Chapter 6

Balanced Incomplete Block Design (BIBD)

The designs like CRD and RBD are the complete block designs. We now discuss the balanced
incomplete block design (BIBD) and the partially balanced incomplete block design (PBIBD) which
are the incomplete block designs.

A balanced incomplete block design (BIBD) is an incomplete block design in which

- b blocks have the same number k of plots each and
- every treatment is replicated r times in the design.
- Each treatment occurs at most once in a block, i.e., nij  0 or 1 where nij is the number of

times the jth treatment occurs in ith block, i  1, 2,..., b; j  1, 2,..., v.

- Every pair of treatments occurs together is  of the b blocks.

Five parameters denote such design as D(b, k , v, r;  ) .

The parameters b, k , v, r and  are not chosen arbitrarily.
They satisfy the following relations:
( I ) bk  vr
( II )  (v  1)  r (k  1)
( III ) b  v (and hence r  k ).

Hence nj
ij k for all i

nj
ij r for all j

nij
and n1 j nij '  n2 j nij '  ...  nb j nb j '   for all j  j '  1, 2,..., v. Obviously cannot be a constant for all
r
j. So the design is not orthogonal.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
1
Example of BIBD
In the design D(b, k ; v, r;  ) : consider b  10 (say, B1 ,..., B10 ), v  6 (say, T1 ,..., T6 ), k  3, r  5,   2
Blocks Treatments
B1 T1 T2 T5
B2 T1 T2 T6
B3 T1 T3 T4
B4 T1 T3 T6
B5 T1 T4 T5
B6 T2 T3 T4
B7 T2 T3 T5
B8 T2 T4 T6
B9 T3 T5 T6
B10 T4 T5 T6

Now we see how the conditions of BIBD are satisfied.

(i ) bk  10  3  30 and vr  6  5  30
 bk  vr
(ii )  (v  1)  2  5  10 and r ( k  1)  5  2  10
  (v  1)  r ( k  1)
(iii ) b  10  6

Even if the parameters satisfy the relations, it is not always possible to arrange the treatments in blocks
to get the corresponding design.

The necessary and sufficient conditions to be satisfied by the parameters for the existence of a BIBD
are not known.

The conditions (I)-(III) are some necessary condition only. The construction of such design depends
on the actual arrangement of the treatments into blocks and this problem is handled in combinatorial
mathematics. Tables are available, giving all the designs involving at most 20 replication and their
method of construction.

Theorem:
( I )bk  vr
( II )  (v  1)  r (k  1)
( III ) b  v.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
2
Proof: (I)
Let N  ( nij ) : b  v the incidence matrix

Observing that the quantities E1b NEv1 and E1v N ' Eb1 are the scalars and the transpose of each other, we

find their values.

Consider
 n11 n21  nb1  1
 
n n22  nb 2  1
E1b NEv1  (1,1,...,1)  12
      
  
 n1v n2 v  nbv  1
  n1 j 
 j 
 
  n2 j 
 (1,1,...,1)  j 
 
 
  nbj 
 j 
k 
 
k
 (1,1,...,1)1b   = bk .
 
 
k 
Similarly,
 n11 n21  nb1  1
 
n n22  nb 2   
E1v N ' Eb1  (1,...,1)  12
      
  
 n1v n2 v  nbv  1
  ni 
 i  r
=(1,1,...,1)    (1,1,...,1)    vr
1v  
 
  niv  r
 
 i 

But
E1b NEv1  E1v N ' Eb1 as both are scalars.
Thus bk  vr.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
3
Proof: (II)
Consider
 n11 n21  nb1   n11
n12  n1v 
  
n n22  nb 2   n21
n22  n2 v 
N ' N   12
     
  
  
 n1v n2 v  nbv   nb1
nb 2  nbv 
  ni21  ni1ni 2   ni1niv 
 i i i 
 
  ni1ni 2  ni 2   ni 2 niv 
2

 i i i

     
 2 
  niv ni1  niv ni 2   niv 
 i i i 
 r   

 r   
 . (1)
   
 
 r   

Since nij2  1 or 0 as nij  1 or 0,

so n
i
2
ij  Number of times  j occurs in the design

= r for all j  1, 2,..., v of times occurs in the design

and n n
i
ij ij '  Number of blocks in which  j and  j ' occurs together

=  for all j  j '.

