TWO STAGE SAMPLING

In cluster sampling, the clusters were considered sampling units and all the elements in the
selected clusters were enumerated completely. This method of sampling is economical under
certain circumstances but it is generally less efficient than sampling of individual units
directly, because it restricts the spread of the sample over the population. It is, therefore,
logical to expect that the efficiency of the estimator will be increased by distributing elements
over a large number of clusters and surveying only a sample of units in each selected cluster
instead of completely enumerating all the elements in the sample of small number of clusters.
This type of sampling which consists in first selecting the clusters and then selecting a
specified number of elements from each selected cluster is known as sub-sampling or two-
stage sampling. The clusters that form the units of sampling at first stage are termed first
stage units ( fsu ) or primary stage units and the elements within clusters which form the units
of sampling at the second stage are called second-stage units (ssu) or secondary units. For
example, in conducting a socio-economic survey in a region, villages or urban blocks may be
considered first stage units and the households the second stage units.

Two stage sampling with equal first stage units

We shall assume that the population under study has NM elements grouped into N first
stage units, each containing M second stage units. Suppose a sample of size nm is drawn by
first selecting n first stage units, and then drawing m second stage units from each of the
selected first stage units. Also we assume that the units at each stage are selected with equal
probability. The following notations are used.
y ij , value of j  th ssu in the i  th fsu , i  1, 2, , N ; j  1, 2, , M .

1 M
Yi.   yij , mean per ssu of the i  th fsu .
M j 1

1 N 1 N M
Y  i. NM   yij , mean per element in the population.
N i 1
Y 
i 1 j 1

1 N
S b2  
N  1 i 1
(Yi.  Y ) 2 , mean square between fsu means.

1 M
S i2  
M  1 j 1
( yij  Yi. ) 2 , mean square between ssu within i  th fsu .

1 N 2
S w2   Si , mean of the mean square between ssu within fsu .
N i 1

1 m
y i.   yij , sample mean per ssu in the i  th fsu or mean based on m selected
m j 1
ssu from the i  th fsu .

1 n 1 n m
y  i. nm   yij ,
n i 1
y  sample mean based on all the nm units.
i 1 j 1
126 RU Khan

Theorem: In two-stage sampling with equal first stage units, the sample mean y is
2 2
 N  n  Sb  M  m  S w
unbiased estimator of Y with variance V ( y )      .
 N  n  M  nm
Proof: Here, the expectation is taken in two stages. In the 1st stage, the expectation is taken
over all possible samples of ssu from the selected fsu (i.e. fixed n fsu ) and second
expectation is taken for all samples of size n fsu from N fsu . We have

 1 n  1 n 
E ( y )  E1  E 2   yi. i   E1   E 2 ( yi. i) , since ssu are srs , then
  n i 1 
  n i 1 

1 n  1 N
 E1   Yi.    Yi.  Y , as fsu are srs .
 n i 1  N i 1
To obtain the variance of y , we have

V ( y )  V1[ E2 ( y i)]  E1[V2 ( y i)] (8.1)

Consider
 1 n  1 n 
V1[ E 2 ( y i )]  V1  E 2   yi. i   V1   E 2 ( yi. i) 
  n i 1 

n
 i 1


1 n 
 V1   Yi.   V1 (Yn ) , where Yn is the mean of n selected fsu under srs .
n 
 i 1 
1 1 
    S b2 . (8.2)
n N 
and
 1 n  1 n 
E1[V2 ( y i)]  E1 V2   yi. i   E1  V2 ( yi. i)

  n i 1 

2
 n i 1 

 1 n  1 1  2  1
n
1 1 
 E1      i 
S    
2
 E1 ( S i )
 n 2 i 1  m M   n 2 i 1 m M 

1 N  1 1  2 1 1 1  2
     Si  n  m  M  S w
nN i 1  m M 
(8.3)

In view of equations (8.2) and (8.3), equation (8.1) reduces as

2
 1 1  2  1 1  Sw
V ( y)     Sb    
n N  m M  n
2
 N  n  2  M  m  Sw
 S 
 b   .
 nN   mM  n
Two stage sampling 127

