7.

4 Latin Square 1

Source of Sum of squares Degree of Mean square F0

variation freedom
Algorithm 938.4 3 312.8 6.50
Error 855.7 8 107
Total 1794.1 11

would have been obtained by chance. The incorrect analysis of the data based on a
completely randomized experiment is shown in Table 7.18. Because F0 > F0.05,3,8
=
4.07 (Table A.11), the test statistic does not fall in the critical region adopted,
hence this does not call for rejection of null hypothesis. It is therefore concluded that
there is no significant difference between the population mean efficiencies of the
algorithms at 0.05 level of significance. It can be thus observed that the
randomized block design reduces the amount of experimental error such that the
differences among the four algorithms are detected. This is a very important
feature of randomized block design. If the experimenter fails to block a factor,
then the experimental error can be inflated so much that the important differences
among the treatment means may remain undetected.

7.4 Latin Square Design

Latin square design is used to eliminate two nuisance sources of variability by system-
atically blocking in two directions. The rows and columns represent two
restrictions on randomization. In general, a p p Latin square is a square
×
containing p rows and p columns. In this Latin square, each of the resulting p2
cells contains one of the p letters that corresponds to the treatment, and each letter
appears only once in each row and column. Some Latin square designs are shown in
Fig. 7.9. Here, the Latin letters A, B, C, D, and E denote the treatments.

Fig. 7.9 Display of Latin

square designs
248 7 Single-Factor Experimental Design

7.4.1
Table 7.18A Incorrect
Practical Problem
analysis of algorithm efficiency based on a completely randomized
design

Suppose a material supplied by three different vendors is required to be analyzed

for its weight in grams. It is known from past experience that the weighing machines
used and the operators employed for measurements have influence on the weights.
The experimenter therefore decides to run an experiment such that it takes into
account of the variability caused by the machine and the operator. A 3 3 Latin square
×
design is chosen where each material supplied by three different vendors is
weighed by using three different weighing machines and three different operators.

7.4.2 Data Visualization

As stated earlier, the material is analyzed for weight in grams from three different
vendors (a, b, c) by three different operators (I, II, III) and using three different
weighing machines (1, 2, 3) in accordance with a 3 3 Latin square design. The
×
results of experiment are shown in Table 7.19. Here, A, B, and C denote the weights
obtained by three different vendors.

7.4.3 Descriptive Model

Let us describe the observations using the following linear statistical model

yijk = μ + αi + τj + βk + εijk ; i = 1, 2 ,..., p; j = 1, 2 ,..., p; k = 1, 2 ,..., p

Here, yijk is a random variable denoting the ijkth observation. μ is a parameter

common to all levels called the overall mean, αi is a parameter associated with the
ith row called the ith row effect, and τj is a parameter associated with the jth treatment
called the jth treatment effect, βk is a parameter associated with kth column called
kth column effect, and εijk is a random error component. The above expression can
also be written as yijk μijk εijk where μijk μ αi τj βk .
= + = + + +
Note that this is a fixed effect model. The followings are the reasonable
estimates of the model parameters

Table 7.19 Experimental

results of material weight Operators Weighing machines
1 2 3
I A = 16 B = 10 C = 11
II B = 15 C=9 A = 14
III C = 13 A = 11 B = 13
 2 7 Single-Factor Experimental

Table 7.20 Data and residuals for material

Operators Weighing machines
1 2 3
I A = 16 A = 0.21 B = 10 B= −0.12 C = 11 C = −0.12
15.79 10.12 11.12
II B = 15 B = −0.13 C = 9 C= 0.21 A = 14 A = −0.13
15.13 8.79 14.13
III C = 13 C = −0.12 A = 11 A= −0.12 B = 13 B = 0.21
13.12 11.12 12.79

p p b
y...
μˆ = where y... = yijk and y ... =
y.. i=1 j=1 k=1 p3
p p
yi..
αˆ i =
yi.. — y... where yi.. = yijk , i = 1, 2 , . . . , p and i.. =
j=1 k=1 y p
p p
y.j.
τˆj =
y.j. — y... where y.j. = yijk , j = 1, 2 , . . . , p and .j. =
i=1 k=1 y p
p p
y..k
β = y — y where y..k = yijk , k = 1, 2 , . . . , pand
..k ... ..k =
ˆk i=1 j=1 y p

μˆ =
ijk
μˆ + αˆ i + τˆj + βˆk = yi.. + y.j. + y..k − 2y...

