Design of experiments

UNIT IV
FACTORIAL EXPERIMENTS
In Randomised Block Design (RBD) or in Latin Square Design (LSD) the testing of a number of
treatments not necessarily related were tested. In industrial applications frequently that several
factors may affect the characteristic in which we are interested, and wish to estimate the effects of
each of the factors and how the effect of one factor varies over the level of other factors.
For example, the yield of a chemical process may be affected by several factors such as the levels
of pressure, temperature, rate of agitation, and proportions of reactants etc. One might try to test
each of the factors separately, holding all other factors constant in a given experiment. The logical
procedure would be to vary all factors simultaneously, i.e., within framework of the same
experiment. This is called as factorial experiment.
The factorial experiments are particularly useful in experimental situations which require the
examination of the effects of varying two or more factors.
In CRD, RBD and LSD the comparison and estimation of the effects of a single set of treatments
like varieties of wheat, manure, of different methods of cultivation etc,. such experiments which
deal with only one factor only may be called simple experiments.
In factorial experiment the effect of several factors of variation are studied and investigated
simultaneously, the treatments being all the combinations of different factors under study. In these
experiments an attempt is made to estimate the effects of each of the factors and also interaction
effects i.e., the variation in the effect of one factor as a result to different levels of other factors.
As a simple illustration let us consider two fertilizers, say, Potash (K) and Nitrogen (N). Let us
suppose that there are p different varieties of Potash and q different varieties of Nitrogen. P and q
are termed as the levels of factors Potash and Nitrogen respectively. To find the effectiveness of
various treatments, viz the different levels of Potash or Nitrogen we might conduct two simple
experiments one for Potash and one for Nitrogen. A series of experiments in which only one factor
is varied at a time would be both lengthy and costly. These simple experiments do not give us any
information regarding the dependence or independence of one factor on the other, i.e., they do not
tell us about the interaction effect (NK). The only alternative is try to investigate the variations in
several factors simultaneously by conducting the above experiment as p x q factorial experiment,
where p and q are the levels of various factors under consideration. In general, if the levels of
various factors are equal then sn factorial experiment means an experiment with n factors each at
s levels where n is any positive integer greater than or equal to 2, for example2 3 experiment
means an experiment with 3 factors each at 2 levels and 32 experiment means an experiment with 2
factorsat 3 levels each.
ADVANTAGES OF FACTORIAL EXPERIMENT
1. It increases the scope of the experiment and its inductive value by giving information not
only on the main factors but on their interactions.
2. The various levels of one factor constitute replication of other factors and increase the
amount of information obtained on all factors.
3. When there are no interactions, the factorial design gives the maximum efficiency in the
estimates of the effects.
4. When interactions exist, their nature being unknown a factorial design is necessary to avoid
misleading conclusions.
5. In the factorial design the effect of a factor is estimated at different levels of other factors
and the conclusions hold over a wide range of conditions.

22 FACTORIAL DESIGN
Here we have two factors each at two levels (0,1), say, so that there are 2 x 2 = 4 treatment
combinations in all. Following the notations due to Yates let the capital letters A and B indicate
the names of two factors under study and let the small letters a and b denote one of the two
levels of each of the corresponding factors and this will be called the second level. The first
level of A and B is generally expressed by the absence of the corresponding letters in the
treatment combinations, the four treatment combinations can be given as follows
a0b0 or ‘1’ : factors A and B both at first level
a1b0 or a : factor A at second level and B at first level
a0b1 or b : factor A at first level and B at first level
a1b1 or ab : factors A and B both at second level
Main effects and interaction effects
Suppose the factorial experiment with 22 = 4 treatments is conducted in r blocks or replicates. Let
[1],[a],[b],and [ab] denote the total yields of the r units(plots)receiving the treatments 1, a, b, and
ab respectively and let the corresponding mean values obtained on dividing these totals by r be
denoted by (1), (a), (b) and (ab) respectively.
The effect of factor A at the first level b0 of B = (a1b0) – (a0b0) = (a) – (1) ...............(I)
Similarly,
the effect of A at the second level b1 0f B = (a1b1) – (a0b1) = (ab) – (b) ............... (II)
These two effects given in (I) and (II) are called simple effects of factor A.
The average observed effect of A over the two levels of B is called the main effect due to A and
is defined by A = ½ [(ab) – (b) + (a) – (1)]
A = ½ [ b(a - 1) + (a–1)] =1/2 [(a – 1) (b + 1)]
Where the right-hand side is to be expanded algebraically and then the treatment combinations are
to be replaced by treatment means.
similarly the main effects of factor B is
B = ½ [(ab) – (a) +(b) – (1)] or
B = ½ [(a + 1) (b – 1)]
Where the right-hand side is to be expanded algebraically and then the treatment combinations are
to be replaced by treatment means.
The interaction of two factors is the failure of the levels of one factor, say, A to retain the same
order and magnitude of performance throughout all levels of the second factor, say, B. if the two
factors act independently of one another, the true effect of one would be the same at either level
of other. If the two factors are not independent, the two expressions in (I) and (II) will not be the
same and the difference of these two numbers is a measure of the extent to which the factors
interact and the two-factor interaction or the first order interaction between the factors A and
B as:
AB = ½ [(ab) – (b) – (a) + (1)] or
AB = ½ [(a – 1) (b – 1)]
Where the right-hand side is to be expanded algebraically and then the treatment combinations are
to be replaced by treatment means.
CONTRAST
Definition: A linear combination ∑𝑖=1
𝑘
𝑐𝑖𝑡𝑖 of k treatment means ti (i = 1,2, … ,k) is called a
contrast(or a comparison) of treatment means if ∑𝑘𝑖=1 𝑐𝑖 = 0. In other words, contrast is a linear
combination of treatment means such that the sum of the co-efficients is zero.
ORTHOGONAL CONTRASTS
Definition: Two contrasts of k-treatment means ti, (I =1,2, … ,k)

