Math 408 - Mathematical Statistics

Lecture 20-21.
Neyman Allocation vs Proportional Allocation
and
Stratified Random Sampling vs Simple Random Sampling

March 8-13, 2013

Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 1/9
Agenda
Neyman Allocation and its properties
∗
Variance of the optimal stratified estimate X n,opt
Drawbacks of Neyman Allocation
Proportional Allocation
Neyman vs Proportional
Stratified vs Simple
Summary

Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 2/9
Neyman allocation
In Lecture 19, we described the optimal allocation scheme for stratified random
sampling, called Neyman allocation. Neyman allocation scheme minimizes
∗ PN
variance V[X n ] subject to k=1 nk = n.

Theorem
The sample sizes n1 , . . . , nL that solve the optimization problem
L L
∗ X σ2 X
V[X n ] = ωk2 k → min s.t. nk = n
nk
k=1 k=1

are given by ωk σk
n̂k = n PL k = 1, . . . , L (1)
j=1 ωj σj

The theorem says that if ωk σk is large, then the corresponding stratum should be
sampled heavily. This is very natural since
if ωk is large, then the stratum contains a large portion of the population
if σk is large, then the population values in the stratum are quite variable
and, therefore, to estimate µk accurately a relatively large sample size must
be used
Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 3/9
Variance of the optimal stratified estimate
In stratified random sampling, an (unbiased) estimate of µ is
L
∗ X (k)
Xn = ωk X nk
k=1

If Neyman (i.e. optimal) allocation is used (nk = n̂k ), then the optimal stratified
∗
estimate of µ, denoted by X n,opt , is
L
∗ X (k)
X n,opt = ωk X n̂k
k=1

Theorem
The variance of the optimal stratified estimate is

L
!2
∗ 1 X
V[X n,opt ] = ωk σk
n
k=1

Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 4/9
Proportional Allocation
There are two main disadvantages of Neyman allocation:
1 Optimal allocations n̂k depends on σk which generally will not be known
2 If a survey measures several values for each population member, then it is
usually impossible to find an allocation that is simultaneously optimal for all
values
A simple and popular alternative method of allocation is proportional allocation:
to choose n1 , . . . , nL such that

n1 n2 nL
= = ... =
N1 N2 NL

This holds if
Nk
ñk = n = nωk k = 1, . . . , L (2)
N

Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 5/9
Proportional Allocation
If proportional allocation is used (nk = ñk = nωk ), then the corresponding
∗
stratified estimate of µ, denoted by X n,p , is
L L ñk L ñk
∗ X (k) X 1 X (k) 1 XX (k)
X n,p = ωk X ñk = ωk Xi = Xi
ñk n
k=1 k=1 i=1 k=1 i=1

∗
Thus, X n,p is simply the unweighted mean of the sample values.

Theorem
∗
The variance of X n,p is given by

L
∗ 1X
V[X n,p ] = ωk σk2
n
k=1

Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 6/9
Neyman vs Proportional
By definition, Neyman allocation is always better than proportional allocation
(since Neyman allocation is optimal).
Question: When is it substantially better?

Proposition
L L
∗ ∗ 1X X
V[X n,p ] − V[X n,opt ] = ωk (σk − σ̄)2 , σ̄ = ωk σk
n
k=1 k=1

Therefore,
if the variances σk of the strata are all the same, then proportional allocation
∗ ∗
is as efficient as Neyman allocation, V[X n,p ] = V[X n,opt ]
the more variable σk , the more efficient the Neyman allocation scheme

Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 7/9
Stratified vs Simple
Let us now compare simple random sampling and stratified random sampling with
proportional allocation.
Question: What is more efficient? (more efficient = has smaller variance)

Proposition
L
∗ 1X
V[X n ] − V[X n,p ] = ωk (µk − µ)2
n
k=1

Thus, stratified random sampling with proportional allocation always gives a

Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 8/9
Summary
The variance of the optimal stratified estimate (Neyman allocation) of µ is

L
!2
∗ 1 X
V[X n,opt ] = ωk σk
n
k=1
Neyman allocation is difficult to implement in practice
Proportional allocation: ñk = n NNk = nωk
The variance of the stratified estimate under proportional allocation:
L
∗ 1X
V[X n,p ] = ωk σk2
n
k=1
By definition, Neyman allocation is better than proportional allocation, but
if the variances σk of the strata are all the same, then proportional allocation
is as efficient as Neyman allocation
Stratified random sampling with proportional allocation is always more
efficient than simple random sampling.

Konstantin Zuev (USC) Math 408, Lecture 9-10 March 8-13, 2013 9/9