Chapter 3

Preliminaries

This section contains the basic theories to provide a good sampling method. These

are based on the books of Cochran (1977) and Lohr (2010). Sampling has not always been

accepted by researcher at the start. Complete enumeration has been favored long time ago.

Although this is no longer perceived now a days, it is good to point out the advantages

of sampling compared with complete enumeration. The advantages are the following (1)

Reduced Cost (2) Greater Speed (3) Greater Scope. The types of information that can

be gathered through surveys that use sampling are more varied and flexible. (4) Greater

Accuracy. Because the volume of labor is reduced, higher quality personnel can be hired,

given intensive training, more careful field supervision of the job, and result processing are

possible.

3.1 Steps in a Design of Sample Survey

The following list are the steps in design of sample survey. The degree of complexity

in surveys varies substantially.The 11 survey main steps are not strictly arranged and are

not limited to it.

10
 1. Objectives of the Study

2. Population to be Sampled

3. Data to be Collected

4. Degree of Precision Desired

5. Methods of Measurement

6. The Frame

7. Selection of the Sample

8. The Pretest

9. Organization of the Field Work

10. Summary and Analysis of the Data

11. Information Gained for Future Surveys

Some of the concepts that worth elaborating are the population to be sampled, data to

be collected, and degree of precision desired.

Definition 1. Population The group from which the sample is drawn is referred to as the

population or in inferential Statistics the target population.

The population should be clearly defined. The enumerator must be able to quickly

determine in the field whether a questionable case is representative of the population or not.

Any additional information that can be acquired about the nature of the differences between

the sampled and target populations may be helpful. The population must be separated into

groups known as sampling units, or Primary Sampling Units (PSU), before the sample

is chosen. These units must encompass the whole population and cannot overlap in any way

such that every component of the population is a member of one and only one unit. In

survey sampling it is important to have a list of members of the target population.

Definition 2. Sampling Frame. A list, map, or other specification of sampling units in the

11
population from which a sample may be selected [1]

Examples of sampling frame are list of every home phone number in the city, list of street

addresses. In agricultural survey, sampling frame may be a list of all farms or a map of the

areas with farms.

Ideally, the target population should coincide with the sampling frame. But this is rare.

For example, the target population is employees in the Philipppines. The sampling frame

excludes employees in institutional places such as prison, dormitories, and home for the

elderly.

For data to be collected, the level of the responses is lowered by a lengthy questionnaire.

For the precision desired, because just a portion of the population has been measured and due

to measurement mistakes, the results of sample surveys are always subject to some degree

of uncertainty.A crucial stage is defining the level of precision desired in the outcomes. The

person who will use the data is responsible for this stage. Since many administrators are not

used to thinking in terms of the amount of inaccuracy that can be tolerated in estimations

while still making wise judgments, this could provide some challenges. At this point, the

statistician can frequently assist.

This subsection is a quick overview of how to design a sample survey. The focus of this

study is one part of the survey design which is the statistical theory that is introduced in

the next subsection, ”Role of Sampling Theory.”

3.2 Role of Sampling Theory

Several types of skills are needed to develop a survey- population definition, determi-

nation of data to be collected, methods of measurement, and the organization of the field

work. Sampling theory has a small part. However, poor work in one phase affects everything

12
else.

Two important factors are prioritize in sampling theory the (1) efficiency at a (2) lowest

possible cost. Sample selection is develop to provide estimates that are precise enough for a

minimum cost possible. The expected precision and cost should be predicted. This is better

understood in the succeeding sections.

3.3 Properties of Probability Sampling

The sampling theories that are discussed in this study are based on probability sam-

pling. The term probability sampling has the following mathematical properties:

1. From the sampling frame the derivation of distinct sets of samples are defined,

S1 , S2 , · · · , Sv . For example, a population of six units numbered as 1 to 6, a sample of

size 2 gives possible set of samples- S1 ∼ (1, 4); S2 ∼ (2, 5); S3 ∼ (3, 6). Note that not all

possible samples of size 2 need be included.

2. Each possible sample Si has assigned to it a known probability of selection, Si .

πi = P (unit i in the Sample)

3. The method for computing the estimate from the sample must be stated and must

lead to a unique estimate for any specific sample. For example, the estimate for a set of

sample is the average of units in the sample.

4. The frequency distribution of the estimates can be generated if sampling procedure

is repeatedly applied to the same population. That is to know how frequent any particular

sample Si will be selected, and its estimate.

