A hidden projection of

Plackett-Burman design
by Yashi Pal
 Fractional factorial design :
Table of 24−1 fractional
factorial  The goal is to create designs that allow us to screen a
Trt. A B C D=AB large number of factors but without having a very
C large experiment.
 Example: in 24 we need 16 observation but in 24−1
(1) − − − − 1
fractional factorial design we need ∗ 16 observation
2
i.e we need only 8 observation.
a + − − +
 We can then construct the levels of D by using the
b − + − + relationship where D = ABC.
ab + + − −  This 24-1 design is a Resolution IV design.
c − − + +  The number of letters in the generator determines
ac + − + − the confounding or aliasing properties in the resulting
design.
bc − + + −
 We can see this best by looking at the expression I =
abc + + + + ABCD. We obtain the alias structure by multiplying A ×
I = A × ABCD = A2BCD which implies A = BCD, B = ACD, C
= ABD, D = ABC, AB = CD, AC = BD, AD = BC
Resolution:
 Resolution is another vital property of fractional factorial design.
which describes the degree to which main effect are aliased or
confounded with two-factor interactions, three-factor interactions,
etc.
 Resolution III : Main effects are aliased with two-factor interactions .
Example: 23−1 with I = ABC.
 Resolution IV :Similarly, main effects are confounded with at worst
three-factor interactions. Or two-factor Interactions are confounded
with certain other two-factor interactions. Example : 24−1 , I =ABCD.
 Resolution V :Main effects are aliased with four-factor interactions. Or
two-factor interactions are aliased with certain three-factor
interactions. Example: 25−1 , I = ABCDE.
PROJECTION:

Factor A B C  Look at the following four design in factor A, B, and

C.
 The design is a 23−1
𝐼𝐼𝐼
− − +
 Here C=AB so I=ABC. If you drop out any one factor
you obtain a 22 factorial In the remaining two
factors. If we drop C then we get 22 factorial
+ − − design i.e we can say that design projects a 22
factorial design is thus said to be of projectivity
P=2. In the similar way in 24−1
IV if we dropped any
− + − one of the four factor A,B,C,D you would have a full
factorial 23 design in remaining three.
 In general for fractional factorial design the
+ + + projectivity p is 1 less than the resolution R that is
every subset of P=R-1 factors produced a complet
factorial in p factors.
Here projectivity has two aspect-
geometric projection given by LD.
Ability to entertain the estimation .
For PB design however the second aspect can be achieved
even when geometric projection does not lead to a full
factorial design or fractional factorial design.
PROJECTION OF 12 RUN PLACKET BURMAN DESIGN INTO K
DIMENSIONS:
K DESIGN DESCRIPTION  Lin and draper show that the
NO. projection of 12 run PB design in
any k of 11 factor dimensions we
2 2.1 22 ∗ 3(22 design with 3 replicates) select k columns and examine the
design that result by ignoring the
3 3.1 1
23 + 2 ∗ 23 (23 𝑑𝑒𝑠𝑖𝑔𝑛 + other 11-k columns.
23−1 𝑑𝑒𝑠𝑖𝑔𝑛)  For k=2,3 we do not need to add
any additional run.
4 4.1 Add one more run to obtain a 24−1
IV
design.  For k=4,5 we need to add
additional runs.
Add five more run to obtain a
24 design.
2 COLUMN OF PB DESIGN

Table-1

Runs 1 2
1 + + 12
2 − + 2 The treatment combination of
3 + − 1
following table is replicated 3
times in table 1
4 − + 2
5 − − (1) 1 2
6 − − (1) − − (1)
7 + − 1 + − 1
8 + + 12 − + 2
9 + + 12 + + 12
10 − + 2
11 + − 1
12 − − (1)
Runs 1 2 3
1 + + − 12
2 − + + 23
• For K=3:we know that in 23 we have 8
3 + − + 13
treatment combination (1),1,2,12,3,13,23,123.
4 − + − 2
• In the below table of 23−1 , 3=12 i.e. is I=123.
5 − − + 3 we take take (1),12,13,23, so 23 +23−1 gives the
6 − − − (1) 12 runs design.
7 + − − 1 (1) 12 13 23
8 + + − 12
9 + + + 123 123 3 2 1

