CHAPTER FOUR

4. STRATIFIED RANDOM SAMPLING

4.1. Basic Concepts of Stratification
Definition: Stratified sampling is a technique, which uses any relevant information, which
might be available, in order to increase efficiency. Stratified sampling is the appropriate method
if the population under consideration consists of heterogeneous units. It involves the division or
stratification of a population by partitioning the sampling frame into non-overlapping and
relatively homogeneous groups called strata. The selection of samples can be performed
independently in each of those strata.
Stratified random sampling is a sampling plan in which a population is divided into L mutually
exclusive and exhaustive strata, and a simple random sample of nh elements is taken separately
and independently within each stratum. Let N1, N2, N3, - - -, NL represent the number of
sampling units within each stratum, and n1, n2, n3, - - -, nL represent the number of randomly
selected sampling units within each stratum. Then the total number of possible stratified random
samples is equal to (𝑁
𝑛
1
) × (𝑁
𝑛
2
) × (𝑁
𝑛
3
) × − − − × (𝑁𝑛 𝐿) ≤ (𝑁𝑛).
1 2 3 𝐿
Stratified random sampling, in particular, involves dividing the population into strata, and then
selecting simple random samples from each of the stratum. Stratification variables may be
geographic (region, zone, district, rural/urban) or non-geographic (income, age, sex, etc.). It
should be kept in mind that stratification is limited only to those items of information, which are
available on the frame.
4.2. Major Advantages Stratification
Stratified sampling is used in certain types of surveys because it combines the conceptual
simplicity of simple random sampling with potentially significant gains in reliability. Basically
there are four major reasons for resorting to stratification:
i) The principal objective of stratification is to reduce sampling error. Under certain
conditions the variances, of the sample estimates may be decreased, which means
precision may be increased over simple random sample.
In a stratified sample, the sampling error depends on the population variance existing
within the strata but not b/n the strata. For this reason, it pays to create strata with low
internal variability. It follows that the more homogeneous the groups, the greater the
precision of the sample estimate.
ii) In some cases, separate estimates are required at the stratum level. For example, in
household surveys estimates may be required by province, income group, occupation, age
group, urban size group, educational category, etc.
iii) Stratified sampling is administratively convenient. It can enable a survey organization to
control the distribution of fieldwork among its regional offices.
iv) Sometimes, different parts of the population may call for different sampling procedures.
For example:
 There may be incomplete frames in which lists may be available for different parts of
the population and in this case supplementing an incomplete list will be required.

Sampling Theory | Chapter Four | Stratified Random Sampling Page 1

  Physical distribution of a population such as private households, institutions (military
camp, prisons, hostel), and densely or sparsely populated areas may call for different
procedures. A different procedure may be used to sample persons in sparsely
populated rural areas than that used for the densely populated urban areas.
 The diverse nature of the sampling units of the population may call for different
procedures. For instance, employee’s study requires stratification by literate and
illiterate; and study on young people between 14 and 18 years requires young people
at school and those who are not at school.
The strategy employed for constructing strata involves, first determining the population
parameter we are interested in estimating and, second stratify the population with respect to
another variable that is thought to be associated with the variable of interest. For example, in the
farm expenditure survey, it is desirable to stratify the population of farms on the basis of
information, which highly related to the cost of running a farm. One may consider size and types
of farming (crops, livestock, etc) as appropriate stratification variables.
The major disadvantage of stratified sampling is that it may take more time to select the sample
than would be the case for simple random sampling. More time is involved because complete
frames are necessary within each of the strata and each stratum must be sampled.
Notations:
For Population Characteristics: Strata can be defined with the same type of notation that we used
to define characteristics of a population in simple random sampling. Consider a population
containing N elements that are grouped exclusively and exhaustively into L strata.
Let L represents the number of strata and h stands for the particular stratum, h = 1, 2, 3, - - -, L
Nh = Total number of non-overlapping units in a stratum h
𝑁 = ∑𝐿ℎ=1 𝑁ℎ =Total population size.
𝑌ℎ𝑖 = value of the study variable for the ith unit in the stratum h, i = 1, 2, 3, - - -, 𝑁ℎ .
𝑁ℎ
Yh = ∑𝑖=1 𝑌ℎ𝑖 = Total amount of study variable for stratum h
𝑁
𝑌 = ∑𝐿ℎ=1 𝑌ℎ = ∑𝐿ℎ=1 ∑𝑖=1
ℎ
𝑌ℎ𝑖 = Total amount of study variable for the entire population.
𝑁 ℎ 𝑌
𝑌 ∑𝑖=1 ℎ𝑖
𝑌̅ℎ = 𝑁ℎ = = the true mean for stratum h,
ℎ 𝑁ℎ
𝑌 𝑁 𝑌̅ 𝑁ℎ
𝑌̅ = 𝑁 = ∑𝐿ℎ=1 ℎ ℎ = ∑𝐿ℎ=1 𝑊ℎ 𝑌̅ℎ = the true mean for the entire population, where 𝑊ℎ = ,
𝑁 𝑁
stratum weight or proportion of the population units falling in the stratum.
𝑁
ℎ (𝑌 −𝑌 2 𝑁 ℎ (𝑌 −𝑌 2
∑𝑖=1 ℎ𝑖 ̅ ℎ ) ∑𝑖=1 ℎ𝑖 ̅ ℎ )
𝑆 2ℎ = or 𝜎 2 ℎ = is the population variance for stratum h.
𝑁ℎ −1 𝑁ℎ
For Sample Characteristics:
nh = Number of sample units in a stratum h, 𝑛 = ∑𝐿ℎ=1 𝑛ℎ =Total sample size
𝑦ℎ𝑖 = The value of the variable in the sample for the ith unit in the stratum h, i = 1, 2, 3, - - -, 𝑛ℎ .
𝑛ℎ 𝑦
𝑦ℎ ∑𝑖=1
yh = ∑𝑛𝑖=1
ℎ
𝑦ℎ𝑖 = Sample total for stratum h, 𝑦̅ℎ = 𝑛 =
ℎ𝑖
= Sample mean for stratum h,
ℎ 𝑛ℎ
𝑛ℎ
𝑓ℎ = 𝑁 = Sampling fraction in stratum h,
ℎ

Sampling Theory | Chapter Four | Stratified Random Sampling Page 2

 𝑛
ℎ (𝑦 −𝑦 2
∑𝑖=1 ℎ𝑖 ̅ ℎ )
𝑠2ℎ = = Sample variance for stratum h.
𝑛ℎ −1
4.3.Estimation of Population Mean and Total
a) Total Estimate: The estimated total for stratum h is given by 𝑌̂ℎ = 𝑁ℎ 𝑦̅ℎ . For the entire
population the estimator can be expressed as 𝒀̂ 𝒔𝒕 = ∑𝑳𝒉=𝟏 𝒀
̂ 𝒉 = ∑𝑳𝒉=𝟏 𝑵𝒉 𝒚
̅𝒉 .
b) Mean Estimate: The estimator used in stratified sampling is denoted by 𝒚 ̅𝒔𝒕 and given as
̂ 𝒔𝒕
𝒀 ∑𝑳𝒉=𝟏 𝑌̂ℎ ∑𝑳𝒉=𝟏 𝑵𝒉 𝒚
̅𝒉
̅𝒔𝒕 =
𝒚 = = = ∑𝑳𝒉=𝟏 𝑾𝒉 𝒚
̅𝒉 .
𝑵 𝑵 𝑵