The estimate of the observations is yijk y y y 2y . Hence, the residual

ˆ = i.. + .j. + ..k − ..
is eijk yijk yijk .
= −
These formulae are used to estimate the observations and the residuals. Table 7.20
compares the estimated observations with the actual ones and reports the residuals.
Here, in each cell, three values are reported for a given operator and for a given
weighing machine. The leftmost value is obtained from experiment, the middle
one is the fitted value, and the rightmost value represents residual.
As in any design problem, it is necessary to check whether the model is
adequate or not. The methods of doing this are described in Sect. 8.2.3. The reader is
instructed to carry out a similar analysis and conclude on the adequacy of the
model.

7.4.4 Test of Hypothesis

The test of hypothesis is carried out as stated below.

Step 1: Statement of Hypothesis
We are interested here in testing the equality of vendor means. Then, the null
hypoth- esis states that the population vendor means are equal.
 7.4 Latin Square 2
H0 : μ.1. = μ.2. = · · · = μ.p..

The alternative hypothesis states that there is a difference between the population
treatment means for at least a pair of populations.

H1 : μ.i. /= μ.j. for at least one i and j.

Step 2: Selection of Level of Significance

Suppose the level of significance is chosen as 0.05. It means that the probability of
rejecting the null hypothesis when it is true is less than or equal to 0.05.
Step 3: Computation of Test Statistic
The test statistic is computed by carrying out an analysis of variance of the data. The
calculations for the sum of squares are shown here.
3 3 3

SSTotal 2
...
(16 + 10 + 11 + 15 +
2 2 2 2 2 2 2
+ 14 + 13 + 11 + 13
2 1122
−3 × = 2 —
= 9 )
3
i=1 j=1 k=1
yi2
9
= 1438 − 1393.78 = 44.22
3 2 2

1 2 y...
=1 (16 + 14 + 11)2 + (10 + 15 + 13)2 + (11 + 9 + 13)2 ]− 112
SSVendors = y .−
3 3×3 9
= 1404.67
1 3 2j − 1393.78
y... = 10.89
1 112 2
2 2 2 2
SSOperators = y − 37 + 38 + 37 ] − = 1394 − 1393.78 = .22
i..
3 = [
i=1 3×3 3 9
3 2 2
1 1 y... 112
SSWeighing machines = y −
2 = 442 + 302 + 382 ] − = 1426.67 − 1393.78 = 32.89
3Total −. SSVendors
SSError = SS 3 × 3 − SSInspectors − SSScales = 44.22
9 − 10.89 − .22 − 32.89 = .22
k

The calculations are summarized in Table 7.21.

Step 5: Specification of Critical Region
The null hypothesis is rejected if F0 > Fα,p−1,(p−1)(p−2) F0.05,2,2 19.00
= =
(Table A.11). The critical region is shown in Fig. 7.10.
Step 6: Decision
As the value of the test statistic falls in the critical region adopted, this calls for
rejection of null hypothesis. It is therefore concluded that there is a difference between

Table 7.21 ANOVA table for

material weight data Source of Sum of Degree of Mean F Value
variation squares freedom square
Vendors 10.89 2 5.445 49.5
Operators 0.22 2 0.11
Weighing 32.89 2 16.445
machines
Errors .22 2 .11
Total 44.22 8
 2 7 Single-Factor Experimental

Fig. 7.10 Critical region

the weight of the materials supplied by the vendors, at least for a pair of weights of
the materials supplied by the vendors.

7.4.5 Multiple Comparison Among Treatment Means

Tukey’s test compares between the pairs of mean values. This test is described in the
earlier section. The numerical value of the studentized range statistic is obtained as
follows.
Here, the studentized range statistic is mentioned below. Using Table A.17,
q0.05(2, 2) = 6.09, we have
, ,
MSError 0.11
T α= q (a,
α f) n = q 0 (2, 2) 3 = 1.166.

The three averages are y.1. 13.67, y.2. 12.67, y.3. 11. The absolute differ-
= = =
ences in averages are

|y.1. − y .2. |= 1.00∗ |y.1. − y .3. |= 2.67∗ |y.2. − y .3. |= 1.67∗.

