92
3 96 95 97 95
4 95 96 99 98

H 0 : 2 0
Step 1:
H1 : 2 0

Step 2: 0.05

Step 3: (declare the appropriate ANOVA model) Completely

Randomized Design, Random-Effects Model

Page 11 of 23
Step 4: Construct your ANOVA table:

ANOVA
Source of Variation SS df MS F P-value F crit
Between Groups 89.19 3 29.73 15.68 0.000187792 3.49
Within Groups 22.75 12 1.896

Total 111.9 15

Step 5: Determine the critical region, i.e. f 0 f 0.05,a1,a n1

*Based on the statistical tables, f 0.05,3,12 3.49 , therefore the
critical region will be f 0 3.49 .

Step 6: Declare the result of your test: since the computed

value is within the critical region, i.e. greater than 3.49, then we
reject the null hypothesis.

Step 7: Declare your conclusion: There is loom-to-loom

variability in terms of tensile strength.

Let us also determine the total variance for the experiment above.

2 MS E 1.90
MStreatments MS E 29.73 1.90
2 6.96
n 4

Total variability is therefore:

Total 2 2
6.96 1.90
8.86

Page 12 of 23
Completely Randomized BLOCK Design

In many experimental design problems, it is necessary to design

the experiment so that the variability arising from a nuisance factor
can be controlled.

For example, suppose that in the process of your Work Study (IE
REVIEW: What is WORK STUDY?) we want to evaluate three
new alternative methods that can be used to test the strength
readings on steel plate girders. A randomized block design will be
used when we choose say n samples and each of those n samples
will undergo all three alternative methods of testing. In this case,
the three alternative methods are our levels or treatments a and our
n samples will be our blocks, which is represented by the letter b.

The sum of squares identity for the randomized complete block

design is:

SST SStreatments SSblocks SS E

The test of hypotheses in a randomized complete block design is

the same as that for a fixed-effects model, that is:
H 0 :1 2 3 4 0
H 1 : i 0 for at least one i

Page 13 of 23
The ANOVA table for a randomized complete block design is
given below:

Source of Sum of Degrees of

Variation Squares Freedom Mean Square F0
SS treatments MS treatments
Treatments SStreatments a-1 a 1 MS E
SS blocks
Blocks SSblocks b-1 b 1
SS E
Error SSE (a - 1)(b - 1) a 1b 1
Total SST ab - 1

EXAMPLE: An experiment was performed to determine the effect

of four different chemicals on the strength of a fabric. These
chemicals are used as part of the permanent press finishing
process. Five fabric samples were selected, and a randomized
complete block design was run by testing each chemical type once
in random order on each fabric sample. Data is shown in the table
below. Test for the differences in means using an ANOVA with
0.01 .

Fabric Sample
Chemical
Type 1 2 3 4 5
1 1.3 1.6 0.5 1.2 1.1
2 2.2 2.4 0.4 2.0 1.8
3 1.8 1.7 0.6 1.5 1.3
4 3.9 4.4 2.0 4.1 3.4

Page 14 of 23
 H 0 :1 2 3 4 0
Step 1: H 1 : i 0 for at least one i

Step 2: 0.01

Step 3: (declare the appropriate ANOVA model) Randomized

Complete Block Design

Step 4: Construct your ANOVA table:

ANOVA

Sum of Degrees of Mean

Source of Variation Squares Freedom Square F0
Chemical Types 18.04 3 6.01 75.17
Fabric Samples 6.69 4 1.67
Error 0.96 12 0.08
Total 25.69 19

Step 5: Determine the critical region, i.e. f 0 f 0.01,a1,a1b1

*Based on the statistical tables, f 0.01,3,12 5.95 , therefore the
critical region will be f 0 5.95 .

Step 6: Declare the result of your test: since the computed

value is within the critical region, i.e. greater than 5.95, then we
reject the null hypothesis.

Step 7: Declare your conclusion: There is a significant

difference in the effect of the chemical types on fabric strength.

Page 15 of 23
EXERCISES:

1. An experiment involving the contact resistance of a brake-only

relay was studied for three different silver-based alloys.

a. Does the type of alloy affect mean contact resistance? Use

0.01.
b. Use Fishers LSD method to determine which means differ.

2. An experiment which aims to test three random different

methods of preparing fish are evaluated on the basis of sensory
criteria to which a quality score is assigned. Assume that these
methods have been randomly selected from a large population of
preparation methods. Data is presented in the table below:

a. Is there any difference in preparation methods? Use

0.05 .

Page 16 of 23
 b. Estimate the components of variance.

3. An experiment to determine the effect of temperature on mean

weight gain of swine is made. The mean weight of each group of
swine at the start of the experiment is considered to be a nuisance
factor. The data from the experiment are as follows:

a. Does housing air temperature affect mean weight gain? Use

0.05 .
b. Use Fishers LSD method to determine which temperature
levels are different.

Page 17 of 23
Latin Squares

The randomized block design is very effective for reducing

experimental errors by removing one source of variation. Another
design that is useful in controlling two sources of variation, while
reducing the required number of treatment combinations, is called
the LATIN SQUARE.

Suppose that we would like to study four varieties of wheat using

four different fertilizers over a period of four years. The total
number of treatment combinations for a completely randomized
design would be 64 (4 varieties x 4 fertilizers x 4 years). A Latin
square design can be used in this situation, as the same number of
categories for all criteria of classification was used. Thus, instead
of analyzing 64 combinations, we can reduce it to 16.

