Statistics 3

Revision Notes

June 2016
2 S3 JUNE 2016 SDB
Statistics 3
1 Combinations of random variables ....................................................................................................3
Expected mean and variance for X ± Y .....................................................................................................3
Reminder .............................................................................................................................................................................. 3
Combining independent normal random variables Y ................................................................................3
2 Sampling ...............................................................................................................................................4
Methods of collecting data .........................................................................................................................4
Taking a census .................................................................................................................................................................... 4
Sampling ............................................................................................................................................................................... 4
Simple random sampling............................................................................................................................5
Using random number tables ................................................................................................................................................ 5
Systematic sampling ...................................................................................................................................5
Stratified sampling .....................................................................................................................................6
Sampling with and without replacement ....................................................................................................6
Quota sampling ..........................................................................................................................................7
Primary data ..............................................................................................................................................7
Secondary data ...........................................................................................................................................7
3 Biased & unbiased estimators .............................................................................................................8
Unbiased estimators of µ and σ 2 .............................................................................................................11
Estimating µ and σ 2 from a sample ................................................................................................................................. 11

4 Confidence intervals and significance tests......................................................................................14

Sampling distribution of the mean ...........................................................................................................14
Central limit theorem and standard error ...............................................................................................15
Confidence intervals ................................................................................................................................16
Central Limit Theorem Example ........................................................................................................................................ 16
Significance testing– variance of population known ...............................................................................18
Mean of normal distribution ............................................................................................................................................... 18
Difference between means of normal distributions ............................................................................................................ 19
Significance testing – variance of population NOT known, large sample ...............................................21
Mean of normal distribution ............................................................................................................................................... 21
Important assumption ......................................................................................................................................................... 21
Difference between means .................................................................................................................................................. 22
5 Goodness of fit, χ2 test ..................................................................................................................... 23
General points.......................................................................................................................................... 23
Discrete uniform distribution................................................................................................................... 23
Continuous uniform distribution.............................................................................................................. 24
Binomial distribution ............................................................................................................................... 24
Poisson distribution ................................................................................................................................. 24
The normal distribution ........................................................................................................................... 26
Contingency tables ................................................................................................................................... 27
6 Regression and correlation................................................................................................................ 29
Spearman’s rank correlation coefficient ................................................................................................. 29
Ranking and equal ranks .....................................................................................................................................................29
Spearman’s rank correlation coefficient ..............................................................................................................................29
Spearman or PMCC.............................................................................................................................................................30
Testing for zero correlation ..................................................................................................................... 31
Product moment correlation coefficient ..............................................................................................................................31
Spearman’s rank correlation coefficient ..............................................................................................................................31
Comparison between PMCC and Spearman........................................................................................................................32

7 Appendix ............................................................................................................................................. 33
Combining random variables .................................................................................................................. 33
E[X + Y] = E[X] + E[Y]......................................................................................................................................................33
Var[X + Y] = Var[X] + Var[Y] ...........................................................................................................................................34
Unbiased & biased estimators ................................................................................................................. 34
Unbiased estimators ............................................................................................................................................................34
Biased Estimators ................................................................................................................................................................35
Unbiased estimates of population mean and variance ............................................................................ 35
Unbiased estimate of the mean ............................................................................................................................................35
Unbiased estimate of the variance of the population ...........................................................................................................36
Bias ......................................................................................................................................................................................37
Probability generating functions ............................................................................................................. 39
Expected mean and variance for a p.g.f...............................................................................................................................39
Mean and variance of a Binomial distribution ....................................................................................................................40
Mean and variance of a Poisson distribution .......................................................................................................................40

Index ............................................................................................................................................................ 41

2 S3 JUNE 2016 SDB

1 Combinations of random variables
Expected mean and variance for X ± Y
Reminder
For any two random variables X and Y
E[aX] = aE[X] and Var[aX] = a2 Var[X]
E[X + Y] = E[X] + E[Y] and E[X – Y] = E[X] – E[Y]
and for two independent random variables
Var[X + Y] = Var[X] + Var[Y] and Var[X – Y] = Var[X] + Var[Y].

Combining independent normal random variables Y

If X1 and X2 are independent normal random variables

X1 ~ N(µ1, σ12) and X2 ~ N(µ2, σ22)
then X1 + X2 and X1 – X2 are also normal random variables
X1 + X2 ~ N(µ1 + µ2, σ12 + σ22)
and X1 – X2 ~ N(µ1 – µ2, σ12 + σ22)

Example: X1 and X2 are independent normal random variables

X1 ~ N(21, 12) and X2 ~ N(9, 6).
Find the expected mean and standard deviation of X1 – 2X2.

Solution: E[X1] =21, Var[X1] = 12 and E[X2] = 9, Var[X2] = 6

⇒ E[2X2] = 2E[X2] = 2 × 9 = 18
and Var[2X2] = 22 × Var[X2] = 4 × 6 = 24
⇒ E[X1 – 2X2] = E[X1] – E[2X2] = 21 – 18 = 3
and Var[X1 – 2X2] = Var[X1] + Var[2X2] = 12 + 24 = 36
⇒ the expected mean and standard deviation of X1 – 2X2 are 3 and √36 = 6. Answer