Note:
n m
i) If f1  and f 2  are the sampling fractions in the first and second stages, the
N M
(1  f1 ) 2 (1  f 2 ) 2
result can be written as V ( y )  Sb  Sw .
n nm
1 2 1  f2  2
ii) When N is large as compared to n , then V ( y )  S   Sw.
n b  nm 
2
 1  f1  2 S w
iii) When M is large as compared to m , the result reduces to V ( y )    Sb  .
 n  nm
iv) When both N and M are large as compared to n and m respectively, then
S2 S2
V ( y)  b  w .
n nm
Remarks
i) If the selected fsu' s are completely enumerated i.e. m  M , then this situation
corresponds to cluster sampling.
ii) If every fsu in the population is included in the sample, i.e. n  N , then this case
corresponds to stratified sampling with fsu' s as strata.
Corollary: If the n fsu' s and m ssu' s from each chosen unit are selected by simple
1 n 
random sampling, wor , the estimator Y  NM   yi.   NM y is an unbiased of the
ˆ
n 
 i 1 
 1  f1  2 2 2 1 f2  2
population total Y and its variance is V (Yˆ )  N 2 M 2   Sb  N M   Sw.
 n   nm 

Estimation of V ( y )
Define,

1 n 2 1 n
2
sw  
n i 1
s i , and s 2
b  
n  1 i 1
( y i.  y ) 2 .

Consider

1 n 2
2
sw   si , then
n i 1

 1 n   1 n  1 N
2
E (s w )  E1   E 2 ( si2 i)  E1   S i2    S i2  S w2 .
 i 1
n   i 1
n  N i 1

2
This shows that s w is an unbiased estimate of S w2 , and

1 n
1  n 2 
sb2 
n 1
 ( yi.  y )  n  1   yi.  n y 2  , so that
2
i 1  i 1 
128 RU Khan

1 n 
E ( sb2 )  
n 1 
 E ( yi2. )  n E ( y 2 ) (8.4)
i 1 
Now
 n  n  n 
E   yi2.   E1  E 2 ( yi2. i )  E1  {Yi.2  V2 ( yi. i)}
 
 i 1  i 1  i 1 
n  1 1  2 
n n
1 1 
 E1   Yi.2     i   1 i.
S  E (Y 2
)    
2
 E1 ( S i )
i 1  m M   i 1 i 1
m M 

n n 1 1  N 2
 [( N  1) S b2  N Y 2 ]      Si
N N  m M  i 1

n 1 1 
 [( N  1) S b2  N Y 2 ]  n    S w2 . (8.5)
N m M 
and
2
1 1   1 1 S
E ( y 2 )  V ( y )  Y 2     S b2     w  Y 2 (8.6)
n N  m M  n
Substitute the values from equations (8.5) and (8.6) in equation (8.4), we get
1 n 1 1 
E ( sb2 )   ( N  1) S b2  n Y 2  n    S w2
n 1  N m M 

1 1  1 1  
 n     S b2    S w2  n Y 2 
n N  m M  
1  n ( N  1) 2 1 1   N  n 2
  S b  (n  1)    S w2    Sb 
n 1  N m M   N  
1  2  1 1  2
 (n  1) S b  (n  1)  m  M  S w 
n 1    
1 1 
 S b2     S w2 (8.7)
m M 
2
Since s w is an unbiased estimate of S w2 , it follows from equation (8.7) that an unbiased
1 1  2
estimate of S b2 is Sˆb2  sb2     s w .
m M 
Therefore, an unbiased estimate of V ( y ) will be
2
 1 1   1 1  2   1 1  sw
Vˆ ( y )  v ( y )      sb2     s w     
 n N  m M   m M  n
2
1 1   1 1   1 1  2  1 1  sw
    sb2        s w   
n N   n N  m M  m M  n
Two stage sampling 129

2
1 1  1 1  2  1  1 1  2 1 1  sw
    sb2    s w       sb   
n N  m M  N  n N  m M  N
2
 N  n  2  M  m  sw
  sb    .
 nN   mM  N

