You are on page 1of 64

Design of Experiments (DOE)

and
Analysis of Variance (ANOVA)
for

Implementing SIX SIGMA

SIX SIGMA , DOE, ANOVA


5 step DMAIC Method essential for success of a breakthrough
project (implementation of SIX SIGMA)
Each step brings a distinct set of tools to bear on the project
objective
DOE & ANOVA - power tool of six sigma for the Analyze and
Improve step
Of all the statistical tools employed within Six Sigma, DOE
offers the most power for making breakthroughs

DOE - First conceived and developed by Sir Ronald A. Fisher in


the 1920s and 1930s.
ANOVA Method for analyzing designed experiments

DOE

What is it?
DOE is a systematic approach to investigation of a system or process. A series of
structured tests are designed in which planned changes are made to the input
variables of a process or system. The effects of these changes on a pre-defined
output are then assessed.

The experimental design is the manner in which the researcher randomly assigns
the treatment to the experimental units. We have considered four experimental
designs:completely randomized design, randomized complete block, latin square
and factorial experiments.

DOE

Why is it important?
DOE is important as a formal way of maximizing information gained while
minimizing resources required. It has more to offer than 'one change at a time'
experimental methods, because it allows a judgment on the significance to the
output of input variables acting alone, as well input variables acting in
combination with one another.

DOE

How to use it?

ANOVA
Definition
ANOVA stands for analysis of variance, a method by which the
source of variability is identified. This method is widely used in
industry to help identify the source of potential problems in the
production process and identify whether variation in measured
output values is due to variability between various manufacturing
processes, or within them. By varying the factors in a
predetermined pattern and analyzing the output, one can use
statistical techniques to make an accurate assessment as to the
cause of variation in a manufacturing process.

Steps in DOE and ANOVA


Choosing appropriate responses (output variables)
Choosing appropriate factors (input variables)
Setting appropriate factor ranges or levels
Creating documentation for the experiment
Managing the experiment as it takes place
Reporting and presenting results (ANOVA)
Deciding whether to further optimize the process or just run a
confirmation experiment.

Types of DOE
Completely Randomized Design
Randomized Block Design
Latin Square Design

Factorial Design

Example 4.2 Alpha Industries Limited has deputed four


different batches of its employees to four different
training organizations (A, B, C and D) for the same
training programme, which aims to train them in improving
their decision-making skills. Each batch consists of five
employees with similar qualifications and work experience.
After the training programme, the company conducted a
common examination to test their level of additional
knowledge gained through the training programme. The
percentage scores of the employees of the batches are
summarized in Table 4.8.

Table 4.8 Percentage Scores of Employees of


Batches in the Examination
Training organization
A

80

70

65

90

90

60

50

89

96

55

58

85

85

85

55

95

70

90

40

80

Perform ANOVA to check whether there is a significant difference


between the training organizations in terms improving the decisionmaking skills of the employees by assuming a significance level of
0.05.

Solution.
The model of this problems is:
Yij = + Tj + eij
where is the overall mean; Yij, the ith observation
under the jth treatment of the factor, training organization;
Tj, the effect of the jth treatment of the factor, training
organization; eij, the random error associated with the ith
observation under the jth treatment of the factor training
organization,
Null hypothesis,

Ho : T1 = T2 = T3 = T4

Alternate hypothesis, H1: Treatment means are not


equal for at least one pair
of treatment means.
The data of Table 4.8 is reproduced below in Table 4.9
along with column totals.

Table 4.9 Column Totals of Table 4.8


Training organization

Yj

A
80

B
70

C
65

D
90

90
96

60
55

50
58

89
85

85
70
421

85
90
360

55
40
268

95
80
439

We have

i 1

j 1

..

Yij 1488

.1

421,Y .2 360,Y .3 268and Y .4 439

Then
5 4
Y ..2
2

SS total Y ij
20
i 1 j 1
(80) 2 (70) 2 (65) 2 (90) 2 (60) 2 .... (40) 2 (80) 2 (1488)
20
5368.8

4 Y 2j Y 2
..
ss treatments
j 1 5 20
(421) 2 (360) 2 (268) 2 (439) 2 (1488) 2

5
3570
SS error SS total SS treatments
5368.8 3570
1798.8
The results of this problem are summarized Randomized Design (Example4.2)

Table 4.10 Results in Completely Randomized Design (Example 4.2)


Source of
variation

Degrees of Sum of
freedom squares

Between
treatments

3570.0

Within
treatments

16

1798.8

Total

19

5368.8

Mean sum of
squares (MSS)

F ratio

3570/3 = 1190 1190/112.43 = 10.58


1798.8/16 = 112.43

In Table 4.10, the value of the calculated F ratio is 10.58. The value of F
ratio from the statistical table for a significance level of 0.05 and degrees
of freedom (3,16) is 3.24. Then
Fcalculated (10.58) >Fa = 0.05 and df = (3.14) (3.24)

Hence , the null hypothesis (Ho) should be rejected.


