STA3030:

Survey Sampling
Introduction and Some Basic
Statistics
Lecturer: Dr. John A. Wright

1
 The Problem
• How to infer facts about a population from
the information within a sample?
– The average height of statistics students in
Shenzhen
– The proportion of statistics students in Shenzhen
who wear glasses
– The number of statistics students in Shenzhen
who support Tottenham Hotspur F.C.

2
 The Problem
• How to infer facts about a population from
the information within a sample, given a
budget?
• Each observation (a measurement, answers on
a questionnaire) contains information
• But each observation also costs something
(time and money)

3
 The Problem
• Our resources are finite
• Too little information and we will be invariably
make a poor estimate
• Too much information means we have wasted
time/money
• What does the amount of information in a
sample depend on?

4
 The Problem
• The amount of information in a sample
depends on the sample size and the amount
of variation in the data
• In this course we will see many methods to
control the amount of variation – these are
called sample survey designs

5
 The Problem
• Survey design and sample size determine the
amount of information in the sample,
assuming that each element in the sample is
measured accurately
– How much do you earn?
– Does it seem possible or does it seem impossible
to you that the Nazi extermination of the Jews
never happened?
– Have you ever copied your homework?
6
 The Problem
– How much do you earn?
• People tend to inflate their true wage to impress the
person conducting the questionnaire
– Does it seem possible or does it seem impossible
to you that the Nazi extermination of the Jews
never happened?
• The double negative is confusing. A clearer version of
this question reduced the ‘Yes’ response from 22% to
1%
– Have you ever copied your homework?
• This question is deeply embarrassing to answer

7
 The Problem
• Problems like these can’t be handled using
statistical techniques, as these process the
data after it has been collected
• To address these kinds of errors, the
questionnaire/survey should be planned
carefully
• We shall see methods to do this

8
 The Problem
• We will want to summarize information about
populations and samples
• We treat the results obtained in sample
surveys as random, because we will randomly
select members of the population to be in our
sample
– Why would we want to do this?
• Therefore different people using the same
methods can obtain different results

9
Example – The Scottish Restaurant
• Example
– Our survey consists of one question: “How many
times a month do you eat food from McDonalds?”
– The population is everyone in this lecture theatre,
now
– Answer the question and write down your answer

10
Example – The Scottish Restaurant
• Example
– Our survey consists of one question: “How many
times a month do you eat food from McDonalds?”
– The population is everyone in this lecture theatre,
now
– Answer the question and write down your answer
– I need 3 volunteers…

11
Example – The Scottish Restaurant
• Each volunteer is interested in how much
McDonalds food this class consumes
• But they do not have enough time or much to
ask everyone
• So they randomly select a sample of 3 each
• And they collect their results…

12
Example – The Scottish Restaurant
• What can they do with their results?
• With the information they have, how can they
summarize their sample?
• Can they infer anything about the population
from their data?

13
Example – The Scottish Restaurant
• They might average the responses they
received
• They might work out the standard deviation of
the responses they received
• Then each volunteer will have estimates of the
population mean and standard deviation
• Let’s compare their estimates to these
population parameters

14
Example – The Scottish Restaurant
• Think about: how could the volunteers
improve their surveys?

15
 Some Basic Statistics
• We casually introduced population mean and
population standard deviation and their
sample versions, the sample mean and sample
standard deviation
• Before we start creating survey designs, we
need to remember some fundamental
definitions and concepts from statistics

16
 Some Basic Statistics
• Population Mean
– Let there be 𝑁 people in the population. Let the
variable of interest be 𝑌 (e.g. height in cm) and let
person 𝑖 have value 𝑦% (e.g. the 1729th person has
height 180 cm)
– Is the average value of all measurements in the
&' (&) (⋯(&+
population, i.e. ≔ 𝜇
,
• We will often denote the population mean by 𝜇
– The population mean in our McDonalds example
is…
17
 Some Basic Statistics
• Finite Population Variance
– The finite population variance is given by
/ , 3 ≔ 𝜎3
∑ 𝑦
%4/ % − 𝜇
,
• We often denote the population variance by 𝜎 3
– The population standard deviation (SD) simply the
square root of the population variance
&6 78 )
,
∑%4/
• ≔𝜎
,
– The population standard deviation in our
McDonalds example is…?