4.3.1. Properties of the estimates

Theorem 4.1: If in every stratum the sample estimate 𝑦̅ℎ is an unbiased estimator of 𝑌̅ℎ , then 𝑦̅𝑠𝑡
is an unbiased estimator of the population mean 𝑌̅; i.e., 𝐸(𝑦̅𝑠𝑡 ) = 𝑌̅
𝑁 𝑁ℎ 𝑌ℎ𝑖
Proof: 𝐸(𝑦̅𝑠𝑡 ) = 𝐸(∑𝐿ℎ=1 𝑊ℎ 𝑦̅ℎ ) = ∑𝐿ℎ=1 𝑊ℎ 𝐸(𝑦̅ℎ ) = ∑𝐿ℎ=1 𝑊ℎ 𝑌̅ℎ = ∑𝐿ℎ=1 ℎ ∑𝑖=1 = 𝑁 𝑁ℎ
1 𝑌
∑𝐿 ∑𝑁 ℎ 𝑌 = 𝑁 = 𝑌̅.
𝑁 ℎ=1 𝑖=1 ℎ𝑖
Since the estimates are unbiased in the individual strata. But the population mean 𝑌̅ may be
written

Corollary: The total estimate 𝒀 ̂ 𝒔𝒕 = 𝑵𝒚̅𝒔𝒕 is an unbiased estimator of the population total, Y, i.e.,
𝐸(𝑌̂𝑠𝑡 ) = 𝑌.
Theorem 4.2: If 𝑦̅ℎ is an unbiased estimate of 𝑌̅ℎ in every stratum, and sample selection is
independent in different strata, then the variance of the estimate is given by: 𝑉𝑎𝑟(𝑦̅𝑠𝑡 ) =
∑𝐿ℎ=1 𝑊ℎ 2 𝑉𝑎𝑟(𝑦̅ℎ ), where 𝑉𝑎𝑟(𝑦̅ℎ ) the variance of 𝑦̅ℎ over repeated samples from stratum h.
Proof: From linear combination of ∑ 𝑎𝑖 𝑦𝑖
𝑉𝑎𝑟 (∑ 𝑎𝑖 𝑦𝑖 ) = ∑ 𝑎𝑖 2 𝑉𝑎𝑟(𝑦𝑖 ) + ∑ ∑ 𝑎𝑖 𝑎𝑗 𝐶𝑜𝑣(𝑦𝑖 𝑦𝑗 )
𝑖≠𝑗
Similarly,
𝐿 𝐿

𝑉𝑎𝑟(𝑦̅𝑠𝑡 ) = 𝑉𝑎𝑟 (∑ 𝑊ℎ 𝑦̅ℎ ) = ∑ 𝑊ℎ 2 𝑉𝑎𝑟(𝑦̅ℎ ) + 2 ∑ ∑ 𝑦̅ℎ 𝑦̅𝑗 𝐶𝑜𝑣(𝑦̅ℎ 𝑦̅𝑗 )

ℎ=1 ℎ=1 ℎ<𝑗
But since the samples are selected independently in different strata, the covariance terms vanish.
Thus, 𝑉𝑎𝑟(𝑦̅𝑠𝑡 ) = ∑𝐿ℎ=1 𝑊ℎ 2 𝑉𝑎𝑟(𝑦̅ℎ )
If 𝑦̅ℎ is an unbiased estimate of 𝑌̅ℎ in every stratum, and sample selection is independent in
different strata, then 𝑦̅𝑠𝑡 is an unbiased estimate of 𝑌̅ with variance of ∑𝐿ℎ=1 𝑊ℎ 2 𝑉𝑎𝑟(𝑦̅ℎ ).
The important point about this result is that the variance of 𝑦̅𝑠𝑡 depends only on the variance of
the estimate of the individual stratum mean 𝑌̅ℎ .
Theorem 4.3: For stratified random sampling, without replacement, the variance of the estimate,
𝑆ℎ 2 𝑆ℎ 2
𝑦̅𝑠𝑡 is 𝑉𝑎𝑟(𝑦̅𝑠𝑡 ) = ∑𝐿ℎ=1 𝑊ℎ 2 (1 − 𝑓ℎ ) = 𝑁12 ∑𝐿ℎ=1 𝑁ℎ (𝑁ℎ − 𝑛ℎ )
𝑛ℎ 𝑛ℎ

Sampling Theory | Chapter Four | Stratified Random Sampling Page 3

 2 2
𝑆 𝑆 𝑁 −𝑛
Proof: since 𝑦̅ℎ is an unbiased estimate of 𝑌̅ℎ ; and 𝑉𝑎𝑟(𝑦̅ℎ ) = 𝑛ℎ (1 − 𝑓ℎ ) = 𝑛ℎ ( ℎ𝑁 ℎ ). Then,
ℎ ℎ ℎ

If the sampling fractions 𝑛ℎ ⁄𝑁ℎ are negligible in every strata,

𝑛ℎ 𝑁ℎ 𝑛ℎ 𝑛
If in every stratum, = 𝑜𝑟 = 𝑁 𝑜𝑟 𝑓ℎ = 𝑓
𝑛 𝑁 𝑁ℎ
This means that the sampling fraction is the same in all strata. This stratification is described as
stratification with proportional allocation of the 𝑛ℎ .
𝑁ℎ
With proportional allocation, we substitute 𝑛ℎ = 𝑛 .
𝑁
Thus, the variance 𝑉𝑎𝑟(𝑦̅𝑠𝑡 ) reduces to

Estimated variance: If an SRS is taken within each stratum, an unbiased estimate of 𝑆ℎ 2 is

Theorem 4.4: With stratified random sampling, an unbiased estimate of the variance of 𝑦̅𝑠𝑡 is

Estimating the population total: Since 𝑦̅𝑠𝑡 is an estimator of the population mean 𝑌̅ in
stratified random sampling, 𝑌̂ =N𝑦̅𝑠𝑡 is an
estimator of the population total 𝑌.

Sampling Theory | Chapter Four | Stratified Random Sampling Page 4

And

𝑆ℎ 2
Corollary: The standard error of 𝑦̅𝑠𝑡 is S.E.(𝑦̅𝑠𝑡 ) = √∑𝐿ℎ=1 𝑊ℎ 2 (1 − 𝑓ℎ ).
𝑛ℎ
4.3.2. Confidence Intervals
For Population Mean: A 100(1 - α)% confidence interval for population mean 𝑌̅ is given as:
𝑌̅ = 𝑦̅𝑠𝑡 ± 𝑍𝛼 𝑆. 𝐸. (𝑦̅𝑠𝑡 )
2
For Population Total: A 100(1 - α)% confidence interval for population total 𝑌 is given as:
𝑌 = 𝑌̂𝑠𝑡 ± 𝑍𝛼 𝑆. 𝐸. (𝑌̂𝑠𝑡 ) 𝑜𝑟 𝑌 = 𝑁𝑦̅𝑠𝑡 ± 𝑍𝛼 𝑁𝑆. 𝐸. (𝑦̅𝑠𝑡 ).
2 2
Where 𝑦̅𝑠𝑡 is assumed to be distributed normally and Z value is obtained from table of normal
distribution. For small sample, 𝑛ℎ , we use the t-value from the tables of the t-distribution.
Example: Consider a population of 14 military families living on three compounds. We can use
the families as elementary units, the compounds as strata, and family size as the characteristics.
The values of population characteristics for each stratum are given below.
Stratum 1 Stratum 2 Stratum 3
Nh: N1= 3 N2 = 5 N3= 6
Yhi: Y11 = 4 Y21 = 4 Y31 = 2
Y12 =3 Y22 = 6 Y32 = 3
Y13 = 4 Y23 = 4 Y33 = 2
Y24 = 7 Y34 = 2
Y25 = 8 Y35 = 2
Y36 = 3
Yh: Y1= 11 Y2= 29 Y3 = 14
̅
𝑌ℎ : ̅
𝑌1 = 3.67 ̅
𝑌2 = 5.8 ̅
𝑌3 = 2.33
2 2 2
𝑆 ℎ: 𝑆 1 = 0.333 𝑆 2 = 3.200 𝑆 2 3 = 0.267
3 3