Inference: This means that there is a significant difference in terms of
percentage scores between the batches trained by different training
organization.

Example 4.4 A business school recently conducted a management


meet titled 'Synergy in which 'stock exchange simulation game'
was an event. Students from 15 different business schools
participated in the meet. Five different teams (T1, T2, T3, T4 and
T5)have been formed from these 15 schools to participate in the
simulation game. Each team consisted of a team leader (L1, L2, L3,
L4 and L5) and four members. The details of the game were as
follows:
Each team was given 10 lakhs rupees.
Daily quotations were displayed in an online network system.
Rumours, political news and industrial news were broadcasted
over a mike. Based on these information, the teams effected
purchases and sales.
At the end of the game, the net worth of each team was
assessed.
Then, team leaders were rotated to serve in all other teams as
team leaders and the game was repeated four more times.

It is clear that the experimentation is a randomized complete block design.


The net worth of the teams (in lakhs of rupees) after the experiment are, as
summarized in Table 4.16.

Table 4.16 Results of Simulation Game


Team Leader

Team

L1

L2

L3

L4

L5

T1

40

30

55

25

35

T2

45

60

10

22

33

T3

38

55

40

55

28

T4

30

27

32

56

17

T5

45

34

20

34

37

Perform the randomized complete block design to check whether


there is a significant difference between the team leaders (L1, L2,
L3, L4 and L5) as well as between different teams (T1, T2, T3, T4 and
T5) in terms of final net worth by assuming a significance level of

0.05.
Solution. In this problem, 'team leader' is considered to be the
factor and 'team' is considered to be the block. The model of

the randomized complete block design applied to this example is:


Yij = + Ti +Lj + eij
where is the overall mean; Yij the observation with respect to jth
treatment of the factor (team leader) and ith block (team); Tj, the
effect of the ith block (team); Li the effect of the jth treatment of the
factor (team leader); and eij the random error associated with the ith
block (team) and the jth treatment of the factor (team leader).

Hypothesis with respect to treatment (team leader): In this case,


Null hypothesis,
Ho : L1 = L2 = L3 = L4 = L5
Alternate hypothesis,
H1 : Treatment means are not equal for at least
one pair of treatment means.
Hypothesis with respect to block (team): Here,
Null hypothesis,
Ho : T1 = T2 = T3 = T4 = T5
Alternate hypothesis,
H1 : Block means are not equal for at least one
pair of block means.
The data of this example is reproduced with row totals and column totals, as
shown in Table 4.17.
Table 4.17 Results of Simulation Game
Team leader

Team

L1

L2

L3

L4

L5

Yi.

T1

40

30

55

25

35

185

T2

45

60

10

22

33

170

T3

38

55

40

55

28

216

T4

30

27

32

56

17

162

T5

45

34

20

34

37

170

Y.j

198

206

157

192

150

903

For the design, the following calculations are done based on the Table 4.17:

5
5
Y .. Y ij 903
i 1 j 1
Y .1 198, Y .2 206, Y .3 157, Y .4 192and Y .5 150
Y .1 185, Y .2 170, Y .3 216, Y .4 162and Y .5 170
Now ,w e get
2
5
5
2 Y ..
SStotal Y ij
25
i 1 j 1
( 40) 2 (30) 2 (55) 2 ( 25) 2 (35) 2 ( 45) 2 .... (34) 2 (37) 2
4118.64
5 Y .2j Y 2
..
SSteam leader
25
i 1 5
(198) 2 ( 206) 2 (157) 2 (192) 2 (150) 2 (903) 2

5
25
514.24
5 Y2 Y2
i. ..
SSteam
25
i 1 5
(185) 2 (170) 2 ( 216) 2 (162) 2 (170) 2 (903) 2

5
25
368.64
SStotal SStotal SStreaments SSblocks
4118.64 514.24 368.64
3235.76
The results of this example are summarized in Table 4.18.

(903) 2
25

Table 4.18 Results in Randomized Complete Block Design (Example 4.4)


Source of

Degrees. of

Sum of

Mean sum of
F ratio

variation

freedom

squares

squares (MSS)

514.24

514.24/4 = 128.56

128.56/202.24 = 0.636

368.64

368.64/4 = 92.16

92.16/202.24 = 0.456

Error

16

3235.76

3235.76/16 = 202.24

Total

24

4118.64

Between
treatments
Between
blocks

In Table 4.18, the value of the calculated F ratio of the treatment is


0.636, whereas its table value with the significance level of 0.05 and
degrees of freedom (4,16) is 3.01. The calculated F ratio for the block is
0.456, whereas its table value with the significance level of 0.05 and
degrees of freedom (4, 16) is 3.01.
Component-treatment (team leader): For this component,
Fcalculated [0.636] < Fa= 0.05 and df =(4,16) [3.01]
Hence, the null hypothesis, Ho should be accepted.
Inference: This means that there is no significant difference in
terms of net worth between the team leaders.
Component-block (team): For this component,

Fcalculated [0.456] < Fa = 0.05 and df = (4,16) [3.01]


Hence, the null hypothesis, Ho should be accepted.
Inference: This meat there is no significant difference in terms of
net worth between different teams.

