Two-stage sampling

Anower Hossain, PhD

Institute of Statistical Research and Training (ISRT)

University of Dhaka

22 August 2020

Anower Hossain Sampling

Expectation and variance

Let θ̂ be an estimator of the parameter θ using a two-stage sampling.

Then
  h i
E θ̂ = E1 E2 (θ̂) = E1 E2 (θ̂)

where E denotes expected or average value over all samples, E2 denotes

averaging over all possible second-stage selections from a fixed set of
units, and E1 denotes averaging over all first-stage selections. and
h  i h  i
V (θ̂) = V1 E2 θ̂ + E1 V2 θ̂

Anower Hossain Sampling

Proof of V (θ̂)

h  i h  i
V (θ̂) = V1 E2 θ̂ + E1 V2 θ̂

Proof:
 2
V (θ̂) = E1 E2 θ̂ − θ , θ̂ is unbiased for θ

 2
E2 θ̂ − θ = E2 (θ̂2 ) − 2θE2 (θ̂) + θ2
h i2
= E2 (θ̂2 ) − θ2 = E2 (θ̂) + V2 (θ̂) − θ2

Average now over first-stage selections and since θ = E1 E2 (θ̂)

h i2 h i h i h i
V (θ̂) = E1 E2 (θ̂) − θ2 + E1 V2 (θ̂) = V1 E2 (θ̂) + E1 V2 (θ̂)

Anower Hossain Sampling

Theorem

Theorem
Suppose that Yi is the total of ith primary sampling unit (PSU) and has
an unbiased estimate Ŷi considering only the second stage sampling.
Moreover, let wi is the weight for the ith PSU depending on the first
stage sampling. Then an unbiased estimate of population total Y can be
obtained based on a two stage sampling as
n1
X
Ŷ2ss = wi Ŷi ,
i=1

where n1 is the number of PSUs in the sample. Further, defining a

0 0
random variable wi for which E1 (wi ) = 1 and

if the ith PSU is in the sample


0 wi
wi = ,
0 otherwise

Anower Hossain Sampling

Theorem

The variance of the estimator Ŷ2ss is given by,

n1
! N
  X X 1  0   
V Ŷ2ss = V1 wi Yi + E1 wi2 V2 Ŷi ,
i=1 i=1

where the subscripts 1 and 2 indicate the expectation and variability are
to be computed for 1st and 2nd stage sampling respectively.

Anower Hossain Sampling

Proof

Since the estimation is based on a two stage sampling expectation has to

be taken over the randomization in both stages. However, the second
stage expectation is to be taken for fixed PSU, then expectation is to be
taken over the randomization in selection of PSUs at first stage.
Therefore,
   
E Ŷ2ss = E1 E2 Ŷ2ss
n1
! "n #
X X 1  
= E1 E2 wi Ŷi = E1 wi E2 Ŷi
i=1 i=1
"n # "N #
1 1
X X 0
= E1 wi Yi = E1 wi Yi , (N1 = # of PSUs in population
i=1 i=1
N1
X  0
= Yi E1 wi
i=1
N1
X
= Yi = Y
i=1

Anower Hossain Sampling

Variance

Due to randomization occurred in two stages, the variance of Ŷ2ss

calculated as
     
V Ŷ2ss = E1 V2 Ŷ2ss + V1 E2 Ŷ2ss

Now,
n1
!
  X
E1 V2 Ŷ2ss = E1 V2 wi Ŷi
i=1
N1
!
X 0
= E1 V 2 wi Ŷi
i=1
(N )
1  
X 0
2
= E1 wi V2 Ŷi ; (subsampling is independent)
i=1
N1
X  0   
= E1 wi2 V2 Ŷi
i=1

Anower Hossain Sampling

Cont...

Again,
n1
!
  X
V1 E2 Ŷ2ss = V1 E2 wi Ŷi
i=1
(n )
X 1  
= V1 wi E2 Ŷi
i=1
n1
!
X
= V1 wi Yi
i=1

Finally we have,
n1
! N1
  X X  0   
V Ŷ2ss = V1 wi Yi + E1 wi2 V2 Ŷi
i=1 i=1

Anower Hossain Sampling

Estimator of V (Ŷ2ss )

Theorem
An unbiased estimate of V (Ŷ2ss ) can be obtained by,
n1
! n1
  X X  
v Ŷ2ss = v1 wi Ŷi + wi v2 Ŷi ,
i=1 i=1

where
 
E2 [v2 (Yi )] = V2 Ŷi , and
" n1
!# n1
!
X X
E1 v1 wi Ŷi = V1 wi Ŷi .
i=1 i=1

Anower Hossain Sampling

Proof.
 