𝑁 = ∑ 𝑁ℎ = 3 + 5 + 6 = 14, 𝑌 = ∑ 𝑌ℎ = 11 + 29 + 14 = 54,
ℎ=1 ℎ=1

Sampling Theory | Chapter Four | Stratified Random Sampling Page 5

 3
𝑌
𝑌̅ = = ∑ 𝑊ℎ 𝑌̅ℎ = 3.857
𝑁
ℎ=1
Now consider a sample of two families from stratum 1, two families from stratum 2, and four
families from stratum 3. Suppose that elements Y12 and Y13 from stratum 1, Y22 and Y25 from
stratum 2, and Y31, Y32, Y34 and Y36 from stratum 3, are selected with simple random sample in each
stratum.
a) What is the estimated mean number of individuals per family and its standard error?
b) Find the total estimated number of individuals along with its standard error.
c) Find a 95% confidence interval of 𝑌̅ and Y.
Solution: a)
Stratum 1 stratum 2 stratum 3
N1= 3 N2 = 5 N3 = 6
n1 =2 n2 = 2 n3 = 4
Y12 =3 Y22=6 Y31=2
Y13 =4 Y25 =8 Y32 =3
Y34= 2
Y35 = 3
yh: y1= 7 y2= 14 y 3 = 10
𝑦̅ℎ : 𝑦̅1 = 3.5 𝑦̅2 = 7.0 𝑦̅3 = 2.5
𝑠2ℎ : 2 2
𝑠 1 = 0.5 𝑠 2 = 2.0 2
𝑠 3 = 0.333
𝐿
∑ℎ=1 𝑁ℎ 𝑦̅ℎ 𝑁1 𝑦̅1 + 𝑁2 𝑦̅2 + 𝑁3 𝑦̅3 3 × 3.5 + 5 × 7 + 6 × 2.5
𝑦̅𝑠𝑡 = = = = 4.321
𝑁 𝑁 14
1 𝑠2 1 0.5 2
𝑣𝑎𝑟(𝑦̅𝑠𝑡 ) = 2 ∑ 𝑁ℎ (𝑁ℎ − 𝑛ℎ ) ℎ = 2 (3(3 − 2) × + 5 × (5 − 2) × + 6 × (6 − 4) ×
𝑁 𝑛ℎ 14 2 2
0.333 16.749
)= = 0.085454, 𝑎𝑛𝑑 𝑠. 𝑒(𝑦̅𝑠𝑡 ) = 0.2923.
4 196
b) 𝑌̂𝑠𝑡 = ∑3ℎ=1 𝑌̂ℎ= ∑3ℎ=1 𝑁ℎ 𝑦̅ℎ = 𝑁1 𝑦̅1 + 𝑁2 𝑦̅2 + 𝑁3 𝑦̅3
= 3 × 3.5 + 5 × 7 + 6 × 2.5 = 60.5 𝑜𝑟 𝑌̂𝑠𝑡 = 𝑁𝑦̅𝑠𝑡 = 14 × 4.321 = 60.5
s.e(𝑌̂𝑠𝑡 ) = N s.e(𝑦̅𝑠𝑡 ) = 14×0.2923 = 4.0922.
c) The 95% confidence interval for the mean and for the total:
𝑌̅ = 𝑦̅𝑠𝑡 ± 𝑍𝛼 𝑠. 𝑒(𝑦̅𝑠𝑡 )
2
𝑦̅𝑠𝑡 ± 𝑡𝛼 , (𝑛 − 1) ∗ 𝑠. 𝑒(𝑦̅𝑠𝑡 )4.321 ± 𝑡𝛼 , (𝑛 − 1) × 0.2923 = 4.321 ± 𝑡0.05 , 7 × 0.2923
2 2 2
= 4.321 ± 2.365 × 0.2923 = (3.6297, 5.0123)
𝑌 = 𝑌𝑠𝑡 ± 𝑡𝛼 , (𝑛 − 1) ∗ 𝑠. 𝑒(𝑌̂𝑠𝑡 ) = 𝑁𝑦̅𝑠𝑡 ± 𝑡𝛼 , (𝑛 − 1) ∗ 𝑁 𝑠. 𝑒(𝑦̅𝑠𝑡 ) = 60.5 ± 2.365 × 4.0922
̂
2 2
= (50.8219, 70.1781)
In each case the interval covers the true population value.
4.4. Allocating Samples to Strata
An important consideration in stratified sampling is the way in which the total sample size is
allocated to each of the strata. A good allocation is one where there is seasonable balanced

Sampling Theory | Chapter Four | Stratified Random Sampling Page 6

between precision and resources. The commonly used methods of sample allocation may include
equal allocation, proportional allocation and optimum allocation.
(a) Equal allocation: In this method the same numbers of elements are sampled from each
stratum. In other words, for each stratum h, the sample size is given by
𝑆 2
𝑛
∑𝐿ℎ=1 𝑁ℎ (𝑁ℎ 𝐿−𝑛) ℎ𝑛
̅𝒔𝒕 )𝑒𝑞𝑢𝑎𝑙 =
𝑛ℎ = 𝑛1 = 𝑛2 = ⋯ = , ℎ = 1,2, − − −, 𝐿 and 𝑉𝑎𝑟(𝒚 .
𝐿 𝑁2
This allocation is used when information about the variation within the strata is not available and
sizes of strata do not differ much from each other.
(b) Proportional Allocation: When the size of stratum, Nh, is the only available information and
there is difference in size between strata, a proportional allocation may be used. With
proportional allocation, the sample allocated to each stratum is proportional to the total
number of units in the stratum. That is, 𝑛ℎ ∝ 𝑁ℎ 𝑜𝑟 𝑛ℎ = 𝛼𝑁ℎ and from this relation we
obtain
𝐿 𝐿
𝑛
=> ∑ 𝑛ℎ = ∑ 𝛼𝑁ℎ 𝑜𝑟 𝑛 = 𝛼𝑁 𝑎𝑛𝑑 𝛼 =
𝑁
ℎ=1 ℎ=1
𝑛ℎ = 𝑁𝑛 𝑁ℎ = 𝑓𝑁ℎ = 𝑛𝑊ℎ 𝑜𝑟 𝑛𝑛ℎ = 𝑁𝑁ℎ 𝑜𝑟 𝑓ℎ = 𝑓 .
Example 1: In a population of 2400 unit’s men and 1600 women proportional allocation with a
10% sample would produce a sample of 240 men and 160 women.
This design yield a self-weighting sample, and the population mean can be estimated with the
𝒏
∑𝒏 𝒚 ∑𝑳 ∑ 𝒉 𝒚
simple mean of the sample cases. That is 𝒚̅𝒔𝒕 = 𝒚̅ = 𝒋=𝟏𝒏 𝒉 𝒐𝒓 𝒚 ̅𝒔𝒕 = 𝒉=𝟏 𝒏𝒊=𝟏 𝒉𝒊. Each sample
element is multiplied by the same constant 1/n and does not require the use of strata weight, and
this simplifies data processing.
Corollary: In proportional allocation, with nh =𝑁𝑛 𝑁ℎ , the variance of 𝒚
̅𝒔𝒕 reduces to
𝐿 𝐿
1−𝑓 2 1−𝑓 2 𝑁 𝑆ℎ 2 𝑁 − 𝑛
̅𝒔𝒕 ) =
𝑉𝑎𝑟(𝒚 ∑ 𝑊ℎ 𝑆ℎ = ∑ 𝑁ℎ 𝑆ℎ = ∑ ( )
𝑛 𝑛𝑁 𝑁ℎ 𝑛ℎ 𝑁
ℎ=1 ℎ=1
Corollary: If sampling is proportional and the variances in all strata have the same value, say
𝑆𝑤 2 𝑆𝑤 2 𝑁−𝑛
𝑆𝑤 2 , then the variance of 𝒚 ̅𝒔𝒕 ) =
̅𝒔𝒕 reduces to 𝑉𝑎𝑟(𝒚 (1 − 𝑓) = ( ).
𝑛 𝑛 𝑁
Example 2: Consider the number of general hospitals located in four strata, where stratum is
composed of one or more geographical regions. Suppose we wish to take a stratified random
sample of 51 hospitals from among the 255 hospitals in the total regions. The distribution of
hospitals by strata is 𝑁1 = 44, 𝑁2 = 116, 𝑁3 = 48, 𝑁4 = 47. Determine the sample size required
in each stratum using proportional allocation.
𝑛 𝑛 51×44 𝑛 51×116 𝑛 51×48
Solution: 𝑛ℎ = 𝑁 𝑁ℎ , 𝑛1 = 𝑁 𝑁1 = ≅ 9, 𝑛2 = 𝑁 𝑁2 = ≅ 23, 𝑛3 = 𝑁 𝑁3 = ≅
255 255 255
𝑛 51×47
10, 𝑛4 = 𝑁 𝑁4 = ≅ 9. Verify that 𝑓1 = 𝑓2 = 𝑓3 = 𝑓4 = 𝑓.
255
(c) Optimum Allocation (Disproportionate Sampling): It refers to the aim of assigning
sampling rates to the strata in such a way as to achieve the least variance for the overall mean
per unit of cost. In other words, it is the method that provides minimum variance for a given

