Chapter 5

Incomplete Block Designs

If the number of treatments to be compared is large, then we need large number of blocks to
accommodate all the treatments. This requires more experimental material and so the cost of
experimentation becomes high which may be in terms of money, labor, time etc. The completely
randomized design and randomized block design may not be suitable in such situations because they will
require large number of experimental units to accommodate all the treatments. In such situations when
sufficient number of homogeneous experimental units are not available to accommodate all the treatments
in a block, then incomplete block designs can be used. In incomplete block designs, each block receives
only some of the selected treatments and not all the treatments. Sometimes it is possible that the available
blocks can accommodate only a limited number of treatments due to several reasons. For example, the
goodness of a car is judged by different features like fuel efficiency, engine performance, body structure
etc. Each of this factor depends on many other factors, e.g., engine consists of many parts and the
performance of every part combined together will result in the final performance of the engine. These
factors can be treated as treatment effects. If all these factors are to be compared, then we need large
number of cars to design a complete experiment. This may be an expensive affair. The incomplete block
designs overcome such problems. It is possible to use much less number of cars with the set up of an
incomplete block design and all the treatments need not to be assigned to all the cars. Rather some
treatments will be implemented in some cars and remaining treatments in other cars. The efficiency of
such designs is, in general, not less than the efficiency of a complete block design. In another example,
consider a situation of destructive experiments, e.g., testing the life of televisionsets, LCD panels, etc. If
there are large number of treatments to be compared, then we need large number of television sets or LCD
panels. The incomplete block designs can use lesser number of television sets or LCD panels to conduct
the test of significance of treatment effects without losing, in general, the efficiency of design of
experiment. This also results in the reduction of experimental cost. Similarly, in any experiment
involving the animals like as biological experiments, one would always like to sacrifice less animals.
Moreover the government guidelines also restrict the experimenter to use smaller number of animals. In
such cases, either the number of treatments to be compared can be reduced depending upon the number of
animals in each block or to reduce the block size. In such cases when the number of treatments to be
compared is larger than the number of animals in each block, then the block size is reduced and the setup
of incomplete block designs can be used. This will result in the lower cost of experimentation. The

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

1

1
incomplete block designs need less number of observations in a block than the observations in a
complete block design to conduct the test of hypothesis without losing the efficiency of design of
experiment, in general.

Complete and incomplete block designs:

The designs in which every block receives all the treatments are called the complete block designs.

The designs in which every block does not receive all the treatments but only some of the treatments are
called incomplete block design.

The block size is smaller than the total number of treatments to be compared in the incomplete block
designs.

There are three types of analysis in the incomplete block designs

 intrablock analysis,
 interblock analysis and
 recovery of interblock information.

Intrablock analysis:
In intrablock analysis, the treatment effects are estimated after eliminating the block effects and then the
analysis and the test of significance of treatment effects are conducted further. If the blocking factor is
not marked, then the intrablock analysis is sufficient enough to provide reliable, correct and valid
statistical inferences.

Interblock analysis:
There is a possibility that the blocking factor is important and the block totals may carry some important
information about the treatment effects. In such situations, one would like to utilize the information on
block effects (instead of removing it as in the intrablock analysis) in estimating the treatment effects to
conduct the analysis of design. This is achieved through the interblock analysis of an incomplete block
design by considering the block effects to be random.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

2

2
Recovery of interblock information:
When both the intrablock and the interblock analysis have been conducted, then two estimates of
treatment effects are available from each of the analysis. A natural question then arises -- Is it possible to
pool these two estimates together and obtain an improved estimator of the treatment effects to use it for
the construction of test statistic for testing of hypothesis? Since such an estimator comprises of more
information to estimate the treatment effects, so this is naturally expected to provide better statistical
inferences. This is achieved by combining the intrablock and interblock analysis together through the
recovery of interblock information.

Intrablock analysis of incomplete block design:

We start here with the usual approach involving the summations over different subscripts of y’s. Then
gradually, we will switch to matrix based approach so that the reader can compare both the approaches.
They can also learn the one –to-one relationships between the two approaches for better understanding.

Notations and normal equations:

Let
- v treatments have to be compared.
- b blocks are available.
- ki : Number of plots in ith block (i = 1,2,…,b).
- rj : Number of plots receiving jth treatment ( j = 1,2,…,v).

- n: Total number of plots.

n  r1  r2  ...  rv  k1  k2  ...  kb .
- Each treatment may occur more than once in each block
or
may not occur at all.
- nij  Number of times the jth treatment occurs in ith block

For example, nij  1 or 0 for all i , j means that no treatment occurs more than once in a block and a

treatment may not occur in some blocks at all. Similarly, nij  1 means that jth treatment occurs in ith

block and nij  0 means that jth treatment does not occurs in ith block.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

3

3
 v

n
j 1
ij  ki i  1,..., b

n
i
ij  rj j  1,..., v

n    nij
i j

Model:
Let yijm denotes the response (yield) from the mth replicate of jth treatment in ith block and

yijm   i   j   ijm i  1, 2,..., b, j  1, 2,.., v , m  1, 2,..., nij

[Note: We are not considering here the general mean effect in this model for better understanding of the
issues in the estimation of parameters. Later, we will consider it in the analysis.]

Following notations are used in further description.

Block totals : B1 , B2 ,..., Bb where Bi    yijm .
j m

Treatment totals: V1 ,V2 ,...,Vv where V j    yijm

i m

Grand total : Y     yijm

i j m

Generally, a design is denoted by D(v, b, r , k , n) where v, b, r, k and n are the parameters of the design.

Normal equations:
Minimizing S      ijm
2
with respect to  i and  j , we obtain the least squares estimators of the
i j m

parameters as follows:
S  ( y
i j m
ijm  i   j )2

S
0
 i
  ( yijm   i   j )  0
j m

or
Bi   i 1   1  0
j m j
j
m
(1)

or
Bi   i ki  ni1 1  ni 2 2  ....  niv v , i  1,..., b
Bi   i ki    j nij [b equations]
j

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

4

4
S
0
 i
  ( yijm  i   j )  0
i m

or  y
i m
ijm   i 1   j 1  0
i m i m

V j   i nij   j  nij  0 (2)

i i

or V j  nij 1  n2 j  2  ...  nbj b  rj j , j  1, 2,..., v

or V j   i nij  rj j [v equations]
i

Equations (1) and (2) constitute (b + v) equations.

Note that

 equation (1)   equation

i j
(2)

 B  V
i
i
j
j

   
   y ijm      yijm .
j  i 
i  j m  m

Thus there are atmost (b + v - 1) degrees of freedom for estimates. So the estimates of only (b + v - 1)
parameters can be obtained out of all (b + v) parameters.

[Note: We will see later that degrees of freedom may be less than or equal to (b + v - 1) in special cases.
Also note that we have not assumed any side conditions like    
i
i
j
j  0 as in the case of complete

block designs.]