 r    1

 r    1

N ' NEv1  
    
  
    r 1
 r   (v  1) 
 
 r   (v  1) 
  [r   (v  1)]Ev1. (2)
 
 
 r   (v  1) 

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
4
Also
 n11 n12  n1v   1 
  
n n22  n2 v   1 
N ' NEv1  N '  21
        
   
 nb1 nb 2  nbv   1 
  n1 j 
 j 
 
  n2 j 
 N ' j 
 
 
  nbj 
 j 
 n11 n21  nb1   k 
 
 n12 n22  nb 2   k 

   
  
 niv n2 v  nbv   k b1
  ni1 
 i 
 n 
  i2 
k i 
 
 n 
  iv 
 i 
r
 
r
k 
 
 
r
 krEv1 (3)

From  2 and  3
[r   (v  1)]Ev1  krEv1
or r   (v  1)  kr
or  (v  1)  r (k  1)

Proof: (III)
From (I), the determinant of N ' N is
det N ' N  [ r   (v  1)]( r   ) v 1
  r  r ( k  1)   r   
v 1

 rk ( r   ) v 1
0

because since if r    from (II) that k = v. This contradicts the incompleteness of the design.
Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
5
Thus N ' N is a v  v nonsingular matrix.
Thus rank ( N ' N )  v.
We know from matrix theory result
rank ( N )  rank ( N ' N )
so rank ( N )  v
But rank ( N )  b , there being b rows in N.
Thus v  b.

Interpretation of conditions of BIBD

Interpretation of (I) bk = vr
This condition is related to the total number of plots in an experiment. In our settings, there are k
plots in each block and there are b blocks. So the total number of plots are bk .
Further, there are v treatments and each treatment is replicated r times such that each treatment
occurs atmost in one block. So total number of plots containing all the treatments is vr . Since both
the statements counts the total number of plots, hence bk  vr.

Interpretation of (II)
 k  k (k  1)
Each block has k plots. Thus the total pairs of plots in a block =    .
2 2
There are b blocks. Thus the total pairs of plots such that each pair consists of plots within a block =
k ( k  1)
b .
2
 v  v(v  1)
There are v treatments, thus the total number of pairs of treatment =    .
 2 2
Each pair of treatment is replicated  times, i.e., each pair of treatment occurs in  blocks.
v (v  1)
Thus the total number of pairs of plots within blocks must be   .
2
k (k  1) v  v  1
Hence b 
2 2
Using bk  vr in this relation, we get r  k  1    v  1 .

Proof of (III) was given by Fisher but quite long, so not needed here.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
6
Balancing in designs:
There are two types of balancing – Variance balanced and efficiency balanced. We discuss the
variance balancing now and the efficiency balancing later.

Balanced Design (Variance Balanced):

A connected design is said to be balanced (variance balanced) if all the elementary contrasts of the
treatment effects can be estimated with the same precision. This definition does not hold for the
disconnected design, as all the elementary contrasts are not estimable in this design.

Proper Design:
An incomplete block design with k1  k2  ....  kb  k is called a proper design.

Symmetric BIBD:
A BIBD is called symmetrical if the number of blocks = number of treatments, i.e., b  v.

Since b  v, so from bk  vr
 k  r.
Thus the number of pairs of treatments common between any two blocks =  .

The determinant of N ' N is

N ' N  [r   (v  1)](r   )v 1
  r  r  k  1   r   
v 1

 rk (r   )v 1.
When BIBD is symmetric, b = v and then using bk  vr , we have k  r. Thus
2
N ' N  N  r 2 (r   )v 1 ,
so
v 1
N   r (r   ) 2
.

Since N is an integer, hence when v is an even number, ( r   ) must be a perfect square. So

N ' N  (r   ) I   Ev1Ev' 1 ,
( N ' N )1  N 1 N '1
1   
  I  2 Ev1Ev' 1  ,
r   r 
1   
N '1   I  2 Ev1 Ev' 1  .
r   r 
Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
7
Post-multiplying both sides by N ' , we get
NN '  ( r   ) I   Ev1 Ev' 1  N ' N .

Hence in the case of a symmetric BIBD, any two blocks have  treatment in common.