S3 JUNE 2016 SDB 3

Example: The weights of empty coffee jars are normally distributed with mean 0.1 kg and standard
deviation 0.02 kg. The weight of coffee in the jars is normally distributed with mean 1 kg and
standard deviation 0.06kg.
Find the distribution of 12 full jars of coffee.
What is the probability that 12 full jars weigh more than 13⋅5 kg?
Solution: Let X1, X2, ... X12 be the weights of 12 empty jars and Y1, Y2, ... Y12 be the weights of coffee
in the jars. X ~ N(0⋅1, 0⋅022) and Y ~ N(1, 0⋅062).
Let W be the total weight of 12 full jars then W = X1 + X2 + ... + X12 + Y1 + Y2 + ... + Y12.
Then E[W] = 12 E[X] + 12 E[Y] = 12 × 0⋅1 + 12 × 1 = 13⋅2
and, assuming independence,
Var[W] = 12 Var[X] + 12 Var[Y] = 12 × 0⋅022 + 12 × 0⋅062 = 0⋅048.
As we are combining normal distributions
the distribution for 12 full jars is N(13.2, 0⋅048). Answer

The probability that 12 full jars weigh more than 13⋅5 kg is

 13 ⋅ 5 − 13 ⋅ 2 
1− Φ   = 1 − Φ (1 ⋅ 37) = 0 ⋅ 0853 to 3 S.F. Answer.
 0 ⋅ 048 

2 Sampling
Methods of collecting data
Taking a census
A census involves observing every member of a population
and is used if
the size of the population is small
or if extreme accuracy is required.
Advantages
it should give a completely accurate result, a full picture.
Disadvantages
very time consuming and expensive
it cannot be used when testing process destroys article being tested
information is difficult to process because there is so much of it.

Sampling
Sampling involves observing or testing a part of the population.
It is cheaper but does not give such a full picture.
The size of the sample depends on the accuracy desired (for a varied population a large sample will be
required to give a reasonable accuracy).
4 S3 JUNE 2016 SDB
Simple random sampling
Every member of the population must have an equal chance of being selected.

Using random number tables

To take a simple random sample of size n from a population of N sampling units first make a list and
give each member of the population a number. Then use random number tables to select the sample.
We ignore any numbers which do not refer to a member of the population – for example using three
figure random numbers for a population numbered from 001 to 659 we would ignore numbers from 660
to 999.
Also we ignore the second occurrence of the same number.
Advantages
the numbers are truly random and free from bias
it is easy to use
each member has a known equal chance of selection
Disadvantages
it is not suitable when the sample size is large.
Lottery sampling
A sampling frame is needed – identifying each member of the population. The name or number of each
member is written on a ticket (all the same size, colour and shape), and the tickets are all put in a
container which is then shaken. Tickets are then drawn without replacement.
Advantages
the tickets are drawn at random.
it is easy to use.
each ticket has a known chance of selection (considered as constant as long as the sample size is
much smaller than the total number of tickets).
Disadvantages
it is not suitable for a large sample
a sampling frame is needed.

Systematic sampling
First make an ordered list, and divide into equal groups each of size 50 (or??).
Second select every 50th (or ??) member from the list.
In order to make sure that the first on the list is not automatically selected random number tables must be
used to select the member in the first group, then select every 50th (or ??) after that.
Used when the population is too large for simple random number sampling.
Advantages
simple to use
suitable for large samples
Disadvantages
only random if the ordered list is truly random.
it can introduce bias

S3 JUNE 2016 SDB 5

Stratified sampling
First divide the population into exclusive (distinct) groups or strata and then select a sample so that the
proportion of each stratum in the sample equals the proportion of that stratum in the population.

Example: How would you take a stratified sample of 50 children from a school of 500 pupils divided
as follows:
Boys Girls
Upper sixth 30 40
Lower sixth 30 30
Fifth form 70 60
Fourth form 60 70
Third form 50 60

Solution: As 50 is 1/10 of the total population, 1/10 of each stratum should be selected in the sample.
Thus the sample would comprise
Boys Girls
Upper sixth 3 4
Lower sixth 3 3
Fifth form 7 6
Fourth form 6 7
Third form 5 6

and simple random number sampling would be used within each stratum.
Used when
the sample is large
the population divides naturally into mutually exclusive groups.
Advantages
it can give more accurate estimates (or a more representative picture) than simple random number
sampling when there are clear strata present.
It reflects the population structure.
Disadvantages
within the strata the problems are the same as for any simple random sample
if the strata are not clearly defined they may overlap.

Sampling with and without replacement

Simple random sampling is sampling without replacement in which each member of population can be
selected at most once.
In sampling with replacement each member of the population can be selected more than once: this is
called unrestricted random sampling.

6 S3 JUNE 2016 SDB

Quota sampling
This is a non-random method.
First decide on groups into which the population is divided and a number from each group to be
interviewed to form quotas.
Then go out and interview and enter each result into the relevant quota.
If someone refuses to answer or belongs to a quota which is already full then ignore that persons reply and
continue interviewing until all quotas are full.
Used when it is not possible to use random methods - for example when the whole population is not
known (homeless in a big city).
Advantages
can be done quickly as a representative sample can be obtained with a small sample size
costs are kept to a minimum
administration is fairly easy.
Disadvantages
it is not possible to estimate the sampling errors (as it is not a random process)
interviewer may not put into correct quota
non-responses are not recorded
it can introduce interviewer bias

Primary data
Primary data is data collected by or on behalf of the person who is going to use the data.
Advantages
collection method is known
accuracy is known
exact data needed are collected
Disadvantages
costly in time and effort

Secondary data
Secondary data is data not collected by or on behalf of the person who is going to use it. The data are
second-hand – e.g. government census statistics.
Advantages
cheap to obtain
large quantity available (e.g. internet)
much has been collected year on year and can be used to plot trends
Disadvantages
collection method may not be known
accuracy may not be known
it can be in a form which is difficult to handle
bias is not always recognised.