Example 4.6. The production manager of a firm is keen in analyzing the


hourly production volume of a product. He feels that operators (A, B, C,
D and E), machines (M1, M2, M3, M4 and M5) and suppliers of raw
materials (S1, S2, S3, S4 and S5) will affect the production volume of the
product. Hence, he designed a Latin square design, as shown in Table
4.23. The corresponding observations are shown in Table 4.24.
Table 4.23 Experimental Combinations in Latin Square Design (Example 4.6)

Supplier

Machine

S1

S2

S3

S4

S5

M1

M2

M3

M4

M5

Table 4.24 Hourly Production Volume as per Latin Square Design (Example 4.6)

Supplier

Machine

S1

S2

S3

S4

S5

M1

12 (A)

8 (B)

9 (C)

15 (D)

18 (E)

M2

14 (B)

7 (C)

13 (D)

10 (E)

12 (A)

M3

7 (C)

15 (D)

12 (E)

8 (A)

6 (B)

M4

16 (D)

17 (E)

13 (A)

11 (B)

12 (C)

M5

9 (E)

14 (A)

8 (B)

10 (C)

8 (D)

Perform statistical analysis to test the following at a significance level of 0.05:


Whether there is significant difference between operators in terms of production
volume of the product;
Whether there is significant difference between machines in terms of production
volume of the product;
Whether there is significant difference between suppliers in terms of production
volume of the product.

Solution. The model of the Latin square design of this


example is as follows:
Yijk = + Mi + Oj + Sk + eijk
Where is the overall mean; Yijk, the observation in the ith row
(machine) and kth column (supplier) for the jth treatment of the
factor (Operator); Mi, the effect of the ith block representing
machine. OJ, the effect of the jth treatment of the factor
(Operator); Sk, the effect of the kth block representing
supplier; and eijk the random error associated with the ith row
(machine) and the kth column (supplier) for the jth treatment of
the factor ( operator).

Hypothesis with respect to treatment (operator):


Null hypothesis, Ho: 01 = O2 = 03 = 04 = 05
Alternate hypothesis, H1: Treatment means are not
equal for at least one pair of treatment means.

Hypothesis with respect to rows (machine):


Null hypothesis, Ho: M1 = M2 = M3 = M4 = M5
Alternate Hypothesis, H1: Row means are not equal for
at least one pair of row means.
Hypothesis with respect to columns (supplier):
Null hypothesis, H0: S1 = S2 = S3 = S4 = S5
Alternate hypothesis, H1 : Column means are not equal
for at least one pair of column means.
The data of Table 4.24 is reproduced in Table 4.25 with
row totals and column totals.

Table 4.25 Hourly Production Volume in Latin Square Design {Example 4.6)
Supplier

Machine

S1

S2

S3

S4

S5

Yi..

M1

12 (A)

8 (B)

9 (C)

15 (D)

18 (E)

62

M2

14 (8)

7 (C)

13 (D)

10 (E)

12 (A)

56

M3

7 (C)

15 (D)

12 (E)

8 (A)

6 (B)

48

M4

16 (D)

17 (E)

13 (A)

11 (B)

12 (C)

69

M5

9 (E)

14 (A)

8 (B)

10 (C)

8 (D)

49

Y..k

58

61

55

54

56

284

For theexample under c onsideration.

Y
Y
Y

...

Y
i

284

ijk

1..

62, Y 2.. 56, Y 3.. 48, Y 4.. 69and Y 5.. 49

..1

58, Y ..2 61, Y ..3 55, Y ..4 54and Y ..5 56

Then, w e get

SS

total

Y
i

2
ijk

25

2
...

(12) (8) (9) (15) (18) (14) .... (10) (8)


271.76
2

Y
5

SS

machine

(62)

i 1

1..

2
...

25

(56) (48) (69) (49)

(284)

25

62.96
Y
5

SS

supplier

(58)

6.16

k 1

2
..k

2
...

25

(61) (55) (54) (56)


5

(284)

25

(284)
25

The data of Table 4.24 is rearranged as shown in Table 4.26, to


compute the sum of squares of 'operator'.
Table 4.26 Rearranged Hourly Production Volume in Latin
Square Design (Example 4.6)
Operator

Y.1.

12

15

18

12

14

13

10

15

12

13

11

12

16

17

14

10

59

47

45

67

66

We get

59,Y .2. 47,Y .3. 45,Y .4. 67,Y .5. 66

.1.