The 1st term in V Ŷ2ss can be expressed as a quadratic form of Yi ’s as,

n1
! N1
!
X X 0
V1 wi Yi = V1 wi Yi
i=1 i=1
N1 N1 X
N1
X 0 X 0 0
= Yi2 V1 (wi ) + 2 Yi Yj cov(wi wj ) (1)
i=1 i=1 j>i

Let us define the two indicator as the following way,

ai if the ith PSU is in the sample


0
ai = ,
0 otherwise

and,

if the ith and j th PSUs are in the sample


0 bij
bij = ,
0 otherwise

Anower Hossain Sampling

Cont...

Using the idea in equation 1, the 1st stage estimated variance can be
written as,
n1
! N1 N1 X
N1
X X 0 X 0
v1 wi Yi = ai Yi2 + 2 bij Yi Yj . (2)
i=1 i=1 i=1 j>i

Comparing the equations (1) and (2) and using the given unbiasedness
condition (second) we can write,
 0  0
E1 ai = V1 wi

Now, using the second stage estimators the equation (2) takes the
following form,
n1
! N1 N1 X
N1
X X 0 X 0
v1 wi Ŷi = ai Ŷi2 + 2 bij Ŷi Ŷj .
i=1 i=1 i=1 j>i

Anower Hossain Sampling

Cont...

Now,
h  i h  i
E v Ŷ2ss = E1 E2 v Ŷ2ss
" n1
! n1
#
X X  
= E1 E2 v1 wi Ŷi + wi v2 Ŷi
i=1 i=1
 
N1 N1 X
N1 n1 h  i
X 0 X 0 X
2
= E1 E2  ai Ŷi + 2 bij Ŷi Ŷj + E1
 wi E2 v2 Ŷi
i=1 i=1 j>i i=1
 
N1   N1
N1 X    
X 0 X 0
= E1  ai E2 Ŷi2 + 2 bij E2 Ŷi E2 Ŷj 
i=1 i=1 j>i
n1
X  
+ E1 wi V2 Ŷi
i=1

Anower Hossain Sampling

Cont...

 
N1     N1 XN1 n1
X 0
n o2 X 0
 X  
= E1 ai V2 Ŷi + E2 Ŷi +2 bij Yi Yj + E1 wi V2 Ŷi
 
i=1 i=1 j>i i=1
 
N1 h  0  0 i   X N1 N1 X
N1 
X 0 X 0
= E1 ai + E1 wi V2 Ŷi + E1 ai Yi2 + 2 bij Yi Yj
 
i=1 i=1 i=1 j>i
N1 h
( n1
!)
X  0  0 i   X
= E1 ai + E1 wi V2 Ŷi + E1 v1 wi Yi
i=1 i=1
N1  n1
!
X  0  n  0 o2    X
= V1 wi + E1 wi V2 Ŷi + V1 wi Yi
i=1 i=1
N1 n1
!
X  0
   X
2
= E1 wi V2 Ŷi + V1 wi Yi
i=1 i=1
 
= V Ŷ2ss

Anower Hossain Sampling

Two-stage sampling

In a two-stage unequal cluster sampling, an unbiased estimator of Yi ,

total of the ith PSU, can be found in Ŷi . Let wi be the weight for the
ith PSU depending on the 1st stage sampling method. Also, let wi0 be a
random variable such that
(
0 wi , if the ith PSU is in the sample
wi =
0, otherwise.

Then an unbiased estimate of the population total Y is given by

n
X
Ŷ2ss = wi Ŷi
i=1

and the variance of the estimate is given by

n
! N
  X X  
V Ŷ2ss = V1 wi Yi + E1 (wi02 )V2 Ŷi
i=1 i=1

E1 (wi0 )
if and only if = 1 for ∀i. Note that n is the number of PSU in
the sample and N is the number of PSU in the population.
Anower Hossain Sampling
Two stage sampling

 
An unbiased estimator of V Ŷ2ss is given by
n
! n
  X X  
v Ŷ2ss = v1 wi Ŷi + wi v2 Ŷi
i=1 i=1

Anower Hossain Sampling

Example 1- Two-stage sampling

Suppose in the first-stage a sample of n PSU is selected using SRS from

a population of N units, and in the second-stage a sample of mi subunits
are also drawn using SRS from the ith selected PSU of size Mi . Let yij
be the observation for the jth subunit within the ith PSU.
Derive an expression for the unbiased estimate, say Ŷ2SRS , of the
population total Y .
 
Derive the variance of Ŷ2SRS , i.e., V Ŷ2SRS .
 
Derive an unbiased estimator of V Ŷ2SRS .