Sampling Theory | Chapter Four | Stratified Random Sampling Page 7

 fixed total cost (C) or minimum cost for a fixed variance (V). The simplest cost function is
given by 𝐶 = 𝐶𝑜 + ∑𝐿ℎ=1 𝑐ℎ 𝑛ℎ , where ch is the cost of sampling an elementary unit in stratum
h, 𝐶𝑜 is the overhead cost. Let us first assume that the cost per unit varies from stratum to
stratum.
Theorem 4.5: In stratified random sampling with linear cost function 𝐶 = 𝐶𝑜 + ∑𝐿ℎ=1 𝑐ℎ 𝑛ℎ , the
variance of the estimated mean 𝒚 ̅𝒔𝒕 is a minimum for a specified cost C, and the cost is a
𝑊ℎ 𝑆ℎ
̅𝒔𝒕 ), when nh is proportional to
minimum for a specified variance, 𝑉𝑎𝑟(𝒚 , i.e.,
√ 𝑐ℎ
𝑊ℎ 𝑆ℎ ⁄√𝑐ℎ 𝑁ℎ 𝑆ℎ ⁄√𝑐ℎ
𝑛ℎ = ∑𝐿 × 𝑛 = ∑𝐿 × 𝑛.
ℎ=1 𝑊ℎ 𝑆ℎ /√𝑐ℎ ℎ=1 𝑁ℎ 𝑆ℎ /√𝑐ℎ

This shows that 𝑛ℎ is directly proportional to 𝑁ℎ and 𝑆ℎ , and inversely proportional to √𝑐ℎ . These
three factors should be considered in sample allocation. As a rule, in a given stratum, take a large
sample if the stratum is larger, the stratum is more variable internally, and sampling is cheaper in
the stratum.
The sample size 𝑛ℎ , is expressed in terms of n, but the value of n is unknown. It can be
determined by using the cost function or variances specification.
(𝐶−𝐶𝑜 )𝑁ℎ 𝑆ℎ ⁄√𝑐ℎ
Corollary: a) If the cost is fixed, the optimum n is given by 𝑛 = ∑ 𝑁ℎ 𝑆ℎ √𝑐ℎ
, from this we
(𝐶−𝐶𝑜 )𝑁ℎ 𝑆ℎ ⁄√𝑐ℎ
can calculate 𝑛ℎ as 𝑛ℎ = ∑ 𝑁ℎ 𝑆ℎ √𝑐ℎ
.
〈∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ √𝑐ℎ 〉(∑𝐿ℎ=1𝑊ℎ 𝑆ℎ ⁄√𝑐ℎ )
b) If the variance, V0, is fixed, the optimum n is: 𝑛ℎ = 1 .
𝑉𝑜 + ∑ 𝑊ℎ 𝑆ℎ 2
𝑁
Assume that the cost per unit is the same for all the strata, i.e., 𝑐ℎ = 𝑐, constant for each stratum.
The cost function reduces to 𝐶 = 𝐶𝑜 + 𝑐 × 𝑛, then the optimum allocation for fixed cost reduce
to optimum allocation for simple fixed sample size.
Theorem 4.6: In stratified random sampling the variance, 𝑉𝑎𝑟(𝑦̅𝑠𝑡 ), is minimized for a fixed
𝑊 𝑆ℎ 𝑁 𝑆ℎ
total sample size n if: 𝑛ℎ = 𝑛 ∑ 𝑊ℎ = 𝑛 ∑ 𝑁ℎ . This is known as Neyman allocation. The
ℎ 𝑆ℎ ℎ 𝑆ℎ
(∑ 𝑊ℎ 𝑆ℎ )2 ∑ 𝑊ℎ 𝑆ℎ 2
minimum variance will be given as 𝑉(𝑦̅𝑠𝑡 )𝑚𝑖𝑛 = −
𝑛 𝑁
Proof:

2. Suppose that a corporation has 260,000 accident reports available over a period of a time. A
sample survey is required for the purpose of estimating the average number of days of work lost
per accident. Of these reports, 150,000 are coded and 110,000 are un-coded. Before processing
the forms on computer directly, un-coded forms must first be coded. Approximately $10,000 is
available for selecting the sample and coding and processing the data. From experience it is

Sampling Theory | Chapter Four | Stratified Random Sampling Page 8

known that the cost of the sampling and processing sample forms is equal to $0.32 for coded
form and $0.98 for un-coded form. In addition, assume that the standard deviation of the
distribution of days lost from work is twice as large among un-coded reports as among coded
reports. Find the best way of allocating the sample elements among coded and un-coded forms.
Solution: let stratum 1 represents coded forms, and stratum 2 represents un-coded forms.
𝑁1 = 150,000, 𝑁2 = 110,000, 𝐶 = $10,000, 𝑐1 = $0.32, 𝑐2 = $0.98, 𝑆2 = 2𝑆1
Assume that the overhead cost 𝑐0 = 0.
(𝐶−𝐶𝑜 )𝑁ℎ 𝑆ℎ ⁄√𝑐ℎ (𝐶−𝐶𝑜 )𝑁1 𝑆1 ⁄√𝑐1 𝑁1 ⁄√𝑐1
𝑛ℎ = ∑ 𝑁ℎ 𝑆ℎ √𝑐ℎ
, 𝑛1 = ∑ 𝑁ℎ 𝑆ℎ √𝑐ℎ
=𝑁 ×𝐶
1 √𝑐1 +2𝑁2 √𝑐2