Note:
n m
i) If f1  and f 2  are the sampling fractions in the first and second stages, an
N M
alternative form of the result is
2
ˆ (1  f1 ) 2 (1  f 2 ) s w (1  f1 ) 2 f1 (1  f 2 ) 2
V ( y)  v ( y)  sb   sb  sw .
n m N n nm
1
ii) If N is large, Vˆ ( y )  v ( y )  sb2 .
n

Optimum allocation of sample to the two stages

As the efficiency of two stage sampling depends very much upon values of m and n ,
because the variance component due to fsu' s decreases with an increase in n only while the
variance component due to ssu' s decreases with an increase in both n and m , therefore, it is
necessary to determine optimum allocation for both n and m such that V ( y ) is minimum for
given cost or cost is minimum for given V ( y ) (the estimated population mean has maximum
precision for given cost or has desired precision for minimum cost). In a two stage sampling
design, a simple cost function of the survey can be written as
C  c0  n c1  nm c2

where c 0 is the overhead cost, c1 is the cost of including an fsu in the sample and c2 is the
cost of including an ssu in the sample.
C  c0  n c1  nm c2  C  (say)
In a two stage sampling, the variance can be written as
2 2 2 2 2
1 1  1 1  S w Sb Sb S w S w
V ( y )     S b2        .
n N  m M  n n N nm nM
So that

S2 S2 S2 S2
V ( y )  b  b  w  w  V  (say).
N n nm nM
where C  and V  are function of n and m . Choosing n and m to minimize V for fixed C
or C for fixed V are both equivalent to minimizing the product.
 S2 S2 S2   S w2 
 (nc  nmc )   S 2  S w  
2
V C   b  w  w  (c1  mc2 )
 n nm nM  1 2

b M  m
  
c 2 mc2 2 c1 2
 c1 Sb2  mc2 Sb2  1 S w  2
S w  S w  c2 S w
M M m
130 RU Khan

 2 S w2   2 S w2  c 2

 c1 S b    mc2  S b   1 S  c S2
 M   M  m w 2 w
   

c S2
 c1   mc2   1 S w2  c2 S w2 , where,   S b2  w (8.8)
m M
Since equation (8.8) is independent of n , so minimizing V  C  will provide the optimum
value of m . We have to consider two cases according to   0 or   0 .
Case i) For optimum m , when   0 , differentiate (8.8) with respect to m and equate it to
zero, we have

 c c S2
V  C   0  c2   1 S w2 ,  m 2  1 w , and hence,
m m2 c2 

c1 c1
mopt  S w  Sw .
c2   2 S w2 
c2  S b  
 M 
 
The optimum value of n is found by solving either the cost equation or the variance
equation, depending on which has been pre-assigned.
a) When cost is fixed, substitute the optimum value of m in the cost function and solve for
n as
 c1   cc 
C  c0  n c1  n  S w  c2  n  c1  S w 1 2  , then
 
 c2     
C  c0
nopt  .
c1c2
c1  S w

Substitute the value of mopt and nopt in the expression for variance, we get the minimum
variance.
b) When variance is fixed, substitute the optimum value of m in the expression for
variance, we get the optimum value of n as

S b2 S b2
S2 S2 1 S2  S2
V ( y)   w  w   S b2  w   w

N n nm nM n  M  nm

1 S w2 1 S c 
      w 2  , then
n c1 n  c1 
n Sw
c2 
  S w c2  / c1
nopt  .
S b2
V
N
Substitute the value of mopt and nopt in the expression for cost, we get the minimum
cost.
Two stage sampling 131

Case ii) For optimum m , when   0 , in equation (7.8), the terms containing m are
c c
mc2  and 1 S w2 , clearly mc2  is negative, as   0 and 1 S w2 is minimum when m
m m
attain the possible attainable value, i.e. m  M . Thus, V  C  will be minimum, when m  M ,
so that, mopt  M . The optimum value of n is found by solving either the cost equation or
the variance equation, depending on which has been pre-assigned.
a) When cost is fixed, substitute the optimum value of m in the cost function and solve for
n as
C  c0
C  c0  n c1  n M c2  n (c1  M c2 )  nopt  .
c1  M c2
Substitute the value of mopt and nopt in the expression for variance, we get the minimum
variance.
b) When variance is fixed, substitute the optimum value of m in the expression for
variance, we get the optimum value of n as

S b2 S w2 S b2
S w2 S b2 S b2
V ( y)      ,  nopt  .
N n nM nM n 1 2
V  Sb
N
Substitute the value of mopt and nopt in the expression for cost, we get the minimum
cost.