Anower Hossain Sampling

Example 2

n
X N
Ŷ2SRS = Ŷi
i=1
n

 N (N − n) XN N
 N X Mi (Mi − mi ) 2
Var Ŷ2SRS = S12 + S2i
n i=1
n i=1 mi
where !2
N N
1 X 1 X
S12 = Yi − Yi
N − 1 i=1 N i=1
and !2
Mi M
2 1 X 1 Xi
S2i = Yij − Yij
Mi − 1 j=1 Mi i=1

Anower Hossain Sampling

Example 1

n n
c Ŷ2SRS = N (N − n) N X Mi (Mi − mi ) 2
  X
Var s21 + s2i
n i=1
n i=1 mi
where !2
n n
1 X 1X
s21 = Ŷi − Ŷi
n − 1 i=1 n i=1
and !2
mi mi
1 X 1 X
s22i = Yij − Yij
mi − 1 j=1 mi i=1

Anower Hossain Sampling

Example1

n
!
  X N
E Ŷ2SRS = E1 E2 Ŷi
i=1
n
n
!
X N
= E1 E2 (Ŷi )
i=1
n
n
!
X N
= E1 Yi
i=1
n
n
!
N X
= E1 Yi
n i=1
N
!
N X n
= Yi
n i=1
N
N
X
= Yi = Y
i=1

Anower Hossain Sampling

Example 2

Two-Stage Sampling with Unequal Probability Without Replacement

Sampling at Both Stages

Anower Hossain Sampling

Example 2- Two-Stage Sampling with Unequal Probability Without
Replacement Sampling at Both Stages

Horvitz-Thompson estimator
n
X yi
ŶHT =
π
i=1 i

N   N XN
  X 1 − πi X Yi Yj
V1 ŶHT = Yi2 + (πij − πi πj )
i=1
πi i=1
πi πj
j6=i

n   n X
n  
  X 1 − πi X πij − πi πj Yi Yj
V
b 1 ŶHT = Yi2 +
i=1
πi2 i=1 j6=i
πij πi πj

Anower Hossain Sampling

Example 2- Two-Stage Sampling with Unequal Probability Without
Replacement Sampling at Both Stages

Horvitz-Thompson estimator
n
X yi
ŶHT =
π
i=1 i

N X
N  2
  X Yi Yj
V2 ŶHT = (πi πj − πij ) −
i=1 j<i
πi πj
n X n   2
  X πi πj − πij Yi Yj
V
b 2 ŶHT = −
i=1 j<i
πij πi πj

Anower Hossain Sampling

Example 2- Two-Stage Sampling with Unequal Probability Without
Replacement Sampling at Both Stages

Let
πi : prob. that ith PSU is in the sample
πij : prob. that both ith and jth PSUs are in the sample
πi,k : prob. that kth subumit of the ith PSU is in the sample
πi,kl : prob. that both kth and lth subunits of ith PSU are in the sample
Then
E1 (wi0 ) = wi × πi + 0 × (1 − πi ) = 1
which imples
1
wi =
πi
Similarly, we get
1
E1 (wi02 ) = wi2 × πi + 02 × (1 − πi ) =
πi

Anower Hossain Sampling

Example 2- Two-Stage Sampling with Unequal Probability Without
Replacement Sampling at Both Stages

The estimator of the population is given by

n
X 1
Ŷ2upwor = Ŷi
π
i=1 i

and it’s variance is given by

n
! N
  X 1 X 1  
V Ŷ2upwor = V1 Yi + V2 Ŷi
π
i=1 i
π
i=1 i

Since the 1st stage sampling is unequal probability without replacement

sampling, we have the following results directly from Horvitz-Thompson
theories
n
! N X N  2
X 1 X Yi Yj
V1 Yi = (πi πj − πij ) −
π
i=1 i i=1 j<i
πi πj

Anower Hossain Sampling

Example 2- Two-Stage Sampling with Unequal Probability Without
Replacement Sampling at Both Stages

The 2nd stage sampling is also unequal probability without replacement

sampling, we have
Mi X
Mi  2
  X Yil Yik
V2 Ŷi = (πi,l πi,k − πi,kl ) −
πi,l πi,k
l=1 k<l

Then
N X
N  2
  X Yi Yj
V Ŷ2upwor = (πi πj − πij ) −
i=1 j<i
πi πj
N M M 2
1 Xi Xi

X Yil Yik
+ (πi,l πi,k − πi,kl ) −
π
i=1 i
πi,l πi,k
l=1 k<l

Anower Hossain Sampling

Example 2- Two-Stage Sampling with Unequal Probability Without
Replacement Sampling at Both Stages