Then,
5

SSoperator
i 1

Y ...
5
25
. j.

(47) (45) (67) (66) (284)


(
59
)

25

85.76

SS

error

SStotal SSm achine SSoperator SSsupplier

271.76 62.96 85.76 6.16


116.88
The ANOVA results of this problem are summarisedin Table 4.27.

Table 4.27 Results in Latin Square Design (Example 4.6)

Source of
variation

Degrees
of
freedom

Sum of
squares

Mean sum of
squares (MSS)

F ratio

Between
operators

85.76

85.76/4 = 21.44

21.44/9.74 = 2.20

Between
machines

62.96

62.96/4 = 15.74

15.74/9.74 = 1.62

Between suppliers

6.i6

6.16/4 = 1.54

1.54/9.74 = 0.16

12

116.88

24

271.76

Error
Total

116.88/12 = 9.74

In Table 4.27, the value of the calculated F ratio for 'operator' is 2.2, whereas its table value
with the significance level of 0.05 and degrees of freedom (4,12) is 3.26. The calculated F ratio
for 'machine' is 1.62, whereas its table value with the significance level of 0.05 and degrees of
freedom (4,12) is 3.26. The calculated F ratio for 'supplier' is 0.16, whereas its table value with
the significance level of 0.05 and degrees of freedom (4,12) is 3.26.
Component-treatment (operator): For this component,
Fcalculated [2.2] < Fa=0.05 and dj=(4,12) [3.26]
Hence, the null hypothesis, Ho should be accepted.
Inference: This means that there is no significant difference between operators in terms
of hourly production volume.
Component-rows (machines): For this component,
Fcalculated [1.62] < Fa=0.05 and dj=(4,12) [3.26]
Hence, the null hypothesis, (Ho) should be accepted.
Inference: This means that there is no significant difference between machines in terms
of hourly production volume.
Component-columns (supplier): For this component,
Fcalculated [0.16] < Fa = 0.05 and 4f= (4,12) [3.26]
Hence, the null hypothesis, (Ho) should be accepted.
Inference: This means there is no significant difference between suppliers in terms
of hourly production volume.

Example 4.7 Consider Example 4.2 again for a quick reference.


(a) Perform ANOVA to check whether there is significant difference
between the . training organizations in terms improving the decisionmaking skills of the employees by assuming a significance level of
0.05.
(b) If there is significant different between the treatment means of the
factor, Training Organization, then perform Duncan's Multiple Range
Test and comment on the results.

Solution.
(a) The ANOVA results has already been discussed earlier (refer to
Table 4.10).
(b) Since the null hypothesis is rejected, for further interpretation on the
difference between different treatment means, Duncan's multiple
range test is performed as follows:

Step-1:

The ascending order of treatment means are:


_

53.6

72.0

84.2

87.8

.1

.2

.3

Step-2

.4

The standard error of each treatment mean is

SE

Y
.j

MSS
n

error

112.43

4.74
5

Step-3: Obtain 3 (Le. a-I) significant ranges from the Duncan's table of
multiple ranges for j = 4 and df = 16 and with a = 0.05 as shown
in Table 4.28.
Table 4.28 Significant Ranges
j

Significant range

3.00

3.15

3.23

Step-4: Find the least significant ranges (LSR) by multiplying the


significant ranges with the standard error as shown in Table
4.29.
Table 4.29 Least Significant Ranges (LSR)

LSR

3.00 x 4.74 = 14.22

3.15 x 4.74 = 14.93

3.23 x 4.74 = 15.31

Table 4.30 Summary of Comparison of Duncan's Multiple Range Test


Treatment
pair

Actual difference
between means

Least significant
range

3 and 4

34.2

15.31

Significant

2 and 4

15.8

14.93

Significant

1 and 4

3.6

14.22

Not significant

3 and 1

30.6

14.93

Significant

2 and 1

12.2

14.22

Not significant

3 and 2

18.4

14.22

Significant

Remark

The comparison of differences between different pairs of treatment


means with the corresponding least significant ranges are summarized
in Table 4.30. From Table 4.30, it is clear that the actual difference
between the means of the treatment pairs (3, 4), 2, 4), (3, 1) and (3, 2)
are significant. For the remaining pairs of treatments [(1, 4) and (2, 1)],
the actual differences between their means are not significant.

