ECO 313A

Variables, distributions, inference, and confidence intervals

– What is a variable?
– Means, variances and distributions
– Theoretical probability distributions
– Inference – using sample stats to estimate population parameters
• Normal / standard normal distributions
• Standardisation of a normally distributed variable
• Sampling distributions
• The Central Limit Theorem
• Confidence intervals
1
 What is a variable?
• Some defined quality/condition that varies between one person and another, one country
and another, over time, etc.
– The price of bread?
– Household income?
– Personality?
– The number of kilos in a ton?
– People’s age?
– ECO 313 students’ age?
– Gross domestic product?
– Kg of fertilizer used per hectare?
– The distance between East London and KWT?
– Well-being?
• It has to vary, and it has to be defined/definable….
• Also note: different types of variables – gender, number of children, vs income [i.e.
categorical, discrete, continuous; ordinal, cardinal…] 2
 What is a variable?
• Some defined quality/condition that varies between one person and another, one country
and another, over time, etc.
– The price of bread? √
– Household income? √
– Personality? √? X?
– The number of kilos in a ton? X
– People’s age? √
– ECO 313 students’ age? √
– Gross domestic product? √
– Kg of fertilizer used per hectare? √
– The distance between East London and KWT? X
– Well-being? √? X?
• It has to vary, and it has to be defined/definable….
• Also note: different types of variables – gender, number of children, vs income [i.e.
categorical, discrete, continuous; ordinal, cardinal…] 3
 How can we ‘know a variable’?
• We can look at its values
• We can calculate its average mean value
• We can calculate its variance
• … and other ‘descriptive statistics’….
• Broadly, we can know/measure/estimate its distribution

• But to begin, we need to look at some of the values of that

variable….
4
Suppose our variable of interest is the Age of
students in ECO 313…
• We can imagine a table:
Observation Age
1 ?
2 ?
3 ?
4 ?
5 ?
6 ?
7 ?
5
• Suppose:
Observation Age
1 21
2 25
3 24
4 22
5 22
6 24
7 23
6
 And we can calculate some descriptive statistics
• The mean:
n
1
ഥ
X = ෍ Xi
n
i=1

• Eish – what does it mean? It means I sum up the values of X (eg the
ages), and then divide by the number of observations.
• So 21 + 25 + 24 + 22 + 22 + 24 + 23 = 161
• 161 / 7 = 23 = ഥ
X

• But what does it mean????! MEASURE OF ‘CENTRAL TENDENCY’ 7

 • And variance?
𝑛
1
𝑆𝑋2 ത 2
= ෍(𝑋𝑖 − 𝑋)
𝑛
𝑖=1

Obs Age Age – AvgAge (Age – AvgAge)2

1 21 -2 4
2 25 2 4
3 24 1 1
4 22 -1 1
5 22 -1 1
6 24 1 1
7 23 0 0
Sum 12

So Var = 12/7 = 1.71; but what does it ‘mean’???! MEASURE OF ‘DISPERSION’8

But we can also summarise using a different
type of table, ie a frequency table
Age Frequency/
count
21 1
22 2
23 1
24 2
25 1

This table is one way of representing the distribution

of age. Another way is graphically…
9
Frequency distribution of age

10
Probability distribution of age

11
 Now look at them together…

Basically the
same picture…
How can we use the probability distribution?

Suppose I select
someone from the class
at random…. What is:
• Prob(Age = 21)?
• Prob(Age = 23)?
• Prob(Age = 20)?

13
What about Prob(22 ≤ Age ≤ 24)? In other words, the probability
that someone selected at random is 22 or 23 or 24 years old)

14
What about Prob(22 ≤ Age ≤ 24)? In other words, the probability
that someone selected at random is 22 or 23 or 24 years old)

Prob(22 ≤ Age ≤ 24) = 0.23 + 0.14 + 0.15 = 0.52 15

What about Prob(21 ≤ Age ≤ 42)?

Prob(21 ≤ Age ≤ 42) = ? 16

What about Prob(22 ≤ Age ≤ 42)?