18
 Some Basic Statistics
• In some textbooks, we may see finite population variance
defined as
/
– ∑, 𝑦% − 𝜇 3 ≔ 𝑆 3 , the so-called “corrected population
,7/ %4/
variance”
• Comparing this definition to our 𝜎 3, notice
/
– 𝑆3 = 1 − 𝜎 3 , so as 𝑁 become large, they are more or less
,
the same
• Why bother with 𝑆 3 ?
/
– Because the “,7/” factor simplifies many formulas, as we’ll see
later.
• I prefer 𝜎 3 to 𝑆 3 as it better reflects the idea of variance as
an expectation
3
– 𝑉𝑎𝑟 𝑌 = 𝐸 𝑌 − 𝜇
19
 Some Basic Statistics
• It will be useful to have some measure of
dependence between two random variables
• Some variables are clearly independent
– Level of the Ulan Bator Stock Index and my shoe
size
• But some obviously have some dependence
– FTSE100 stock index and UK gilt prices
• Covariance and Correlation help us quantify
dependence
20
 Some Basic Statistics
• Covariance
– We have a finite population of size 𝑁. Each person in
the population has two variables of interest, 𝑌 and 𝑋
– Let the population means of 𝑌 and 𝑋 are 𝜇A and 𝜇B
respectively. Then the population covariance of 𝑋 and
𝑌 is given by
/
• ∑,
%4/ 𝑦% − 𝜇 A 𝑥% − 𝜇 B ≔ 𝐶𝑜𝑣(𝑋, 𝑌)
,7/
– If 𝑥 and 𝑦 are independent, then 𝐶𝑜𝑣(𝑋, 𝑌) is zero
– The converse is not necessarily true
– It is tricky to decide whether a covariance is large or
small because the scales of 𝑦 and 𝑥 are combined
– It would be better to have a dimensionless number to
describe dependence…

21
 Some Basic Statistics
• Correlation
– Is that dimensionless quantity
– The correlation between 𝑋 and 𝑌 is just the
covariance of 𝑋 and 𝑌 divided by the standard
deviations of 𝑋 and 𝑌
JKL B,A
– That is, M M ≔𝜌
N O
– This must lie between −1 and 1 and it has no
units
– We often denote it by 𝜌 (rho)

22
 Some Basic Statistics
• Population mean, variance and correlation are
examples of parameters
– Parameters are numerical descriptive measures of
a population (mean, total, variance, proportion,
population size, etc.)
– Unless we have measured every member of the
population, parameters will always be unknown

23
 Some Basic Statistics
• Much of this course will be spent on
investigating how we can best estimate the
parameters of a population from a sample of
the population
– We’ll take the observations from our sample, plug
them into a formula that gives our guess for the
value of the parameter
– That guess from our sample is called the estimate.
The general formula is called the estimator.

24
 Some Basic Statistics
• Example
– From our McDonald’s surveys, what were the parameters,
estimates and estimators?
• We denote a generic parameter by 𝜃 (theta)
• We denote an estimator of that parameter by putting a
hat on it - 𝜃R
• Notice that, before we conduct the survey, 𝜃R is a
random variable
– Because the elements in the sample were chosen
randomly
• Thus 𝜃R will have a probability distribution
– This distribution is called the sampling distribution

25
 Some Basic Statistics
• Why do we care about the sampling distribution?
– Because it is not enough to just present a single
number as your best guess for the value of 𝜃
– This gives no idea about how good or accurate your
guess is
• Your estimate for 𝜃 is very unlikely to be exactly
correct and indeed, we wouldn’t know if it was,
because we’ll never know the true value of 𝜃
– But it would be nice to know that the chances of it
being wildly incorrect are quite small

26
 Some Basic Statistics
• That is, we would like our guess to be close to 𝜃, most
of the time
– i.e. we would like |𝜃R − 𝜃| to be smaller than, say, some
number 𝑑, most of the time
– 𝜃R − 𝜃 is called the error of estimation
• We can write this in terms of a probability statement
– We would like Pr 𝜃R − 𝜃 < 𝑑 = 1 − 𝛼, where 𝛼 is a
small number like 0.05 or 0.01 and 𝑑 is called the margin
of error
• The only random variable in the statement above is 𝜃R.
In order to choose a sensible number 𝑑, we must have
some idea what the sampling distribution is

27
 Some Basic Statistics
• The distribution of 𝜃R will depend crucially on 𝑛,
the size of the sample
• If 𝑛 is small, the distribution of 𝜃R is usually
complicated and unknown
• If 𝑛 is large, the distribution of 𝜃R is much more
regular and is well approximated by distributions
we know well
• If 𝑛 is large and close to 𝑁, the distribution of 𝜃R
tends to contentrate more and more around the
true value of 𝜃, but is not well approximated by
distributions we know well
28
 Some Basic Statistics
• Imagine we are confident enough to say
^ 7_ ]
] ^
– ~𝑁 0,1 , where 𝜎]^ is the standard
M`^
deviation of 𝜃b
• Then, we may write
^ 7_ ]
] ^
– 𝑃 𝑧e < < 𝑧/7e = 1 − 𝛼
) M`^ )
^ 7_ ]
] ^
– ⇒ 𝑃 −𝑧/7e < < 𝑧/7e = 1 − 𝛼
) M`^ )