FACTORIAL DESIGN
If the number of factors are more than one, then there is a
need for a generalized design of experiment. Such design
is called as factorial design. Analysis of many of the real
life situations will require this type of design.
The concept of factorial design is demonstrated with a
problem involving as in Example 4.8.
Example 4.8 The sales manager of Krishna Engineering
Company wants to study effect of sales regions as well as
salesmen on quarterly sales revenue. So, he has designed
a factorial experiment involving two factors as shown in
Table 4.31, with two replications for each experimental
combination:

Table 4.31 Quarterly Sales of Krishna Engineering Company

Salesman (B) (Rupees in lakhs)

East

65

30

50

20

70

35

65

40

30

60

20

50

35

55

25

40

Sales region (A)

West

As per this design each salesman is assigned to each


region for two different quarters. Perform ANOVA and
check the following at the significance level of 0.05.
a) Whether there is a significant difference in terms of
sales revenue between sales regions,
b) Whether there is a significant difference in terms of

sales revenue between salesmen.


c) Whether there is a significant difference in terms of
sales revenue between different combinations of sales

region and salesman.


Solution. The data in Table 4.31 are presented in Table
4.32 along with row totals and column totals.

Table 4.32 Quarterly Sales of Krishna Engineering Company

Salesman (B) (Rupees in lakhs)

A1

B1

B2

B3

B4

65

30

50

20

70

35

65

40

30

60

20

50

35

55

25

40

200

180

160

150

375

Sales region (A)


A2

Y.j.

315

690

This problem has two factors, namely sales region (A) and salesman (B).

The number of treatments for the factor A is 2 (A1 and A2) and that of the
factor B is 4 (B1, B2, B3 and B4). The number of replications for each of the
eight experimental combinations (A1B1, A1B2, A1B3, A1B4, A2B1, A2B2, A2B3
and A2B4) is 2.

The generalized model of the factorial design is as follows:


Yijk = + Ai + BJ + ABij + eijk
where a is the number of treatments of the factor A; b, the number of
treatments of the factor B; n, the number of replications under each of the
experimental combinations of the factor A and the factor B; , the overall
mean; Yijk the kth replication under the ith treatment of the factor A and the jth
treatment of the factor B; Ai, the effect of the ith treatment of the factor A; Bj'
is the effect of the jth treatment of the factor B. ABij is the effect of the ith
treatment of the factor A and the jth treatment of the factor B (interaction
effect); and eijk, the random error associated with the kth replication under
the ith treatment of the factor A and the jth treatment of the factor B.

Hypothesis with respect to treatment of the factor A:


Null hypothesis,

Ho: A1 = A2 = A3 =... = Aa

Alternate hypothesis,

H1: Treatment means are not equal for at least


one pair of the treatment means of the
factor A.

Hypothesis with respect to treatment of the factor B:


Null hypothesis,

Ho: B1 = B2 = B3 =... = Bb

Alternate hypothesis,

HI: Treatment means are not equal for at least


one pair of the treatment means of the

factor B.
Hypothesis with respect to interaction component, AB:
Null hypothesis,

Ho: A1B1 = A1B2 = A1B3 =... = A2B3 = A2B4

Alternate hypothesis,

H1: Interaction means are not equal for at least

one pair of interaction means.


The relationship between different sum of squares of this model is:
Total sum of squares = (Sum of squares of the factor A)
+ (Sum of squares of the factor B)

+ (Sum of squares of the interaction AB)


+ (Sum of squares of errors)

This is,
SStotal SSA SSB SSAB SSerror
for theexample problem,
SStotal SSregion SSsalesman SSregion x salesman SSerror
The generalized shortcutformulas to compute the sum of squares of different
componentsof the model are as follows :
2
a b
n
Y ...
2
SStotal Y ijk
N
i 1 j 1k 1
a 2
2
Y
SSrows i.. Y ...
N
i 1 bn
b
2
Y2
SScolumns .i. Y ...
j 1 an N
a b Y2
2
SSsubtotals ij. Y ...
i 1 j 1 n
N
SSinteraction SSsubtotals SSrows SScolumns
SSerror SStotal SSrows SScolumns SSinteraction
where Y... is the sum of Yij overall values of i, j and k; Y.j ., the sum of Yij k overall values of i
and k for a given j; Yi.., the sum of Yij k overall vaues of j and k for a given i; Y..k,
the sum of Yij k overall values of i and j for a given value of k; and N is the total numner of
observations in the experiments (abn 2 x 4 x 2 16).

For theexample under consideration, w e have


2
4
2

YY ...
Y ijk 690
i 1 j 1k 1
Y1.. 375, Y1.. 375,
Y.1. 200, Y.1. 180, Y.3. 160, and Y.4. 150,
Then,
2
4
2
2
2
SStotal Y ijk Y ...
i 1 j 1k 1
i 1 16
(6902
2
2
2
2
2
2
2
2
(65) (30) (50) (20) (70) (35) ....(25) ( 40)
16
4093.75
2
2 Y2
i.. Y ...
SSA
16
i 1 8
(375) 2 (315) 2 (690) 2

8
16
225
4 Y2
2
SSB . j. Y ...
16
i 1 4
(200) 2 (180) 2 (160) 2 (150) 2 (690) 2

4
16
368.75

Table 4.32 is rearranged as in Table 4.33 to compute the


sum of squares of subtotals.
Table 4.33 Data to Compute Sum of Squares of Subtotals
A 1B 1