150,000⁄√0.32 265,165.0429
𝑛1 = × 10,000 = × 10,000 ≅ 8762.
150,000√0.32 + 2 × 110,000√0.98 302,641.7023
(𝐶 − 𝐶𝑜 ) ∑ 𝑁2 𝑆2 ⁄√𝑐2 2𝑁2 ⁄√𝑐2
𝑛2 = = ×𝐶
∑ 𝑁ℎ 𝑆ℎ √𝑐ℎ 𝑁1 √𝑐1 + 2𝑁2 √𝑐2
2 × 110,000⁄√0.98 222,233.5598
𝑛2 = × 10,000 = × 10,000 ≅ 7343
150,000√0.32 + 2 × 110,000√0.98 302,641.7023
Therefore, a sample of 8762 coded reports and 7343 un-coded reports could be taken.
d) Allocation Requiring More Than 100% Sampling :
In optimal allocation the sample size 𝑛ℎ may be greater than the total number of elements in
stratum, 𝑁ℎ, i.e., 𝑛ℎ > 𝑁ℎ . When this occurs set 𝑛ℎ = 𝑁ℎ for each stratum having optimal
allocation greater than 𝑁ℎ and reallocate remaining sample to other strata.
𝑁ℎ 𝑆ℎ
If 𝑛1 > 𝑁1 , set 𝑛1 = 𝑁1 and for the remaining strata calculate 𝑛̃ℎ = ∑𝐿 × (𝑛 − 𝑁1 ),
ℎ=2 𝑁ℎ 𝑆ℎ
provided that 𝑛̃ℎ ≤ 𝑁ℎ for ℎ ≥ 2.
Similarly, if 𝑛1 > 𝑁1 and 𝑛2 > 𝑁2 , then set 𝑛1 = 𝑁1 and 𝑛2 = 𝑁2 . For the remaining strata
𝑁ℎ 𝑆ℎ
calculate 𝑛̃ℎ = ∑𝐿 × (𝑛 − 𝑁1 − 𝑁2 ), provided that 𝑛̃ℎ ≤ 𝑁ℎ for ℎ ≥ 3.
ℎ=3 𝑁ℎ 𝑆ℎ
Example: Consider the following data from three strata. Stratum 1, 𝑁1 = 100, 𝑆1 = 50, stratum
2, 𝑁2 = 110, 𝑆2 = 10, stratum 3, 𝑁3 = 120, 𝑆3 = 5. Allocate a total sample of 140 elements to
each stratum by using optimum allocation.
𝑁ℎ 𝑆ℎ 𝑁1 𝑆1 100×50
Solution: 𝑛ℎ = ∑𝐿 × 𝑛, 𝑛1 = ∑𝐿 × 𝑛 ⟹ 𝑛1 = 100×50+110×10+120×5
ℎ=3 𝑁ℎ 𝑆ℎ ℎ=3 𝑁ℎ 𝑆ℎ
5000×140 110×10×140 120×5×140
⟹ 𝑛1 = = 104.477 ≈ 104, 𝑛2 = = 23, 𝑛3 = = 13.
6700 6700 6700
Since 𝑛1 > 𝑁1 , we allocate 𝑛1 = 𝑁1 = 100. For the remaining two strata we calculate as
follows.
𝑁2 𝑆2 110 × 10 × (140 − 100) 44000
𝑛̃2 = 𝐿 × (𝑛 − 𝑁1 ) ⟹ 𝑛2 = = = 26
∑ℎ=2 𝑁ℎ 𝑆ℎ 110 × 10 + 120 × 5 1700
𝑁3 𝑆3 120 × 5 × (140 − 100) 24000
𝑛̃3 = 𝐿 × (𝑛 − 𝑁1 ) ⟹ 𝑛3 = = = 14
∑ℎ=2 𝑁ℎ 𝑆ℎ 110 × 10 + 120 × 5 1700

Sampling Theory | Chapter Four | Stratified Random Sampling Page 9

 4.5. Relative Precision of Stratified Random sampling and SRS(Reading Assignment)
If intelligibly used, stratification nearly always results in smaller variance for the estimated mean
or total than is given by a comparable simple random sampling. For the estimated mean or total,
stratification almost always results in smaller variance, i.e., 𝑉𝑎𝑟(𝑦̅𝑠𝑡 ) ≤ 𝑉𝑎𝑟(𝑦̅). Consider the
following variances.
𝑆2 1−𝑓
𝑉𝑆𝑅𝑆 = 𝑉𝑎𝑟(𝑦̅) = (1 − 𝑓),𝑉𝑝𝑟𝑜𝑝 = 𝑉𝑎𝑟(𝑦̅𝑠𝑡 )𝑝𝑟𝑜𝑝 = ∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ 2, 𝑉𝑜𝑝𝑡 = 𝑉𝑎𝑟(𝑦̅𝑠𝑡 )𝑜𝑝𝑡 =
𝑛 𝑛
(∑ 𝑊ℎ 𝑆ℎ )2 ∑ 𝑊ℎ 𝑆ℎ 2
− , for fixed n.
𝑛 𝑁
Theorem 4.7: If terms 1⁄𝑁ℎ are ignored relative to unity, 𝑉𝑜𝑝𝑡 ≤ 𝑉𝑝𝑟𝑜𝑝 ≤ 𝑉𝑆𝑅𝑆 where the
optimum allocation is for fixed n.
Where
𝑉𝑆𝑅𝑆 is variance of estimated mean in SRS
𝑉𝑝𝑟𝑜𝑝 is variance of estimated mean in stratified random sampling with proportional allocation
𝑉𝑜𝑝𝑡 is variance of estimated mean in stratified random sampling with optimum allocation for
fixed sample size n.
Proof: we have

From the standard algebraic identity for the analysis of variance of stratified population, we have

Sampling Theory | Chapter Four | Stratified Random Sampling Page 10

If terms of order 1⁄𝑁ℎ (and hence 1⁄𝑁) are negligible as compared to unity, then

4.6. Relative Gain in Precision (Reading Assignment)

i) Gain due to proportional allocation as compared to SRS:
𝑉𝑆𝑅𝑆 − 𝑉𝑝𝑟𝑜𝑝 1−𝑓 ∑ 𝑊ℎ (𝑌̅ℎ −𝑌̅)2
Relative Gain = =
𝑉𝑝𝑟𝑜𝑝 𝑛 𝑉𝑝𝑟𝑜𝑝
The greater the heterogeneity among stratum means, the higher the relative gain.
ii) Gain from optimum allocation as compared to proportional allocation:
𝑉𝑝𝑟𝑜𝑝 − 𝑉𝑜𝑝𝑡 1 ∑ 𝑊ℎ (𝑆ℎ −𝑆̅)2
Relative Gain = =𝑛 , where 𝑆̅ = ∑ 𝑊ℎ 𝑆ℎ
𝑉𝑜𝑝𝑡 𝑉𝑜𝑝𝑡
The larger the difference among stratum standard deviation, the higher is the gain.
iii) Gain from optimum allocation as compared to SRSR:
𝑉𝑆𝑅𝑆 − 𝑉𝑜𝑝𝑡 1−𝑓 ∑ 𝑊ℎ (𝑌̅ℎ −𝑌̅)2 1 ∑ 𝑊ℎ (𝑆ℎ −𝑆̅)2
Relative Gain = = +𝑛
𝑉𝑜𝑝𝑡 𝑛 𝑉𝑜𝑝𝑡 𝑉𝑜𝑝𝑡
If the difference among the stratum means and stratum standard deviations are great, then the
gain will be high.
4.7. Stratified sampling for proportion
If the characteristic under study is qualitative in nature, then its values will fall into one of the two
mutually exclusive complementary classes C and C’. Ideally, only two strata are needed in which all the
units can be divided depending on whether they belong to C or its complement C’.
In stratification, for any element in the population or in the sample, we define an observation Yhi
1, 𝑖𝑓 𝑡ℎ𝑒 𝑒𝑙𝑒𝑚𝑒𝑛𝑡 𝑖𝑠 𝑖𝑛 𝑐𝑙𝑎𝑠𝑠 𝐶 𝑓𝑜𝑟 𝑠𝑡𝑟𝑎𝑡𝑢𝑚 ℎ
as 𝑌ℎ𝑖 = {
0, 𝑖𝑓 𝑡ℎ𝑒 𝑒𝑙𝑒𝑚𝑒𝑛𝑡 𝑖𝑠 𝑛𝑜𝑡 𝑖 𝐶,
where C represent the class of interest.
a) Notation:
𝑁ℎ
𝐴ℎ = ∑𝑖=1 𝑌ℎ𝑖 = 𝑌ℎ , the number of elements in the hth stratum that belongs to C.
𝐴
𝑃ℎ = 𝑁ℎ = 𝑌̅ℎ , the proportion of units in the hth stratum that belongs to C.
ℎ
𝑁 ℎ 𝑃ℎ 𝑄 ℎ
𝑄ℎ = 1 − 𝑃ℎ , 𝑆ℎ 2 = , the population variance in stratum h,
𝑁ℎ −1