An unbiased estimate of the variance is

n X n 
!2
Yˆj

  X πi πj − πij Ŷi
Vb Ŷ2upwor = −
i=1 j<i
πij πi πj
n mi Xmi   2
X 1 X πi,l πi,k − πi,kl Yil Yik
+ −
π
i=1 i
πi,kl πi,l πi,k
l=1 k<l

Anower Hossain Sampling

Three stage sampling

In a three stage unequal cluster sampling, an unbiased estimator of Yij ,

total of the jth secondary sampling unit (SSU) in the ith primary
sampling unit (PSU), can be found in Ŷij . Let wi be the weight for the
ith PSU depending on the 1st stage sampling method, and uij be the
weight for the jth SSU depending on the 2nd stage sampling method.
Let wi0 and u0ij be the random variables such that
(
wi , if the ith PSU is in the sample
wi0 =
0, otherwise.

and
(
uij , if the jth SSU in the ith PSU is in the sample
u0ij =
0, otherwise.

Anower Hossain Sampling

Three stage sampling

Then an unbiased estimate of the population total Y is given by

n
X mi
X
Ŷ3ss = wi uij Ŷij
i=1 j=1

and the variance of the estimate is given by

" n #  
  X N
X mi
X
V Ŷ3ss = V1 wi Yi + E1 (wi02 )V2  uij Yij 
i=1 i=1 j=1
N
X Mi
X  
+ E1 (wi02 ) E2 (u02
ij )V3 Ŷij
i=1 i=1

if and only if E1 (wi0 ) = 1 and E2 u0ij = 1 for ∀i, j. Note that n is the

number of PSU selected in the sample, N is the number of PSU in the

population, Mi is the number of SSUs in the ith selected PSU, mi is the
number of SSUs selected from the ithe PSU selected in the first-stage.
Anower Hossain Sampling
Three stage sampling

Unbiasedness:
 
Xn mi
X
E(Ŷ3ss ) = E1 E2 E3  wi uij Ŷij 
i=1 j=1
 
XN Mi
X
= E1 E2 E3  wi0 u0ij Ŷij 
i=1 j=1
N
X Mi
X
= E1 (wi0 ) E2 (u0ij )E3 (Ŷij )
i=1 j=1
Mi
N X
X
= Yij
i=1 j=1

=Y

Anower Hossain Sampling

Three stage sampling

Variance:
Var(Ŷ3ss ) = E1 E2 V3 (Ŷ3ss ) + E1 V2 E3 (Ŷ3ss ) + V1 E2 E3 (Ŷ3ss )
1st Term:
 
Xn mi
X
E1 E2 V3 (Ŷ3ss ) = E1 E2 V3  wi uij Ŷij 
i=1 j=1
 
XN Mi
X
= E1 E2 V3  wi0 u0ij Ŷij 
i=1 j=1
 
XN Mi
X
= E1 E2  wi02 u02
ij V3 (Ŷij )

i=1 j=1
N
X Mi
X
= E1 (wi02 ) E2 (u02
ij )V3 (Ŷij )
i=1 j=1

Anower Hossain Sampling

Three stage sampling

2nd Term:
 
Xn mi
X
E1 V2 E3 (Ŷ3ss ) = E1 V2 E3  wi uij Ŷij 
i=1 j=1
 
n
X mi
X
= E1 V2  wi uij E3 (Ŷij )
i=1 j=1
 
n
X mi
X
= E1 V2  wi uij Yij 
i=1 j=1
  
XN mi
X
= E1  wi02 V2  uij Yij 
i=1 j=1
 
N
X Xmi
= E1 (wi02 )V2  uij Yij 
i=1 j=1

Anower Hossain Sampling

Three stage sampling

3rd Term:
 
Xn mi
X
V1 E2 E3 (Ŷ3ss ) = E1 V2 E3  wi uij Ŷij 
i=1 j=1
 
n
X mi
X
= V1 E2  wi uij E3 (Ŷij )
i=1 j=1
 
n
X mi
X
= V1 E2  wi uij Yij 
i=1 j=1
   
Xn Mi
X Xn Mi
X
= V1  wi E2 (u0ij )Yij  = V1  wi Yij 
i=1 j=1 i=1 j=1
n
!
X
= V1 wi Yi
i=1

Anower Hossain Sampling

Three stage sampling with SRS at each stage

Suppose in the first-stage a sample of n PSU is selected using SRS from a

population of N PSUs, in the second-stage a sample of mi SSUs is drawn
using SRS from the ith selected PSU of size Mi , and in the final stage a
sample of dij is drawn also using SRS from the (ij) unit of size Dij .

Anower Hossain Sampling