10 − + + 23
11 + − + 13
12 − − − (1)
 • From the table of 24−1 we see that
1 2 3 4 I=1234,so 14=23,34=12,13=24. so
Runs need to add only one run.
1 + + − + 124
2 − + + − 23
3 + − + + 134
1 2 3 4=123
4 − + − + 24
− − − − (1)
5 − − + − 3
+ − − + 14
6 − − − + 4
− + − + 24
7 + − − − 1
+ + − − 12
8 + + − − 12
− − + + 34
9 + + + − 123
+ − + − 13
10 − + + + 234
− + + − 23
11 + − + + 134
+ + + + 1234
12 − − − − (1)
Hidden projection property of PB :

 24−1 design obtain by adding one run has the defining

relation I=1234 allow three out of the six 2fi’s to be
estimated because the six 2fi’s are confounded in three
pairs.
 In the hidden projection property of PB design given by
J.C Wang take some no of factor like four and show that
all the six two factor interaction among the four factor
can be estimated without adding runs.
 Hidden projection of 12 run placket-Burman design :
 In hidden projection of 12 run placket-Burman design we take some main effect with
their different two factor interaction and then estimate their efficiency for
comparison.
 To measure the estimation efficiency of a set of effects we can use the following 𝐷𝑆
efficiency criterion. Let 𝑥1, 𝑥2, … … … . , 𝑥𝑛 denote the standardized contrast vector of n
effects then the 𝐷𝑆 efficiency for the estimate associated with 𝑥𝑖 is-
 𝐷𝑆 = {𝑥𝑖𝑡 𝑥𝑖 − 𝑥𝑖𝑡 𝑋 𝑖 𝑋 𝑡𝑖 𝑋 𝑖
−1
𝑋 𝑡𝑖 𝑥𝑖 }/𝑥𝑖𝑡 𝑥𝑖 =1-𝑥𝑖𝑡 𝑋 𝑖 (𝑋 𝑡𝑖 𝑋 𝑖 ) −1 𝑋 𝑡𝑖 𝑥𝑖 ……(1)
 Where 𝑋 𝑖 ={𝑥1, 𝑥2, ………, 𝑥𝑖−1, 𝑥𝑖+1,…………., 𝑥𝑖 } i.e. ignore the𝑖𝑡ℎ column.
 If 𝑥𝑖 were orthogonal to the remaining n-1 vectors, the second term in the above
equation would be zero and the 𝐷𝑆 value would be 1. Therefore 1 is the upper bound
for the 𝐷𝑆 efficiency.
 The closer the 𝐷𝑆 efficiency is to 1the more accurately the effect for 𝑥𝑖 can be
estimated.
 Then we calculate D criterion for measuring the overall efficiency for estimating the
collection of effects:
 D= 𝑋 𝑡 𝑋 1/𝐾 where X=[𝑥1 / 𝑥1 ,…………, 𝑥𝑘 / 𝑥𝑘 ] and 𝑥𝑖 is the coefficient vector of 𝑖 𝑡ℎ
effect. The maximum value of D is 1.
 Now we choose n factor the choice of n factors is equivalent to
the choice of a 12 X n sub matrix of the matrix from the table
of 12 run PB design. Two such matrices are said to be
equivalent if one can be obtain from the permutations of rows,
columns and sign and changes.
 We choose here four factors because it is know that, expect for
n=5 and 6, any 12 X n sub matrix are equivalent.
 For n=4, we calculate the value of D and 𝐷𝑆 for 10 cases. The
last one consisting of four main effects and six 2fi’s is the most
comprehensive. From their 𝐷𝑆 values we can see that all the
2fi’s and the main effects can be estimated without adding
runs.
 We see that first nine cases represent all possible submodalels
of the comprehensive model of case 10.We see this point by
the use of graphs.
Now estimation of efficiency for four factors and h=1using R:
>a=c(1,1,-1,1,-1,1,1,-1,1,-1,1,1,-1,1,-1,1,-1,-1,1,-1,-1,-1,-1,1,1,-1,-1,-
1,1,1,-1,-1,1,1,1,-1,-1,1,1,1,1,-1,1,1,-1,-1,-1,-1)
>x1=matrix(a,ncol=12,nrow=5,byrow=TRUE) # matrix of 4 main and 1 interaction

