021521

Cluster Sampling
The divisions of population into a finite number of distinct and identifiable units are called the sampling units. The
smallest unit into which the population can be divided is called an element of the population and the groups of
elements are called the clusters.
Thus cluster sampling is a sampling technique which consists in forming suitable clusters of elements and surveying
all the elements in a sample of clusters selected according to an appropriate sampling scheme. In cluster sampling
the sampling unit is a cluster.

Example
To estimate the milk production in a Thana, village can be taken as cluster which consists of elements as
households. In order to save time and expenditure we can choose some villages as clusters from the Thana and take
all households belonging to the villages and surveying them, such a procedure, to estimate the milk production in a
Thana, is called cluster sampling and it can be extended to a district or the whole country.

Area Sampling
If the entire area containing the population under study is divided into smaller area segments and each element in
the population is associated with one and only one such area segment, the procedure is called area sampling.

Conditions in Cluster Sampling

1. Every unit of the population under study must correspond to one and only one unit of the cluster or area
segment. So that the total number of sampling units in the frame may cover all the units of the population under
study with no omission or duplication.

2. To increase the efficiency of cluster estimates the number of cluster should be large and the number of
elements in a cluster should be small.

Discuss the Reasons for Conducting Cluster Sampling With Some Example.
There are two main reasons for the widespread application of cluster sampling:
1. When the population under study is large enough so that no reliable list of the elements in the population is
available or even impossible to obtain a frame. For example, cluster sampling.

2. When time, money and labor is beyond our capacity respect to survey.

Example
A simple random sample of 600 houses covers a town more evenly than 20 city blocks containing an average of 30
houses apiece. But greater field costs are incurred in locating 600 houses and in travel between them than in
locating 20 blocks and visiting all the houses in these blocks.

Advantages of Cluster Sampling

In cluster sampling, collection of data for neighboring elements is easier, cheaper, faster and more convenient than
observing units scattered over a region.

1. Cluster sampling is operationally more convenient and less costly than SRS and stratified sampling due to
saving of listing cost and traveling cost. It also save time in journeys, identifications, contacts etc.

2. When the sampling frame of elements may not be readily available or costly to make up, cluster sampling
procedure can be adopted.

1
 021521

Disadvantages of Cluster Sampling

1. Cluster sampling is generally less efficient than SRS and stratified sampling due to usual tending of units in a
cluster to be similar.

2. The efficiency of cluster sampling is likely to decrease with increase of cluster size.

3. In most practical situations, the loss of efficiency may be balanced by the cost. Therefore cost per unit in cluster
sampling is more than simple random sampling.

Assumptions of Cluster Sampling

1. Cluster should be as small as possible consistent with the cost and limitations of the survey.
2. The number of sampling units in each cluster should be approximately same.

Applications of Cluster Sampling

1. Psychology: Classifying individuals on the bases of personality types.

2. Regional analysis: Classifying cities of region into typologist based on demographic or fiscal variables.

3. Marketing analysis: Classifying customers into segment on the basis of product use.

4. Chemistry: Classifying the components based on their performance properties.

Also cluster sampling is applied in agriculture and business etc.

Difference between Cluster Sampling and Simple Random Sampling

Simple random sampling Cluster sampling

SRS is the ultimate element of the observations. Cluster sampling is a unit which consist a number of
elements.

Frame is essential for drawing a simple random sample. Frame is not essential for drawing a cluster sample.

When the population under study is large enough then When the population under study is large enough then
SRS is more expensive and more time consuming. cluster sampling is less expensive and less time
consuming.

If each cluster of size then there is no difference between The efficiency of cluster sampling increases at mean
SRS and cluster sampling. square between cluster decreases.

SRS is more efficient than cluster sampling. Cluster sampling is less efficient than SRS.

2
 021521
Difference between Cluster and Strata

Cluster Strata

The smallest unit into which the population can be Relatively homogenous (non-overlapping) subgroup of a
divided is called an element of the population. A group of population is known as strata.
such elements is known as cluster.

The elements within each cluster are heterogeneous and The elements within each stratum are homogeneous and
between clusters are homogeneous. between strata are heterogeneous.