Sampling Theory | Chapter Four | Stratified Random Sampling Page 11

 𝐿 𝐿 𝑁ℎ 𝐿 𝐿

𝐴 = ∑ 𝐴ℎ = ∑ ∑ 𝑌ℎ𝑖 = ∑ 𝑌ℎ = 𝑌, 𝑃 = 𝐴⁄𝑁 = ∑ 𝑊ℎ 𝑃ℎ
ℎ=1 ℎ=1 𝑖=1 ℎ=1 ℎ=1
𝑛ℎ th
𝑎ℎ = ∑𝑖=1 𝑦ℎ𝑖 = the number of sample units in the h stratum.
𝑝ℎ = 𝑎ℎ ⁄𝑛ℎ = 𝑦̅ℎ , the sample proportion in hth stratum.
𝑛 𝑝ℎ 𝑞 ℎ
𝑞ℎ = 1 − 𝑝ℎ , 𝑠ℎ 2 = 𝑛ℎ −1 = sample variance in stratum h.
ℎ
b) Properties of Estimates
For the proportion in the whole population, the estimate appropriate to stratified random
sampling is

For stratified random sampling: 𝑝ℎ is an unbiased estimate of 𝑃ℎ , i.e., 𝐸(𝑝ℎ ) = 𝑃ℎ

𝑉𝑎𝑟(𝑝ℎ ) = 𝑃ℎ𝑛𝑄ℎ (𝑁𝑁ℎ−𝑛ℎ ), for stratum h. The over all estimate of the population proportion P is
−1
ℎ ℎ
∑ 𝐿
= ℎ=1𝑁 ℎ ℎ = 𝑦̅𝑠𝑡 and the estimate of the population total 𝐴̂ is 𝐴̂𝑠𝑡 = 𝑁𝑝𝑠𝑡 =
𝑁 𝑝
𝑝𝑠𝑡 = ∑𝐿ℎ=1 𝑊ℎ 𝑝ℎ
∑𝐿ℎ=1 𝑁ℎ 𝑝ℎ
Theorem 4.8: With stratified random sampling, the variance of 𝑝𝑠𝑡 is given by
1 ∑ 𝑁ℎ 2 (𝑁ℎ −𝑛ℎ ) 𝑃ℎ 𝑄ℎ 𝑁ℎ −𝑛ℎ
𝑉𝑎𝑟(𝑝𝑠𝑡 ) = ∑𝐿ℎ=1 𝑊ℎ 2 (𝑁𝑁ℎ−𝑛ℎ ) 𝑃 ℎ 𝑄ℎ =
, where 𝑆ℎ 2 = 𝑃ℎ 𝑄ℎ
−1 𝑛 ℎ𝑁2 ℎ 𝑁ℎ −1 𝑛ℎ 𝑁ℎ −1
𝑉𝑎𝑟(𝑝𝑠𝑡 ) = ∑𝐿ℎ=1 𝑊ℎ 2 𝑃 ℎ 𝑄ℎ
𝑛ℎ
(1 − 𝑓ℎ ).
Corollary: If fpc is ignored, 𝑉𝑎𝑟(𝑝𝑠𝑡 ) = ∑𝐿ℎ=1 𝑊ℎ 2 𝑃ℎ𝑛𝑄ℎ .
ℎ

Since the population proportion 𝑃ℎ is unknown, we use the population total, 𝐴̂𝑠𝑡 = 𝑁𝑝𝑠𝑡 =
∑ 𝑁ℎ 𝑝ℎ , the variance is given by 𝑉𝑎𝑟(𝐴̂𝑠𝑡 ) = 𝑁 2 𝑉𝑎𝑟(𝑝𝑠𝑡 )
c) Sample Allocation
With proportional allocation, 𝑛ℎ = 𝑛𝑁𝑁ℎ, and the variance of 𝑝𝑠𝑡 is

1−𝑓 1
= ∑𝐿ℎ=1 𝑊ℎ 𝑃ℎ 𝑄ℎ = ∑𝐿ℎ=1 𝑊ℎ 𝑃ℎ 𝑄ℎ (1 − 𝑓), if 𝑁ℎ is large.
𝑛 𝑛
𝑝 𝑞 𝑃 𝑄
For the sample estimate of the variance, substitute 𝑛 ℎ−1ℎ for the unknown 𝑁ℎ −1ℎ in any of the
ℎ ℎ
formulas above.
𝑁ℎ
With optimum allocation: For fixed total sample size n, 𝑛ℎ ∝ 𝑁ℎ √𝑁 𝑃ℎ 𝑄ℎ ≅ 𝑁ℎ √𝑃ℎ 𝑄ℎ
ℎ −1

𝑁ℎ √𝑁ℎ 𝑃ℎ 𝑄ℎ ⁄𝑁ℎ −1 𝑁ℎ √𝑃ℎ 𝑄ℎ

𝑛ℎ = ∑𝑁𝑁ℎ 𝑆𝑆ℎ × 𝑛 = ∑ × 𝑛 ⇒ 𝑛ℎ ≐ ∑𝐿ℎ=1 𝑁ℎ √𝑃ℎ 𝑄ℎ
× 𝑛, if 1⁄𝑁ℎ is negligible. In
ℎ ℎ 𝑁ℎ √𝑁ℎ 𝑃ℎ 𝑄ℎ ⁄𝑁ℎ −1
this case the optimum variance will be

Sampling Theory | Chapter Four | Stratified Random Sampling Page 12

 〈∑𝐿ℎ=1 𝑊ℎ √𝑁ℎ 𝑃ℎ 𝑄ℎ ⁄𝑁ℎ −1〉2 2 〈∑𝐿ℎ=1 𝑊ℎ √𝑃ℎ 𝑄ℎ 〉2
𝑉𝑎𝑟(𝑝𝑠𝑡 )𝑜𝑝𝑡 = − ∑ 𝑊 𝑁ℎ 𝑃−1
ℎ 𝑄ℎ
≐ − ∑ 𝑊ℎ 𝑃𝑁ℎ 𝑄ℎ , if 𝑁ℎ is large.
𝑛 ℎ 𝑛
〈∑𝐿ℎ=1 𝑊ℎ √𝑃ℎ 𝑄ℎ 〉2
If fpc is ignored, then it reduces to 𝑉𝑎𝑟(𝑝𝑠𝑡 )𝑜𝑝𝑡 = .
𝑛
𝑁
𝑁ℎ √𝑁 ℎ 𝑃ℎ 𝑄ℎ ⁄𝐶ℎ
ℎ −1 𝑁ℎ √𝑃ℎ 𝑄ℎ ⁄𝐶ℎ
For fixed cost, C, 𝑛ℎ = ×𝑛 = ∑ × 𝑛, if 𝑁ℎ is large.
𝑁
∑ 𝑁ℎ √𝑁 ℎ 𝑃ℎ 𝑄ℎ ⁄𝐶ℎ 𝑁ℎ √𝑃ℎ 𝑄ℎ ⁄𝐶ℎ
ℎ −1