In order to obtain the estimates of the parameters, there are two options-
1. Using equation (1), eliminate  i from equation (2) to estimate  j or

2. Using equation (2), eliminate  j from equation (1) to estimate  i .

We consider first the approach 1., i.e., using equation (1), eliminate  i from equation (2).
From equation (1),

1  v 
i   i  nij j 
B 
ki  j 1 
Use it in (2) as follows.
Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur
5

5
V j  nij 1  ...  nbj  b  rj j
1 
 n1 j  ( B1  n11 1  ...  n1v 1v ) 
 k1 
1 
 n2 j  ( B2  n21 1  ...  n2 v v )   ...
 k2 
1 
 nbj  ( Bb  nb1 1  ...  nbv v )   rj j
 kb 
or
n1 j B1 n2 j B2 nbj Bb
Vj    ... 
k1 k2 kb
 n n n n n n 
  1   11 1 j  21 2 j  ....  b1 bj   ...
 k1 k2 kb 
 n n n n n n 
  v   1v 1 j  2 v 2 j  ....  bv bj   rj j , j  1,..., v
 k1 k2 kb 
or
b nij Bi  n n n n   n n n n 
Vj     1   11 1 j ...  b1 bj   ...   v   1v 1 j ...  bv bj   rj j
i 1 ki  k1 kb   k1 kb 
or
 n n n n   n n n n 
Q j   1   11 1 j ...  b1 bj   ...   v   1v 1 j ...  bv bj   rj j , j  1....v
 k1 kb   k1 kb 
where
n B n B 
Q j  V j   1 j 1  ...  bj b  , j  1, 2,..., v
 k1 kb 

are called adjusted treatment totals.

b nij Bi
[Note: Compared to the earlier case, the jth treatment total V j is adjusted by a factor 
i 1 ki
, that it why

it is called “adjusted”. The adjustment is being made for the block effects because they were eliminated to
estimate the treatment effects.]
Note that
ki : Number of plots in ith block.
Bi
is called the average (response) yield per plot from ith block.
ki

nij Bi
is considered as average contribution to the jth treatment total from the ith block.
ki
Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur
6

6
Q j is obtained by removing the sum of the average contributions of the b blocks from the jth treatment

total V j .

Write
 n n n n   n n n n 
Q j   1   11 1 j ...  b1 bj   ...   v   1v 1 j ...  bv bj   rj j , j  1, 2,..., v.
 k1 kb   k1 kb 
as
Q j  C j1 1  C j 2 2  ...  C jv v
where
n12j n22 j nbj2
C jj  rj    ... 
k1 k2 kb
n1 j n1 j ' n2 j n2 j ' nbj nbj '
C jj '     ...  ; j  j ', j  1, 2,.., v.
k1 k2 kb
The v  v matrix C  ((C jj ' )), j  1, 2,..., v; j '  1, 2,..., v with C jj as diagonal elements and C jj ' as off-

diagonal elements is called the C-matrix of the incomplete block design.

C matrix is symmetric. Its row sum and column sum are zero. (proved later)
Rewrite
 n11n1 j nb1nbj   n1v n1 j nbv nbj 
Q j  1  ...    ...   v   ...    rj j , j  1, 2,..., v.
 k1 kb   k1 kb 
as
Q  C .

This equation is called as reduced normal equations where Q '  (Q1 , Q2 ,.., Qv ),  '  ( 1 , 2 ,..., v )

Equations (1) and (2) are EQUIVALENT.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

7

7
Alternative presentation in matrix notations:
Now let us try to represent and translate the same algebra in matrix notations.
Let
Emn : m  n matrix whose all elements are unity.
N  ( nij ) is b  v matrix called as incidence matrix.
v
ki   nij
j 1
b
rj   nij
i 1

n   nij
i j

E1b N  ( r1 , r2 ,..., rv )  r '

NEv1  (k1 , k2 ,..., kb ) '  k .

For illustration, we verify one of the relationships as follows.

E1b  (1,1,...,1)1b
 n11 n12  n1v 
 
n n22  n2 v 
N   21
   
 
 nb1 nb 2  nbv 
 n11 n12  n1v 
 
n n22  n2 v 
E1b N  (1,1,...,1)1b  21
   
 
 nb1 nb 2  nbv bv
 b b b

   ni1 ,  ni 2 ,...,  niv 
 i 1 i 1 i 1 
 (r1 , r2 ,..., rv )
= r '.
It is now clear that the treatment and blocks are not estimable as such as in the case of complete block
designs. Note that we have not made any assumption like    
i
i
j
j  0 also.

Now we introduce the general mean effect (denoted by  ) in the linear model and carry out the further
analysis on the same lines as earlier.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

8

8
Consider the model
yijm     i   j   ijm , i  1, 2,..., b; j  1, 2,..., v; m  0,1,.., nij .

The normal equations are obtained by minimizing S      ijm

2
with respect to the parameters
i j m

 ,  i and  j and solving them, we can obtain the least squares estimators of the parameters.

Minimizing S      ijm
2
with respect to the parameters  ,  i and  j , the normal equations are
i j m

obtained as
nˆ   nio ˆi   nojˆ j  G
i j

ni ˆ  nio ˆi   nijˆ j  Bi i  1,..., b

j

noj ˆ  nojˆ j   nij ˆi  V j j  1,..., v.

i

Now we write these normal equations in matrix notations. Denote

  Col ( 1 ,  2 ,...,  b )
  Col ( 1 , 2 ,..., v )
B  Col ( B1 , B2 ,..., Bb ) \
V  Col (V1 , V2 ..., Vv )
N  (( nij )) : incidence matrix of order b  v

where Col (.) denotes the column vector.

Let
K  diag (k1 ,.., kb ) : b  b diagonal matrix
R  diag (r1 ,.., rv ) : v  v diagonal matrix.
Then the (b + v +1) normal equations can be written as

G   n E1b K E1v R   ˆ 
    
 B    KEb1 K N   ˆ  . (*)
 V   RE R  
   v1 N '  ˆ 
Since we are presently interested in the testing of hypothesis related to the treatment effects, so we
eliminate the block effects ˆ to estimate the treatment effects. For doing so, multiply both sides on the
left of equation (*) by

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

9

9
 1 0 
0
  1
0 Ib  NR 
  N K 1 
0 Iv 

where
1 1 1 1 1 1
R 1  diag  , ,...,  , K 1  diag  , ...,  .
 r1 r2 rv   k1 k2 kb 
Solving it further, we get 3 sets of equations as follows:
G  nˆ  E1b K ˆ  Eiv Rˆ
B  NR 1V  [ K  NR 1 N ]ˆ
V  N K 1 B  [ R  N K 1 N ]ˆ
These are called as ‘reduced normal equations’ or ‘reduced intrablock equations’.

The reduced normal equation in the treatment effects can be written as

Q  Cˆ
where
Q  V  N K 1 B
C  R  N K 1 N .
The vector Q is called as the vector of adjusted treatment totals, since it contains the treatment totals
adjusted for the block effects, the matrix C is called as C-matrix.

The C matrix is symmetric and its row sums and columns sums are zero.

To show that row sum is zero in C matrix, we proceed as follows:

Row sum:
CEv1  REv1  N ' K 1 NEv1
= (r1 , r2 ,..., rv ) ' N ' K 1k 1
= (r1 , r2 ,..., rv ) ' N ' Eb1
= r r
=0
Similarly, the column sum can also be shown to be zero.