∑𝑘𝑖=1 𝑐𝑖𝑡𝑖 , ∑𝑘𝑖=1 𝑐𝑖 = 0.

and ∑𝑘𝑖=1 𝑑𝑖𝑡𝑖 , ∑𝑘𝑖=1 𝑑𝑖 = 0.

are said to be orthogonal if ∑𝑘𝑖=1 𝑐𝑖𝑑𝑖 = 0.

Let M denote the mean yield of the four treatment combinations. Then
M = ¼ [(ab) + (a) + (b) + (1)] = ¼ [(a+1)(b+1)]
The general mean M and three mutually orthogonal contrasts viz., the main effects A and B and
the interaction AB are summarized in terms of treatment combinations in the following table.
The signs for the mean effects are written according to the following rule. “ give a
plus sign to each of the treatments means whenever the corresponding factor is at
the second level, otherwise a negative sign. For a two-factor interaction, the signs to
be attached to various treatment means are obtained by multiplying the signs of the
corresponding main effects.
TABLE OF SIGNS AND DIVISORS GIVING M, THE MAIN EFFECTS AND
INTERACTIONS IN TERMS OF INDIVIDUAL TREATMENT MEANS

Factorial Treatments Divisor

experiments (1) (a) (b) (ab)
M + + + + 4
A - + - + 2
B - - + + 2
AB + - - + 2

YATES METHOD OF COMPUTING FACTORIAL EFFECT TOTALS

For calculation of various effect totals for 2n factorial experiment F.Yates developed
a special computational rule. Yates’ method consists of the following steps:
1. In the first column we write the treatment combinations. It is an essential that
the treatment combinations should be written in a systematic order as
explained below:
“ starting with the treatment combination 1, each factor is introduced in turn
and then it is followed by all the combinations of itself with the treatment
combinations previously written. For example, for 22 experiment with factors
A and B the order of treatment combinations will be 1, a, b, ab and for 2 3
factorial experiment with factors A,B and C, the order of treatment
combinations will be 1, a, b, ab, c, ac, bc, abc and so on.
2. Against each treatment combination, write the corresponding total yields from
all the replicates.
3. The entries in the third column can be split into two halves. The first half is
obtained by writing down in order, the pairwise sums of the values in column
2 and the second half is obtained by writing in the same order the pairwise
differences of the values in the second column. The first member is to be
subtracted from the second member of a pair.
 4. To complete the next (4th) column, the whole of the whole of the procedure as
explained in step 3 is repeated on column 3, and for 23 design, the 5th column
is derived from 4th in a similar manner.
Thus for a 2n factorial experiment there will be n cycles of this ‘sum and
difference’ procedure. The first term in the last, viz., (n+2)th column always gives
the grand total (G) while the other entries in the last column are the totals of the
main effects or the interactions corresponding to the treatment combinations in
the first column of the table.
YATES’METHOD FOR A 22 EXPERIMENT
Treatment Total yield from
Combination all replicates Effect
(1) (2) (3) (4) totals
‘1’ [1] [1]+[a] [1]+[a]+ [b]+[ab] Grand total
a [a] [b]+[ab] [ab] – [b] + [a] – [1] [A]
b [b] [a] – [1] [ab] + [b] - [a] – [1] [B]
ab [ab] [ab] – [b] [ab] – [b] - [a] + [1] [AB]