Clusters are generally made up on the basis of Stratum is generally made up on the basis of some
geographical location. characteristics of the population.

Obtaining sample from clusters generally costs less. Obtaining sample from strata generally costs more.

Difference between Cluster Sampling and Stratified Sampling

Cluster sampling Stratified sampling

In cluster sampling elements within each cluster are In stratified sampling elements within each stratum are
heterogeneous and between clusters are homogeneous. homogeneous and between strata are heterogeneous.

Clustering is often done on the basis of geographical Stratification is done on the basis of some characteristics
location. of the population.
In cluster sampling clusters are subjected to complete In stratified sampling strata are subjected to sampling.
enumeration.

It is less costly than stratified sampling. It is more costly than cluster sampling.

It gives less efficient result than stratified sampling. It gives more efficient result than cluster sampling.

Difference between Two stage Sampling and Two phase Sampling

Two stage Sampling Two phase Sampling

In two stage sampling auxiliary information is not In two phase sampling auxiliary information is
necessary. necessary.

Sampling unit is not same for both stages. Sampling unit is same for both stages.

It is less costly than two phase sampling.

It is more costly than two stage sampling.

It is not more advantageous when the gain in precision It is more advantageous when the gain in precision of
of the estimates increases. the estimates increases.
A sampling frame of the second stage units is necessary It is necessary to have a complete sampling frame of the
for the selected first stage units. units.

Design Effect

3
 021521
L. Kish defined design effect as the ratio of the variance of the estimate obtained from the more complex sample to
the variance of the estimate obtained from a simple random sample of same number of units (elements).
The design effect has two primary uses:

1. The sample size estimation and

2. In appraising (increasing) the efficiency of more complex plan.

Justify Cluster Sampling Can be Regarded as a Generalization of Any System of Sampling.

In a general sense any system of sampling may be regarded as a kind of cluster sampling, since in every sampling
scheme the units are conceptually grouped to from samples and one of them is selected with a certain probability.

For instance systematic sampling may be considered a particular case of cluster sampling, since in this case the
population is divided into a number of cluster, each cluster consists of units distributed at a fixed interval
(systematically) over the whole population and one such cluster is selected at random.
Other sampling procedures, viz. SRS, PPS and stratified sampling can be applied to sampling of clusters by treating
the clusters themselves as sampling units.

Theorem
th
For the u type of unit, let,

M u  relative size of unit

Su2  variance among the unit totals
Cu  relative cost of measuring one unit.
Cu Su2
Then relative cost for specified variance or relative variance for specified cos  t   .
M u2
Proof
Suppose V is the specified variance of the estimated population total. For the u th type of unit the estimate is Nu yu

with variance

Nu2 Su2
V
nu
Nu2 Su2
 nu  i 
V

Cu Nu2 Su2
The cost of taking nu units is Cu nu 
V
Since Nu M u  Constant for different types of unit,

Cu Su2
The cost is proportional to
M u2
C
On the other hand, if the cost C is specified, nu  and equation in 1 .
Cu
Cu Su2
V  Proved  .
M u2

Cluster Sampling of Equal Cluster Size

Let us consider the following table

1 2 … i … N

4
 021521
y11 y21 … yi1 … yN1
y12 y22 … yi 2 … yN 2

   
y1M y2M … yiM … y NM

i.e. the population consists of N clusters, each of M elements and a simple of n cluster is drawn by the method of
simple random sampling, let

yij  the value of the characteristics under study for the jth element  j  1, 2,..., M  in the i th cluster  i  1, 2,..., N  .
M
 yij
j 1
Yi   the mean per element of the i th cluster.
M
1 n N M
 Yi  the mean of cluster sample means of n cluster  NM   yij .
1
yi 
n i 1 i 1 j 1
N M

N  Yi
 Yi 
1 i 1 j 1
Y   the mean per element in the population,
N i 1 NM

M
  yij  Yi 
2

j 1
Si2   the mean square among the elements within the i th cluster.
M 1

S w2 
 Si2
 the mean square within clusters,
N

  yi  Y 
N 2

i 1
Sb2   the mean square between cluster means in the populaton.
N 1

  yij  Y 
N M 2

i 1 j 1
S2   the mean square between the elements in the population.
NM  1

  yij  Y  yik  Y 

N M M



E yij  Y  y ik Y  
i 1 j 1 j  k

   M  1 NM  1 S 2
2
E yij  Y
where  is theintra cluster correlation coefficient between elements within clusters.