13
3.4 Nonprobability Sampling

The term probability sampling is better understood by its opposite, nonprobability

sampling. Here are common types of nonprobability sampling:

1. Only a portion of the population that is easily available is included in the sample.

The top 6 to 9 inches of coal in an open wagon can be sampled.

2. The sample is conveniently chosen. The investigator may choose 10 rabbits at random

from a huge cage in the lab by just resting his hands on them.

3. With a small but diverse population, the sampler examines the entire population and

chooses a small sample of ”typical” units—that is, units that are reasonably near to his

perception of the population’s average.

4. In studies when the individual being tested finds the measurement process uncomfort-

able or difficult, the sample is mostly made up of volunteers.

With the right conditions, these methods provide useful results. However, they are not

applicable to sampling theory.

3.5 Normal Distribution

Assumption of Normality

A part of sampling survey theory is finding formulas for means and variances. This is

derived from frequency distribution of estimates from each Si that is repeated sampled from

the population. A simplification of this can be observed in the size of the sample. In practice,

sampling is only done once for big sample size. This large sample size is a good reason to

assume that the frequency distribution of the sample estimates is normally distributed.

Setting aside the type of frequency distribution of the estimates. It is important to

establish notations used in every sampling survey design. As well as the definitions. Most

14
of the following notations are used in survey books such as Cochran and Lohr.

Definition 3. Estimator. The estimator, denoted by µ̂, denotes the rule by which an

estimate of some population parameter, µ , is calculated from the sample results.

Definition 4. Estimate. It is the value obtained for the estimator from a sample, Si .

Definition 5. Expected Value. An estimator µ̂ of µ taken over all possible samples provided

by the plan has an expected value denoted by

v
X
E(µ̂) = πi µ̂i (3.1)
i=1

where µ̂i is the estimate given by the i th sample, and πi is the probability of selecting Si .

The symbol E stands for ”the expected value of,”.

Definition 6. Unbiased Estimator. a method of estimation is unbiased if the average value

of the estimate, taken over all possible samples of given size n, is exactly equal to the true

population value. An estimator µ̂ is called unbiased if the mean value of µ̂, taken over all

possible samples provided by the plan, is equal to the true value µ(Cochran, 1977).

E(µ̂) = µ (3.2)

The amount of bias is

Bias(µ̂) = E(µ̂) − µ (3.3)

It was mentioned earlier that the role of sampling theory is to have efficient estimator.

One factor in efficiency is precision.

Definition 7. Precision. Precision refers to the size of deviation from the mean E(µ̂)

obtained by repeated application of the sampling procedure. This is computed using variance

15
of estimator.

Definition 8. Variance of estimator. The variance of µ, denoted as σµ̂ , is defined as

σµ̂ 2 = E(µ̂ − µ)2 (3.4)

Definition 9. Standard deviation. Standard deviation of the estimator µ̂, denoted as SE(µ̂)
q
or σµ̂ , is derived from σµ̂2 . It is also known as standard error.

How Good is an Unbiased Estimate?

Suppose a sampling procedure gives an unbiased estimator. A sample Si is selected

and have the following: estimate µ̂ and its standard deviation σµ̂ (for a sample set it is

alternatively termed as standard error ).

We cannot know the exact value of the error of estimate (µ̂ − µ) from the population

parameter µ but, from the properties of the normal curve, the chances are

0.32 (about 1 in 3 ) that the absolute error |µ̂ − µ| exceeds σµ̂

0.01 (1 in 100) that the absolute error |µ̂ − µ| exceeds 2.58σµ̂

Definition 10. Confidence Limits. Lower limit

µ̂L = µ̂ − tσµ̂ (3.5)

Upper limit

µ̂U = µ̂ + tσµ̂ (3.6)

The symbol t is the value of the normal deviate corresponding to the desired confidence

probability. The most common values are

16
 Confidence probability (%) 50 80 90 95 99

0.67 1.28 1.64 1.96 2.58

For example, if a probability sample of the records of batteries in routine use in a large

factory shows an average life µ̂ = 394 days, with a standard error σµ̂ = 4.6 days, the chances

are 99 in 100 that the average life in the population of batteries lies between

µ̂L = 394 − (2.58)(4.6) = 382 days

and

µ̂U = 394 + (2.58)(4.6) = 406 days

First, it is not correct to interpret it as a single estimate from a single survey such as, ”µ

lies between 382 and 406 days” . The correct way to interpret it is, ”using the same sampling

plan repeatedly, multiple samples provides their corresponding estimates and confidence

limit, about 99% of these statements would be correct and 1% wrong. The example above

is just one of the estimates and confidence limit.