S3 JUNE 2016 SDB 7

3 Biased & unbiased estimators

Example: A bag contains a large number of coins, of which 25 are 2p coins and 35 are 5p coins.
(a) X is the value of a single coin draw from the bag. Find the expected mean of all coins in
the bag, µ = E[X].
Samples of size 3 are now drawn from the bag.
(b) Find the sampling distribution of and the expected value of (i) the median, and (ii) the
mean.

(c) (i) The median, Q2, is used as an estimator of the mean of all the coins, µ. Show that
Q2 is a biased estimator of µ, and find the bias.

(ii) The mean, 𝑋, is used as an estimator of the mean of all the coins, µ. Show that 𝑋 is
an unbiased estimator of µ.

(d) kQ2 is now used as an unbiased estimator of the mean of all the coins. Find the value of k.

Solution:

(a) µ = E[X] = ∑ 𝑥𝑖 𝑝𝑖 = 2
5
× 2 + 35 × 5 = 3 ∙ 8

(b) Sample Probability median mean

3
(2,2,2) �25� = 8
125
2 2
2 2
(2,2,5), (2,5,2), (5,2,2) 3 × �5� × 35 = 36
125
2 3
2
(2,5,5), (5,2,5), (5,5,2) 3 × 25 × �35� = 54
125
5 4
3 3 27
(5,5,5) �5 � = 125
5 5

Sampling distribution

(i) Median (ii) Mean

Q2 = xi pi xipi 𝑋 = xi pi xipi
8+36 88 8 16
2 125 125
2 125 125
54+27 405 36 108
5 125 125
3 125 125
493 54 216
125
4 125 125
27 135
5 125 125
475
125
493 475
⇒ E[Q2] = 125 = 3⋅944 E[𝑋] = 125 = 3⋅8

8 S3 JUNE 2016 SDB

 Unbiased estimator:
If X (usually found from a sample) is used to estimate the value of a population parameter,
t, then X is an unbiased estimator of t if E[X] = the true value of the parameter t.
Bias: If an estimator, X, is biased, then the bias is the difference between E[X] and the true value
of the parameter t.
(c) (i) The median Q2 is used as the estimate of the mean.
From part (a) we know that the true value of the mean µ is 3⋅8, and in part (b) we have
shown that E[Q2] = 3⋅944
⇒ Q2 is a biased estimator of µ,
and the bias is |E[Q2] – true value of µ| = |3⋅944 − 3⋅8| = 0⋅144
(ii) The mean 𝑋 is used as the estimate of the mean.
From part (a) we know that the true value of the mean µ is 3⋅8, and in part (b) we have
shown that E[𝑋] = 3⋅8
⇒ E[𝑋] = the true value of the mean
⇒ 𝑋 is an unbiased estimator of µ,

(d) If we now use kQ2 as an unbiased estimator of the mean value of all the coins.
493
E[kQ2] = k E[Q2] = 125 𝑘
475
But the true mean µ = 125

If kQ2 is an unbiased estimator of µ, E[kQ2] = true value of µ

493 475 475
⇒ 125
𝑘 = 125 ⇒ k = 493.

Example: A sample of size 3 is drawn from a binomial distribution B(10, 0⋅25) and the mean, 𝑋, is
calculated.
1
The probability of success, p, is estimated by 𝑝̂ = 10 𝑋. Show that 𝑝̂ is an unbiased estimator of p.

Solution: E[X] = np = 10 × 0⋅25 = 2⋅5

1
For a sample {X1, X2, X3}, 𝑋 = 3 (𝑋1 + 𝑋2 + 𝑋3 )
1 1
⇒ E� 𝑋 � = E �3 (𝑋1 + 𝑋2 + 𝑋3 )� = 3 (E[𝑋1 ] + E[𝑋2 ] + E[𝑋3 ])
1
⇒ E� 𝑋 � = 3 × 3 × 2 ∙ 5 = 2 ∙ 5 since E[𝑋𝑖 ] = E[𝑋] = 2 ∙ 5 , for i = 1, 2, 3
1 1 1
⇒ E[𝑝̂ ] = E �10 𝑋� = 10 E�𝑋� = 10 × 2 ∙ 5 = 0 ∙ 25, which is the true value of p
1
⇒ 𝑝̂ = 10
𝑋 is an unbiased estimator of p.

S3 JUNE 2016 SDB 9

 Example: A sample of size 4 is drawn from a continuous uniform distribution, U[3, β ]. The mean of
the sample, 𝑋, is calculated.
The upper limit, β, is estimated by 𝛽̂ = 2𝑋 − 3. Show that 𝛽̂ is an unbiased estimator of β.

1
Solution: E[X] = 2 (3 + β )
1
For a sample {X1, X2, X3, X4}, 𝑋 = 4 (𝑋1 + 𝑋2 + 𝑋3 + 𝑋4 )
1 1
⇒ E� 𝑋 � = E �4 (𝑋1 + 𝑋2 + 𝑋3 + 𝑋4 )� = 4 (E[𝑋1 ] + E[𝑋2 ] + E[𝑋3 ] + E[𝑋4 ])
1
⇒ E� 𝑋 � = 4 × 4 × E[𝑋] = E[𝑋] since E[𝑋𝑖 ] = E[𝑋], for i = 1, 2, 3, 4
1
⇒ E� 𝑋 � = 2
(3 + β )
1
⇒ E�𝛽̂� = E�2𝑋 − 3� = 2E� 𝑋 � − 3 = 2 × 2 (3 + β ) − 3 = 𝛽,

⇒ E�𝛽̂� = 𝛽 which is the true value of the (unknown) upper limit, β.