A typical latin square, selected at random from all possible 4x4

squares, is the following:

Column
Row 1 2 3 4
1 A B C D
2 D A B C
3 C D A B
4 B C D A

The four letters, i.e. A, B, C, and D, represent the four varieties of

wheat, which are correspondingly referred to as the treatments.
The rows and columns, represented by the four fertilizers and the
four years respectively, are the two sources of variation that we
wish to control. Note that each treatment occurs exactly once in
each row and in each column.

Page 18 of 23
CAUTION: When there is interaction present between any of the
sources of variation, the f-values in the ANOVA of the Latin
square will become invalid. This makes the Latin square design
inappropriate for the experiment.

ANOVA
Source of Variation SS df MS F0
Rows SSR r-1 SSR/(r-1)
Columns SSC r-1 SSC/(r-1)
Treatments SSTr r-1 SSTr/(r-1) MSTr/MSE
(r-1)(r-
Error SSE 2) SSE/[(r-1)(r-2)]
Total SST r2-1

EXAMPLE: The data presented in the table below are the yields
for the four varieties of wheat that was previously described,
measured in kilograms per plot. It is assumed that the various
sources of variation do not interact. Use a 0.05 level of
significance to test whether or not there is a difference in the
average yields of the four varieties of wheat.

Fertilizer Year
Treatment 2001 2002 2003 2004
A B C D
F1
70 75 68 81
D A B C
F2
66 59 55 63
C D A B
F3
59 66 39 42
B C D A
F4
41 57 39 55

Page 19 of 23
 H 0 :1 2 3 4 0
Step 1: H 1 : i 0 for at least one i

Step 2: 0.05

Step 3: (declare the appropriate ANOVA model) Latin Squares

Design

Step 4: Construct your ANOVA table:

ANOVA
Source of Variation SS df MS F0
Rows 1557.19 3 519.06
Columns 417.69 3 139.23
Treatments 263.69 3 87.90 2.02
Error 261.38 6 43.56
Total 2499.94 15

Step 5: Determine the critical region, i.e. f 0 f 0.05,r 1,r 1r 2

*Based on the statistical tables, f 0.05,3,6 4.76 , therefore the
critical region will be f 0 4.76 .

Step 6: Declare the result of your test: since the computed

value is outside the critical region, i.e. less than 4.76, then we fail
to reject the null hypothesis.

Step 7: Declare your conclusion: There is no significant

difference in the average yields of the four varieties of wheat.

Page 20 of 23
EXERCISE: The mathematics department of a large university
wishes to evaluate the teaching capabilities of four professors. In
order to eliminate any effects due to different mathematics courses
and different times of the day, it was decided to conduct an
experiment using a Latin square design in which the letter A, B, C,
and D represent the four different professors. Each professor taught
one section of each of four different courses scheduled at each of
four different times during the day. The data presented below show
the grades assigned by these professors to 16 students of
approximately equal ability. Use a 0.05 level of significance to test
the hypothesis that different professors have no effect on the
grades.

Time Subject
Period Algebra Geometry Statistics Calculus
A B C D
1
84 79 63 97
B C D A
2
91 82 80 93
C D A B
3
59 70 77 80
D A B C
4
75 91 75 68

Page 21 of 23
PROBLEM SET:

1. The following data show the effect of four operators chosen

randomly, on the output of a particular machine:

Operator
1 2 3 4
175.4 168.5 170.1 175.2
171.7 162.7 173.4 175.7
173.0 165.0 175.7 180.1
170.5 164.1 170.7 183.7

Compute for the estimate of the total variance.

2. An experiment is conducted in which four treatments are to be

compared in five blocks. The following data are generated:

Block
Treatments 1 2 3 4 5
1 12.8 10.6 11.7 10.7 11.0
2 11.7 14.2 11.8 9.9 13.8
3 11.5 14.7 13.6 10.7 15.9
4 12.6 16.5 15.4 9.6 17.1

Test the hypothesis at the 0.05 level of significance that there is no

difference between treatment means. This time, instead of using a
fixed-effects model, assume a random-effects model.

3. An analysis was conducted by the Statistics Consulting Center at

Virginia Polytechnic Institute and State University in which a
certain treatment was applied to a set of tree stumps. The chemical
Garlon was used the with the purpose of regenerating the roots of
the stumps. A spray was used with four levels of Garlon

Page 22 of 23
concentration. After a period of time the height of the shoots was
observed. Does the concentration of Garlon has a significant
impact on the height of the shoots? Use 0.05 .

1 2.87 2.31 3.91 2.04

Garlon 2 3.27 2.66 3.15 2.00
level 3 2.39 1.91 2.89 1.89
4 3.05 0.91 2.43 0.01

4. In 1983, the Department of Dairy Science at the Virginia

Polytechnic Institute and State University conducted an experiment
to study the effect of feed rations, differing by source of protein, on
the average daily milk production of cows. There were 5 rations
used in the experiment. A 5x5 Latin square was used in which the
rows represented different cows and the columns were different
lactation periods. The following data, recorded in kilograms, were
analyzed by the Statistical Consulting Center at Virginia Tech. At
the 0.01 level of significance, can we conclude that rations with
different sources of protein have an effect on the average daily
milk production of cows?
Lactation Cows
Period 1 2 3 4 5
A B C D E
1
33.1 34.4 26.4 34.6 33.9
C D E A B
2
30.7 28.7 24.9 28.8 28.0
D E A B C
3
28.7 28.8 20.0 31.9 22.7
E A B C D
4
31.4 22.3 18.7 31.0 21.3
B C D E A
5
28.9 22.3 15.8 30.9 19.0

Page 23 of 23