The starred values indicate the pairs of means that are significantly different.
Some- times, it is useful to draw a graph, as shown below, underlining pairs of
means that do not differ significantly.

y.3. = 11, y.2. = 12.67, y.1. = 13.67.

 7.4 Latin Square 2
It can be observed that the material supplied by vendor A is on an average the
heaviest, followed by those supplied by vendors B and C, respectively. Further,
the differences between the materials supplied by vendors A and C and B and C
are statistically significant.
There is another test, Fisher’s test, that compares whether the difference
between the averages is statistically significant or not. In this test, the least
significance dif- ference (LSD) is obtained as follows (Using Table A.10, t0.05,2 =
4.3027).
, ,
2MSError 2 × .11
LSD = t = t =
2 n 3

The three averages are y.1. 13.67, y.2. 12.67, y.3. 11. Then, the differences
= = =
in averages are

|y.1. − y .2. |= 1.00 |y.1. − y .3. |= 2.67∗ |y.2. − y .3. |= 1.67∗.

The starred values indicate the pairs of means that are significantly different.
Example 7.7 An agricultural scientist wishes to examine the effect of five varieties
of barley (A, B, C, D, and E) on their yield (kg per plot). An experiment is
conducted according to a Latin Square Design so that the effects of plot and
season are sys- tematically controlled. Yield of barley (kg per plot) is shown in Table
7.22. Suppose we wish to know if the five varieties of barley exhibit same yield or
not at a level of significance of 0.05. Let us carry out the test of hypothesis,
described in Sect. 8.3.4.
Step 1: Statement of Hypothesis: Here, the null hypothesis states that there is no
difference among the population mean yields of barley produced from five differ-
ent varieties (treatment). The alternative hypothesis states that there is a difference
between the population mean yields of barley for at least a pair of populations.
Step 2: Selection of Level of Significance: Suppose the level of significance is
chosen as 0.05.
Step 3: Computation of Test Statistic: The test statistic can be computed by carrying
out an analysis of variance of the data. The calculations for ANOVA are summarized
in Table 7.23.
Step 4: Specification of Critical Region: The null hypothesis is rejected if F0 >
F0.05,3,6 = 4.76 (Table A.11).

Table 7.22 Experimental

data barley yield as per Latin Plot Season
I II III IV
Square
1 B = 40 C = 26 D = 26 A = 52
2 D = 15 A = 50 B = 38 C = 38
3 C = 30 B = 45 A = 48 D = 24
4 A = 55 D = 20 C = 35 B = 42
 2 7 Single-Factor Experimental
Table 7.23 ANOVA for
barley yield Source of Sum of Degree of Mean F0
variation squares freedom square
Variety of 1963 3 654.33 25.83
barley
Plot 16.40 3 5.5
Season 40.50 3 13.5
Errors 152.00 6 25.33
Total 2172.00 15

Step 5: Take a Decision: As the value of the test statistic falls in the critical
region adopted, this calls for rejection of null hypothesis. It is therefore concluded
that there is a significant difference between the population mean yields of barley
for at least a pair of populations at 0.05 level of significance.

Example 7.8 Consider Example 7.7. Let us find out which of the mean barley yields
are significantly different by applying Tukey’s test, as described in Sect. 8.4.5. Here,
the mean barley yields are as follows.

y.1. = 51.25, y.2. = 41.25, y.3. = 32.25, y.4. = 4.08.

The numerical value of the studentized range statistic is T0.05 q0.05(4, 15) 4.08
= =
(Table A.17). The differences in mean barley yields are

|y.1. − y.2. | = 10∗ , |y.1. − y.3. | = 19∗ , |y.1. − y.4. | = 30∗

|y.2. − y.3. | = 9∗ , |y.2. − y.4. | = 19∗
|y.3. − y.4. | = 11∗ .

The starred values indicate the pairs of means that are significantly different. Clearly,
all the pairs of means are significantly different in terms of yielding barley at 0.05
level of significance.

7.5 Balanced Incomplete Block Design

Sometimes, in a randomized block design, it is not possible to carry out runs with
all treatment combinations in each block. This type of situation occurs because of
shortage of apparatus or unavailability of sufficient materials or lack of time or
limited facilities. In such situation, it is possible to use a block design in which
every treatment may not be present in every block. Then, the block is said to be
incomplete, and a design constituted of such blocks is called incomplete block
design. However, when all treatment combinations are equally important as it
ensures equal precision