Since BIBD is an incomplete block design. So every pair of treatment can occur at most once is a
block, we must have v  k .

If v  k , then it means that each treatment occurs once in every block which occurs in case of RBD.
So in BIBD, always assume v > k .

Similarly   r.
[If   r then  (v  1)  r (k  1)  v  k  which means that the design is RBD]

Resolvable design:
A block design of
- b blocks in which
- each of v treatments is replicated r times
is said to be resolvable if b blocks can be divided into r sets of b / r blocks each, such that every
treatment appears in each set precisely once. Obviously, in a resolvable design, b is a multiple of r.

Theorem: If in a BIBD D(v, b, r , k ,  ), b is divisible by r, then

b  v  r  1.
Proof: Let b  nr (where n  1 is a positive integer).

For a BIBD,  (v  1)  r (k  1)
 because vr  bk 
 (v  1) or
or r   vr  nrk 
(k  1)
 or v  nk 
 (nk  1)

(k  1)
 n 1 
     n.
 k 1 
Since n  1 and k  1, so  n  1 is an integer. Since r has to be an integer.
(n  1)
 is also a positive integer.
k 1

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
8
Now, if possible, let
b  v  r 1
 nr  v  r  1
or r (n  1)  v  1
r (k  1) r (k  1)
or r (n  1)  (because v  1  )
 
 (n  1)
  1which is a contradiction as integer can not be less than one
k 1
 b  v  r  1 is impossible. Thus the opposite is true.
 b  v  r  1 holds correct.

Intrablock analysis of BIBD:

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

where
 is the general mean effect;
i is the fixed additive i th block effect;
j is the fixed additive j th treatment effect and
 ij is the i.i.d. random error with  ij ~ N (0,  2 ).

We don’t need to develop the analysis of BIBD from starting. Since BIBD is also an incomplete block
design and the analysis of incomplete block design has already been presented in the earlier module,
so we implement those derived expressions directly under the setup and conditions of BIBD. Using
v
the same notations, we represent the blocks totals by Bi   yij , treatment totals by V j   yij ,
b

j 1 i 1

b v
adjusted treatment totals by Q j and grand total by G   yij The normal equations are obtained
i 1 j 

by differentiating the error sum of squares. Then the block effects are eliminated from the normal
equations and the normal equations are solved for the treatment effects. The resulting intrablock
equations of treatment effects in matrix notations are expressible as
Q  Cˆ .

Now we obtain the forms of C and Q in the case of BIBD. The diagonal elements of C are given
by
b

n 2
ij
c jj  r  i 1
( j  1, 2,..., )
k
r
r  .
k

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
9
The off-diagonal elements of C are given by
1 b
c jj '    nij nij ' ( j  j '; j, j '  1, 2,..., )
k i 1

 .
k
The adjusted treatment totals are obtained as
1 b
Qj  Vj   nij Bi
k i 1
( j  1, 2,..., )

1
 Vj   Bi
k i( j)

where 
i( j)
denotes the sum over those blocks containing jth treatment. Denote

T j   Bi , then
i( j)

Tj
Qj  Vj  .
k
The C matrix is simplified as follows:
N 'N
C  rI 
k
1
 rI  (r   ) I   Ev1 Ev' 1 
k
 k 1  
 r  I  ( I  Ev1 Evi )
'

 k  k
 v 1  
  I  ( I  Ev1 Evi )
'

 k  k
V  Ev1 Ev' 1 
 I  .
k  v 

Since C is not as a full rank matrix, so its unique inverse does not exist. The generalized inverse of C
is denoted as C  which is obtained as
1
 E E' 
C   C  v1 v1  .
 v 

Since
v 
Ev1 Ev' 1 
C  v
I  
k  v 
kC E E'
or  I v  v1 v1 ,
v v

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
10
 k
the generalized inverse of C is
v
1 1
 k    Ev1 Ev' 1 
  C   C  ,
 v   v 
1
 E E' E E' 
  I v  v1 v1  v1 v1 
 v v 
 Iv .
v
Thus C   Iv .
k
Thus an estimate of  is obtained from Q  C as

ˆ  C Q
v
 Q.
k
The null hypothesis of our interest is H 0 : 1   2  ...   v against the alternative hypothesis H1 : at

least one pair of  j ' s is different. Now we obtain the various sum of squares involved in the

development of analysis of variance as follows.