⇒ 𝛽̂ = 2𝑋 − 3 is an unbiased estimator of β.

There is more on biased and unbiased estimators in the Appendix.

10 S3 JUNE 2016 SDB

Unbiased estimators of µ and σ 2

Estimating µ and σ 2 from a sample

We usually do not know the mean, µ, and the variance, σ 2, of a population.

To estimate these values we take a sample {X1, X2, X3, …, Xn} of size n and calculate
1
the sample mean, 𝑋� = 𝑛
∑ 𝑋𝑖 , and
1 1
(𝑠𝑑)𝑥 2 = 𝑛
∑ 𝑋𝑖 2 − 𝑋� 2 =
𝑛
∑(𝑋 − 𝑋�)2
these can be compared with the formulae for population variance from the S1 module.

It can be shown that E� 𝑋 � = the true value of µ

⇒ 𝑋� = 𝜇̂ is an unbiased estimator of the population mean µ.
𝑛−1
It can be shown that E�(𝑠𝑑)𝑥 2 � = 𝑛
σ2 ….… I

⇒ (𝑠𝑑)𝑥 2 is a biased estimator of σ 2.

𝑛 𝑛 𝑛−1
I ⇒ E�𝑛−1 (𝑠𝑑)𝑥 2 � = 𝑛−1 × 𝑛
σ 2 = σ 2, the true value of the variance
𝑛
⇒ 𝑛−1
(𝑠𝑑)𝑥 2 = 𝜎� 2 is an unbiased estimator of the population variance, σ 2.

(Proofs of these results are given in the Appendix.)

Note: the Edexcel course uses both the letters S2 and sx2 to mean the unbiased estimate of σ 2.
Also, the term Sample Variance is used to denote the unbiased estimate of σ 2, the variance of the
population.
1 1
In these notes I shall always think of the variance, (𝑠𝑑)𝑥 2, as 𝑛
∑ 𝑋𝑖 2 − 𝑋� 2 =
𝑛
∑(𝑋 − 𝑋�)2
To find S2 or sx2, the unbiased estimator for σ 2:–

𝒏
Calculate (𝒔𝒅)𝒙 𝟐, and then multiply by 𝒏−𝟏

S3 JUNE 2016 SDB 11

 Example:
The weights of a sample of five chocolate bars produced by a machine were 56, 53, 57, 51 and 54
grams. Find unbiased estimators for the weight of all chocolate bars produced by that machine.

Solution:
X 𝑋 − 𝑋� (𝑋 − 𝑋�)2
56 1⋅8 3⋅24
53 -1⋅2 1⋅44
57 2⋅8 7⋅84
51 -3⋅2 10⋅24
54 -0⋅2 0⋅04
271 22⋅8

1 271
⇒ 𝑋� = 𝑛
∑𝑋 =
5
= 54⋅2
1 22∙8
⇒ (𝑠𝑑)𝑥 2 = 𝑛 ∑(𝑋 − 𝑋�)2 = 5
= 4 ∙ 56
𝑛 5
⇒ 𝜎� 2 = 𝑛−1
(𝑠𝑑)𝑥 2 = 4
× 4 ∙ 56 = 5 ∙ 7

Answer Unbiased estimators for the mean and variance of all chocolate bars are 54⋅2 grams
and 5⋅7 grams2.

Example: The volume of water in each of a sample of 14 litre bottles of water from a day’s
production is taken. The results are shown below, in ml.
1023, 1019, 1004, 1011, 1023, 1014, 1017, 1020, 1020, 1010, 1025, 1007, 1016, 1019
Find unbiased estimates for the mean and variance of all bottles produced on that day.

14228
Solution: First find the sample mean, 𝑋�, = 14
= 1016⋅286….
(finding 𝑋 − 𝑋� each time) would give unpleasant arithmetic,
1
so use (𝑠𝑑)𝑥 2 = 𝑛
∑ 𝑋 2 − 𝑋� 2
∑ 𝑋 2 = 14460232
1446032 14228 2
⇒ (𝑠𝑑)𝑥 2 = 14
− � 14
� = 37⋅06122…
𝑛 14
⇒ S2 = sx2 = 𝑛−1
(𝑠𝑑)𝑥 2 = 14−1
× 37 ∙ 06122 … = 39 ∙ 91209 …

Answer Unbiased estimators for the mean and variance of the whole day’s production are
1016⋅3 ml and 39⋅91 ml2.

12 S3 JUNE 2016 SDB

 Example: The weights of a sample of 15 packets of biscuits are recorded and give the following
results.
Σ X = 3797 grams, and Σ X 2 = 973692.
Find unbiased estimators for the mean and variance of all biscuits produced by this
process.
3797
Solution: µ = 𝑋� = 15
= 253 ∙ 133 … = 252 ∙ 1 grams.
1 973692 3797 2
(𝑠𝑑)𝑥 2 = 𝑛
∑ 𝑋 2 − 𝑋� 2 =
15
− � 15
� = 836⋅3156…
15
⇒ 𝜎� 2 = 14 × 836 ∙ 3156 … = 896 ∙ 0524 … = 896 ∙ 1 grams2.