YATES’METHOD FOR A 23 EXPERIMENT

Treatment Total yield
Combination from all Effect totals
(1) replicates (3) (4)
(2)
‘1’ [1] [1]+[a]= u1(say) u1+u2=v1 v1+v2=w1 Grand total
a [a] [b]+[ab]= u2(say) u3+u4=v2 v3+v4=w2 [A]
b [b] [c] + [ac]= u3(say) u5+u6=v3 v5+v6=w3 [B]
ab [ab] [bc] + [abc]= u4(say) u7+u8=v4 v7+v8=w4 [AB]
c [c] [a] - [1]= u5(say) u2 - u1=v5 v2 - v1=w5 [C]
ac [ac] [ab]-[ b]= u6(say) u4 - u3=v6 v4 - v3=w6 [AC]
bc [bc] [ac] - [c]= u7(say) u6 - u5=v7 v6 - v5=w7 [BC]
abc [abc] [abc] - [bc]= u8(say) u8 - u7=v8 v8 - v7=w8 [ABC]

STATISTICAL ANALYSIS OF 22-DESIGN.

Factorial experiments are conducted either using CRD or RBD or LSD and thus they
can be analyzed in the usual manner except that in this case treatments S.S. is split
into three orthogonal components each with 1 d.f. . The main effects A and B and
the interaction AB are mutually orthogonal contrasts of the treatment means.
Null Hypothesis:
H0 : the blocks as well as treatments are homogeneous.
H1 : the blocks as well as treatments are not homogeneous.

Total S.Sq. = ∑𝑖 ∑𝑗(𝑦𝑖𝑗 − 𝑦̿.. )2 – C.F.

Block sum of squares = 1/𝑘 ∑𝑖 𝑩2𝑖 – C.F.
Treatment sum of squares = 1/𝑟 ∑𝑗 𝑻𝑗2 – C.F.
Error Sum of Squares = Total S.Sq. – block sum of squares – treatments sum
of squares
The S.S. due to any factorial effect is obtained on multiplying the square of the effect
total by the factor (1/4r), where r is the common replication number. Thus,
S.S due to main effect of A = [A]2 / 4r , where [A] = [ab] – [b] + [a] – [1]
S.S due to main effect of B = [B]2 / 4r, where [B] = [ab] + [b] – [a] – [1]
S.S due to main effect of AB = [AB]2 / 4r, where [AB] = [ab] – [b] – [a] + [1]
Degrees of freedom
For blocks, we have (r – 1) d.f.
For main effects A, B and interaction AB each 1 d.f.
For error 3(r – 1) d.f.
ANOVA table for fixed effect model two factor (22) experiment in RBD in r replicates
Sources of d.f. S.S. M.S.S. Variance Ratio
variation ‘F’
Blocks(replicates) r–1 S2R s2R = S2R/(r-1) FR = s2R/ s2E
Main effect A 1 S2A=[A]2 / 4r s2A= S2A FA = s2A/ s2E
Main effect B 1 S2B=[B]2 / 4r s2B= S2B FB = s2B/ s2E
Interaction A x B 1 S2AB=[AB]2 / 4r s2AB= S2AB FAB = s2AB / s2E
Error 3(r-1) S2E= by subtraction s2E= S2E/[3(r-1)]
Total 4r – 1 ∑ ∑(𝑦𝑖𝑗 − 𝑦̿.. )2
𝑖 𝑗