Yi

A 1B 2

A 1B 3

A 1B 4

A 2B 1

A 2B 2

A 2B 3

A 2B 4

65

30

50

20

30

60

20

50

70

35

65

40

35

55

25

40

115

60

65

115

45

90

135

65

Then, we get
2
2 4 Y ij
Y ..2

SSsubtotals
i 1 j 1 2 16
(690) 2
(135) 2 (65) 2 (115) 2 (60) 2 (65) 2 (115) 2 (45) 2 (90) 2

2
16
3668.75
SSAB SSsubtotals SSA SSB
3668.75 225 368.75
3075
SSerror SStotals SSA SSB SSAB
4093.75 225 368.75 3075
425
The generatlised results and results applied to the example problem of
this design are summarized in Table 4.34 and Table 4.35, respectively.

Table 4.34 Generalized Results in Factorial Design


Source of
variation
Between
rows
Between
columns

Degrees
of
freedom

Sum of
squares

a-1

SSrows

b-1

SScolumns

Between
Rows x

(a-1)(b-1)

SSrows x columns

Error

ab(n -1)

SSerror

Total

abn-1

SStotal

columns

Mean sum of
squares
(MSS)

F ratio

SSrows

MSSrows

a-1

MSSerror

SScolumns

MSScolumns

b -1

MSSerror

SSrows x columns

MSSrows x columns

(a -1)(b - 1)

M SSerror

SSerror
ab(n-1)

Table 4.35 Results in Factorial Design (Example 4.8)


Source
of
variation

Degrees of
freedom

Sum of
squares

Mean sum of
squares (MSS)

F ratio

SSA

225.00

225.00/1=225.00

225.00/53.13 =4.24

SSB

368.75

368.75/3=122.92

122.92/53.13 =2.31
1025.00/53.13 =

SSAB

3075.00

SSerror

425.00

Total

15

4093.75

3075/3 = 1025.00

19.29
425/8 = 53.13

In Table 4.35, the value of the calculated F ratio for the factor A is 4.24,
whereas its table value with the significance level of 0.05 and degrees of
freedom (I, 8) is 5.32. The calculated f ratio for the factor B is 2.31,
whereas its table value with the significance level of 0.05 and degrees of
freedom (3, 8) is 4.07. The calculated f, ratio for the interaction AB is
19.29, whereas its table value with the significance level of 0.05 and
degrees of freedom (3, 8) is 4.07.
Component-factor A: For this component,
Fcalculated[4.24] < Fa=O.O5 and df=(1,8) [5.32]
Hence, the null hypothesis (Ho) should be accepted.
Inference: This means that there is no significant difference in terms
of quarterly sales between different sales regions.
Component-factor B: For this component,
Fcalculated[4.24] < Fa=O.O5 and df=(1,8) [5.32]
Hence, the null hypothesis (Ho) should be accepted.

Inference: This means that there is no significant. difference in


terms of quarterly sales between different salesmen.

Component - interaction, AXB: For this component,


Fcalculated [19.29] > Fa = 0.05 and df= (3.8) [4.07]
Hence, the null hypothesis (H0 should be rejected.

Inference: This means that there is a significant difference in


quarterly sales because of the interaction between the factor
sales region' and the factor 'salesman'. Alternatively, this means

that the interaction between 'sales region' and 'salesman' has


significant effect on quarterly sales.

Example 4.9 The quality control department of a fabric finishing plant is studying
the effect of several factors on the dyeing of cotton-synthetic cloth used to
manufacture men's shirts. Two operators, three cycle times and two
temperatures were selected and two small specimens of the cloth were
dyed under each combination of experimental conditions. The finished cloth
was compared to a standard and a numerical score was assigned. The results
are shown in Table 4.36. Analyze the data and draw conclusions at the
significance level of 0.05.
Table 4.36 Numerical Scores of Specimens of Cloth

Temperature
2500

3000

Operator-1

Operator-2

Operator-1

Operator-2

23

27

24

38

25

26

28

35

36

34

37

34

36

39

35

36

28

35

26

36

27

34

25

34

40

Cycle time

50

60

Solution : The data in Table 4.36 is presented in Table 4.37 along with row totals and ; column
totals. This problem has three factors: cycle time (A), temperature (B) and operator (C).
Let a be he number of treatments for the factor A (a=3); b, the number of treatments 1 for the
factor B, (b=2); c be the number of treatments for the factor C (c=2); and n be the number of
replications under each experimental combination (n = 2).
Table 4.37 Numerical Scores of Specimens of Cloth
Temperature (B)
2500
3000
1
2
1
2
40

Cycle time (A)

50

60

23

27

24

38

25

26

28

35

36

34

37

34

36

39

35

36

28

35

26

36

27

34

25

34

370

388

Y.j..
Y..k.