Answer Unbiased estimators are 𝜇̂ =252⋅1 g, and 𝜎� 2 = 896⋅1 g2.

Example: The lengths of 10 rods are measured, and the sample has mean, 𝑋� = 26⋅7 cm and variance
s2 = 76⋅9 cm2. An eleventh rod has length 30 cm.
Find (a) the mean and (b) the variance of the sample of 11 rods.
Solution: (a) With the sample mean there are no complications.
10 10
1
For 𝑛 = 10, 𝑋�10 = � 𝑋𝑖 = 26 ∙ 7 ⇒ � 𝑋𝑖 = 267
10
𝑖=1 𝑖=1
11 11
1 297
For 𝑛 = 11, � 𝑋𝑖 = 267 + 30 = 297 ⇒ X11 = � 𝑋𝑖 = = 27 𝑐𝑚
11 11
𝑖=1 𝑖=1

(b) WARNING: The question refers to the variance of the sample, which means the unbiased
estimate of the variance of the population.
10 9
𝑠10 2 = 76⋅9 = (10−1)
× (sd)10 2 ⇒ (sd)10 2 = 10
× 76 ∙ 9 = 69 ∙ 21
10
2 1 2
⇒ (sd)10 = � 𝑋𝑖 2 − 𝑋�10 = 69 ∙ 21
10
𝑖=1
10

⇒ � 𝑋𝑖 2 = 692 ∙ 1 + 10 × 26 ∙ 72 = 7821
𝑖=1
11

with extra rod ⇒ � 𝑋𝑖 2 = 7821 + 302 = 8721

𝑖=1
11
2 1 2 8721 702
⇒ (sd)11 = � 𝑋𝑖 2 − 𝑋�11 = − 272 =
11 11 11
𝑖=1
11 11 702
⇒ 𝑠11 2 = (sd)11 2 = × = 70 ∙ 2 𝑐𝑚2
(11 − 1) 10 11

For 11 rods, sample mean is 27 cm, and sample variance is 70⋅2 cm2.

S3 JUNE 2016 SDB 13

4 Confidence intervals and significance tests

Sampling distribution of the mean

X is a random variable draw from a population with mean µ and standard deviation σ.
X 1 + X 2 + ... + X n
If {X1, X2, ... , Xn} is a random sample of size n with mean 𝑋 =
n
then E[Xi] = µ, and Var[Xi] = σ 2, for i = 1, 2, 3, …, n

and the expected mean of the population of sample means is

 X + X 2 + ... + X n 
E� 𝑋 � = E  1 
 n 

=
1
(E[ X 1 ] + E[ X 2 ] + ... E[ X n ]) = 1 (µ + µ + ... µ )
n n
= µ.

Also the expected variance of the population of sample means is

 X + X 2 + ... + X n 
Var� 𝑋 � = Var  1  assuming that all the Xi are independent
 n 
σ2
1
(
= 2 (Var[ X 1 ] + Var[ X 2 ] + ... Var[ X n ]) = 2 σ 2 + σ 2 ... σ 2
1
) =
𝑛𝜎2
𝑛2
=
n
n n

This means that if very many samples were taken and the mean of each sample calculated then the
σ2
mean of these means would be µ and the variance of these means would be .
n
It can also be shown that the sample means form a Normal distribution (provided that n is ‘large
enough’).
We can then say that for samples drawn from a population with mean µ and variance σ 2,
𝜎2
the sampling distribution of the mean is N(µ, ).
𝑛

14 S3 JUNE 2016 SDB

Central limit theorem and standard error

The central limit theorem states that

If {X1, X2, ... , Xn} is a random sample of size n drawn from any population with mean µ and
variance σ 2 then the population of sample means
(i) has expected mean µ
σ2
(ii) has expected variance
n
(iii) forms a normal distribution if n is ‘large enough’.
 σ2
i.e. X ~ N µ , .
 n 
The central limit theorem is used for sampling when the sample size is ‘large’ (> about 50) as the
population of sample means is then approximately normal whatever the distribution of the original
population.
𝜎
The standard error of the sample mean is .
√𝑛

Example: A sample of size 50 is taken from a population of eggs with mean 23⋅4 grams and variance
36 grams2.
(i) Find the probability that a single egg weighs more than 25 grams.
(ii) Find the probability that the sample mean is larger than 25.
(iii) What assumptions did you make?
Solution:
(i) The weight of a single egg, X ∼ N(23⋅4, 62)
25−23∙4
⇒ P(X > 25) = Φ � 6
� = Φ(0 ∙ 27) = 0 ∙ 6064

(ii) µ = 23⋅4, σ 2 = 36
6 2
The sample mean 𝑋 ∼ N �23⋅4, � � �
√50
𝜎 6
⇒ standard error is = = 0⋅848528137
√𝑛 √50
𝑋� ~ 𝑁(23 ∙ 4, 0 ∙ 8485 …2 )
25−23∙4
⇒ P(𝑋� > 25) = 1 – Φ �0∙8485…�

= 1 – Φ(1⋅89) = 1 – 0⋅9706 (from Normal tables)

= 0⋅0294.

(iii) We have assumed the Central Limit Theorem: in particular that the sample means form a
normal distribution .

S3 JUNE 2016 SDB 15

Confidence intervals
Central Limit Theorem Example
Example:
A biscuit manufacturer makes packets of biscuits with a nominal weight of 250 grams. It is known
that over a long period the variance of the weights of the packets of biscuits produced is
25 grams2. A sample of 10 packets is taken and found to have a mean weight of 253⋅4 grams.
Find 95% confidence limits for the mean weight of all packets produced by the machine.