The adjusted treatment sum of squares is

SSTreat ( adj )  ˆ ' Q
k
 Q 'Q

v
k


Q ,
j 1
2
j

The unadjusted block sum of squares is

b
Bi2 G2
SSBlock (unadj )    .
i 1 k bk
The total sum of squares is
b v
G2
SSTotal   yij2 
i 1 j 1 bk
The residual sum of squares is obtained by
SS Error ( t )  SSTotal  SS Block ( unadj )  SSTreat ( adj )
.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
11
 A test for H 0 : 1   2  ...   v is then based on the statistic

SSTreat ( adj ) /(v  1)

FTr 
SS Error (t ) /(bk  b  v  1)
v

 Q 2j
k bk  b  v  1 j 1
 . .
v v 1 SS Error (t )

If FTr  F1 ,v 1,bk  b  v 1; then H 0( t ) is rejected.

This completes the analysis of variance test and is termed as intrablock analysis of variance. This
analysis can be compiled into the intrablock analysis of variance table for testing the significance of
the treatment effect given as follows.

Intrablock analysis of variance table of BIBD for H 0 : 1   2  ...   v

Source Sum of squares Degrees of Mean squares F
freedom
Between treatment SSTreat ( adj ) v 1 SSTreat ( adj ) MSTreat
MStreat 
(adjusted) v 1 MS E

Between blocks SS Block ( unadj ) b-1

(unadjusted)
SS Error (t )
SS Error (t ) MS E 
Intrablock error bk – b – v + 1 bk  b  v  1
(by substraction)

Total SSTotal  bk  1
G2
 y
i j
2
ij 
bk

In case, the null hyperthesis is rejected, then we go for a pairwise comparison of the treatments. For
that, we need an expression for the variance of the difference of two treatment effects.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
12
The variance of an elementary contrast ( j   j ' , j  j ') under the intrablock analysis is

V *  Var (ˆ j  ˆ j ' )

 k 
 Var  (Q j  Q j ' ) 
 v 
2
k
 2 2 [Var (Q j )  Var (Q j )  2Cov(Q j Q j ' )]

k2
 (c jj  c j ' j '  2c jj ' ) 2
 2 2
k 2   1  2  2
 2r 1   
 2 2   k  k 
2k
  2.
v
This expression depends on  2 which is unknown. So it is unfit for use in the real data applications.
One solution is to estimate  2 from the given data and use it in the place of  2 .

An unbiased estimator of σ 2 is
SS Error ( t )
ˆ 2  .
bk  b   1
Thus an unbiased estimator of V * can be obtained by substituting ̂ 2 in it as
2k SS Error (t )
Vˆ*  . .
 v bk  b   1
If H 0 is rejected, then we make pairwise comparison and use the multiple comparison test. To test

H 0 :  j   j ' ( j  j '), a suitable statistic is

k (bk  b  v  1) Q j  Q j '
t .
v SS Error ( t )

which follows a t-distribution with (bk  b  v  1) degrees of freedom under H 0 .

A question arises that how a BIBD compares to an RBD. Note that BIBD is an incomplete block
design whereas RBD is a complete block design. This point should be kept in mind while making such
restrictive comparison.

We now compare the efficiency of BIBD with a randomized block (complete) design with r
replicates. The variance of an elementary contrast under a randomized block design (RBD) is
2 *2
VR*  Var (ˆ 2j  ˆ j ' ) RBD 
r
where Var ( yij )   *2 under RBD.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
13
Thus the relative efficiency of BIBD relative to RBD is
 2 *2 
Var (ˆ j  ˆ j ' ) RBD  
 
r 
Var (ˆ j  ˆ j ' ) BIBD  2k 2 
 
 v 
 v   *2 
  .
rk   2 
v
The factor  E (say) is termed as the efficiency factor of BIBD and
rk
v v  k 1 
E   
rk k  v 1 
1
 1  1 
 1  1  
 k  v 
 1 (since v  k ).
The actual efficiency of BIBD over RBD not only depends on the efficiency factor but also on the
ratio of variances  *2 /  2 . So BIBD can be more efficient than RBD as  *2 can be more than  2

because k  v.