>x1
[,1] [,2] [,3] [,4] [,5]=12
[1,] 1 1 -1 1 1
[2,] -1 1 1 -1 -1
[3,] 1 -1 1 1 -1
[4,] -1 1 -1 1 -1
[5,] -1 -1 1 -1 1
[6,] -1 -1 -1 1 1
[7,] 1 -1 -1 -1 -1
[8,] 1 1 -1 -1 1
[9,] 1 1 1 -1 1
[10,] -1 1 1 1 -1
[11,] 1 -1 1 1 -1
[12,] -1 -1 -1 -1 1
> x=1/sqrt(12)*x1 #multiplication of 1/ 12 to x1
>x
[,1] [,2] [,3] [,4] [,5]
[1,] 0.2886751 0.2886751 -0.2886751 0.2886751 0.2886751
[2,] -0.2886751 0.2886751 0.2886751 -0.2886751 -0.2886751
[3,] 0.2886751 -0.2886751 0.2886751 0.2886751 -0.2886751
[4,] -0.2886751 0.2886751 -0.2886751 0.2886751 -0.2886751
[5,] -0.2886751 -0.2886751 0.2886751 -0.2886751 0.2886751
[6,] -0.2886751 -0.2886751 -0.2886751 0.2886751 0.2886751
[7,] 0.2886751 -0.2886751 -0.2886751 -0.2886751 -0.2886751
[8,] 0.2886751 0.2886751 -0.2886751 -0.2886751 0.2886751
[9,] 0.2886751 0.2886751 0.2886751 -0.2886751 0.2886751
[10,] -0.2886751 0.2886751 0.2886751 0.2886751 -0.2886751
[11,] 0.2886751 -0.2886751 0.2886751 0.2886751 -0.2886751
[12,] -0.2886751 -0.2886751 -0.2886751 -0.2886751 0.2886751
> x𝑡 =t(x)
>x𝑡
[,1] [,2] [,3] [,4] [,5] [,6]
[1,] 0.2886751 -0.2886751 0.2886751 -0.2886751 -0.2886751 -0.2886751
[2,] 0.2886751 0.2886751 -0.2886751 0.2886751 -0.2886751 -0.2886751
[3,] -0.2886751 0.2886751 0.2886751 -0.2886751 0.2886751 -0.2886751
[4,] 0.2886751 -0.2886751 0.2886751 0.2886751 -0.2886751 0.2886751
[5,] 0.2886751 -0.2886751 -0.2886751 -0.2886751 0.2886751 0.2886751

[,7] [,8] [,9] [,10] [,11] [,12]

[1,] 0.2886751 0.2886751 0.2886751 -0.2886751 0.2886751 -0.2886751
[2,] -0.2886751 0.2886751 0.2886751 0.2886751 -0.2886751 -0.2886751
[3,] -0.2886751 -0.2886751 0.2886751 0.2886751 0.2886751 -0.2886751
[4,] -0.2886751 -0.2886751 -0.2886751 0.2886751 0.2886751 -0.2886751
[5,] -0.2886751 0.2886751 0.2886751 -0.2886751 -0.2886751 0.2886751
>x𝑡 x=x𝑡 %*%x
>x𝑡 x

[,1] [,2] [,3] [,4] [,5]

[1,] 1.000000e+00 0.000000e+00 0.000000e+00 -2.775558e-17 0.0000000
[2,] 0.000000e+00 1.000000e+00 2.775558e-17 0.000000e+00 0.0000000
[3,] 0.000000e+00 2.775558e-17 1.000000e+00 2.775558e-17 -0.3333333
[4,] -2.775558e-17 0.000000e+00 2.775558e-17 1.000000e+00 -0.3333333
[5,] 0.000000e+00 0.000000e+00 -3.333333e-01 -3.333333e-01 1.0000000