CONCLUSION
For a certain level of α
If FR ≤ F(r – 1), 3(r-1) accept H0, otherwise reject H0
If FA, FB and FAB ≤ F1, 3(r-1), the presence of the factorial effect is accepted otherwise
rejected.
23 FACTORIAL EXPERIMENT
In 23 factorial experiment we consider three factors, say, A,B, and C each at two
levels, say, (a0,a1), (b0,b1) and (c0,c1) respectively, so that there are 23 = 8 treatment
combinations in all.
The eight treatment combinations in a standard order are
‘1’, a, b, ab, c, ac, bc, abc
where 1=a0b0c0, a= a1b0c0, b= a0b1c0, ab= a1b1c0, c= a0b0c1, ac= a1b0c1, bc = a0b1c1,
abc= a1b1c1
In 23 factorial experiment we split up the treatment sum of squares with 7d.f. into
orthogonal components corresponding to the three main effects A, B and C, three
first order (or two factor) interactions AB, AC and BC and one second order
interaction (or three factor interaction) ABC, each carrying 1 d.f.
YATES’ TABLE FOR 23 FACTORIAL EFFECTS
Factorial Treatment mean Divisor
effect (1) (a) (b) (ab) (c) (ac) (bc) (abc)
M + + + + + + + + 8
A - + - + - + - + 4
B - - + + - - + + 4
C - - - - + + + + 4
AB + - - + + - - + 4
AC + - + - - + - + 4
BC + + - - - - + + 4
ABC - + + - + - - + 4

MAIN EFFECTS AND INTERACTIONS

Following the same notations for treatment totals and treatment means as in 2 2
factorial experiment, the simple effect of A,(say), is given by the difference in the
mean yields of A as a result of increasing the factor A from the level a0 to a1 at other
levels of the factors B and C.
 Level of B Level of C Simple effect of A
b0 c0 (a1 b0 c0) - (a0 b0 c0) = (a) – (1)
b1 c0 (a1 b1 c0) - (a0 b1 c0) = (ab) – (b)
b0 c1 (a1 b0 c1) - (a0 b0 c1) = (ac) – (c)
b1 c1 (a1 b1 c1) - (a0 b1 c1) = (abc) – (b)

The main effect of A is defined as the average of these 4 simple effects. Thus
A = ¼ [(abc) – (bc) + (ac) – (c) + (ab) – (b) +(a) – (1)]
=¼ [{(abc) + (ac) + (ab) +(a)} –{ (bc) + (c) + (b) + (1)}]
= ¼ (a -1) (b+1) (c+1)
Similarly, the main effect of the factors B and C can be obtained as
B = ¼ [{(abc) + (ab) + (bc) +(b)} –{ (a) + (c) + (ac) + (1)}]
= ¼ (a +1) (b-1) (c+1)

C = ¼ [{(abc) + (bc) + (ac) +(c)} –{ (a) + (b) + (ab) + (1)}]

= ¼ (a +1) (b+1) (c-1)
The interaction effects AB,AC and BC is given by
AB = ¼ [{(abc) - (bc) + (ab) - (b)} –{ (ac) - (c) + (a) - (1)}]
= ¼ (a -1) (b-1) (c+1)
AC = ¼ (a -1) (b +1) (c-1)
BC = ¼ (a +1) (b -1) (c-1)
The interaction effect of AB with C i.e., the interaction ABC is given by
ABC = ¼ [ (abc) - (bc) - (ac) + (c)} –{ (ab) + (b) + (a) - (1)]
= ¼ (a -1) (b-1) (c-1)
Model of 22 design
If yijkl = µ + αi + βj + γk +(αβ) ij + (αγ) ik + (βγ) jk +(αβγ) ijk+ (ρ) i + (ϵ) ijkl

(i,j,k =0,1: l=1,2,…r)

where µis the general mean
αi is the effects of ith level of A
βj is the effects of jth level of B
γk is the effects of kth level of C
(αβ) ij is the interaction effect of ith level of A with jth level of B

(αγ) ik is the interaction effect of ith level of A with kth level of C

(βγ) jk is the interaction effect of jth level of B with kth level of C

(αβγ) ijk is the interaction effect of ith level of A with jth level of B and
kth level of C
(ρ) i is the effect due to ith replicate

(ϵ) ijkl is the error effect due to chance and are i.i.d. N(0, σ2)
The above parameter are subject to the following restrictions
∑1𝑖=0 α𝑖 = ∑ 1𝑗=0 βj = ∑ 1𝑘=0 γ𝑘 = ∑𝑖=0
1 𝜌𝑖 = 0 and
the sums
Σ (αβ) ij, Σ (αγ) ik ,Σ (βγ) jk , Σ (αβγ) ijk are equal to zero respectively.