Efficiency balanced design:

A block design is said to be efficiency balanced if every contrast of the treatment effects is estimated
through the design with the same efficiency factor.

If a block design satisfies any two of the following properties:

(i) efficiency balanced,
(ii) variance balanced and
(iii) an equal number of replications,
then the third property also holds true.

Missing observations in BIBD:

The intrablock estimate of missing (i, j)th observation yij is

vr (k  1) Bi  k (v  1)Q j  (v  1)Q'j
yij 
k (k  1)(bk  b  v  1)

Q 'j : the sum of Q value for all other treatment (but not the jth one) which are present in the

ith block.
All other procedures remain the same.
Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
14
Interblock analysis and recovery of interblock information in BIBD
In the intrablock analysis of variance of an incomplete block design or BIBD, the treatment effects
were estimated after eliminating the block effects from the normal equations. In a way, the block
effects were assumed to be not marked enough and so they were eliminated. It is possible in many
situations that the block effects are influential and marked. In such situations, the block totals may
carry information about the treatment combinations also. This information can be used in estimating
the treatment effects which may provide more efficient results. This is accomplished by an interblock
analysis of BIBD and used further through the recovery of interblock information. So we first conduct
the interblock analysis of BIBD. We do not derive the expressions a fresh but we use the assumptions
and results from the interblock analysis of an incomplete block design. We additionally assume that
the block effects are random with variance  2 .

After estimating the treatment effects under interblock analysis, we use the results for the pooled
estimation and recovery of interblock information in a BIBD.

In case of BIBD,
  ni21 i ni1ni 2   ni1niv 
 i i 
 
  ni1ni 2  ni 2   ni 2 niv 
2

N 'N   i i i

     
 
  niv ni1  niv ni 2   niv  2

 i i i 
 r   

  r   

    
 
   r 
 ( r   ) I v   Ev1 Ev' 1

1   Ev1 Ev' 1 
( N ' N ) 1   Iv  
r   rk 

The interblock estimate of  can be obtained by substituting the expression on  N ' N 

1
in the earlier

obtained interblock estimate.

GEv1
  ( N ' N ) 1 N ' B  .
bk

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
15
Our next objective is to use the intrablock and interblock estimates of treatment effects together to find
an improved estimate of treatment effects.

In order to use the interblock and intrablock estimates of  together through pooled estimate, we
consider the interblock and intrablock estimates of the treatment contrast.

The intrablock estimate of treatment contrast l ' is

l 'ˆ  l ' C  Q
k
 l 'Q
v
k
 l Q
v j j j
  l jˆ j , say.
j

The interblock estimate of treatment contrast l ' is

l 'N 'B
l '  (since l ' Ev1  0)
r 
1 v  b 
  l j   nij Bi 
r   j 1  i 1 
1 v
  l jT j
r   j 1
v
  l j j .
j 1

The variance of l 'ˆ is obtained as

 k   
Var (l 'ˆ)    Var   l j Q j 
 v   j 
2
 k   
     l 2j Var (Q j )  2  l j l j 'Cov(Q j , Q j ' )  .
 v   j j j '(  j ) 
Since
 1
Var (Q j )  r 1    2 ,
 k

Cov(Q j , Q j ' )    2 , ( j  j '),
k
so

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
16
 2   
 
2
 k    1 2 
Var (l ' )    r 1     l j    l j    l 2j   2 
ˆ 2

 v    k  k  j 
  

 j j

2
 k   r (k  1)  
  
 v   k

j
l 2j   l 2j   2 (since   j  0 being contrast)
k j  j

2
 k  1
     (v  1)     l 2j (using r (k  1)   (v  1))
 v  k j

 k 
    2  l 2j .
 v  j

Similarly, the variance of  'ˆ is obtained as

2
 1   
Var (l ' )      l j Var (T j )  2  l j l j 'Cov(T j , T j ' ) 
2

 r    j j j '(  j ) 
2   2

 1   2 2   2
  f  j   j j j j  
2
 r l  l  l
 r   
f

 j 
  
 2f

r  j
 l 2j .