Solution:
First assume that the machine is still producing packets with the same variance, 25.
Suppose that the mean weight of all packets of biscuits is µ grams then the population of all
packets has mean µ and standard deviation 5.

From the central limit theorem we can assume that the sample means form an approximately
σ 5
normal population with mean µ and standard error (standard deviation) = = 1⋅5811
n 10
f(x)

95% of the samples will have a mean in the

region
–1⋅96 < Z < 1⋅96 95%

We assume that the mean of this sample, 253⋅4,

x
lies in this region −1.96 1.96

253∙4−𝜇
⇒ –1⋅9600 < 1∙5811
< 1⋅9600
253∙4−𝜇 253∙4−𝜇
⇒ –1⋅9600 < 1∙5811
and 1∙5811
< 1⋅9600

⇒ µ – 1⋅9600 × 1⋅5811 < 253⋅4 and 253⋅4 < µ + 1⋅9600 × 1⋅5811

⇔ µ < 253⋅4 + 1⋅9600 × 1⋅5811 and 253⋅4 – 1⋅9600 × 1⋅5811 < µ
⇔ 253⋅4 – 1⋅9600 × 1⋅5811 < µ < 253⋅4 + 1⋅9600 × 1⋅5811
⇔ 250⋅3 < µ < 256⋅5

This means that 95% of the samples will give an interval which contains the mean
and we say that [250⋅3 g, 256⋅5 g] is a 95% confidence interval for µ.

This means that there is a 0⋅95 probability that this interval contains the true mean.
It does not mean that there is a probability of 0⋅95 that the true mean lies in this interval - the true
mean is a fixed number, and either does or does not lie in the interval so the probability that the
true mean lies in the interval is either 1 or 0.
16 S3 JUNE 2016 SDB
In practice we go straight to the last line of the example:
𝜎
95% confidence limits are µ ± 1⋅9600 × since P(Z –1⋅9600 < z < 1⋅9600) = 0⋅95
√𝑛
tables give P(Z > 1⋅9600) = 0⋅025
𝜎
90% confidence limits are µ ± 1⋅6449 × since P(Z –1⋅6449 < z < 1⋅6449) = 0⋅90
√𝑛
tables give P(Z > 1⋅6449) = 0⋅05

Other confidence limits can be found using the Normal Distribution tables.

Example: A sample of 64 packets of cornflakes has a mean weight X = 510 grams and a variance
S 2 = 36 grams2. Find 90% confidence limits for the mean weight of all packets.
(Note that the ‘sample variance’ is taken as the unbiased estimate of σ 2.)

Solution: We assume that the sample variance = the variance of the population of all packets
⇒ S 2 = 36 = σ 2.
Now find standard deviation (standard error) of the sampling distribution of the mean (population
σ 6
of sample means), standard error = = = 0 ⋅ 75
n 64

For 90% confidence limits z = ± 1⋅6449 (remember to use the 4 D.P. tables after the Normal Dist. tables),
using the sample mean X = 510 grams
⇒ 90% confidence limits are 510 ± 1⋅6449 × 0.75 = 510 ± 1⋅234
⇒ a 90% confidence interval is [508⋅8, 511⋅2] to 4 S.F.

Note that we have assumed that the unbiased estimate, S 2 (=36), is the actual variance, σ 2, of the
population.
This is a reasonable assumption as the number in the sample, 64, is large and the error introduced is
therefore small.

S3 JUNE 2016 SDB 17

Significance testing– variance of population known
Mean of normal distribution
Example:
A machine, when correctly set, is known to produce ball bearings with a mean weight of 84 grams
with a standard deviation of 5 grams. The production manager decides to test whether the machine
is working correctly and takes a sample of 120 ball bearings. The sample has mean weight
83.2 grams. Would you advise the production manager to alter the setting of his machine? Use a
5% significance level.
Solution:
1) H0: µ = 84 grams
2) H1: µ ≠ 84 grams ⇒ 2 tail test
(Note that the machine is not working correctly if the test result is too high or too low)
3) 5% Significance level
4) The Test
We assume that the machine is still working with a standard deviation of σ = 5 g.
From H0, the mean weight of all ball bearings is assumed to be µ = 84 g.
These are the parameters for the population of all ball bearings.
We want to test a sample mean and therefore need the mean and standard deviation of the
population of sample means (the sampling distribution of the sample mean, 𝑋�).
Expected mean of the sample means = µ = 84 g. and
σ 5
expected standard deviation of the sample means = standard error = = = 0 ⋅ 456435...
n 120
.
We have an observed mean of 83⋅2
For a two-tailed test at 5%, we take 2⋅5% at
each end
P(𝑋� < 83⋅2) = Φ ( 0⋅83456435...
⋅2 −84
) =Φ( −1 ⋅ 7527)
= (1 – Φ(1⋅75)) = 0⋅0401 x

= 4⋅01% > 2⋅5%

and so not significant at the 5% level. 83⋅2 84

5) Conclusion
Do not reject H0 at the 5% level and advise the production manager that there is evidence that
he should not change his setting, or that there is evidence that the machine is working correctly,
etc.