5
 021521

6
 021521

7
 021521

Intra-Cluster Correlation Coefficient

The correlation coefficient between elements in the same cluster, known as intra-cluster correlation coefficient and is
denoted by  .

th th
Let yij be the observed value for the j element within the i unit and let Yi be the unit total. The intra-cluster

correlation coefficient  is then defined by

8
 021521


E yij  Y  y ik Y  2 yij  Y
i 1 j  k
  y ik Y 
 
   M  1 NM  1 S 2
2
E yij  Y

Where, Y 
 yi and . Y 
 yi  Y
N NM M
The number of terms (cross product) in the numerator E is NM  M  1 2 and in the denominator E is

 NM  1 S 2
NM

9
 021521

10
 021521

Theorem
A simple random sample of n clusters, each containing M elements, is drawn from the N clusters in the

population. Then the sample mean per element y is an unbiased estimate of Y with variance

1 f NM  1 1 f 2
var  y    2 S 2 1   M  1    S 1   M  1   .
n M  N  1  NM
Where  is the intra-cluster correlation coefficient.

Proof
N
 yi
i 1
Let yi denote the total for the i th cluster and y  .
n
Now we have,
N M
 yij
1
y
nM i 1 j 1
n
 Yi
1

n i 1

11
 021521
n
E y    1
 E Yi
n i 1
 
1 N Yi  N
Yi 
  n 

 
E Yi   
n i 1 N i 1 N 
N M yij
   Y.
i 1 j 1 NM

Since mean per element y is an unbiased estimate of Y . Now

Y  Y 
2

1 f N
1 f 2
  
i
var y   Sb
n i 1 N 1 n
Here,
2
 M yij 
 Yi  Y 
N 2 N
  Y 

i 1  j 1 M

i 1 
2
 M 
 
N
 2   yij  MY 
1
M i 1  j 1 

  yi1  Y    yi 2  Y   ...   yiM  Y 

N 2
1

M2 i 1

M 
   yij  Y     yij  Y  yik  Y 
N 2 M M
1
 2
M i 1  j 1  j 1 j  k

 
    
N M 2 N M M
 2   yij  Y   yij  Y yik  Y
1
M  i 1 j 1 i 1 j 1 j  k

1
 2  NM  1 S 2   M  1 NM  1 S 2  
M  
 NM  1 S 2 1 
   M  1  
M2

1 f 1  NM  1 2
 var y   
n N 1 M 2
S 1   M  1  

1 f NM 2
 S 1   M  1    for large N 
n NM 2
1 f
 S 2 1   M  1    proved 
nM
From the above function, it is easily observed that the variance in cluster sampling depends on the number of
2
clusters, the size of the cluster, the intra-cluster correlation coefficient and S .
Again we know that,

y  My
 var  y   M 2 var y  
Also,

12
 021521
yˆ  NMy
 var  yˆ   N 2 M 2 var y  
1 f 2
 N 2M 2 S 1   M  1  
nM
N 2 M 1  f  2
 S 1   M  1   .
n

If M is large and  is positive then  M  1  is positive which will increase the value of var y   i.e. the cluster

sampling is less precise, but if   0 the cluster sampling is more precise. Though usually  decrease with

increase in M the efficiency of cluster sampling decline, because the factor  M  1  greatly increases with

increasing cluster size.

Thus, the importance behind the intra-cluster correlation coefficient  is that, it helps us to make decision whether

we adopt cluster sampling design or not.

According to Hansen, Hurwitz and Madow 1953  ,  is a “measure of homogeneity” of the cluster. If M  1 , the

cluster sampling design and simple random sampling design are equally efficient i.e. both process are equally good
in this situation.