Prob(22 ≤ Age ≤ 42) = 0.23 + 0.14 + 0.15 + …. = ?? 17

‘Theoretical probability distributions’

18
 Uniform distribution – example 1

We can write Prob(x=heads) = 0.5; but it so happens that we can

also write Prob(x=tails) = 0.5
But in what sense is this distribution ‘theoretical’? 19
 Uniform distribution – example 2

We can write Prob(x=1) = Prob(x=2) = … =1/6

So now what is Prob(3 ≤ x ≤ 5)?
20
Normal distribution (but we’ll come back
to this just now…)

21
 The question of inference
• Suppose you want to know the mean age of every person who
lives in Alice
– Alice is your population, therefore…
– you want to know the population mean, μ (‘mu’)
• There are 15 143 people (supposedly)
• How are you going to do it?
• What are your options?

22
 The question of inference…
• Basically there are three options:
• First, you can collect data from everyone (i.e. conduct a census)
so that you can calculate the population mean, μ, directly.
We can write:
N
1
μ = ෍ Xi
N
i=1
where N represents the size of the population. (Compare this to our
formula for the sample mean:
n
1
ഥ
X = ෍ Xi )
n
i=1
23
• Second, we can stop one person on the street and ask him/her
his/her age. Problem – how ‘representative’ is one person?
• Third, we can draw a sample, calculate the sample mean, then
declare that our sample mean is an estimate of the (unknown)
population mean, ie X ഥ is an estimate of μ.
• (You can see that the so-called second option is a special case
of the 3rd option in which n=1)
• The big hairy question in statistics is, how good an estimate is
𝑋ത of μ?
– However, we don’t yet have the tools to answer that…

24
The standard

Standardising
variable (2)
distributed
a normally
normal
distribution
(1)

Confidence
intervals
etc.
distributions
Sampling

The Central
(3)

Limit
Theorem (4)

25
 The normal distribution
• The normal distribution plays a central role in the answer, but we
have to get there through a circuitous route.
• First, the so-called normal distribution is actually a family of
distributions which follow a common distribution function:
1 − 𝑋−𝜇 2 /2𝜎2
𝑓 𝑋 = ∙𝑒
𝜎 2𝜋
• Don’t worry about the detail or try to ‘understand’ this function. All
it means is that if you have a random variable X which is ‘normally
distributed’ with mean μ and variance σ2, then f(X) tells you the
height of the graph for different possible values of X.
26
Pause: but where does the normal distribution ‘come from’?
[see Lane, n.d. (on BlackBoard)]
• Different explanations, eg the normal distribution “occurs naturally
in many situations”. An early case was measurement errors in
astronomy, which Galileo (17th c.) studied and characterised.
• Abraham de Moivre, an “18th century statistician and consultant to
gamblers”, in effect discovered the idea of the Central Limit Theorem
(see below), and then sought a formula that closely matched what
he observed.
• Independently, mathematicians Ardrian and Gauss developed
formulae for the normal distribution, in 1808 and 1809, respectively.
• However, why some variables in the world follow a normal
distribution is somewhat contentious (see e.g. Lyon, 2014).
27
Welcome to the family…

28
• In general we write 𝑋~ℕ(𝜇, 𝜎 2 )
• Which we read as ‘X follows a normal distribution with mean μ
and variance σ2 ’.
• (The convention is that, for any distribution, the Greek letter μ
(‘mu’) represents the population mean, and the Greek letter σ
(‘sigma’), when squared, represents the population variance.)
• So we can imagine an infinite number of normal distributions
by supposing different values of μ and σ2, eg 𝑋1 ~ℕ(4, 5),
𝑋2 ~ℕ(−250, 12), 𝑋3 ~ℕ(78, 99), etc. etc. etc.

29
 The Standard Normal Distribution
• But there is one member of the family of normal distributions that is
regarded as special, namely the ‘standard normal distribution’, which
is X~ℕ(0, 1), that is, where the mean of X is 0 and the variance is 1.
• The convention is to write 𝐙~ℕ(0, 1)

30
• A few useful facts about normal distributions:
– They are infinite in both directions
– They are symmetric
– The total area ‘under them’ is 1 (as with any probability
distribution, by definition)
• We can apply these to the standard normal distribution…

31
 Prob(-∞ ≤ Z ≤ ∞) = ? Prob(0 ≤ Z ≤ ∞) = ?

Prob(-∞ ≤ Z ≤ 0) = ? Prob(0 ≤ Z ≤ 1.51) = ? 32

 Prob(-∞ ≤ Z ≤ ∞) = 1 Prob(0 ≤ Z ≤ ∞) = ?