350

408

Yi
226

287

245

The generalized model of the factorial design with three factors is as


follows:
Yijkl = + A1+ Bj + ABij + Ck + ACik + BCjk + ABCijk + eijkl
where is the overall mean; Yijkl, the lth replication for the ith treatment
of the factor A, jth treatment of the factor B and kth treatment of the
factor C, A1, the effect of the ith treatment of the Factor A; Bj, the effect
of the jth treatment of the factor A; ABij, the effect of the ith treatment of
the factor A and the jth treatment of the factor B (two-w interaction
effect); Cb the effect of the kth treatment of the factor C; ACik, the effect
of the ith treatment of the factor A and the kth treatment of the factor C
(two-way interaction); BCjk the effect of the jth treatment of the factor B
and the kth: treatment of the factor C (two-way interaction); ABCijk, the
effect of the ith treatment of the factor A, jth treatment of the factor B
and the kth treatment of the factor C (three-way- interaction); and eijkl
the random error associated with the lth replication under the ith
treatment of the factor jth treatment of the factor Band kth treatment of
the factor C.

Hypotheses:- As; in the previous example, one can define H0 and H1


for each of the seven components (Ai + Bj + ABij + Ck + ACik + BCjk +
ABCijk) in the model.
The relationship between different sum of squares of this model is:
Total sum of squares = (sum of squares of the factor A)
+ (sum of squares of the factor B)
+ (sum of squares of the interaction AB)
+ (sum of squares of the factor C)
+ (sum of squares of the interaction AC)
+ (sum of squares of the interaction BC)
+ (sum of squares of the interaction ABC)
+ (sum of squares of the errors)

That is,
SS total SS A SS B SS AB SS C SS AC SS BC SS ABC SS error
The generalized shortcutformulas to compute the sum
of squares of different componentsof the model are as follows :
a

2
SS totali1 j1k1 l1Y ijkl
a Y2

i ...
SS A i1 bcn

SS

2
Y ....
N

2
Y ....
N

b Y .2j .. Y 2

....
B
N
j 1 acn
2
2
b Y ij
.. Y ....

cn
N

SS AB(subtotals) i1 j1

SSAB SSAB(subtotals) SSA SSB


SS

c Y2
Y2
.. k . ....
C
abn
N
k 1
2
c Y2
i . k . Y ....
bn
N

SS AC(subtotals) i1k1

SSAC SSAC(subtotals) SSA SSC


c Y .2jk . Y 2
....
an
N

SS BC(subtotals) j1k1

SSBC SSBC(subtotals) SSB SSC


a

2
2
c Y ijk
. Y ....
N

SS ABCsubtotals i1 j1k1 an

SS ABC SS ABC(subtotals) SS A SS B SS AB SS C SS AC SS BC
SS error SS total SS A SS B SS AB SS C SS AC SS BC SS ABC

Where N is the total number of obervations in the experiments (abcn 3x2x2x2 24)
For theexample under consideration, w e have
3
2
2
2
Y ....
Y ijkl 758
i 1 j 1 k 1l 1
Y 1... 226, Y 2... 287, Y 3... 245,
Y .1.. 370, Y .2.. 388
Y ..1. 350, Y ..2. 408
Therefore,
2
3
2
2
2
2
SStotal
Y ....
Y ijkl
24
i 1 j 1 k 1l 1
2

23 27 24 38 25 26 ....... 25 34
609.834
2
3 Y2
i... Y ....
SS A
24
i 1 8
2262 2872 2452

8
243.584
2
3 Y2
Y ....
.
i
..

SS B
24
j 1 12

370 388 758


2

12

24

13.5
2
2 Y2
..k . Y ....
SS C
24
k 1 12
3502 4082 7582

12
24

140.166

758

24

758
24

Computation of SSAB: To compute the sum of squares of the interaction AB,

Table 4.37 is rearranged, as shown in Table 4.38.


Table 4.38 Rearranged Data to compute sum of Squares of AB
Temperature (B)
2500
40

23

101

25
36

Cycle time (A)

50

145

36
60

28
27

124

3000

27

24

26

28

34

37

39

35

35

26

34

25

125

38
35

142

34
36

121

Note : Subtotals of AB are given in the respective cells.

36
34

We also have,
SS

2
2
3 2 Y ij
.. Y ....

AB(subtotals)
24
i 1 j 1 4

2
2
2
2
2
125 1452 142 124 121
758

4
24
317.834
SSAB SSAB(subtotals) SSA SSB

101

317.834 243.584 13.5


60.75
Computationof SSAC : To compute the sum of squares of the interaction AC,
Table 4.37 is rearranged, as shownin Table 4.39

Table 4.39 Rearranged Data to compute sum of Squares of AC


Temperature (C)
1
40

23

100

25
Cycle time (A)

50

36

144

36
60

28
27

106

27

24

26

28

34

37

39

35

35

26

34

25

Note : Subtotals of AC are given in the respective cells.