In order to obtain the reduced normal equation for treatment effects, we first estimated the block effects
from one of the normal equation and substituted it into another normal equation related to the treatment
effects. This way the adjusted treatment total vector Q (which is adjusted for block effects) is obtained.
Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur
10

10
Similarly, the reduced normal equations for the block effects can be found as follows. First estimate the
treatment effects from one of the normal equations and substitute it into another normal equation
related to the block effects. Then we get the adjusted block totals (adjusted for treatment totals).

So, similar to Q  Cˆ, we can obtain another equation which can be represented as

D ˆ  P
where
1 1 1
D  diag (k1 , k2 ,.., kb )  N diag  , ,...,  N 
 r1 r2 rv 
 K  NR 1 N 
1 1 1
P  B  N diag  , ,...,  V
 r1 r2 rv 
1
 B  NR V
ˆ  ( ˆ , ˆ ,.., ˆ ) '
1 2 b
where P is the adjusted block totals which are obtained after removing the treatment effects

Analysis of variance table:

Under the null hypothesis H 0 :   0 , the design is one way analysis of variance set up with blocks as
classifications. In this set up, we have the following:
b
Bi2 G 2
sum of squares due to blocks   
i 1 ki n
 B1 
 
1 1 1   B2  G 2
 ( B1 , B2 ,..., Bb ) ' diag  , ,...,  
 k1 k2 kb     n
 
 Bb 
G2
 BK 1 B 
n
If y is the vector of all the observations, then

Error sum of squares ( Se )   ( yijm  ˆ  ˆi  ˆ j ) 2

i j m

  yijm ( yijm  ˆ  ˆi  ˆ j ) [Using normal equations, other terms will be zero]
i j m

  yijm
2
 ˆ G   ˆ jV j   ˆi Bi
i j m j i

 y ' y  ˆ G  V 'ˆ  B ' ˆ .

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

11

11
Using original normal equations given by
B  KEb1ˆ  K ˆ  Nˆ ,

we have
ˆ  K 1 B  Eb1ˆ  K 1 Nˆ.

Since G  V ' E1b  B ' Eb1 , substituting ˆ in Se gives

Se  y ' y  G ˆ  B '[ K 1 B  Eb1ˆ  K 1 Nˆ]  V 'ˆ
 y ' y  G ˆ  B '[ K 1 B  Eb1ˆ  K 1 Nˆ]  V 'ˆ
 y ' y  G ˆ  B ' K 1 B  G ˆ  B ' K 1 Nˆ  V 'ˆ
 y ' y  B ' K 1B  ( B ' K 1 N  V ')ˆ
 G2   G2 
  (V  N ' K B) 'ˆ
1 1
 y' y  
  B ' K B 
 n   n 
 G  
2
G2 
Se   y' y  
  B ' K 1
B    Q 'ˆ
 n   n 
   
Error SS = TotalSS Block SS Adjusted treatment SS
(unadjusted) (adjusted for blocks)
The degrees of freedom associated with the different sum of squares are as follows:

Block SS (unadjusted) : b – 1
Treatment SS (adjusted) : v – 1
Error SS : n – b – v + 1
Total SS : n – 1

The adjusted treatment sum of squares and the sum of squares due to error are independently distributed
and follow a Chi-square distribution with (v  1) and (n  b  v  1) degrees of freedom, respectively.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

12

12
The analysis of variance table for H 0 :   0 is as follows:

Source of Degrees of Sum of Mean sum F

variation freedom squares of squares
Q 'ˆ Q 'ˆ Q 'ˆ / (v  1)
Treat v–1 F
(Adjusted) v 1 S e / ( n  b  v  1)
G2
B ' K 1 B 
Blocks b–1 n
(Unadjusted)
Se
Error n–b–v+1 Se
n  b  v 1
G2
Total n–1 y' y 
n

Q 'ˆ /(v  1)
Under H 0 , ~ F (v  1, n  b  v  1) .
S e /( n  b  v  1)

Thus in an incomplete block design, it matters whether we are estimating the block effects first and then
the treatment effects are estimated
or
first estimate the treatment effects and then the block effects are estimated.

In complete block designs, it doesn’t matter at all. So the testing of hypothesis related to the block and
treatment effects can be done from the same estimates.

A reason for this is as follows: In an incomplete block design, either the

 Adjusted sum of squares due to treatments, unadjusted sum of squares due to blocks and
corresponding sum of squares due to errors are orthogonal

or

 Adjusted sum of squares due to blocks, unadjusted sum of squares due to treatments and
corresponding sum of squares due to errors are orthogonal .

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

13

13
Note that the adjusted sum of squares due to treatment and the adjusted sum of squares due to blocks are
not orthogonal. So
either
Error S.S =Total SS– SS block (Unadjusted) – SS treat (Adjusted)
holds true
or
Error S.S = Total SS – SS block (Adjusted) – SS treat (Unadjusted)
holds true due to Fisher Cochran theorem .
Since CEv1  0, so C is a rank deficient matrix. Also, since

Q ' Ev1  V ' Ev1  ( N ' K 1 B ) Ev1

 (V1 ,...,Vv ) Ev1  B ' K 1 NEv1
 ( Vi )  B ' K 1k '
i

 k1 
 
1 1 1 k
  Vi  ( B1 ,..., Bb )diag  , ,...,   2 
i  k1 k2 kb    
 
 kb 
  Vi  ( B1.....Bb )Eb1
i

  Vi   B j
i j

 G G
0
so the intrablock equations are consistent.

We will confine our attention to those designs for which rank(C) = v - 1. These are called connected
designs and for which all contrasts in the treatments, i.e., all linear combinations  ' where l ' Ev1  0
have unique least squares solutions. This we prove now as follows.

Let G* and H* be any two generalized inverses of C by which we mean that they are square matrix of
order v such that G*Q and H*Q are both the solution vectors to the intrablock equation, i.e.,
ˆ  G * Q and ˆ  H * Q , respectively.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

14

14
Then
Q  Cˆ
 Q  CG * Q
and Q  CH * Q for all Q
so that C (G *  H *)Q  0.

It follows that (G* - H*)Q can be written as a * Ev1 where a is any scalar which may be zero.

Let  be a vector such that  ' Ev1  0. The two estimates of  ' are l ' G * Q and l ' H * Q but
 ' G * Q   ' H * Q   '(G *  H *)Q
  ' a * Ev1
 a *  ' Ev1
0
 l ' is unique.

Theorem: The adjusted treatment totals are orthogonal to the block totals.
Proof: It is enough to prove that
Cov ( Bi , Q j )  0 for all i , j.