Comparison of two stages sampling with simple random sampling

In sampling of nm elements from the population by simple random sampling, wor , the
variance of the sample mean y is given by

 1 1  2 1 2
V ( ysr )    S  S , for large N , and
 nm NM  nm
1 1  1 1 1 
V ( y )     S b2     S w2 . (8.9)
n N  nm M 

From empirical relation between S 2 , S b2 , and S w2 , we have

1
M S b2  [( NM  1) S 2  N ( M  1) S w
2
] (8.10)
N 1
Note that
1
 [ M ( N  1) S b2  N S w2 ] , then
2
( NM  1) S
2
N Sw  M ( N  1) Sb2  ( NM  1)  S 2

 ( NM  1) S 2  N (M  1) S w
2
 ( NM  1)  S 2 , from equation (8.10)
2
or N S w 2
 N (M  1) S w  ( NM  1) S 2  ( NM  1)  S 2
132 RU Khan

NM  1
2
or NM S w  ( NM  1) (1   ) S 2 or S w2  (1   ) S 2 (8.11)
NM
Substituting equation (8.11) in equation (8.10), we get
2
1   NM  1 
S b2  ( NM  1) S 2  N ( M  1)  (1   ) S 2 
M ( N  1)   NM 
NM  1  M  1  NM  1  M  ( M  1) (1   )  2
 1 (1   ) S 2  S

M ( N  1)  M  M ( N  1)  M 
NM  1 1  ( M  1)   2
 S (8.12)
M ( N  1)  M 
Substituting equation (8.11), and (8.12) in equation (8.9), we get

 N  n   NM  1 1  ( M  1)   2  1  M  m   NM  1 
V ( y)     S     (1   ) S 2 
 nN   M ( N  1)  M   n  mM   NM 

 N  n   NM  1 NM  1 
  2 S2  ( M  1)  S 2 
 nN   M ( N  1) M 2 ( N  1) 
1  M  m  NM  1 2 NM  1 
   S   S2
n  mM  NM NM 
 N  n  NM  1 1  M  m  NM  1 2
  2 S2    S
 nN  M ( N  1) n  mM  NM
 N  n  NM  1 1  M  m  NM  1
  2 ( M  1)  S 2     S2
 nN  M ( N  1) n  mM  NM

S 2  NM  1   N  n  m  M  m   N  n  m ( M  1)  M  m   
             
nm  NM   N  1  M  M   N  1  M  M  
If fpc is ignored at both the stages, then
S2
V ( y)  [1  (m  1)  ] .
nm
Thus, the relative efficiency of two stages sampling compared with simple random sampling
is given by
V ( y sr ) 1
RE   .
V ( y) [1  (m  1)  ]
It can be seen that the relative efficiency depends on the value of  , if
i)   0, then V ( y sr )  V ( y ) , i.e. both methods are equally precise.
ii)   0, then V ( y sr )  (m  1)V ( y sr )  V ( y ) , and V ( y sr )  V ( y ) i.e. two stage sampling
is more precise.
iii)   0, then V ( y sr )  (m  1) V ( y sr )  V ( y ) , and V ( y sr )  V ( y ) , i.e. simple random
sampling is more precise.
Two stage sampling 133

Two stage sampling with unequal first stage

Let the population under consideration consist of N first stage units. The i  th fsu consists
N
M i second stage units such that  M i  M 0 . Let a sample of n fsu is selected and from
i 1
n
i  th selected fsu a sample of mi ssu is selected such that  mi  m0 . ( srswor is
i 1
employed at each stage). The following notations are used.
yij , value of j  th ssu in the i  th fsu . i  1, 2,, N ; j  1, 2,, M i .

i M
1
Yi.   yij , mean of the observations in the i  th fsu .
M i j 1

1 N
YN   Yi. , the over all mean of fsu means.
N i 1

1 N M
M  
N i 1
M i  0 , average number of ssu .
N

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

1 N
S b2  
N  1 i 1
(Yi.  YN ) 2 , mean square between fsu means.

Mi
1
S i2 
Mi 1
 ( yij  Yi. ) 2 , mean square between ssu within i  th fsu .
j 1

m
1 i
y i.   yij , sample mean per ssu of i  th fsu .
mi j 1

There are several estimators of the population mean Y , here a few estimators may be
considered from the practical point of view.