Expression for 
2
Let Sb denote the variance among cluster totals, on a single unit basis then

 Yi  Y    N  1 MSb2
2

We have,

 Yi  Y    NM  1 S 2 1   M  1  
2

  N  1 MSb2   NM  1 S 2 1   M  1  

 
 N  1 MSb2   NM  1 S 2 
Sb2  S 2
 NM  1 M  1 S 2  M  1 S 2
1
When terms in are negligible.
N
The value of within-cluster, mean square
N M
Sw2   Yij  Yi   N  M  1
2

i 1 j 1

By the one-way analysis of variance

Yi  Y 
2
N N M
 NM  1 S   Yij  Yi 
2
2
 
i 1 M i 1 j 1

 NM  1
 S 2 1   M  1    N  M  1 S w2
M

13
 021521
M  NM  1 S   NM  1 S 1   M  1    NM  M  1
2 2
S w2
 M  NM  1 S 2   NM  1 S 2 1   M  1    NM  M  1 S w2
 NM  M  1 S w2   NM  1 S 2  M  1   M  1  
 NM  M  1 S w2   NM  1 S 2  M  11    
 NMS w2   NM  1 S 2 1   
NM  1 2
 S w2  S 1   
NM
 S 2 1   
S 2  S w2
  
S2
Relative Efficiency of Cluster Sampling
In sampling of nM elements from the population by SRSWOR, we have,

1 f 2
var  y   S
nM
1 f 2
and in cluster sampling  
var y 
n
Sb .

V y 1 S2
 Relative efficiency  R.E   
V y   M Sb2

This shows that the efficiency of cluster sampling increases as the mean square between clusters sb2   decreases.
Again we have,

 NM  1 S 2   Yij  Y 
N M 2

i 1 j 1

 Yij  Yi  Yi  Y 
N M 2

i  j 1

 Yij  Yi    Yi  Y 
N M N M 2
2

i 1 j 1 i j
M
   M  1 Si2  M  N  1 Sb2
i

  M  1 NS w2  M  N  1 Sb2
1  NM  1 S 2  N  M  1 S w2 
 Sb2 
M  N  1  

1 S2 S2
R.E  
M  NM  1 S 2  N  M  1 Sw2 1 
 NM  1 S 2  N  M  1 Sw2 
M  N  1  N  1 
2
i.e. if S w i.e. mean square within cluster increases, the efficiency of cluster sampling will increases.

Again,

14
 021521
 Yi  Y 
2
  NM  1 S 2  
 
2
Sb2    Yi  Y  1   M  1   
N 1  M 2

 NM  1 S 2 1  M  1  
   
M 2  N  1 
S2
 1   M  1  
M 
1 f 2
 var y    nM
S 1   M  1  

1  f  S 2
Then, R.E  nM
1 f 2
S 1   M  1  
nM
1

1   M  1 
1
In case of complete homogeneity of cluster S w  0 and so   1 and R.E 
2
i.e. cluster sampling is not efficient.
M
1
In case of complete heterogeneity, S w  S , so sb  0 and   
2 2 2
, i.e. cluster sampling is very effective.
 M  1
Estimate the Sampling Variance in Case of Cluster Sampling
We know,

1 f 2 M  N  1 Sb2  N  M  1 S w2
 
var y 
nM
S 1   M  1   where, S2 
NM  1
2 2 2 2
If we select n clusters from the clusters we get, Sb and S w . In this case, Sb is an unbiased estimate of Sb and

S w2 is an unbiased estimate S w2 .

2 2
But s is not an unbiased estimate of S because frmp population of size NM , a simple random sample of size n

can not be drawn. So if we put Sˆb2  sb2 and Sˆw2  sw2 then we get Ŝ which is an unbiased estimate of S .
2 2

1 f ˆ2
 
var y 
nM
S 1   M  1 ˆ 

Also ̂ can be estimated from the sample, where

ˆ 
 n  1 Msb2  nsw2 .
 n  1 Msb2  n  M  1 sw2
Variance Function

It is of interest to examine how the variances var y   behave with the cluster size M . This involves investigative the
2
relation between Sb and M .

2
By the analysis of variance, Sb can be found if we know
2
 The variance S between all elements in the population and

 S w2 between elements that lie in the same unit.