Now
  nij  
Cov( Bi , Q j )  Cov  Bi ,V j     Bi 
 i  ki  
nij
 Cov( Bi ,V j )  Var ( Bi )
ki
because the block totals are mutually orthogonal, see how:
v
For y11 , y12 ,..., y1v , the block total B1   y1 j .
j 1
v
For y21 , y22 ,..., y2 v , the block total B2   y2 j .
j 1
v
Var ( B1 )   Var ( y1 j )  v 2 as Cov( y1 j , y1k )  0 for j  k
j 1
v
Var ( B2 )   Var ( y2 j )  v 2 as Cov ( y2 j , y2 k )  0 for j  k
j 1

Var ( B1 )  Var ( B2 )  2v 2 as Cov( y1 j , y2 k )  0 for j  k

 B1 and B2 are mutually orthogonal as all yij 's are independent.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

15

15
As Bi and V j have nij observations in common and the observations are mutually independent, so

Cov( Bi , V j )  nij 2
Var ( Bi )  ki 2
nij
Thus Cov( Bi , Q j )  nij 2  ki 2  0
ki

Hence proved.

Theorem:
E (Q)  C
Var (Q)   2C
Proof:
n B n B 
Q j  V j   1 j 1  ...  bj b 
 k1 kb 
b n B
 Vj  
ij i

i 1 ki
b nij
E (Q j )  E (V j )   E ( Bi )
i 1 ki
E (V j )   E ( yijm )
i m

  E (    i   j   ijm )
i m

   nij    i nij  j  nij

i i i

  rj    i nij  j rj
i

E ( Bi )   E ( yijm )
j m

  E (   i   j   ijm )
j m

  (  
j m
i  j )

  ki  i ki   j nij
j

b nij b nij  
 k E(B )   k i   ki   i ki   j nij 
i 1 i i 1 i  j 
n
  rj    i nij   ij (  j nij ).
i i ki j

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

16

16
Thus substituting these expressions in E (Q j ), we have

nij
E (Q j )  rj j   (  j nij )
i ki j

 n2  n
  ri   ij  j   ij n
 
i ki  i ki
i 
 j ()

 c jj j   cj ' j '

j '(   )

Further, substituting E (Q j ) in E (Q)  ( E (Q1 ), E (Q2 ),..., E (Qb )) ' , we get

E (Q)  C
Next
 Var (Q1 ) Cov(Q1 , Q2 ) ... Cov(Q1 , Qv ) 
 
Cov(Q2 , Q1 ) Var (Q2 ) ... Cov(Q2 , Qv ) 
Var (Q)  
     
 
 Cov(Qv , Q1 ) Cov(Qv , Q2 ) ... Var (Qv ) 
nij
Var (Q j )  Var[V j   Bi ]
i ki
2
n  nij
 Var (V j )    ij  Var ( Bi )  2 Cov(V j , Bi ),
i  ki  i ki

Note that
Var (V j )  Var ( yijm )  rj 2
i m

Var ( Bi )  Var ( yijm )  ki 2

j m

 
Cov(V j , Bi )  Cov   yijm ,  yijm 
 i m j m 
 nij 2

2
n  n 
Var (Q j )  rj    ij  ki 2  2  ij
2
 nij
2

i  ki 

i  ki 
nij2  nij2  2
 rj 2    2  2   
i ki i k
  i

 n2 
 rj 2    ij   2
 
i  ki 

 c jj 2 .

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

17

17
  n n 
Cov(Q j , Q )  Cov V j   ij Bi , V   i Bi 
 i ki i ki 
n n nn
 Cov(V j , V )   i Cov(V j , Bi )   ij Cov( Bi ,V )   ij 2i Cov( Bi , Bi )
i ki i ki i ki
ni n nn
 0 Cov(V j , Bi )  ij Cov( Bi , V )   ij 2i Var ( Bi )
i ki i ki i ki
n n  n n  nn
   i ij   2    ij i   2   ij 2i ki 2
i  ki  i  ki  i ki
 c j 2

Substituting the terms of Var (Q j )  c jj 2 and Cov (Q j , Ql )  c jl 2 in Var (Q ), we get Var (Q )  C 2 .

Hence proved.
[Note: We will prove this result using matrix approach later].

Covariance matrix of adjusted treatment totals:

Consider
V 
Z    with b  v variable.
 B
We can express
Q  V  N ' K 1 B
V 
 [I  N ' K 1 ]  
B
 [I  N ' K 1 ]Z .
So
I ' 
Cov(Q )  [ I  N ' K 1 ] Cov ( Z )  1 
.
 (  N ' K ) '
Now we find
 Var (V ) Cov(V , B ) 
Cov( Z )   
 Cov ( B, V ) Var ( B ) 
Since Bi and V j have nij observations in common and the observations are mutually independent, so

Cov( Bi , V j )  nij 2
Var ( Bi )  ki 2
Var (V j )  rj 2 .

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

18

18
Thus
 R N ' 2
Cov( Z )   
N K 
 R N '  I  2
Cov(Q )  [ I  N ' K 1 ]    1  
 N K   K N 
I 
 [ R  N ' K 1 N N ' N ']  1   2
 K N 
 ( R  N ' K 1 N ) 2
 C 2 .

Next we show that Cov( B, Q)  0

Cov( B, Q)  Cov( B, V )  Cov( B, V  N ' K 1 B)

 Cov( B, V )  Var ( B) K 1 N
 N 2  KK 1 N 2
 0.

Alternative approach to find/ prove E (Q)  C , D(Q)  C 2

Now we illustrate another approach to find the expectations etc in the set up of an incomplete block
design. We have now learnt three approaches- the classical approach based on summations, the approach
based on matrix theory and this new approach which is also based on the matrix theory. We can choose
any of the approaches. The objective here is to let the reader know these different approaches.

Rewrite the linear model

yijm     i   j   ijm , i  1, 2,..., b; j  1, 2,..., v; m  0,1,.., nij .

as
y   En1  D1'  D2'   

where
  (1 , 2 ,..., v ) '
  ( 1 ,  2 ,..., b ) '

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

19

19
Since Bi and Vj have nij observations in common and the observations are mutually independent, so

denote
D1 : v  n matrix of treatment effect versus N, i.e., (i, j)th element of this matrix is given by

1 if j th observation comes from i th treatment

D1  
0 otherwise
D2 : b  n matrix of block effects versus N, i.e., (i, j)th element of this matrix is given by

1 if j th observation comes from i th block

D2  
0 otherwise.
Following results can be verified:
D1 D1'  R  diag (r1 , r2 ,..., rv ),
D2 D2'  K  diag (k1 , k2 ,..., kb ),
D2 D1'  N or D1 D2'  N '
D1 En1  (r1 , r2 ,..., rv ) '
D2 En1  (k1 , k2 ,..., kb ) '
D1' Ev1  En1  D2' Eb1.
In earlier notations,
V  (V1 ,V2 ,...,Vv ) '  D1 y
B  ( B1 , B2 ,..., Bb ) '  D2 y

Express Q in terms of D1 and D2 as

Q  V  N ' K 1 B
  D1  D1 D2' ( D2 D2' ) 1 D2  y
E (Q)   D1  D1 D2' ( D2 D2' ) 1 D2  E ( y )
  D1  D1 D2' ( D2 D2' ) 1 D2  (  En1  D1'  D2'  )
  D1 En1  D1 D2' ( D2 D2' ) 1 D2 En1     D1 D1'  D1 D2' ( D2 D2' ) 1 D2 D1'  
  D1 D2'  D1 D2' ( D2 D2' ) 1 D2 D2'  
 (r1 , r2 ,..., rv ) ' N ' K 1 (k1 ,..., kb ) '    R  N ' K 1 N     N ' N ' K 1 K   .