126

38
35

143

34
36

139

36
34

We have,
SS

3 2 Y2
Y2
i .k . ....
AC(subtotals)
24
i 1 j 1 4

100

2
2
2
2
126 (144) 2 143 106 139
4

758

24

464.334
SSAC SSAC(subtotals) SSA SSC
464.334 243.584 140.166
80.584
Computationof SSBC : To compute the sum of squares of the interaction BC,
Table 4.37 is rearranged, as shown in Table 4.40
Table 4.40 Rearranged Data to Compute Sum of Squares of BC
Temperature (B)
2500
Operator(C )

Y.jk.

3000
Operator(C )

23

27

24

38

25

26

28

35

36

34

37

34

36

39

35

36

28

35

26

36

27

34

25

34

175

195

175

213

Then,
3

2
2 Y2
.ik . Y ....
6
24

SS BC(subtotals) i1 j1

175

2
2
195 (175) 2 213
6

758

24

167.166
SSBC SSBC(subtotals) SSB SSC
167.166 13.5 140.166
13.5
Computationof SSABC : To compute the sum of squares of the interaction ABC,
Table 4.37 is rearranged, as shownin Table 4.41

Table 4.41 Rearranged Data to compute sum of Squares of ABC


Temperature
Operator-2
Operator-1
40

23

48

25
Cycle time (A)

50

36

28
27

53

26
72

36
60

27

34

35
34

24

52

28
73

39
55

Operator-1

37

26
25

38

73

35
72

35
69

Operator-2

34

70

36
51

36
34

70

Then,
SS

2
2
3 2 2 Y ijk
. Y ....

BC(subtota ls)
24
i 1 j 1k 1 2

2
2
2 73 2 72 2 73 2 .... 51 2 70 2
48

53

(
52
)



2

574.834
SSABC SSABC(subtotals) SSA SSB SSC SSAB SSAC SSBC
574.834 243.584 13.5 140.166 60.75 80.584 13.5
22.75
SSerror SStotal (SSA SSB SSAB SSC SSAC SSBC SSABC
609.834 243.584 13.5 60.75 140.166 80.584 13.5 22.75
35

758
24

Example 4.10 The surface finish of products produced in a


machine shop is suspected to be affected by a factor
operator and another factor machine. The data of
this experiment with two replications in different treatment
combinations are summarized in Table 4.44. perform
ANOVA and check the significance of the components of
the related model when a is 0.05.
Table 4.44 Data (Example 4.10)
Machine (B)
1

65
70

20
40

30
35

50
40

Operator (A)

Solution. Let, the factor, operator be A and the another factor machine
be B. here, all the factors have been assumed as fixed factors. Numbers
of levels for each of the factors, n=2. Number of replications in each
treatment combination, k=2. significance level, a=0.05.
The model of this problem is as shown below:
Yijk = + Ai+ Bj + ABij + eijk
The definitions of the components of this model are as already explained
in this section. The data in Table 4.44 are reproduced in table 4.45 with
necessary notations and the totals of different treatment
combinations(1,a,b,ab)
Table 4.45 Data with Necessary Information
Machine (B)
1
2
1

65
70

135

20
40

60

30
35

65

50
40

ab

90

Operator (A)

Then,
Sum of squares of A

(ContrastA) 2
4k

135 65 60 90

(1 a a ab) 2
4 x2

200
Sum of squares of B

(ContrastB) 2

135 65 60 90

4k

(1 a b ab) 2
4 x2

312.5
Sum of squares of AB

(ContrastAB) 2
4k

(1 a b ab) 2 135 65 60 90

4 x2
8
1250
2
2 Y ...
Total sum of squares Y ijk
k 22
(350) 2
17350
8
2037.5

Sum of squares of error SS total SS A SS B SS AB


2037.5 200 312.5 1250
275
The ANOVA table for this problem is shownin Table 4.46.

Table 4.46 Results of ANOVA (Example 4.10)


Source of
variation

Degrees of
Freedom

Sum of
squares

Mean sum of
squares (MSS)

F Ratio

SSA

200.0

200/1 = 200

2.910

SSB

312.5

312.5/1 = 321.5

4.676

SSAB

1250.0

1250/1 = 1250

18.182

SSerror

275.0

275/4 = 68.75

Total

2037.5

The F value for (1,4) degrees of freedom at a=0.05 is 7.71. based on


this value, one can conclude that the factor a a sell as factor B, which is
not having effect on the response variable, (surface4 finish), but the
interaction effect AB has a significant effect on the response variable.
So, the operator as well as machine is not having effect on the
surface finish; but the interaction between operator and machine
has effect on the surface finish.

You might also like