18 S3 JUNE 2016 SDB

 Difference between means of normal distributions
Suppose that X and Y are two independent random variables from different normal distributions –
X ~ N(µx, σx2) and Y ~ N(µy, σy2).
If samples of sizes nx and ny are drawn from these populations
then the distributions of the sample means, 𝑋� and 𝑌� will be normal
𝜎2 𝜎𝑦 2
𝑋� ~ 𝑁 �𝜇𝑥 , 𝑛𝑥 � and 𝑌� ~ 𝑁 �𝜇𝑦 , �
𝑥 𝑛𝑦

𝐸[𝑋� − 𝑌�] = 𝐸[𝑋�] − 𝐸[𝑌�] and Var[𝑋� − 𝑌�] = Var[𝑋�] + Var[𝑌�]

⇒ the differences of the sample means, 𝑋� − 𝑌� , will be normal
𝜎𝑥 2 𝜎𝑦 2
(𝑋� − 𝑌�) ~ 𝑁 �𝜇𝑥 − 𝜇𝑦, + �
𝑛𝑥 𝑛𝑦

Example: The weights of chocolate bars produced by two machines, A and B, are known to be
normally distributed with variances σA2 = 4 and σB2 = 3 grams2. Samples are taken from each
machine of sizes nA = 25 and nB = 16 which have means X A = 123 ⋅ 1 and X B =124 ⋅ 4
grams. Is there any evidence at the 5% significance level that the bars produced by machine B are
heavier than the bars produced by machine A?

Solution:
Suppose that the mean weights for all bars from the two machines are µA and µB
H0: µA = µB
H1: µB > µA one-tail test at 5% level

The test statistic is the observed difference between sample means,

X B – X A = 124⋅4 – 123⋅1 = 1⋅3,

and we must find the variance of this population of differences of sample means (the sampling
distribution of differences of sample means).

Consider the population of differences of sample means X B – X A .

Firstly, for the population of sample means for machine B
σ B2 3
expected variance Var[ X B ] = =
n B 16

S3 JUNE 2016 SDB 19

 and secondly, for the population of sample means for machine A
σ A2 4
expected variance Var[ X A ] = =
nA 25

and so for the population of differences of sample means

expected mean = E[ X B – X A ] = µA - µB = 0 (from H0)

and Var[ X B – X A ] = Var[ X B ] + Var[ X A ]

𝜎𝐵 2 𝜎𝐴 2
= + = 3/16 + 4/25 = 0⋅3475.
𝑛𝐵 𝑛𝐴

The observed difference, the test statistic, is 124⋅4 – 123⋅1 = 1⋅3

and the standard error is 0 ⋅ 3475)

The Central Limit Theorem tells us that we have a Normal distribution
 1⋅ 3 − 0 
⇒ P(difference > 1⋅3) = 1 – Φ   = 1 − Φ (2 ⋅ 2053)
 0 ⋅ 3475) 
= 1 − Φ (2 ⋅ 20) = 1– 0⋅9861 = 0⋅0134
= 1⋅34% < 5%
⇒ significant at 5% level so reject H0 and conclude that there is evidence that machine B is
producing bars of chocolate with a heavier mean weight than machine A.

Fortunately (!) the formula for testing the difference between sample means
𝑋�−𝑌� – (𝜇𝑥 −𝜇𝑦 )
Z = is in your formula booklet.
𝜎 2 𝜎𝑦 2
� 𝑥 +
𝑛𝑥 𝑛𝑦

20 S3 JUNE 2016 SDB

Significance testing – variance of population NOT known, large sample

When the variance of the population, σ 2, is not known and when the sample is large, we assume that the
variance of the sample (meaning the unbiased estimate of σ 2), S 2, is the variance of the population, σ2.
As the sample is large, the error introduced is small.

Mean of normal distribution

Example: A machine usually produces steel rods with a mean length of 25⋅4 cm. The production
manager wants to test 80 rods to see whether the machine is working correctly. The sample has
mean 25⋅31 cm and variance 0⋅332 cm2. Advise the production manager, using a 5% level of
significance.

Important assumption

� 𝟐 , the unbiased estimate of the variance of the

The sample variance, S2, is taken as, 𝝈
population, σ , and we then assume that the population variance equals the unbiased estimate .
2

Solution:
H0: µ = 25⋅4.
H1: µ ≠ 25⋅4 two-tail test, 2⋅5% in each tail

We assume that population variance σ2 = the sample variance S2 = 0⋅332

⇒ σ = 0⋅33
For the population of sample means (the sampling distribution of the sample means)
expected mean = 25⋅4 from hypothesis

σ 0 ⋅ 33
and standard error = = = 0⋅036895121.
n 80
The observed sample mean is 25.31 and for a two-tail test at 5% we consider
25∙31−25∙4
Φ� � = Φ(−2 ∙ 4393) = 1 − Φ(2 ∙ 44) = 0⋅0073 < 2⋅5%
0∙036895121
⇒ reject H0 and conclude that there is evidence that that the machine is not producing rods of
mean length 25⋅4 cm.

S3 JUNE 2016 SDB 21

Difference between means
Example:
A firm has two machines, A and B, which make steel cable. 40 cables produced by machine A
have a mean breaking strain of 1728 N and variance of 752 N2, whereas 65 cables produced by
machine B have a mean breaking strain of 1757 N and a variance of 632 N2. Is there any evidence,
at the 10% level, to suggest that machine B is producing stronger cables than machine A?
Solution:
Let μA and μB be the mean breaking strengths of all cables produced by machines A and B.
1) H0: μA = μB
2) H1: μB > μA 1 tail test

3) Significance Level 10%.