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

20

20
  k1 
 
1 1 1 k
Since N ' K 1 (k1 ,..., kb ) '  N ' diag  , ,...,   2 
 k1 k2 kb    
 
 kb 
1  n11 n21... nb1  1
   n n ... n   
1 b 2   1
 N '     12 22
        
    
1  niv n2 v ...nbv  1
'
 b b b

   ni1 ,  ni 2 ,...,  niv   (r1 , r2 ,..., rv ) '.
 i 1 i 1 i 1 

Thus
N ' N ' K 1 K   r1 , r2 ,..., rv  ' N ' K 1 (k1 ,..., kb )  0
and so
E (Q)   R  N ' K 1 N  
 C .
Next
Var (Q)  D1  I  D2' ( D2 D2' ) 1 D2  Var ( y )  I  D2' ( D2 D2' ) 1 D2  D1'
  2 D1  I  D2' ( D2 D2' ) 1 D2  D1'
  2  D1 D1'  D1 D2' ( D2 D2' ) 1 D2 D1' 
  2  R  N ' K 1 N 
  2C.
Note that  I  D2' ( D2 D2' ) 1 D2  is an idempotent matrix.

Similarly, we can also express

P  B  NR 1V
 [ D2  D2 D1' R 1D1 ] y.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

21

21
Theorem: E ( P)  D , Var ( P)   2 D
Proof:
D  K  NR 1 N '
P  B  NR 1V
 D2 [ I  D1R 1D1 ] y
 D2 [ I  D1' ( D1D1' )1 D1 ] y

E ( P )  D2 [ I1  D1' ( D1 D1' ) 1 D1 ](  En1  D1'  D2'  )

 [ D2 En1  D2 D1' R 1 D1 En1 ]  [ D2 D1'  D2 D1' R 1 D1 D1' ]
 [ D2 D1'  D2 D1' R 1 D1 D2' ]
 [(k1 , k2 ,..., kb ) ' NR 1 (r1 , r2 ,..., rv ) ']  [ N  NR 1 R]  [ K  NR 1 N ']
 [(k1 , k2 ,..., kb ) ' NEv1 ]  0  D
 [(k1 , k2 ,..., kb ) ' ( k1 , k2 ,..., kb ) ']  D
 D
Next
Var ( P)   2 D2 [ I  D1' ( D1 D1' ) 1 D1 ]D2'
  2 [ D2 D2'  D2 D1' ( D1 D1' ) 1 D1 D2' ]
  2 [ K  NR 1 N ']   2 D

Note that [ I  D1' ( D1 D1' ) 1 D1 ] is an idempotent matrix.

Alternatively, we can also find Var(P) as follows:

B
P  (I  NR 1 )    ( I  NR 1 ) Z where Z  ( B, V ) '
V 
I 
Var ( P ) =( I  NR 1 )Cov( Z )  1 
 R N '
 K N  I  2
= ( I  NR 1 )    1  
 N ' R   R N '
I 
 ( K  NR 1 N ' N  NR 1 R)  1   2
 R N '
  K  NR 1 N '   2
 D 2
Now we consider some properties of incomplete block designs.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

22

22
 Lemma: b + rank(C) = v + rank(D) .
Proof: Consider (b + v)x(b + v) matrix
K N
A
N ' R 
Note that A is a submatrix of C.
Using the result that the rank of a matrix does not changes by the pre-multiplication of nonsingular
matrix, consider the following matrices:
 Ib 0
M  1 
 N ' K Iv 
and
 Ib 0
S   1 .
R N ' Iv 
M and S are nonsingular, so we have
rank(A) = rank(MA) = rank(AS).
Now
 Ib 0  K N  K N
MA    
 N ' K
1
Iv   N ' R  0 C 
D N
AS  
0 R 
Thus
K N D N
rank    rank 
0 C 0 R 
or
rank ( K )  rank (C )  rank ( D)  rank ( R )
or
b  rank (C )  v  rank ( D)
Remark:
C : v  v and D : b  b are symmetric matrices.
One can verify that
CEv1  0

and DEb1  0
Thus rank (C )  v  1
and rank ( D )  b  1.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

23

23
Lemma:
If rank(C) = v - 1, then all blocks and treatment contrasts are estimable.
Proof:
If rank(C) = v – 1, it is obvious that all the treatment contrasts are estimable.
Using the result from the lemma b + rank(C) = v + rank(D), we have

rank(D) + v = rank(C) + b
=v-1+b
Thus
rank(D) = b - 1.
Thus all the block contrasts are also estimable.

Orthogonality of Q and P:
Now we explore the conditions under which Q and P can be orthogonal.
Q  V  N ' K 1 B
 ( D1  D1 D2' K 1 D2 ) y
P  B  NR 1V
 ( D2  D2 D1' R 1 D1 ) y
Cov(Q, P )  ( D1  D1 D2' K 1 D2 )( D2  D2 D1' R 1 D1 ) ' 2
= (D1 D2'  D1 D2' R 1 D1 D2'  D1 D2' K 1 D2 D2'  D1 D2' K 1 D2 D1' R 1 D1 D2' ) 2
 ( N ' RR 1 N ' N ' K 1 K  N ' K 1 NR 1 N ') 2
 ( N ' K 1 NR 1 N ' N ') 2

Q and P (or equivalently Qi and Pj ) are orthogonal when

Cov(Q, P )  0
or N ' K 1 N R 1 N ' N '  0 (i)
 ( R  C ) R N ' N '  0 (Using C  R  N ' K 1 N )
1

 CR 1 N '  0 (ii)
or equivalently
N ' K 1 NR 1 N ' N '  0
 N ' K 1 ( K  D )  N '  0 (Using D  K  NR 1 N ')
 N ' K 1 D  0 (iii)

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

24

24
Thus Qi and Pj are orthogonal if

NR 1 N ' K 1 N  N '
or equivalently
NR 1C  0
or equivalently
DK 1 N  0.

Orthogonal block design:

A block design is said to be orthogonal if Qi ' s and Pj ' s are orthogonal for all i and j. Thus the

condition for the orthogonality of a design is NR 1 N ' K 1 N  N , NR 1C  0 or DK -1 N  0.

nij nij
Lemma: If is constant for all j, then is constant for all i and vice versa. In this case, we have
rj ki

ki r j
nij  .
n
nij nij
Proof. If is constant for all j then = ai , say.
rj rj

 nij  ai rj
or  nij   ai rj  ai  rj  ai n
j j j

or ki  ai n
ki
or ai 
n
Thus
nij ki

rj n
ki r j
or nij 
n
nij ri
So  : independent of i.
kj n
Hence proved.