1 n
1st estimate: It is defined by y I   yi. .
n i 1

By definition,
1 n  1 n 
E ( y I )  E1   E 2 ( yi. i)  E1   Yi.  , as ssu are under srs .
 n i 1   n i 1 

1 n 1 N
 
n i 1
E1 (Yi. )   Yi.  YN  Y , as fsu are under srs .
N i 1
134 RU Khan

Which shows that y I is a biased estimate of Y . The bias of the estimator is given as

1 N 1 N 1  N N 
B  E ( y I )  Y   Yi.   M i Yi.  M  Yi.   M i Yi. 
N i 1 NM i 1 N M  i 1 i 1 

1 N
  (M i  M ) Yi.
NM i 1

To obtain the variance of y I , we have

V ( y I )  V1[ E2 ( y I i)]  E1[V2 ( y I i)] (8.13)

Consider
 1 n  1 n  1 n 
V1[ E 2 ( y I i)]  V1  E 2   yi. i   V1   E 2 ( yi. i )   V1   Yi. 
  n i 1 

n
 i 1


n
 i 1 


 V1 (Yn ) , where Yn is the mean of n selected fsu under srs .

1 1 
    S b2 . (8.14)
n N 
Consider
 1 n   1 n  1 1  2
E1[V2 ( y I i)]  E1 
 n2  V 2 ( y i. i ) 

 E1  2     S i 
 i 1   i 1 
n mi M i  

1 n  1 1  1 N 1 1  2
2  m 
    E1 ( S i2 )     S i . (8.15)
n i 1  i M i  nN i 1  mi M i 
Substituting the values from (8.14) and (8.15) in (8.13), we get

1 1  2 1 N 1 1  2
V ( y I )     Sb 
n N 
  
nN i 1  mi M i
 S i .


1  f  2 1 N  1  fi  2 n m

 n 
 Sb  
nN i 1  mi
 S i , where f  , and f i  i .
 N Mi

Estimation of V ( y I )
We have, for the sample

1  n 2 
n mi
1 1
sb2 
n 1
 ( yi.  y I )  n  1   yi.  n y I  , and si  m  1  ( yij  yi. ) 2 .
2 2 2
i 1  i 1  i j 1

Further, we know that

N   2
1 1 1
E ( sb2 )  S b2    m   S i . (8.16)
i 1 
N i Mi
Two stage sampling 135

Also,
1 n  1 1  2 1  n  1 1  
E      si   E1      E2 ( si2 i)
 n i 1 mi M i   n  i 1 mi M i  

1  n  1 1  2 1 n  1 1  
 E1      S i        E1 ( S i2 )
n  i 1 mi M i   n i 1   mi M i  

1 N 1 1  2
     Si .
N i 1  mi M i 

1 n  1 1  2 1 N 1 1  2
This shows that     si is an unbiased estimate of     S i , thus, it
n i 1 mi M i  N i 1 mi M i 
n 
2 1 1 1  2
follows from (8.16) that an unbiased estimate of S b is S b  sb   
2 ˆ 2
  si .
n i 1 mi M i 

Therefore,

1 1  1 n  1 1  2  1 1 n  1 1  2
v( y I )     sb2      si        si 
 n N   n i 1 mi M i   n  n i 1 mi M i  

1 1  1 n  1 1  2 1 n  1 1  2
    sb2 
n N 
    si      si
n 2 i 1 mi M i  nN i 1 mi M i 

1 n  1 1  2
2  m
    si .
n i 1  i M i

1 1  1 n  1 1  2 1 f  2 1 n  1  fi  2
    sb2 
n N 
  
nN i 1 mi M i
 si    sb 
 n 

nN i 1 mi
 si .
 