What to do with a bias estimator?

It is ideal to have an unbias estimator, but in sampling theory there are reasons to

consider a bias estimator. These are the following:

1. In common cases, especially estimation of ratios, the estimators derived are biased.

2. Errors in measurement and non-response may lead to biases in the data we may

derive from, even using estimators that are unbiased in probability sampling.This occurs,

for example, when those who decline to be questioned are almost all against a certain public

funding expenditure, but people who are interviewed are equally divided between those who

support and oppose it.

Unbiased estimator are measured by precision, or its variance. For comparison of a

17
biased estimator with an unbiased estimator, or two bias estimators, a useful measurement

is accuracy, or the mean square error (MSE) of the estimate

Definition 11. Accuracy. It refers to the size of deviations from the true mean µ. Accuracy

of an estimator is the sum of the precision, and the square of bias of the estimator.

Definition 12. Mean Square Error.

MSE(µ̂) = E(µ̂ − µ)2 = E[(µ̂ − m) + (m − µ)]2

= E(µ̂ − m)2 + 2(m − µ)E(µ̂ − m) + (m − µ)2 (3.7)

= ( variance of µ̂) + ( bias )2

where E(µ̂), and the cross-product term vanishing since E(µ̂ − m) = 0.

Survey design

Survey design defines and plans the tools needed in survey sampling such as the primary

sampling unit (PSU). The objective of the survey should also be in line with the question-

naire design. Choosing PSU from the sampling frame affects the precision and accuracy of

estimates gathered from the sample.

The simplest design to consider is Simple Random Sampling (SRS). If the sampling frame

can be arranged in an order that is useful for the variable of interest Systematic Sampling

has better representation than SRS, although there is no theory that supports it. If more

details are available in the sampling frame then Stratified Sampling has better precision than

SRS. To manage the cost for face-to-face interviews Cluster Sampling is the most practical.

Cluster sampling may increase variance compared to SRS or stratified sampling, but it might

be considerable compared to the cost of data gathering. Details of the

Methodology of data gathering is also part of the design because it considers the budget

for the survey. The cheapest way of data gathering is through mail or email, but it will

18
also result in a low response rate (Lohr, 2010) [1]. There are several ways to increase

the response rate through telephone interviews such as sending mail to inform about the

interview. The face-to-face interview has the highest response rate compared to other forms

of data collection, but it is also very costly.

Another factor of survey design is the number of samples. Finding the number of samples

needed for data collection is very important, and will only be possible to find if there is an

existing population count. In finding the sample size it is very important to ask the following:

(1) How much precision is needed (Variance of population, S)? This can be based on previous

studies, conduct a pilot study first, or an expert’s guess. (2) How much error is tolerable (e)?

Usually e=0.03. We want to use a 95% Confidence interval with width at most 2e. That is

Z( /2) = 1.96(Lohr, 2010).

Even in the 1950’s sampling is not confined to partial coverages: the census of popu-

lation, agriculture, commerce, and industry in the United States and Canada has included

concurrent and supplementary samples, not only to broaden the scope of the information

but to study and evaluate the errors and biases of the census so that the data may be made

more useful (Deming, 1950). The chief concern of any government statistical agency should

be an assurance of the necessity, the success, and the economy of any survey that is autho-

rized at public expense. Such obligations cannot be met in the absence of knowledge and

research in statistical theory. Statistical research is constantly lowering costs and enhancing

the reliability and usefulness of statistical information.

Probability samples are frequently adjusted to solve specific research and policy questions,

but they come with a variety of drawbacks. Response rates are declining globally, with a

typical telephone survey response rate of less than 10% (Kohut et al., 2012)—far below the

95 percent response rate for mail surveys predicted by Deming (1950), page 35. Even face-

to-face surveys, such as the U.S. National Health Interview Survey (NHIS) response rate fell

19
from 92 percent in 1997 to 70 percent in 2015 (National Center for Health Statistics, 2016),

with additional nonresponse happening among individuals inside sampled households.

At least for some statistics, investigations have not identified strong connections between

response rate and bias (Groves, 2006), but lowering response rates have contributed to greater

data collection expenses. Because conducting probability samples is more expensive, sample

sizes are limited. As a result, if credible estimates for subpopulations of interest can be

estimated at all, they may require multiple years of data, and the estimates may be out of

date when they are created. (Lohr Raghunathan, 2017).