29
 Some Basic Statistics
• Carrying on, we have
– ⇒ 𝑃 −𝜎 ^] 𝑧/7e < 𝜃R − 𝐸 𝜃R < 𝜎 ^] 𝑧/7e = 1 − 𝛼
) )

– ⇒ 𝑃 −𝜃R − 𝜎 ]^ 𝑧/7e < −𝐸 𝜃R < −𝜃R + 𝜎]^ 𝑧/7e = 1 − 𝛼

– ⇒ 𝑃 𝜃R − 𝜎^] 𝑧/7e < 𝐸 𝜃R < 𝜃R + 𝜎 ^] 𝑧/7e = 1 − 𝛼

• And if 𝜃R is unbiased, i.e. if 𝐸 𝜃R = 𝜃, then

– ⇒ 𝑃 𝜃R − 𝜃 < 𝜎]^ 𝑧/7e = 1 − 𝛼, meaning we have found our

30
 Some Basic Statistics
• The interval 𝜃R − 𝜎]^ 𝑧/7e, 𝜃R + 𝜎 ]^ 𝑧/7e is random –
) )
R
recall 𝜃 is a random variable
– If we plug in our observation of 𝜃R from our sample, we
have a 100× 1 − 𝛼 % confidence interval for parameter
𝜃
• Having a normal sampling distribution is very handy.
We will discuss conditions for its existence in the next
chapter.
• Clearly we want our estimator 𝜃R to have the property
“𝐸 𝜃R = 𝜃”. What else do we want from our
estimator?

31
 Some Basic Statistics
• We want our observation of 𝜃R from our sample to
be close to the parameter we’re interested in
• This can happen if
– The estimator generates estimates are correct on
average
– And these estimates are not too spread out
• Statistically, we want
– 𝐸 𝜃b = 𝜃, as before
3
–𝐸 𝜃b − 𝜃 is small

32
 Some Basic Statistics
• The quantity 𝐸 𝜃R − 𝜃 is called the bias of the
estimator 𝜃R
– We want our estimators to be unbiased, i.e. to
have zero bias
3
• The quantity 𝐸 𝜃R − 𝜃 is called the mean
square error (MSE) of 𝜃R
• There is a relationship between MSE, bias and
variance

33
 Some Basic Statistics
3 3
• 𝐸 𝜃R − 𝜃 =𝐸 𝜃R − 𝐸 𝜃R + 𝐸 𝜃R − 𝜃
3
–=𝐸 𝜃b − 𝐸 𝜃b + 2 𝐸 𝜃b − 𝜃 𝐸 𝜃b − 𝐸 𝜃b +
3
𝐸 𝜃b − 𝜃
3
– ⇒ 𝑀𝑆𝐸(𝜃b ) = 𝑉𝑎𝑟 𝜃b + 𝐵𝑖𝑎𝑠 𝜃b
• Therefore, if estimator 𝜃R is unbiased (most of
our estimators will be unbiased), then its MSE
equals its variance
34
 Some Basic Statistics
• Here are three useful words to describe an estimator
– Unbiased if 𝐸 𝜃R = 𝜃
– Precise if 𝐸 𝜃R − 𝐸(𝜃R ) ↑2 is small
3
– Accurate if 𝐸 𝜃R = 𝜃 and 𝐸 𝜃R − 𝜃 is small
• Thus a precise estimator may not be accurate because
it is biased.
• Precision considers how close estimates from different
samples are to each other.
• Accuracy considers how close estimates are to the true
value.
35
 Definitions
• We will go through some unavoidable
technical terms which we will use repeatedly
throughout the course
• Element
– An element is an object on which a measurement
is taken
• E.g. a Shenzhen statistics student, an orange, anything

36
 Definitions
• Population
– A collection of elements about which we want to
make inferences
• E.g. Shenzhen statistics students, oranges grown in
California

37
 Definitions
• Sampling Units
– Non-overlapping collections of elements from the
population that cover the entire population (i.e.
they partition the population)
• E.g. CUHK statistics students, HKU stats students,
HKUST stats students… etc. (assuming that a student
cannot study statistics in two university in Hong Kong at
the same time)
• Often sampling units and elements are the same

38
 Definitions
• Frame
– A frame is a list of sampling units
• E.g. Lists of registered CUHK statistics students, of HKU
statistics students, etc.
• ‘List’ can be loosely interpreted: any representation of
the population that allows random selection of units
can be a frame

39
 Definitions
• Sample
– A collection of sampling units drawn from a single
or from multiple frames

40
 Definitions
• Example: I want to infer the proportion of Hong
Kong high school students who support
Tottenham Hotspur F.C.
• I decide I will ask 200 students
• I have a list of all the high schools in Shenzhen. I
select 25 of these
• For each of the 25, I collect a name list from the
school and I select 8 to interview
• What are the element, population, sampling
units, frame(s) and sample in this example?

41