The information on  'ˆ and  'ˆ can be used together to obtain a more efficient estimator of  ' by
considering the weighted arithmetic mean of  'ˆ and  ' . This will be the minimum variance
unbiased and estimator of  ' when the weights of the corresponding estimates are chosen such that
they are inversely proportional to the respective variances of the estimators. Thus the weights to be
assigned to intrablock and interblock estimates are reciprocal to their variances as v /(k 2 ) and

( r   ) /  2f , respectively. Then the pooled mean of these two estimators is

v r  v r 
2  j j
l 'ˆ  2 l 'ˆ l ˆ  2  l j j
k 2
f k j f j
L*  
v r   v r  
 2 
k 2
f k 2  2f
 v1
k
 l ˆ j j  (r   )2  l j j

j j
v
1  (r   )2
k
 v1  l jˆ j  k (r   )2  l j j

j j

 v1  k (r   )2
  v1ˆ j  k (r   )2 j 
 lj  
j   v1  k (r   )2 
  l j *j
j

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
17
 1ˆ j  k (r   )2ˆ j 1 1
where  *j  , 1  2 , 2  2 .
1  k (r   )2  f

Now we simplify the expression of  *j so that it becomes more compatible in further analysis.

Since ˆ j  ( k /  )Q j and  j  T j / ( r   ), so the numerator of  *j can be expressed as

1ˆ j   2 k ( r   ) j  1kQ j   2 kT j

Similarly, the denominator of  *j can be expressed as

1 v  2 k ( r   )
 vr ( k  1)    r ( k  1)  
 1   2  k  r   (using  (v  1)  r ( k  1))
 v  1 
   v  1  
1
 1vr (k  1)  2 kr (v  k ) .
v 1
Let
W j*  (v  k )V j  (v  1)T j  ( k  1)G

where Wj
*
j  0. Using these results we have

(v  1) 1kQ j  2 kT j 
 *j 
1rv(k  1)  2 kr (v  k )
(v  1) 1 (kV j  T j )  2 kT j  Tj
= (using Q j  V j  )
r 1v(k  1)  2 k (v  k ) k
1k (v  1)V j  (k2  1 )(v  1)T j
=
r 1v(k  1)  2 k (v  k )
1k (v  1)V j  (1  k2 ) W j*  (v  k )V j  (k  1)G 

r 1v(k  1)  2 k (v  k )
1k (v  1)  (1  k2 )(v  k )V j  (1  k2 ) W j*  (k  1)G 

r 1v(k  1)  2 k (v  k )
1 1  k2 
 V j 
r 1v(k  1)  2 k (v  k )
W j*  (k  1)G

1
 V j   W j*  (k  1)G
r
where
1  k2 1 1
 , 1  2 , 2  2 .
1v(k  1)  2 k (v  k )  f

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
18
Thus the pooled estimate of the contrast l ' is
l ' *   l j *j
j

1
 
r j
l j (V j  W j* ) (since l
j
j  0 being contrast)

The variance of l ' * is

k
Var (l ' *) 
 v1  k (r   )2
l
j
2
j

k (v  1)
 
r  v(k  11  k (v  k )2  j
l 2j (using  (v  1)  r (k  1)

l 2
j

  E2 j

r
where
k (v  1)
 E2 
v(k  1)1  k (v  k )2
is called as the effective variance.

Note that the variance of any elementary contrast based on the pooled estimates of the treatment
effects is
2
Var ( i*   *j )   E2 .
r
The effective variance can be approximately estimated by
ˆ E2  MSE 1  (v  k ) *
where MSE is the mean square due to error obtained from the intrablock analysis as
SS Error (t )
MSE 
bk  b  v  1
and
1  2
*  .
v ( k  1)1  k (v  k ) 2

The quantity  * depends upon the unknown  2 and  2 . To obtain an estimate of  * , we can

obtain the unbiased estimates of  2 and  2 and then substitute them back in place of  2 and  2 in

 * . To do this, we proceed as follows.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
19
An estimate of 1 can be obtained by estimating  2 from the intrablock analysis of variance as

1
ˆ1   [ MSE ]1 .
ˆ 2
The estimate of 2 depends on ˆ 2 and ˆ 2 . To obtain an unbiased estimator of  2 , consider

SS Block ( adj )  SSTreat ( adj )  SS Block ( unadj )  SSTreat ( unadj )

for which
E ( SS Block ( adj ) )  (bk  v) 2  (b  1) 2 .