One of the example of survey is The Gallup World Poll. It monitors the most pressing

global concerns, including food security, employment, leadership performance, and well-

being. (Gallup, Inc., 2022). Gallup utilizes a random-digit-dial (RDD) approach or a

nationally representative list of phone numbers in countries where telephone interviewing

is practiced. In the United States, Canada, Western Europe, Japan, Australia, and other

countries, telephone methodology is commonplace.

20
Chapter 4

Alternative Sampling Strategies

4.1 Development of Sampling Schemes

This chapter discusses the overview of the alternative sampling strategies considered

in this study. It will start with a recap of the design and data collection of 2018 FIES. It is

followed by the common problems in the history of FIES. A short recap of FIES as sampling

on successive occasions. Then the discussion of alternative sampling strategies using Simple

Random Sampling.

As a quick recap from Chapter 1, the design of the 2018 FIES makes use of the 2013 Mas-

ter Sample (2013 MS), (PSA, 2020). The sampling frame of 2018 FIES is groups of replicates

from 2013 MS. These replicates are groups of Primary Sampling Unit (PSU) with 100-400

households. The details on how replicates are formed are in FIES 2018 technical report,

(PSA, 2020). A summary is provided in the introduction under the Philippine Household

Survey. For the 2018 FIES, there are samples of around 12 replicates in Highly Urbanized

Cities (HUC) and 16 replicates in the province domain. The sample size of replicate sample

allocation per domain was deemed sufficient for the following: 1 replicate for the national-

21
level estimate, 4 replicates for regional level estimates, and 16 replicates for provincial level

estimates. (PSA, 2021). The 16 replicates for province-level estimates leads to 170, 917

sampled households.

The data collection of 2018 FIES is on June 2018 and January 2019, where the 170, 917

households are interviewed twice. This is illustrated in Figure 3.1. The reference period for

both interviews is within the past 6 months. , FIES conducted the data collection twice per

survey year to minimize respondent memory bias, and at the same time capture seasonality

(PSA, 2020).

Figure 4.1: Illustration of the coverage and duration of the two visit data collection of 2018
FIES

4.2 Theoretical Framework

The purpose of this chapter is to illustrate how the alternative sampling strategies are

compared with the FIES 2018 strategy using theories of simple random sampling (SRS).

Assume a simple random sample of size n. There are four sampling schemes which are

denoted as: S∅, S1, S2 and S3, wherein,

Strategy ∅ : represents the complete overlapping sample between the two visits of FIES

2018 where all n samples interviewed in both visits

Strategy 1: represents the alternative sampling strategy of non-overlapping samples be-

22
tween the two visits where the n samples are divided into two.

Strategy 2: represents the alternative sampling strategy that uses the assumption that

the first visit values are equal to the second visit values where all n samples are interviewed

in the first visit only.

Strategy 3: represents the alternative sampling strategy of partial-overlapping samples

between the two visits where the n samples are divided into three sets. First and second sets

are interviewed in first visit. Second and third sets are interviewed in second visit.

In this section, we begin by assuming SRS and the measurements SRS made. The primary

objective is to estimae annual parameters.

4.3 Simple Random Sampling

It has been discussed in Chapter 3 the role of survey design in survey sampling. Simple

random sampling (SRS) is simplest design to start with, and provides the theoretical basis

for the more complicated forms. Given a population with N sampling units, and n units

are to be selected. SRS is the method of choosing the samples with equal chance of being

selected. Each unit in the population are numbered as i = 1, 2, 3, ...., N . A simple method

to selected n units with equal probability is by using fish bowl method or by means of a

computer program.

There are N Cn set of n samples from N population. In probability, the first sampling

unit has is n/N probability of selection. The second sampling unit has (n − 1)/(N − 1)

probabiity of selection, and so on. Hence the the probability that all n specified units are

selected in n draws is

n (n − 1) (n − 2) 1 n!(N − n)! 1
· · ··· = =
N (N − 1) (N − 2) (N − n + 1) (N )! N Cn

23
4.4 Definition and Notations

The primary objective of this study is to estimate annual parameters for example the

average annual income of a Filipino household based on FIES 2019. Suppose each household

sample be denoted as i = 1, 2, 3, . . . , N , where N has yi as the characteristic of interest.

Capital letter refer to the characteristics form the population and lowercase letters refer

to the characteristics from the sample. Note that the sample will not consist of the first n

units in the population. This is to avoid confusion.