Prob(-∞ ≤ Z ≤ 0) = ? Prob(0 ≤ Z ≤ 1.51) = ? 33

 Prob(-∞ ≤ Z ≤ ∞) = 1 Prob(0 ≤ Z ≤ ∞) = 0.5

Prob(-∞ ≤ Z ≤ 0) = ? Prob(0 ≤ Z ≤ 1.51) = ? 34

 Prob(-∞ ≤ Z ≤ ∞) = 1 Prob(0 ≤ Z ≤ ∞) = 0.5

Prob(-∞ ≤ Z ≤ 0) = 0.5 Prob(0 ≤ Z ≤ 1.51) = ? 35

 Prob(-∞ ≤ Z ≤ ∞) = 1 Prob(0 ≤ Z ≤ ∞) = 0.5

Prob(-∞ ≤ Z ≤ 0) = 0.5 Prob(0 ≤ Z ≤ 1.51) = ?????

36
• Conceptually, this is similar to what we’re doing when we ask
what is Prob(22 ≤ Age ≤ 24)? In effect, we add the ‘areas’ within
the distribution so long as they conform to the condition that
22 ≤ Age ≤ 24:

Prob(22 ≤ Age ≤ 24) =

= 0.23 + 0.14 + 0.15 = 0.5

37
• Or in respect of the die, what is Prob(3 ≤ x ≤ 5)?

Prob(3 ≤ x ≤ 5) = 1/6 + 1/6 + 1/6 = 3/6 = 0.5 38

• One big difference between these normal distributions and the
distributions presented previously (e.g. Age, coin, die) is that this
assumes a continuous variable whereas those assumed a discrete
variable.
• This creates a weird situation. For a discrete distribution, we can ask
what is the probability that a person chosen at random from a
distribution is aged 22, and write Prob(Age=22) = 0.23.
• But for a continuous distribution such as that of Z, Prob(Z=2) = 0.
Why? What about Prob(Z=2.0000000000000001)? See the problem?
There are an infinite number of values that Z could take on between
any two other numbers; if they each have a positive probability, then
the total probability would necessarily exceed 1. By a lot.
• So in general, for a continuous distribution we can ask what is
Prob(0 ≤ G ≤ 0.005), but we cannot ask what is Prob(G = 0.005)….
39
Let’s return to the situation below – how do we answer it?

Prob(0 ≤ Z ≤ 1.51) = ?

• Almost every statistics textbook in the world has a probability table

for the standard normal distribution. The one we use in this course
happens to look as follows:

40
41
In other words, the
probability that a
std normal variable
Z is between 0 and
1.51 is 0.4345.

42
In other words, the
probability that a
std normal variable
Z is between 0 and
1.51 is 0.4345.

Prob(0 ≤ Z ≤ 1.51) =
0.4345

43
• Let’s mess around a bit. If this is true, what
about…

Prob(Z ≤ 1.51) = ?

Prob(Z ≥ 1.51) = ?

44
• Let’s mess around a bit. If this is true, what
about…

Prob(Z ≤ 1.51) =
0.4345 + 0.5 =
0.9345

Prob(Z ≥ 1.51) = ?

45
• Let’s mess around a bit. If this is true, what
about…

Prob(Z ≤ 1.51) =
0.4345 + 0.5 =
0.9345

Prob(Z ≥ 1.51) =
1 – 0.9345 =
0.5 – 0.4345 =
0.0646
46
 Let’s try one more, but the other way around...
• Prob(-k ≤ Z ≤ k) = 0.95; in other words we want to find the value k (and
thus -k) such that there is a 95% chance that a randomly selected value
Z is between -k and k.