Contrast:
v v
A linear function  c j j  C '
j 1
where c1 , c2 ,..., cv are given number such that c
j 1
j  0 is called a

contrast of  j ' s.
Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur
25

25
Elementary contrast:
v
A contrast 
j 1
c j j  C ' with C  (c1 , c2 ,..., cv ) ' in treatment effects   (1 , 2 ,..., v ) ' is called an

elementary contrast if C has only two non-zero components 1 and -1.

Elementary contrasts in the treatment effects involve all the differences of the form  i   j , i  j.

It is desirable to design experiments where all the elementary contrasts are estimable.

Connected Design:
A design where all the elementary contrasts are estimable is called a connected design otherwise it is
called a disconnected design.

The physical meaning of connectedness of a design is as follows:

Given any two treatment effects  i1 and  i 2 , it is possible to have a chain of treatment effects like

 i1 , 1 j , 2 j ,..., nj , i 2 , such that two adjoining treatments in this chain occur in the same block.

Example of connected design:

In a connected design, within every block, all the treatment contrasts are estimable and pair-wise
comparison of estimators have similar variances .

Consider a disconnected incomplete block design as follows:

b = 8 (Block numbers: I, II,…,VIII), k = 3, v = 8 (treatment numbers: 1,2,…,8), r = 3

Blocks Treatments
I 1 3 5
II 2 4 6
III 3 5 7
IV 4 6 8
V 5 7 1
VII 6 8 2
VII 7 1 3
VIII 8 2 4

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

26

26
The blocks of this design can be represented graphically as follows:

3 5

1 7

No

8 2

6 4
Note that it is not possible to reach the treatment, e.g., 7 from 2, 3 from 4 etc. So the design is not
connected.

Moreover if the blocks of the design are given like in the following figure, then any treatment can be
reached from any treatment. So the design in this case is connected. For example, treatment 2 can be
reached from treatment 6 through different routes like
6  5  4  3  2, 6  3  2, 6  7  8  1  2, 6  7  2 etc.
2 3

1 4

8 5

7 6
A design is connected if every treatment can be reached from every treatment via lines in the
connectivity graph.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

27

27
Theorem: An incomplete block design with v treatments is connected if and only if rank(C) = v - 1.

Proof: Let the design be connected. Consider a set of (v - 1) linearly independent contrasts  i   j (j = 2,

3,…,v). Let these contrasts be C 2'  , C3' ,..., Cv' where   ( .1 , .2 ,..... v ) '. Obviously, vectors

C2 , C3 ,..., Cv form the basis of vector space of dimension (v - 1). Thus any contrast p ' is

expressible as a linear combination of the contrasts Ci' (i  2, 3,..., v ). Also, p ' is estimable if and only
if p belongs to the column space of C-matrix of the design.

Therefore, dimension of column space of C must be same as that of the vector space spanned by the
vectors Ci (i  2,3,..., v), i.e., equal to (v - 1).
Thus rank(C) = v – 1.

Conversely, let rank(C) = v – 1 and let 1 ,  2 ,..., v 1 be a set of orthonormal eigen vectors corresponding

to the (not necessarily distinct) non-zero eigenvalues 1 , 2 ,..., v 1 of C .

Then
E (1'Q)  i'C
 ii' .

i'Q
Thus an unbiased estimator of  i' is .
i
Also, since each i is orthogonal to Ev1 and  i' s are mutually orthogonal, so any contrast p '
v 1
belongs to the vector space spanned by { i , i  1...v}, i.e., p   ai i .
i 1

 v 1  Q 
So E  ai i   p ' .
 i 1 i 
Thus p ' is estimable and this completes the proof.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

28

28
 rk '
Lemma: For a connected block design Cov(Q, P)  0 if and only if N '  .
n
Proof: “if” part
rk '
When N '  , we have
n
1
Cov(Q, P )  N ' K 1 NR 1 N ' N '
2
rk ' K 1kr ' R 1 N '
 N'
n2
 k1 
 
1 1 1 1   k2 
Since k ' K k  (k1 , k2 ,..., kb ) diag  , ,..., 
 k1 k2 kb    
 
 kb 
 k1  b
 
 (1,...,1)      ki  n
 k  i 1
 b
and
1 1 1
r ' R 1  (r1 , r2 ,..., rv ) diag  , ,...,   E1v .
 r1 r2 rv 

Then
1 rnE1v N '
Cov(Q, P)  N'
 2
n2
rE N ' rk '
 1v  N '  N'
n n
 N ' N '  0.

“Only if part”
Let Cov(Q, P)  0
 N ' K 1 NR 1 N ' N '  0 (Since C  R  N ' K 1 N )
or ( R  C ) R 1 N ' N '  0
or CR 1 N '  0.

Let
R 1 N '  A  ( a1 , a2 ,..., ab )
where a1 , a2 ,..., ab are the columns of A.
Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur
29

29
Since the design is connected, so the columns of A are proportional to Ev1 . Also all row/column sums of
C are zero.
So (CEv1 , CEv1 ,..., CEv1 )  0
and
CA  0
or C ( a1 , a2 ,..., ab )  0
 ai  Ev1
or ai   i Ev1 ; i  1, 2,..., b

where  i are some scalars.

This gives
A  R 1 N '  Ev1 ' where   (1 .... b ) '.

So we have
N '  REv1  (r1 , r2 ,.., rv ) ' '
 r ' where r  (r1 , r2 ,.., rv ) '.

Premultiply by E1v gives

E1v N '  (k1 , k2 ,..., kb ) '  E1v r '  n '

or k  n '
k'
 ' = where k  (k1 , k2 ,..., kb ) '
n
Thus
rk '
N '  r '  .
n
Hence proved.

Definition: A connected block design is said to be orthogonal if and only if the incidence matrix of the
rk '
design satisfies the condition N '  .
n
Design which do not satisfy this condition are called non-orthogonal. It is clear from this result that if at
least one entry of N is zero, the design cannot be orthogonal.

A block design with at least one zero entry in its incidence matrix is called an incomplete block design.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

30

30
 nij
Theorem: A sufficient condition for the orthogonality of a design is that is constant for all j.
rj

Conclusion: It is obvious from the condition of orthogonality of a design that a design which is not
connected and an incomplete design even though it may be connected cannot have an orthogonal
structure.

Now we illustrate the general nature of the incomplete block design. We try to obtain the results for a
randomized block design through the results of an incomplete block design.

Randomized block design:

Randomized block design is an arrangement of v treatment in b blocks of v plots each, such that
every treatment occurs in every block, one treatment in each plot.
The arrangement of treatment within a block is random and in terms of incidence matrix,
nij  1 for all i  1, 2,..., b; j  1, 2,..., v.

Thus we have
ki   nij  v for all i
j

rj   nij  b for all j.

j

nij 1
We have  constant for all j.
rj b
b
C jj  b 
v
b
C jj '  
v
G
Qj  Vj  .
v
Normal equations for  ' s are

 b b v G


b 
v

 j  
v i  j '1
 j '  V j  ; j  1....v
v
 1   2 ....   v  0.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

31

31
Thus
b G
b j  
v j
 j  Vj 
v
1 G
or ˆ j   V j    yoj  yoo .
b v
The sum of squares due to treatments adjusted for blocks is
  ˆ j Q j
j
2
1  G
  V j  
b j  v
V j
2

G2
 
j
,
b bv
which is also the sum of squares due to treatments which are unadjusted for blocks because the design is
orthogonal.