15
 021521
Our approach is to predict S w2 and S 2
to find Sb2 by the analysis variance. The sample data produce estimate S 2

and S w2 . Since S 2 is the variance among elements, it is not affected by the size of the unit. However S w2 will be

affected, Jessen 1942  , Mahalanobis, 1944  , Hendricks 1944  to attempts to develop a general law to predict

2
how S w changed with the size of unit. On the basis of several agricultural surveys within cluster variance formula

Sw2  AM g ;g  0 1
Where A, g are constants that do not depend on M and g begins a small positive quantity.

If this formula fits, log S w should plot as a straight line against log M . Values of S w2 far at least two values of M
2

are needed in order to estimate the constants log A and g .

From the analysis of variance we find

 NM  1 S 2  N  M  1 Sw2
Sb2 
 N  1
 NM  1 S 2  N  M  1 AM g
  MS 2   M  1 AM g .
 N  1
Hendricks has painted out that the complete population might be regarded as a single large sampling unit containing

NM elements. Since 1 holds then

S 2  A  NM 
g

Now A and g can be estimated by using the value of M . The two equations that lead to the estimates are

log sw2  log A  g log M

log S 2  log A  g log  NM 
Thus Sb2  AM g  MN g   M  1 
 
This method might hold well enough for small values of M but fail for a value as large as large as NM .

Determination of Optimum Cluster Size

For a given sample size, the sampling variance increases with decreasing the number of cluster and decreases with
increasing the number of cluster. On the other hand, the cost decreases with cluster size and it increases with the
number of cluster.

Hence it is necessary to determine a balancing point by finding out the optimum cluster size and the number of
clusters in the samples which can minimize the sampling variance for a given cost or alternatively, minimizing the
cost for a fixed variance.

A cost function developed by Jessen 1942  is given as

c  c1nM  c2 n

where c1 is the cost of enumerating an element, including the cost of travel between units within the cluster and c2

is the cost per unit distance traveled between clusters.

16
 021521
Assume that c1 is expected to be, considerably less than c2 . The variance of the estimator y based on a sample of

n clusters of size M each is given by

1 f 2
 
var y 
n
Sb

S2
 b ; if f . p.c. is ignored.
n
 NM  1 S 2  N  M  1 Sw2
where, Sb2   
M  N  1

It is of interest to examine how the variance var y   behave with the cluster size M , This involves investigative the

2
relationship between Sb and M .

Jessen 1942  , Mahalanobis 1944  , Hendricks 1944  empirically demonstrated on the basis of several

2
agricultural survey that S w is related to M by the relation

Sw2  AM g 
where A and g are positive constants, are to be determined from the survey data and are independent of M .

By using    and   we have,

 NM  1 S 2  N  M  1 AM g
Sb2 
M  N  1
NMS  N  M  1 AM g
2
 ; for large N
MN
 S   M  1 AM g 1
2

S 2   M  1 AM g 1
 
 var y 
n

To determine the optimum size of cluster, we find M and incidentally n to minimize V for fixed c .

To minimize   c  V , where  being Lagrange’s multiplier. Differentiate with respect to n and M respectively

and equating to zero we get,

17
 021521
 

n n

c1nM  c2 n  V 
c V
 c1M  2  
2 n n

 c1M 
c2


 S 2   M  1 AM g 1 
2
2 n n
 S   M  1 AM
2 g 1
c2
 c1M  
2 n n n
c2 
 c1M   V 0
2 n n
c2 V
 c1M   i 
2 n n
And
 V
 c1n   0
M M
V
 c1n    ii 
M

Diving  ii  by  i  we get,

V

c1n
 M
c2 
c1M  V
2 n n
c1n n V
 
c V M
c1M  2
2 n
1 V 1
 
V M  c2 
M 1  
 2c M n
 1 
M V 1
   iii 
V M c2
1
2c1M n

Solving

c  c1nM  c2 n
 cc1M  c12 M 2 n  c1c2 M n
 4cc1M  4c12 M 2 n  4c1c2 M n
 4cc1M  c22  4c12 M 2 n  4c1c2 M n  c22