Statistical analysis of 23 factorial experiment

Null Hypothesis:
H0: The blocks as well as treatments are homogeneous
Alternative Hypothesis:
H1: The blocks as well as treatments are not homogeneous
By Using the Yates’ table of divisors and signs of 23 factorial experiment the various
factorial effect totals can be expressed as mutually orthogonal contrasts of the 8
treatment totals. Thus,
[A] = [abc] – [bc] + [ac] – [c] + [ab] – [b] +[a] – [1]
 [AB] = [abc] – [bc] +[ab] – [b] – [ac] + [c] – [a] + [1]
And so on
S.S. due to main effect A=[A]2 / 8r with 1 d.f.
S.S. due to first order interaction BC = [BC]2/8r with 1 d.f.
S.S. due to second order interaction ABC = [ABC]2/8r with 1 d.f.
And so on
ANOVA TABLE For 23 factorial experiment in r randomized blocks
Sources of d.f. S.S. M.S.S. Variance Ratio
variation ‘F’
Blocks(replicates) r–1 S2R s2R = S2R/(r-1) FR = s2R/ s2E
Main effect A 1 S2A=[A]2 / 8r s2A= S2A FA = s2A/ s2E
Main effect B 1 S2B=[B]2 / 8r s2B= S2B FB = s2B/ s2E
Main effect B 1 S2C=[C]2 / 8r s2C= S2C FB = s2C/ s2E
First order 1 S2AB=[AB]2 / 8r s2AB= S2AB FAB = s2AB / s2E
Interaction A x B
First order 1 S2AC=[AC]2 / 8r s2AC= S2AC FAC = s2AC / s2E
Interaction A x C
First order 1 S2BC=[BC]2 / 8r s2Bc= S2BC FBC = s2AB / s2E
Interaction B x C
Error 7(r-1) S2E= by subtraction s2E= S2E/[7(r-1)]
Total (r.22 – 1) ∑ ∑(𝑦𝑖𝑗 − 𝑦̿.. )2
= 8r – 1 𝑖 𝑗

Conclusion
For a certain level of α if FR ≤ F(r – 1),7(r-1), accept H0, otherwise reject H0 for factorial
effects
If FA, FB, FC, FAB, FAC, FBC, FABC ≤ F1,7(r-1) the presence of the factorial effects is
accepted otherwise rejected.
CONFOUNDING IN FACTORIAL DESIGN
In factorial experiments as the number of factors and the levels at which they are
employed increase, the total number of treatment combinations increases rather
rapidly and consequently the block size has to be enlarged. For example, for a 210
factorial experiment a complete factorial will require 1024 units. As a consequence
of increase in the block size or handling such a huge experiment the purpose of local
control is defeated due to the following two reasons.
1. It is sometimes impracticable to get one complete replicate which are
relatively homogeneous and
2. The greater heterogeneity is introduced in the experimental error and reduces
the discriminating power of the test of significance.
In order to maintain homogeneity within blocks the experimenter must either cut
down the number of factors or use an incomplete design. In order to avoid that
one the concept of confounding is introduced.
Confounding definition: confounding may be defined as the technique of
reducing the size of replication over a no. of blocks at the cost of loosing some
information on some effects which is not of much practical importance.
Confounding in 23 experiment
In 23 experiment, the eight treatment combinations required 8 units of
homogeneous material each to form a block. If we decide to use blocks of 4 units
each then a full replication will require only two blocks.
In this case 8 treatment combinations are divided into groups of 4 treatments each
in a special way so as to confound any one of the less important interactions with
blocks and these groups are allocated at random inn the two blocks.
For example, let us consider confounding the highest order interaction ABC. The
interaction effect ABC is given by
ABC = ¼ [ (abc) - (bc) - (ac) + (c) – (ab) + (b) + (a) - (1)]
= ¼ [ (abc) + (a) + (b) + (c) – (ab) + (bc) - (ac) - (1)]
In order to confound the interaction ABC with blocks all the treatment combinations
with positive sign are allocated at random in one block and those with negative signs
in another block. This arrangement gives ABC confounded with blocks and hence
we loose less information on ABC.
ABC CONFOUNDED WITH BLOCKS
Replicates Block I (1) (ab) (ac) (bc)
Block II (abc) (a) (b) (c)
 For carrying out the statistical analysis the various factorial effects and their sum
of squares are estimated in the usual manner but the sum of squares due to
confounded interaction is not computed and it is also not included in the ANOVA
table.
Sources of variation d.f.
blocks 2r - 1
A 1
B 1
C 1
AB 1
AC 1
BC 1
Error 6(r – 1)
total 8r – 1

The error sum of squares is obtained as usual

1.e., S2E = T.S.S. - S2 A - S2 B - S2 C - S2 AB- S2AC - S2 BC - S2ABC