B i
2
G2
Sum of squares due to blocks  i

v bv
2
 B V G
Sum of squares due to error    yij  i  j   .
i j  v b bv 

These expressions are the same as obtained under the analysis of variance in the set up of a randomized
block design.

Interblock analysis of incomplete block design

The purpose of block designs is to reduce the variability of response by removing the part of the
variability as block numbers. If in fact this removal is illusory, the block effects being all equal, then the
estimates are less accurate than those obtained by ignoring the block effects and using the estimates of
treatment effects. On the other hand, if the block effect is very marked, the reduction in the basic
variability may be sufficient to ensure a reduction of the actual variances for the block analysis.

In the intrablock analysis related to treatments, the treatment effects are estimated after eliminating the
block effects. If the block effects are marked, then the block comparisons may also provide information
about the treatment comparison. So a question arises how to utilize the block information additionally to
develop an analysis of variance to test the hypothesis about the significance of treatment effects.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

32

32
Such an analysis can be derived by regarding the block effects as random variables. This assumption
involves the random allocation of different blocks of the design to be the blocks of material selected (at
random from the population of possible blocks) in addition to the random allocation of treatments
occurring in a block to the units of the block selected to contain them. Now the two responses from the
same block are correlated because the error associated with each contains the block number in common .
Such an analysis of incomplete block design is termed as interblock analysis.

To illustrate the idea behind the interblock analysis and how the block comparisons also contain
information about the treatment comparisons, consider an allocation of four selected treatments in two
blocks each. The outputs ( yij ) are recorded as follows:

Block 1: y14 y16 y17 y19

Block 2 : y23 y25 y26 y27.

The block totals are

B1  y14  y16  y17  y19 ,
B2  y23  y25  y26  y27 .

Following the model yij     i   j   ij , i  1, 2, j  1, 2,..., 9, we have

y14    1   4  14 ,
y16    1   6  16 ,
y17    1   7  17 ,
y19    1   9  19 ,
y23     2   3   23 ,
y25     2   5   25 ,
y26     2   6   26 ,
y27     2   7   27 ,
and thus
B1  B2  4( 1   2 )  ( 4   6   7   9 )  ( 3   5   6   7 )
 (14  16  17  19 )  ( 23   25   26   27 ).

If we assume additionally that the block effects 1 and  2 are random with mean zero, then

E ( B1  B2 )  ( 4   9 )  ( 3   5 )
which reflects that the block comparisons can also provide the information about the treatment
comparisons.
Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur
33

33
The intrablock analysis of an incomplete block designs is based on estimating the treatment effects (or
their contrasts) by eliminating the block effects. Since different treatments occur in different blocks, so
one may expect that the block totals may also provide some information on the treatments. The interblock
analysis utilizes the information on block totals to estimate the treatment differences. The block effects
are assumed to be random and so we consider the set up of mixed effect model in which the treatment
effects are fixed but the block effects are random. This approach is applicable only when the number of
blocks are more than the number of treatments. We consider here the interblock analysis of binary proper
designs for which nij  0 or 1 and k1  k2  ...  kb  k in connection with the intrablock analysis.

Model and Normal Equations

Let yij denotes the response from the jth treatment in ith block from the model

yij   *   i*   j   ij , i  1, 2,..., b; j  1, 2,..., v

where
 * is the general mean effect;
 i* is the random additive ith block effect;
 j is the fixed additive jth treatment effect; and

 ij is the i.i.d. random error with  ij ~ N (0,  2 ) .

Since the block effect is now considered to be random, so we additionally assume that  i* (i  1, 2,.., b )

are independently distributed following N (0,  2 ) and are uncorrelated with  ij . One may note that we

can not assume here i

i
*
 0 as in other cases of fixed effect models. In place of this, we take

E (  i* )  0 . Also, yij ' s are no longer independently distributed but

Var ( yij )   2   2 ,

  if i  i ', j  j '

2

Cov ( yij , yi ' j ' )  

0 otherwise.

In case of interblock analysis, we work with the block totals Bi in place of yij where

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

34

34
 v
Bi   nij yij
j 1
v
  nij (  *  i*   j   ij )
j 1

 k  *   nij j  fi
j

where f i   i* k   nij  ij , (i  1, 2,..., b ) are independent and normally distributed with mean 0 and
j

Var ( f i )  k 2 2  k 2   2f .

Thus
E ( Bi )  k  *   nij j ,
j

Var ( Bi )   ; 2
f i  1, 2,.., b,
Cov( Bi , Bi ' )  0; i  i '; i, i '  1, 2,..., b.

In matrix notations, the model under consideration can be written as

B  k  * Eb1  N  f

where f  ( f1 , f 2 ,..., fb ) '.

Estimates of  *and  in interblock analysis:

In order to obtain the estimates of  * and  , we minimize the sum of squares due to error

f  ( f1 , f 2 ,..., fb ) ', , i.e., minimize ( B  k  * Eb1  N ) '( B  k  * Eb1  N ) with respect to  * and  .
The estimates of  * and  are the solutions of following normal equations:

 kEb' 1      kEb' 1 
 N '   b1      N ' 
  kE N  B
     

 k 2 Eb' 1 Eb1 kEb' 1 N      kG 

or      
 kN ' E '      N ' B 
 b1 N N    

 k 2b kEv' 1 R      kG 
or       (using N ' Eb1  r  REv1 )
 kRE
 v1 N ' N      N ' B 

Premultiplying both sides of the equation by

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur
35

35
  1 0
 ,
  REv1 I v 
 b 
we get
 bk Ev' 1   G 
       
0 REv1 Ev1 R    
'
REv1G  .
 N 'N      N ' B  b 
 b 
Using the side condition Ev' 1 R  0 and assuming N ' N to be nonsingular, we get the estimates of

 * and  as  and  given by

G
  ,
bk
REv1G
  ( N ' N )1 ( N ' B  )
b
 kGN ' Eb1 
 ( N ' N ) 1  N ' B   (using REv1  r  N ' Ev1 )
 bk 
 G 
 ( N ' N ) 1  N ' B  N ' NEv1 
 bk 
GEv1
 ( N ' N ) 1 N ' B  .
bk

The normal equations can also be solved in an alternative way also as follows.
 k 2b kEv' 1 R      kG 
The normal equations       can be written as
 kREv1 N ' N     N '  
k 2b  kEv' 1 R  kG
kREv1  N ' N  N ' B.

Using the side condition Ev' 1 R  0 (or equivalently  r 

j
j j  0) and assuming N ' N to be nonsingular,

G
the first equation gives   . Substituting  in the second equation gives .
bk

REv1G 
   N ' N   N ' B 
1

 b 
GEv1
 N 'N N 'B 
1
.
bk

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

36

36
Generally, we are not interested merely in the interblock analysis of variance but we want to utilize the
information from interblock analysis along with the intrablock information to improve upon the statistical
inferences.