e) Relative precision
𝑉𝑎𝑟(𝑝𝑠𝑡 )𝑜𝑝𝑡 ≤ 𝑉𝑎𝑟(𝑝𝑠𝑡 )𝑝𝑟𝑜𝑝 ≤ 𝑉𝑎𝑟(𝑝)𝑆𝑅𝑆
f) Confidence Limit:
A 100(1-𝛼)% confidence interval for population proportion P is
𝑃 = 𝑝𝑠𝑡 ± 𝑍𝜶 𝑠. 𝑒(𝑝𝑠𝑡 )
2
Example: A chain of department stores is interested in estimating the proportion of accounts
receivable that are delinquent. The chain consists of four stores. To reduce the cost of sampling,
stratified random sampling is used with each store as a stratum. Since no information on
population proportions is available before sampling, proportional allocation is used. From the
table given below, estimate P, the proportion of delinquent accounts for the chain, find its
sampling error and 95% CI for population proportion.
Stratum I II III IV
Number of accounts receivable N1 = 65 N2 = 42 N3 = 93 N4 = 25
Sample size n1 = 14 n2 = 9 n3 = 21 n4 = 6
Sample Proportion of delinquent accounts
Solution: 𝑁 = 𝑁1 + 𝑁2 + 𝑁3 + 𝑁4 = 65 + 42 + 93 + 25 = 225
𝑛 = 𝑛1 + 𝑛2 + 𝑛3 + 𝑛4 = 14 + 9 + 21 + 6 = 50
𝑁ℎ 𝑝ℎ 65 42 93 25
𝑝𝑠𝑡 = ∑𝐿ℎ=1 𝑊ℎ 𝑝ℎ = ∑𝐿ℎ=1 = 225 ∗ 0.3 + 225 ∗ 0.2 + 225 ∗ 0.4 + 225 ∗ 0.1 = 0.3
𝑁
𝐿 𝐿
𝑝ℎ 𝑞ℎ 𝑁ℎ − 𝑛ℎ
2 𝑁ℎ 2 𝑝ℎ 𝑞ℎ
𝑉𝑎𝑟(𝑝𝑠𝑡 ) = ∑ 𝑊ℎ ( )=∑ 2 (1 − 𝑓ℎ )
𝑛ℎ − 1 𝑁ℎ − 1 𝑁 𝑛ℎ − 1
ℎ=1 ℎ=1
2
65 0.3 ∗ 0.7 42 2 0.2 ∗ 0.8 93 2
=( ) ∗( ) ∗ 0.78 + ( ) ∗( ) ∗ 0.79 + ( )
225 13 225 8 225
0.4 ∗ 0.6 25 2 0.1 ∗ 0.9
∗( ) ∗ 0.77 + ( ) ∗( ) ∗ 0.76
20 225 5
65 2 42 2 93 2 25 2
=( ) ∗ 0.0126 + ( ) ∗ 0.0158 + ( ) ∗ 0.009 + ( ) ∗ 0.0137
225 225 225 225
= 0.001 + 0.006 + 0.0015 + 0.0002 = 0.00326
𝑆. 𝑒(𝑝𝑠𝑡 ) = √𝑉𝑎𝑟(𝑝𝑠𝑡 ) = √0.00326 = 0.057
95% confidence interval for the population proportion is:
𝑃 = 𝑝𝑠𝑡 ± 𝑍𝜶 𝑠. 𝑒(𝑝𝑠𝑡 ) = 0.3 ± 1.96 ∗ 0.057 = (0.3 − 0.112,0.3 + 0.112)
2
= [0.188,0.412]

Sampling Theory | Chapter Four | Stratified Random Sampling Page 13

We are 95% sure that the proportional of delinquent accounts for the chain is between 0.188 and
0.412.
4.8. Sample Size Estimation in Stratification
A. Sample Size in Estimation of the population Mean 𝒀 ̅:
consider any allocation (proportional or optimum). It is assumed that the estimate has variance
V. Suppose the margin error d has been specified.
2
𝑑
𝑑 = 𝑍𝛼⁄2 √𝑉 => 𝑉 = ( )
𝑍𝛼⁄2
i) General Case: The population standard deviation, 𝑆ℎ , is estimated by 𝑠ℎ . Let 𝜋ℎ = 𝑛ℎ ⁄𝑛 ⇒
𝑛ℎ = 𝑛𝜋ℎ
𝐿 𝐿 𝐿 𝐿 𝐿
𝑆 2
𝑆 2
𝑆 2 1 𝑆 2 1
𝑉 = 𝑉(𝑦̅𝑠𝑡 ) = ∑ 𝑊ℎ 2 𝑛ℎ (1 − 𝑓ℎ ) = ∑ 𝑊ℎ 2 𝑛ℎ − ∑ 𝑊ℎ 2 𝑁ℎ = ∑ 𝑊ℎ 2 𝜋ℎ − ∑ 𝑊ℎ 𝑆ℎ 2
ℎ ℎ ℎ 𝑛 ℎ 𝑁
ℎ=1 ℎ=1 ℎ=1 ℎ=1 ℎ=1
𝑁ℎ
For a specified variance V, a general formula for n is given as, where 𝑊ℎ = 𝑁
𝑆 2
∑𝐿ℎ=1 𝑊ℎ 2 𝜋ℎ 𝑛𝑜 1 2
𝑛= ∑𝐿
ℎ
𝑊ℎ 𝑆ℎ 2
⇒ 𝑛= ∑𝐿 𝑊 𝑆 2
, where 𝑛𝑜 = 𝑉 ∑𝐿ℎ=1 𝑊ℎ 2 𝑆𝜋ℎℎ used as a first approximation
𝑉+ ℎ=1 1+ ℎ=1 ℎ ℎ
𝑁 𝑉𝑁

by ignoring fpc. Instead of V we can use an equivalent value 𝑑 2 ⁄𝑍𝛼⁄2 2 = (𝜀𝑌̅)2 ⁄𝑍𝛼⁄2 2 .
ii) In proportional allocation the sample size n is obtained as follows.
∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ 2 ∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ 2 ⁄𝑉 𝑛 ∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ 2 𝑍 2 ∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ 2
𝑛= ∑𝐿 𝑊 𝑆 2
= ∑𝐿 𝑊 𝑆 2
in which 𝑛 = 1+𝑜𝑛𝑜 , where 𝑛𝑜 = = , and
𝑉 𝜀 2 𝑌̅ 2
𝑉+ ℎ=1 ℎ ℎ 1+ ℎ=1 ℎ ℎ 𝑁
𝑁 𝑁𝑉
can be used as a first approximation for sample size and V is the defined variance, i.e.,
̅2
𝑉 = 𝑉𝑎𝑟(𝑦̅𝑠𝑡 ) = 𝑑 2 ⁄𝑍𝛼⁄ 2 = (𝜀𝑌 2) .
2 𝑍𝛼⁄
2
iii) In optimum allocation the sample size calculation is as follows.
2 2 2
(∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ ) (∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ ) ⁄𝑉 𝑍𝛼⁄ 2 (∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ ) ⁄𝜀2 𝑌̅ 2
 For fixed n, it is given as 𝑛 = ∑𝐿 𝑊 𝑆 2
= ∑𝐿 𝑊 𝑆 2
𝑜𝑟 𝑛 = 2
𝑍𝛼⁄ 2 ∑𝐿 2
𝑉+ ℎ=1 ℎ ℎ 1+ ℎ=1 ℎ ℎ 2 ℎ=1 𝑊ℎ 𝑆ℎ
𝑁 𝑁𝑉 1+
𝑁𝜀2 𝑌
̅2
(𝐶−𝐶𝑜 ) ∑𝑁ℎ 𝑆ℎ ⁄√𝑐ℎ
 For fixed cost, 𝐶 = 𝐶𝑜 + ∑ 𝑐ℎ 𝑛ℎ , the sample size is given as: 𝑛 = ∑ 𝑁ℎ 𝑆ℎ √𝑐ℎ
(∑ 𝑊ℎ 𝑆ℎ √𝐶ℎ )(𝑊ℎ 𝑆ℎ ⁄√𝐶ℎ )
 For fixed variance, V, Optimum n is: 𝑛 = 1
𝑉+ ∑ 𝑊ℎ 𝑆ℎ 2
𝑁
B. Estimation of population total
If V is a desired 𝑉𝑎𝑟(𝑌̂𝑠𝑡 ) the principal formulas are:
𝑆 2 𝑆 2
∑𝐿ℎ=1 𝑊ℎ 2 𝜋ℎ ∑𝐿ℎ=1 𝑁ℎ 2 𝜋ℎ
ℎ ℎ
General: 𝑛 = ∑𝐿 2 = 𝑛 = 𝑉+∑𝐿 2
𝑉+ ℎ=1 𝑊ℎ 𝑆ℎ ℎ=1 𝑁ℎ 𝑆ℎ
𝑁
𝑁 𝐿
∑𝐿ℎ=1 𝑊ℎ 𝑆ℎ 2 ∑ 𝑁 𝑆 2 𝑛0 𝑁
𝑉 ℎ=1 ℎ ℎ
Proportional: 𝑛 = ∑𝐿 𝑊 𝑆 2
= 𝑁 𝐿 = 𝑛 , 𝑤ℎ𝑒𝑟𝑒 𝑛0 = 𝑉 ∑𝐿ℎ=1 𝑁ℎ 𝑆ℎ 2
∑ 𝑁 ℎ 𝑆ℎ 2 1+ 0
𝑉+ ℎ=1 ℎ ℎ 1+ 𝑉 ℎ=1 𝑁
𝑁 𝑁