1 2
Note: For large N , v( y I )  s .
n b

1 n M
2nd estimate: It is defined by y II  
n i 1
ui yi. , where ui  i .
M

By definition,
 1 n  1  n  Mi
E ( y II )  E1  E2   ui yi. i   E1   ui E2 ( yi. i) , as ui  is constant for
  n i 1  n 
  i 1  M
fixed i .

1 n  1 n  1 N
 E1   ui Yi.     E1 (ui Yi. )   ui Yi.
n i 1  n i 1  N i 1

1 N
  M i Yi.  Y , which shows that y II is an unbiased estimate of Y .
NM i 1
136 RU Khan

To find the variance of y II , we have

V ( y II )  V1[ E2 ( y II i)]  E1[V2 ( y II i)] .

Consider
1 n  1 n 
V1[ E2 ( y II i )]  V1   ui E2 ( yi. i)  V1   ui Yi. 
n 
 n i 1   i 1 

 V1 (Yn ) , where Yn is the mean of n selected fsu under srs .

 1 1  2 1 N
    S b , where S b 
n N 
2

N  1 i 1
(ui Yi.  Y ) 2 .

Consider
1 n  1  n  1 1  2
E1[V2 ( y II i)]  E1   ui2 V2 ( yi. i )  E1   ui2    S i 
 n 2 i 1  n 2  i 1  mi M i  

1  n 2 1 1   1 N 2 1 1  2
2  i m 
  u    E1 ( S i2 )  ui    S i .
n i 1  i M i   nN i 1  mi M i 
Therefore,

1 1  1 N 2 1 1  2
V ( y II )     S b 2 
n N 
 ui  
nN i 1  mi M i
 S i


1 f  2 1 N 2  1  fi  2

 n 

 Sb   ui 
nN i 1  mi
 S i .


Estimation of V ( y II )
We have, for the sample

1 n 1  n 2 2 
sb 2  
n  1 i 1
2
(ui yi.  y II )  
n  1  i 1
ui yi.  n y II 2  .


Further, we know that

1 N 2 1 1  2
E ( sb 2 )  S b 2   ui  
N i 1  mi M i
 S i . (8.17)

Also
1 n  1 1  2 1  n 2  1 1  
E   ui2    si   E1   ui    E2 ( si2 i)
 n i 1  mi M i   n  i 1  mi M i  

1  n 2 1 1  2 1 n  2  1 1  
 E1   ui    S i    ui    E1 ( S i2 )
n  i 1  mi M i   n i 1   mi M i  
Two stage sampling 137

1 N 2 1 1  2
  u i    Si .

N i 1  m i M i 

1 n 2 1 1  2 1 N 2 1 1  2
This shows that  ui  
n i 1  mi M i
 si is an unbiased estimate of  ui    S i .
 N i 1  mi M i 
Therefore, it follows from (8.17) that an unbiased estimate of S b 2 is

1 n  1 1  2
Sˆb 2  sb 2   ui2    si .
n i 1  mi M i 
Hence, an unbiased estimate of V ( y II ) will be

1 1  2 1  1  2  1 1 n 2  1  2
n
1 1
v ( y II )     sb   ui2 
   si     ui    si 
 n N   n i 1  mi M i   n  n i 1  mi M i  

1 1  1 n 2 1 1  2 1 n 2 1 1  2
    sb 2 
n N 
 ui    si   ui    si
n 2 i 1  mi M i  nN i 1  mi M i 

1 n  1 1  2

2  ui2  m   si
M i 
n i 1  i

1 1  1 n 2 1 1  2 1  f  2 1 n 2 1  fi  2
    sb 2 
n N 
 ui  
nN i 1  mi M i
 si  
 n
 b

s    ui 
nN i 1  mi
 si .
 

1 2
Note: For large N , v( y II )  s .
n b
n
1 
Corollary: Show that the estimator Y   NM i  y i.  estimates unbiased the population
ˆ
i 1 n 
2  1  fi  2
N
1 f  2 N
total Y and its variance is V (Yˆ )  N 2   S b   M i   S i .
 n  n i 1  mi 