Thus an unbiased estimator of  2 is

1
ˆ 2   SS Block ( adj )  (b  1)ˆ 2 
bk  v 
1
  SS Block ( adj )  (b  1) MSE 
bk  v 
b 1
  MS Block ( adj )  MSE 
bk  v 
b 1
  MS Block ( adj )  MSE 
v(r  1) 
where
SS Block ( adj )
MS Block ( adj )  .
b 1
Thus
1
ˆ 2 
kˆ  ˆ 2
2

1
 .
v(r  1)  k (b  1) SS Block ( adj )  (v  k ) SS Error (t ) 

Recall that our main objective is to develop a test of hypothesis for H 0 :  1   2  ...   v and we now

want to develop it using the information based on both interblock and intrablock analysis.

To test the hypothesis related to treatment effects based on the pooled estimate, we proceed as follows.

Consider the adjusted treatment totals based on the intrablock and the interblock estimates as
T j*  T j   *W j* ; j  1, 2,..., v

and use it as usual treatment total as in earlier cases.

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
20
The sum of squares due to T j* is
2
 v *
  Tj 
ST2*   T j*2    .
v
j 1

j 1 v

Note that in the usual analysis of variance technique, the test statistic for such hull hypothesis is
developed by taking the ratio of the sum of squares due to treatment divided by its degrees of freedom
and the sum of squares due to error divided by its degrees of freedom. Following the same idea, we
define the statistics
ST2* / [(v 1)r ]
F* 
MSE[1  (v  k )ˆ *]
where ˆ * is an estimator of  * . It may be noted that F * depends on ˆ *. The value of ˆ * itself
depends on the estimated variances ˆ 2 and ˆ 2f . So it cannot be ascertained that the statistic F *

necessary follow the F distribution. Since the construction of F * is based on the earlier approaches
where the statistic was found to follow the exact F -distribution, so based on this idea, the distribution
of F * can be considered to be approximately F distributed. Thus the approximate distribution of F *
is considered as F distribution with (v  1) and (bk  b  v  1) degrees of freedom. Also, ˆ * is an

estimator of  * which is obtained by substituting the unbiased estimators of 1 and 2 .

An approximate best pooled estimator of l 

j 1
j j is

v V j  ˆW j
lj
j 1 r
and its variance is approximately estimated by
k  l 2j
j
.
 vˆ1  (r   )kˆ 2

In case of the resolvable BIBD, ˆ 2 can be obtained by using the adjusted block with replications sum

of squares from the intrablock analysis of variance. If sum of squares due to such block total is SS Block
*

and corresponding mean square is

*
SS Block
*
MS Block 
br
then

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
21
 (v  k )(r  1) 2
*
E ( MS Block ) 2  
br
(r  1)k 2
2  
r
and k (b  r )  r (v  k ) for a resolvable design. Thus

E  rMS Block
*
 MSE   ( r  1)( 2  k 2 )

and hence
1
 rMSblock
*
 MSE 
ˆ 2    ,
 r 1 
ˆ1   MSE  .
1

The analysis of variance table for the recovery of interblock information in BIBD is described in the
following table:

Source Sum of squares Degrees of Mean square F*

freedom
Between treatment S T2* v-1 MS Blocks ( adj )
F* 
(unadjusted) MSE

Between blocks
(adjusted) SS Block ( adj )  b-1 SS Block ( adj )
MS Blocks ( adj ) 
SSTreat ( adj )  b 1
SS Block (unadj ) 

SSTreat (unadj )
Intrablock error
SSError (t )
SS Error (t ) MSE 
bk – b – v + 1 bk  b  v  1
(by substraction)

Total SSTotal bk - 1

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
22
The increase in the precision using interblock analysis as compared to intrablock analysis is
Var (ˆ)
1
Var ( *)
 v1  2 k (r   )
 1
 v1
2 ( r   ) k
 .
 v1
Such an increase may be estimated by
ˆ 2 ( r   ) k
.
 vˆ1

Although 1  2 but this may not hold true for ˆ1and ˆ 2 . The estimates ˆ1 and ˆ 2 may be
negative also and in that case we take ˆ1  ˆ 2 .

Analysis of Variance | Chapter 6 | Balanced Incomplete Block Design | Shalabh, IIT Kanpur
23