After obtaining the interblock estimate of treatment effects, the next question that arises is how to use this
information for an improved estimation of treatment effects and use it further for the testing of
significance of treatment effects. Such an estimate will be based on the use of more information, so it is
expected to provide better statistical inferences.

We now have two different estimates of the treatment effect as

- based on intrablock analysis ˆ  C Q and
GEv1
- based on interblock analysis   ( N ' N ) 1 N ' B  .
bk
Let us consider the estimation of linear contrast of treatment effects L  l ' . Since the intrablock and
interblock estimates of  are based on Gauss-Markov model and least squares principle, so the best
estimate of L based on intrablock estimation is
L1  l 'ˆ
 l ' C Q
and the best estimate of L based on interblock estimation is
L2  l '
 GEv1 
 l ' ( N ' N ) 1 N ' B  
 bk 
1
 l '( N ' N ) N ' B (since l ' Ev1  0 being contrast.)

The variances of L1 and L2 are

Var ( L1 )   2 l ' C  l

and
Var ( L2 )   2f l '( N ' N ) 1 l ,

respectively. The covariance between Q (from intrablock) and B (from interblock) is

Cov(Q, B)  Cov(V  N ' K 1 B*, B)
 Cov(V , B)  Cov( N ' K 1 B*, B)
 N ' 2f  N ' K 1 K 2f
 0.
Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur
37

37
Note that B* denotes the block total based on intrablock analysis and B denotes the block totals based
on interblock analysis . We are using two notations B and B* just to indicate that the two block totals are
different. The reader should not misunderstand that it follows from the result Cov(Q, B)  0 in case of
of intrablock analysis.

Thus
Cov( L1 , L2 )  0
irrespective of the values of l .

The question now arises that given the two estimators ˆ and  of  , how to combine them and obtain a
minimum variance unbiased estimator of  . It is illustrated with following example:

Example:
Let ˆ1 and ˆ2 be any two unbiased estimators of a parameter  with Var (ˆ1 )   12 and Var (ˆ 2 )   22 .

Consider a linear combination ˆ  1ˆ1   2ˆ2 with weights 1 and 2 . In order that ̂ is an unbiased

estimator of 1 , we need

E (ˆ )  
or 1 E (ˆ1 )   2 E (ˆ2 )  
or 1   2  
or 1   2  1.

    ˆ
So modify ̂ as 1 1 2 2 which is the weighted mean of ̂1 and ̂2 .
1   2
Further, if ˆ1 and ˆ2 are independent , then

Var (ˆ )  12 12   22 22 .

Now we find 1 and 2 such that Var (ˆ ) is minimum such that 1   2  1 .

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

38

38
 Var (ˆ )
 0  21 12  2(1  1 ) 22  0
1
or 1 12   2 22  0
1  22
or 
 2  12
1
or weight  .
variance
Alternatively, the Lagrangian function approach can be used to obtain such result as follows. The
Lagrangian function with  * as Lagrangian multiplier is given by
  Var (ˆ )   *(1   2  1)

    2
and solving  0, and  0 also gives the same result that 1  22 .
1  2  * 2 1

We note that a pooled estimator of  in the form of weighted arithmetic mean of uncorrelated
L1 and L2 is the minimum variance unbiased estimator of  when the weights 1 and  2 of L1 and L2 ,
respectively are chosen such that
1 Var ( L2 )
 ,
 2 Var ( L1 )
i.e., the chosen weights are reciprocal to the variance of respective estimators, irrespective of the values of
l . So consider the weighted average of L1 and L2 with weights 1 and  2 , respectively as

1 L1   2 L2
* 
1   2
l '(1ˆ   2 )

1   2
with
11  l ' C 1l 2
 21  l '( N ' N ) 1 l 2f .

The linear contrast of  * is

L*  l ' *
and its variance is

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

39

39
 12Var ( L1 )   22Var ( L2 )
Var ( L*)  l 'l (since Cov( L1 , L2 )  0)
(1   2 ) 2
l 'l

(1   2 )
because the weights of estimators are chosen to be inversely proportional to the variance of the respective
estimators. We note that  * can be obtained provided 1 and 2 are known. But 1 and 2 are known

only if  2 and  2 are known. So  * can be obtained when  2 and  2 are known. In case, if

 2 and  2 are unknown, then their estimates can be used. A question arises how to obtain such

estimators ? One such approach to obtain the estimates of  2 and  2 is based on utilizing the results from

intrablock and interblock analysis both and is as follows.

From intrablock analysis

E ( SS Error ( t ) )  ( n  b  v  1) 2 ,

so an unbiased estimator of  2 is
SS Error (t )
ˆ 2  .
n  b  v 1

An unbiased estimator of  2 is obtained by using the following results based on the intrablock analysis:

v V j2 G2
SSTreat (unadj )    ,
j 1 j n
b
Bi2 G 2
SS Block (unadj )    ,
i 1 ki n
v
SSTreat ( adj )   Q jˆ j ,
j 1
b v
G2
SSTotal   yij2  ,
i 1 j 1 n

where
SSTotal  SSTreat ( adj )  SS Block (unadj )  SS Error (t )
 SSTreat (unadj )  SS Block ( adj )  SS Error (t ) .
Hence
SS Block ( adj )  SSTreat ( adj )  SS Block ( unadj )  SSTreat (unadj ) .

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

40

40
Under the interblock analysis model
E[ SS Block ( adj ) ]  E[ SSTreat ( adj ) ]  E[ SS Block ( unadj ) ]  E[ SSTreat ( unadj ) ]

which is obtained as follows:

E[ SS Block ( adj ) ]  (b  1) 2  (n  v) 2
or
 b 1 
E  SS Block ( adj )  SS Error (t )   (n  v) 2 .
 n  b  v 1 

Thus an unbiased estimator of  2 is

1  b 1 
ˆ 2   SS Block ( adj )  SS Error ( t )  .
nv  n  b  v 1 

Now the estimates of weights 1 and 2 can be obtained by replacing  2 and  2 by ˆ 2 and ˆ 2

respectively. Then the estimate of  * can be obtained by replacing 1 and 2 by their estimates and can

be used in place of  * . It may be noted that the exact distribution of associated sum of squares due to
treatments is difficult to find when
 2 and  2 are replaced by ˆ 2 and ˆ 2 , respectively in  * . Some approximate results are possible

which we will present while dealing with the balanced incomplete block design . An increase in the
precision using interblock analysis as compared to intrablock analysis is measured by

1/variance of pooled estimate

 1.
1/variance of intrablock estimate

In the interblock analysis, the block effects are treated as random variable which is appropriate if the
blocks can be regarded as a random sample from a large population of blocks. The best estimate of the
treatment effect from the intrablock analysis is further improved by utilizing the information on block
totals. Since the treatments in different blocks are not all the same, so the difference between block totals
is expected to provide some information about the differences between the treatments. So the interblock
estimates are obtained and pooled with intrablock estimates to obtain the combined estimate of  . The
procedure of obtaining the interblock estimates and then obtaining the pooled estimates is called the
recovery of interblock information.

Analysis of Variance | Chapter 5 | Incomplete Block Designs | Shalabh, IIT Kanpur

41

41