 
2
 4cc1M  c22  2c1M n  c2

 2c1M n  c2  4cc1M  c22

 2c1M n  c22  4cc1M  c2

18
 021521

2c1M n c22  4cc1M

  1
c2 c2
2c1M n c22  4cc1M
  1
c2 c22
12
c1  4cc1M 
 2 M n  1   1
c2  c22
 
12
 4cc1M 
1   1
c22 
 n   iv 
2c1M
c2

Substituting the value of n in  iii  , we get,

M V 1 1
 
V M c2
1
1
1
4cc1M 4cc1M
1 1 1 1
c22 c22
2c1M 
2c1M
c2
1

4cc1M
1 1 1
c22
4cc1M
1 1
c22
4cc1M
1 1
c22

4cc1M
1
c22

  
 12 
M V 4cc1M
   1  1   
V M   c22 
   
 12
M   S   M  1 AM
g 1 
2
 4cc1M 
    1   1
V  M  n   c22 
 12
M A   4cc1M 
 
V n M

M g  M g 1  1 
 c22
  1
 
 12
AM   4cc1M 
  gM g 1   g  1 M g  2   1   1
n V    c22
 
 12
AM 1   4cc1M 
   g   g  1   1   1
n V  M   c22 
 12
A M g 1  gM   g  1   4cc1M 
  M    1   1
n V  M   c22 
 12
AM g 1  gM   g  1   4cc1M 
  1  1 
  v
S 2   M  1 AM g 1  c22 

19
 021521
The above equation can be solved by iterative method and gives the optimum value of M . Using this optimum value
of M we can find optimum value of n .

c1cM
From the relation  iv  , it follows that, M will change according to changes c1 , c2 and c such that is nearly
c22
constant. This leads to the conclusion that the optimum size of the unit will be smaller when

1. The cost of enumeration of an element increases.

2. The cost of travel between units decreases and

3. The cost of the survey is sufficiently large.

Cluster Sampling for Proportions

Suppose that it is required to estimate the proportion of elements belonging to a specified class when the population
consists of N clusters each of size M and a random sample WOR of n elements in any cluster can be classified
into two classes.

Assuming that,

yij  1 ; if the jth element of the i th cluster belongs to the class.

1 ; otherwise
ai
and also, Pi 
M
th th
is the proportion in the i cluster, where, ai is the number of elements in the i cluster belonging to the specified

class.

Then the population proportion is given by

N N Pi
 N
ai
P 
i NM i

The estimate of the population proportions is

N
Pˆc   i  P
P
i n

 1 n
 E Pˆc   E  Pi 
n i
N N
 Pi  Pi
1 1 1
 n 
n N i N i

Thus the estimate is unbiased.

The sampling variance of Pˆc is given by

 
var Pˆc 
1 f 2
n
Sb

2
where Sb is the variance between cluster proportions and is given by

20
 021521
 Pi  P 
2
N N
 
PQ
MSb2   PQ  i i

i N i N
1 f 1  
 
N

PQ
 var Pˆc   PQ 
i i

n M  i N 

For large N we have,

S2 Sb2  S w2  PQ
2
and the within variance S w is given by

N

PQ
Sw2  i i

i N

Hence the intra cluster correlation coefficient  is given by

N
 MPQ
i i
  1 i
 M  1 PQ

Therefore, the sampling variance in terms of the intra cluster correlation coefficient can be expressed as

1  f  NPQ 1  M  1  
 
var Pˆc 
 N  1 nM 
  

If a SRS of nM elements could be taken, the variance of the sample proportion P̂ would be given by

 
var Pˆ  1  f 
NPQ
nM  N  1

The efficiency of cluster sampling as compared to SRSWOR can be obtained as

R.E 
 N  1 NPQ
 NM  1 NPQ  N
 PQ
i i
i

  is given by
An estimator of var Pˆc

 
2
s
var Pˆc  1  f  b
n

  Pi  P 
2

 1  f   
n  n  1

An estimator for the variance of the total number of units belonging to a specified class can be obtained by

to the value in the relation    .

2 2
multiplying N M

21