47
• Given how our table is constructed, we change this into an
equivalent question:
Prob(0 ≤ Z ≤ k) = 0.95/2 = 0.475

48
Now we work
backwards; we find
the value in the table
as close as possible to
0.475, and determine
the value of k from
that…

49
Now we work
backwards; we find
the value in the table
as close as possible to
0.475, and determine
the value of k from
that, i.e. 1.96

50
• So we can write: Prob(-1.96 ≤ Z ≤ 1.96) = 0.95

51
 ‘Standardising’ a normally distributed variable
• Suppose you have a variable X that is normally distributed X~ℕ(3,14)
• The problem is, we can’t determine Prob(0 ≤ X ≤ 1.42) because there is
no probability table for this distribution, nor for X~ℕ(−2, 5), nor for
X~ℕ 301, 7.5 , etc., etc.
• However, if we know the mean and variance of X, in principle we can
standardise the distribution of X to create a standard normal variable.
This is how:
X − μX
= Z~ℕ(0,1)
σX
(But what is this σX thing? It’s the ‘standard deviation’ of X = the square
root of the variance of X.)
• Standardisation will become very useful to us in a short while….
52
 ഥ
The Sampling Distribution of 𝐗
• Apart from being the sample mean, X ഥ is a random variable!
Why? Because every time I draw a different sample I end up
ഥ.
with a different X
ഥ
• So for a given sample size n, one can imagine that, like X, X
has its own distribution. We call this a ‘sampling distribution’.
• What may not be at all obvious is that the sampling
distribution may look very different from the distribution of X
itself.

53
• Let’s recall our frequency distribution of ECO 313 students’ Age:

• What if I draw 5000 samples of size n=5 from this distribution, and
calculate a ഥ
X for each one?
54
 ഥ, n=5
Sampling distribution of X

55
 ഥ, n=15
Sampling distribution of X

56
 ഥ, n=30
Sampling distribution of X

57
 Original distribution of Age Sampling distribution of mean Age, n=5

What do you observe?

Sampling distribution of mean Age, n=30 Sampling distribution of mean Age, n=15
• The larger the sample size, the less the sampling distribution
ഥ resembles the original distribution of Age, and the
of X
more it resembles … (wait for it…) a normal distribution.
• What is not so visible from these graphs is that the variance
of the sampling distribution declines as the sample size goes
up:
Sample size (n) 𝐕𝐚𝐫(ഥ
X)
5 2.31
15 0.61
30 0.22
These two features help explain why, if you can afford it,
you would prefer a larger sample size!
59
 The Central Limit Theorem
• As the sample size increases, the distribution of the sample mean
(ഥ
X) approaches (‘→’) the normal distribution with mean μX and
variance σX 2/n. Or,
σ2X
ഥ
X → ℕ(μX , )
n

• [Hmm, why all this fuss about the sample mean? Because the same
principles apply to econometric estimates, which we’ll get to
soon…]
• [For a proof of the CLT, see Filmus, 2010 (on BlackBoard).]

60
 That’s nice, but recall, we only have probability tables for
the standard normal distribution!
• True, but just as we standardised a normally distributed
variable, we can standardise an almost-normally distributed
sample mean.
• Perform the same trick as before – subtract the mean (in this
ഥ), and divide by the standard deviation (again, of X
case of X ഥ):
(𝑋ത − 𝜇𝑋ത ) (𝑋ത − 𝜇𝑋 ) 𝑛
= = (𝑋ത − 𝜇𝑋 ) ∙
𝜎𝑋ത 𝜎𝑋 Τ 𝑛 𝜎𝑋

= 𝑍ഥ → ℕ(0, 1)
61
• Question: the previous slide implied that 𝜇𝑋ത = 𝜇𝑋 , but
𝜎𝑋ത ≠ 𝜎𝑋 ; what’s going on?

62
Putting it all together – Confidence Intervals!
• Recall that Prob(-1.96 ≤ Z ≤ 1.96) = 0.95
• This holds true approximately for our Zത because our Zത only
approximates the standard normal distribution, so:
0.95 ≈ Prob(-1.96 ≤ Zത ≤ 1.96)
• Now, let’s substitute:
n
ഥ − μX ) ∙
0.95 ≈ Prob(-1.96 ≤ (X ≤ 1.96)
σX