> x𝑡 x =det(x𝑡 x) # determinant of matrix r

> x𝑡 x
[1] 0.7777778
>D= x𝑡 x ^(1/5)
>D
[1] 0.9509794
> 𝑋(1) =x[ ,-1]
>𝑋(1)
[,1] [,2] [,3] [,4]
[1,] 1 -1 1 1
[2,] 1 1 -1 -1
[3,] -1 1 1 -1
[4,] 1 -1 1 -1
[5,] -1 1 -1 1
[6,] -1 -1 1 1
[7,] -1 -1 -1 -1
[8,] 1 -1 -1 1
[9,] 1 1 -1 1
[10,] 1 1 1 -1
[11,] -1 1 1 -1
[12,] -1 -1 -1 1
>𝑋(1) 𝑡 =t(𝑋(1) )
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10] [,11] [,12]
[1,] 1 1 -1 1 -1 -1 -1 1 1 1 -1 -1
[2,] -1 1 1 -1 1 -1 -1 -1 1 1 1 -1
[3,] 1 -1 1 1 -1 1 -1 -1 -1 1 1 -1
[4,] 1 -1 -1 -1 1 1 -1 1 1 -1 -1 1
>C1=𝑋(1) 𝑡 %*%𝑋(1)
>C1
[,1] [,2] [,3] [,4]
[1,] 12 0 0 0
[2,] 0 12 0 -4
[3,] 0 0 12 -4
[4,] 0 -4 -4 12
>(𝑋(1) 𝑡 𝑋(1) )−1 =solve(C1)
>(𝑋(1) 𝑡 𝑋(1) )−1
[,1] [,2] [,3] [,4]
[1,] 0.08333333 0.00000000 0.00000000 0.00000000
[2,] 0.00000000 0.09523810 0.01190476 0.03571429
[3,] 0.00000000 0.01190476 0.09523810 0.03571429
[4,] 0.00000000 0.03571429 0.03571429 0.10714286
>r3=c(1,-1,1,-1,-1,-1,1,1,1,-1,1,-1)
>𝑥1 =matrix(r3,ncol=1,nrow=12,byrow=TRUE)
>𝑥1 𝑡 =t(𝑥1 )
>𝑥1 𝑡 𝑥1 =𝑥1 𝑡 %*%𝑥1
>𝑥1 𝑡 𝑥1
[,1]
[1,] 12
>𝑥1 𝑡 𝑋(1) =𝑥1 𝑡 %*%𝑋(1)
[,1] [,2] [,3] [,4]
[1,] 0 0 0 0 ; 𝐷𝑆 =1 # by equation (1)
Table for estimation efficiency for four factors and h=1,2,…..
Graphical :we can represent any model consisting of main effects and
select 2fi’s by graph.
For example, case 10 can be represented by the following graph.

1 4

2 3
Graphical :we can represent any model consisting of main effects and select 2fi’s by
graph.
For example, case 10 can be represented by the following graph.

1 4

2 3
The first nine cases correspond to all possible non isomorphic subgroups of this
complete graph.
 We define a graph model to be the class of models .All the models within the same
class are said to be graphically equivalent. For example, the model for case4
{1,2,3,4,12,13,14} and the model M={1,2,3,4,14,24,34} are different but both can be
represented by the following graph.
1 4 1 4

2 3 2 3
and therefore they are graphically equivalent.
An important question is whether graphically equivalent models have the same
statistical efficiencies. For example : the two models consider the above. The D value
for model M IS .89 and the 𝐷𝑆 value for the seven effects are .74,.74,.74,1,.74,.74,.74
which are equivalent to the D and 𝐷𝑆 value for case 4 after changing 1 to 4 and 4 to 1.
we call two graphically equivalent models efficiency equivalent if the D and 𝐷𝑆 values
of one model are the same as the other model after relabeling the factors names.
In general graphical equivalence does not imply efficiency equivalence. For each of 10
graph models in above table, graphical equivalence. Therefore one set of D and 𝐷𝑆
values represent all the models the same graph.
Estimation efficiency for five factor and h=1,…….,6.the 𝐷𝑆 efficiency is given for each
effect:
For n=5,there are two non-isomorphic 12 X 5 sub matrices: design 5.1 and design 5.2
in the notation of LD. Design 5.1 has two repeated runs i.e. two runs with the same
level combination. For example, in design consisting of columns 1,2,3,4 and 10, runs
3 and 11 are identical.

Because design 5.1 has two repeated runs, there are only 10 degrees of freedom
for estimating effect so we have at most five 2fi’s.In this case there are altogether 4
graphically non equivalent models for estimating five 2fi’s and five main effects.

C
In the above table of 5.2 for n=5 there are 24 cases the last 5 cases
contain 6 2fi’s in addition to 5 main effect .the 2fi’s for these five cases
can be represented graphically as shown in the figure. The remaining 19
cases can be represented as the subgroup of these five graph.in the
following figure we do not give the column number of the factor because
there can be many choice for the same graph. EX- case 9.
note that design 5.1 is inferior to 5.2 because 5.1 has two repeated runs,
there are only 10 degree of freedom for estimating effect.it can
entertain at most 5 2fi’s it can entertain model represented by 15 non
isomorphic graph all of which are however sub graph of 24 graph for
5.2.among the 15 only 4 has 2fi’s .
Figure-1
Thank you