Population Mean
PN
x1 + x2 + · · · + xN i=1 xi
X̄ = =
N N
Sample Mean
Pn
x1 + x2 + · · · + xn i=1 xi
x̄ = =
n n

Suppose {πi ≥ 0 : i = 1, 2, . . . , N } is a set of known selection probabilities for the units

24
4.5 Strategy ∅ (S∅)

This section discuss the first strategy which represents the data collection method of FIES

2018 where, t1 and t2 , are visit one and visit two, respectively. Suppose a simple random

sample of size n is drawn from a final population of N units. From this same sets of sample

measurement, are taken at two time intervals t1 and t2 . Let x1i and x2i be measurements

about variable X at time 1 and 2 respectively. Let xi = x1i + x2i . In the context of FIES,

x1i represents income/expenditure measurement incured by household i for the period of

January to June, and x2i for the period of July to December. In this context, xi represents

annual Value for household i = 1, 2, 3, . . . , N . In figure 4.2 shows an illustration of complete

overlapping data collection of Strategy ∅.

Figure 4.2: Strategy ∅ with complete overlap of interviews for two visits. Each household
are interviewed for both visits, visit one (t1 ) and visit two (t2 ).

Supposed one is interested in estimation of the population mean X̄ defined as

N
1 X
X̄ = xi (4.1)
N i=1

25
 This can be interpreted as annual values for a specific/income or expenditure variable of

FIES. Further equation 4.1 can be expressed in terms of the average of first visit and second

visit.
N N N N
1 X 1 X 1 X 1 X
X̄ = xi = (x1i + x2i ) = x1 + x2
N i=1 N i=1 N i=1 i N i=1 i

X̄ = X̄1 + X̄2 (4.2)

where, X̄1 is the true mean for the first visit; and X̄2 is the true mean for second visit.

Random sampling of n samples gives an unbiased estimator of the parameter X̄ as

n n n n
1X 1X 1X 1X
x̄ˆ = xi = (x1i + x2i ) = x1i + x2 = x̄ˆ1 + ¯ˆx2
n i=1 n i=1 n i=1 n i=1 i

x̄ˆ = x̄ˆ1 + x̄ˆ2 (4.3)

where x̄ˆ1 = sample mean for first semester; and x̄ˆ2 = sample mean for second semester.

To show that x̄ is unbiased, note that, the bias of an estimator is computed using the

difference of the parameter and expected value of the estimator. If the expected value of the

estimator is equal to its corresponding parameter, then the estimator is unbiased.

From, SRS theory

E(x̄ˆ) = E x̄ˆ1 + x̄ˆ2 = E x̄ˆ1 + E x̄ˆ2 = X̄1 + X̄2 = X̄ (4.4)

This shows that the estimator x̄ˆ of Strategy ∅ is unbiased. Now the variance of Strategy

∅. Using the theories in SRS the variance of ȳ is

26





V (x̄ˆ) = V x̄ˆ1 + x̄ˆ2 = V x̄ˆ1 + V x̄ˆ1 + 2 cov x̄ˆ1 , x̄ˆ1
  (4.5)
1 1
s2x1 + s2x2 + 2ρx1 x2 sx1 sx2


V (x̄ˆ) = −
r N
where ρx1 x2 is correlation between the value from first visit and the value form second

visit in the population; s2x1 and s2x2 are sample variance of values from first visit and second

visit.

Note that in SRS x̄ is an unbiased estimator of X̄ with unbiased estimator of its variance

as

 
1 1
v(x̄) = s2x̄ =  −  s2x
n N

where

1 N
X
s2x = (xi − X̄N )2
N −1 i=1

It is important to look into the relationship of variance between the two visits. Since

variances are constant values they can be written as a ratio of the other. This ratio may be

represented by a constant k. Let s2x̄2 = ks2x̄1 for some constant k. From here we will take

note of the variance of Strategy ∅ as

√
 
1 1  2 
V (x̄ˆ) = − sx1 + ks2x1 + 2 kρx1 x2 s2x1
n N
 
√ (4.6)
1 1  
V (hatx̄) = − s2x1 1 + k + 2 kρx1 x2
n N
Using STATA as the statistical software, empirical illustration using FIES 2018 data of

the values for k and ρx̄1 x̄2 are observed in Chapter 5. For illustration purposes FIES 2009 is

used because FIES 2018 data is not yet available. Take note that these values assume simple

27
random sampling for this chapter of the study.

28