63
• And now we simplify:
n
0.95 ≈ Prob(−1.96 ≤ (ഥ
X − μX ) ∙ ≤ 1.96)
σX
σX σX
= Prob(−1.96 ∙ ഥ
≤ X − μX ≤ 1.96 ∙ )
n n
σX σX
= Prob(−ഥ
X − 1.96 ∙ ≤ −μX ≤ −ഥ
X + 1.96 ∙ )
n n
σX σX
= Prob(ഥ
X + 1.96 ∙ ≥ μX ≥ ഥ
X − 1.96 ∙ )
n n
σX σX
= Prob(ഥ
X − 1.96 ∙ ≤ μX ≤ ഥ
X + 1.96 ∙ )
n n
≈ 0.95
64
• And now we simplify:
n
0.95 ≈ Prob(−1.96 ≤ (ഥ
X − μX ) ∙ ≤ 1.96)
σX
σX σX
= Prob(−1.96 ∙ ഥ
≤ X − μX ≤ 1.96 ∙ )
n n
σX σX
= Prob(−ഥ
X − 1.96 ∙ ≤ −μX ≤ −ഥ
X + 1.96 ∙ )
n n
σX σX
= Prob(ഥ
X + 1.96 ∙ ≥ μX ≥ ഥ
X − 1.96 ∙ )
n n
σX σX
= Prob(ഥ
X − 1.96 ∙ ≤ μX ≤ ഥ
X + 1.96 ∙ )
n n
≈ 0.95 But please note: in practice you don’t need to do this
derivation over and over again. All you need to do is apply
the formula at the end. 65
• But what does it mean?
σX σX
ഥ ഥ
Prob X − 1.96 ∙ ≤ μX ≤ X + 1.96 ∙ ≈ 0.95
n n

• It means that we are approximately 95% certain

that the true, unknown population mean, is
within the interval indicated, i.e.
𝛔𝐗 𝛔𝐗
ഥ − 𝟏. 𝟗𝟔 ∙
𝐗 ഥ
𝐗 ഥ + 𝟏. 𝟗𝟔 ∙
𝐗
𝐧 𝐧

Approx 95% confident that 𝛍𝐗 is here

Finally! This is how we answer the question of how good 𝑿 ഥ

is as an estimate of 𝝁𝑿 ! This also helps explain why you
would tend to want a large n…. (How so?) 66
 Question: how do we know if our ഥX’s & Zത ’s approach a
normal distribution ‘enough’?
• Generally we do not know; remember, we usually don’t observe
the sampling distribution, we simply know that our particular
sample must come from such a distribution.
• There is a common rule of thumb in statistics that a minimum
sample size should be n = 30, implying that this is enough for the
benefits of the Central Limit Theorem to kick in; but this does not
have a solid theoretical basis.
• However, when we get to econometrics, we will discuss a couple of
ways of determining whether the sampling distribution is
adequately close to ‘normal’.

67
 Brief aside for future use…
• Question 1: If ‘a’ is a constant, and μX is the mean of X, then
what is the mean of aX?
• Answer:
𝑁
1 1
𝜇𝑎𝑋 = ෍ 𝑎𝑋𝑖 = 𝑎𝑋1 + 𝑎𝑋2 + ⋯ + 𝑎𝑋𝑁
𝑁 𝑁
𝑖=1
1
= 𝑎 𝑋1 + 𝑋2 + ⋯ + 𝑋𝑁
𝑁
𝑁
1
= 𝑎 ∙ ෍ 𝑋𝑖 = 𝑎𝜇𝑋 . (𝐿𝑖𝑘𝑒 𝑑𝑢ℎ!)
𝑁
𝑖=1
68
• Question 2: If ‘a’ is a constant, and 𝜎𝑋2 is the variance of X,
then what is the variance of aX?
• Answer:
𝑛
2
1
Define 𝜎𝑎𝑋 = ෍(𝑎𝑋𝑖 − 𝑎𝜇)2
𝑁
𝑖=1

𝑁𝑜𝑡𝑒: (𝑎𝑋1 − 𝑎𝜇)2 = 𝑎2 𝑋12 + 𝑎2 𝜇2 − 2𝑎2 𝑋1 𝜇

= 𝑎2 (𝑋1 − 𝜇)2
𝑛 𝑛
2
1 𝑎2
So 𝜎𝑎𝑋 = ෍ 𝑎2 (𝑋𝑖 − 𝜇)2 = ෍(𝑋𝑖 − 𝜇)2 = 𝑎2 𝜎𝑋2
𝑁 𝑁
𝑖=1 𝑖=1
69