Sampling Theory | Chapter Four | Stratified Random Sampling Page 14

 2
ℎ=1 (∑𝐿 𝑁 ℎ 𝑆ℎ )
Optimum: = 𝑉+∑𝐿 2
ℎ=1 𝑁ℎ 𝑆ℎ
Example:
The following is data on sample of colleges and universities drawn in 1946 by the office of
education to estimate enrollments for the 1946-47 academic year. The illustration is for the
population of N=196 teachers’ colleges and normal students. These were divided into six strata.
Estimates of were computed from results for the 1943-1944 academic year. An “optimum”
stratification based on these was employed. The objective was a coefficient of variation of 5% in
the estimated total enrollment. In 1943 the total enrollment for these group of colleges was
Y=56,472.
Thus the desired standard error is 𝑠. 𝑒(𝑌̂) = 𝐶𝑉𝑌̂ 𝑌 = 0.05 ∗ 56,472 = 2824
2
the desired variance is 𝑉𝑎𝑟(𝑌̂) = (𝑠. 𝑒(𝑌̂)) = (2824)2 = 7,974,976
The following table shows values 𝑁ℎ ,𝑠ℎ , and 𝑁ℎ 𝑠ℎ which were known before determining .
Stratum 𝑁ℎ 𝑠ℎ 𝑁ℎ 𝑠ℎ 𝑛ℎ
1 13 325 4,225 9
2 18 190 3,420 7
3 26 189 4,914 10
4 42 82 3,444 7
5 73 86 6,278 13
6 24 190 4,560 10
Total 196 26,841 56
With only 196 units in this population, it is improbable that the fpc will be negligible.
B. Sample size in Estimation of the Population Proportion P:
i) For proportional allocation: the overall sample can be obtained by
∑𝐿ℎ=1 𝑊ℎ 𝑃ℎ 𝑄ℎ ∑𝐿ℎ=1 𝑊ℎ 𝑃ℎ 𝑄ℎ ⁄𝑉 1−𝑓
𝑛= 1 = 𝐿 , where 𝑉 ≐ ∑ 𝑊ℎ 𝑃ℎ 𝑄ℎ , for large 𝑁ℎ .
𝑉+ ∑𝐿ℎ=1 𝑊ℎ 𝑃ℎ 𝑄ℎ 1∑ 𝑊 𝑃 𝑄
1+ ℎ=1 𝑉ℎ ℎ ℎ
𝑛
𝑁
𝑁

𝑛 ∑ 𝑊ℎ 𝑃ℎ 𝑄ℎ 𝑍𝛼⁄ 2 ∑ 𝑊ℎ 𝑃ℎ 𝑄ℎ
𝑛 = 1+𝑜𝑛𝑜, where 𝑛𝑜 = or 𝑛𝑜 = 2
, where 𝑉 = 𝑑 2 ⁄𝑍𝛼⁄2 2 and 𝑑 = 𝜀𝑃.
𝑁
𝑉 𝜀2 𝑃2
ii) For optimum allocation: for fixed n, the variance can be expressed as
2
(∑ 𝑊ℎ √𝑃ℎ 𝑄ℎ ) ∑ 𝑊ℎ 𝑃ℎ 𝑄ℎ
𝑉= − , for large 𝑁ℎ . From this we obtain the sample size
𝑛 𝑁
2 2
(∑ 𝑊ℎ √𝑃ℎ 𝑄ℎ ) (∑ 𝑊ℎ √𝑃ℎ 𝑄ℎ ) ⁄𝑉
𝑛= =
𝑉 + ∑ 𝑊ℎ𝑁𝑃ℎ 𝑄ℎ 1 + ∑ 𝑊𝑉𝑁
ℎ 𝑃ℎ 𝑄 ℎ

In all cases, if the stratum population variance is unknown, substitute their respective estimates.
That is, substitute 𝑠ℎ 2 for 𝑆ℎ 2 and 𝑝ℎ for 𝑃ℎ .

4.9. Design Effect (Deff) (Reading Assignment)

The design effect is the ratio of the variance of an estimator obtained from the more complex
sample or of a particular design to the variance of the estimate obtained from simple random
sample of the same number of units.

Sampling Theory | Chapter Four | Stratified Random Sampling Page 15

 𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒 𝑓𝑜𝑟 𝑐𝑜𝑚𝑝𝑙𝑒𝑥 𝑠𝑢𝑟𝑣𝑒𝑦 𝑉(𝑦̅𝑐 )
For the main: 𝐷𝑒𝑓𝑓 = = 𝑉(𝑦̅
𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒 𝑓𝑜𝑟 𝑆𝑅𝑆 𝑆𝑅𝑆 )
Where 𝑉(𝑦̅𝑐 ) and 𝑉(𝑦̅𝑆𝑅𝑆 ) is the variance from complex design and the variance obtained from
simple random sample, respectively.
𝑉(𝑦̅𝑐 ) 1−𝑓
𝐷𝑒𝑓𝑓 = 𝑉(𝑦̅ in which 𝑉(𝑦̅𝑐 ) = 𝐷𝑒𝑓𝑓 × 𝑉(𝑦̅𝑆𝑅𝑆 ) = 𝐷𝑒𝑓𝑓 × 𝑆 2 . From this we get,
𝑆𝑅𝑆 ) 𝑛
1−𝑓
𝑛 = 𝐷𝑒𝑓𝑓 × 𝑉(𝑦̅ ) 𝑆 2 , if we know 𝑉(𝑦̅𝑐 ) we can determine the sample size. Therefore, the design
𝑐
effect has two main uses.
 To determine the sample size
 To evaluate more complex designs.
Where there is a problem of obtained the 𝑉(𝑦̅𝑐 ), it is possible to guess fairly will be both the
element variances S2and the Deff, from experience with similar past data. The sample variance s2
computed from large samples yield a good approximation of S2

Sampling Theory | Chapter Four | Stratified Random Sampling Page 16