4) The Test
For Machine A
We assume that the population variance, σA2 = the sample variance, S A 2 = 752
σ A2 752
⇒ variance of sample means Var[ X A ] = = = 140 ⋅ 625 .
n 40
For Machine B
We assume that the population variance, σB2 = the sample variance, S B 2 = 632
σ B 2 632
⇒ variance of sample means Var[ X B ] = = = 61 ⋅ 0615... .
n 65

For differences in sample means X B – X A

Expected mean = 0 from hypothesis

Expected variance is Var[ X B – X A ] = Var[ X B ] + Var[ X A ]

= 140⋅625 + 61.0615… = 201⋅6865…
⇒ standard deviation or standard error = √201 ∙ 6865 … = 14⋅2016….
We have an observed difference in means,

test statistic, X B – X A = 1757 – 1728 = 29

and for a 1-tail test that B is stronger
we need the area to the right of 29 mean is treated as continuous, so do not use 28.5

29 − 0
= 1 – Φ� � = 1 – Φ(2 ∙ 04) = 0⋅0207 < 10%
14∙2016…
which is significant at 10%.

5) Conclusion
Reject H0 at the 10% level and conclude that there is evidence that machine B produces
cables with a greater mean strength than machine A.

22 S3 JUNE 2016 SDB

5 Goodness of fit, χ2 test
General points
 The χ2 test can only be used to test two lists of frequencies – the observed and the expected
frequencies calculated from the hypothesis.
 The expected frequencies do not need to be integers (give 2 D.P.)
(Oi − Ei ) 2
 χ2 = ∑ E , where Oi and Ei are the observed and expected frequencies.
i

 If the expected frequency for a class is less than 5, then you must group this class with the next
class (or two …).
 The number of degrees of freedom, ν, is
the number of cells (after grouping if necessary) minus the number of linear equations connecting
the frequencies.

Discrete uniform distribution

Example: A die is rolled 300 times and the frequency of each score recorded.
Score: 1 2 3 4 5 6
Frequency: 43 49 54 57 46 51
Test whether the die is fair at the 2⋅5% level of significance.
Solution: H0: The die is fair, the probability of each score is 1/6.
H1: The die is not fair, the probability of each score is not 1/6.
The expected frequencies are all 1/6 × 300 = 50 and we have

Score Observed Expected (Oi − Ei ) 2

frequency frequency Ei

1 43 50 0⋅98
2 49 50 0⋅02
3 54 50 0⋅32
4 57 50 0⋅98
5 46 50 0⋅72
6 51 50 0⋅02
Totals 300 300 3⋅04

⇒ χ2 = 3⋅04
and ν = number of degrees of freedom = n – 1 = 6 – 1 = 5
since the total is a linear equation connecting the frequencies and is fixed.

From tables we see that χ 5 (2 ⋅ 5%) = 12 ⋅ 832 > 3 ⋅ 04 , so our observed result is not significant.
2

We do not reject H0 and conclude that the die is fair.

S3 JUNE 2016 SDB 23
Continuous uniform distribution
This is very similar to the discrete uniform distribution – pay attention to the class boundaries and find the
expected frequencies.

Binomial distribution
For H0 The Binomial distribution is a good fit
we use the mean of the Observed frequencies to calculate the Expected frequencies, and so both Oi and Ei
give the same mean and total: thus there are 2 linear equations connecting the frequencies and ν = n – 2
but For H0 The Binomial distribution, B(30, 0⋅3), is a good fit
the means using Oi and Ei will be different: thus there is only 1 linear equation, the total, connecting the
frequencies and so ν = n – 1.

Poisson distribution
For H0 The Poisson distribution is a good fit
we use the mean of the Observed frequencies to calculate the Expected frequencies, and so both Oi and Ei
give the same mean and total: thus there are 2 linear equations connecting the frequencies and ν = n – 2
but For H0 The Poisson distribution, Po(3), is a good fit
the means using Oi and Ei will be different: thus there is only 1 linear equation, the total, connecting the
frequencies and so ν = n – 1.
Example: A switchboard operator records the number of new calls in 69 consecutive one-minute
periods in the table below.
number of calls 0 1 2 3 4 5 ≥6
frequency 6 9 11 15 13 9 6
a) Say why you think that a Poisson distribution might be suitable.
b) Find the mean and variance of this distribution. Do these figures support the view that they
might form a Poisson distribution?
c) Test the goodness of fit of a Poisson distribution at the 5% level.
Solution:
a) Telephone calls are likely to occur singly, randomly, independently and uniformly which
are the conditions for a Poisson distribution.
b) Treating ≥ 6 as 7 we calculate the mean and variance

2
x f xf xf
0 6 0 0
1 9 9 9
2 11 22 44
3 15 45 135
4 13 52 208
5 9 45 225
7 6 42 294
69 215 915

24 S3 JUNE 2016 SDB

 ⇒ mean = 215
/69 = 3⋅12
915
and variance = /69 – (215/69)2 = 3⋅55.
From these figures we can see that the mean and variance are approximately equal: since the
mean and variance of a Poisson distribution are equal this confirms the view that the
distribution could be Poisson.

c) H0: The Poisson distribution is a suitable model

H1: The Poisson distribution is not a suitable model.
λ r e−λ
The Poisson probabilities can be calculated from P(r) = where λ = 3⋅12, and the
r!
expected frequencies by multiplying by N = 69.
Note that the probability for ≥ 6 is found by adding the other probabilities and subtracting
from 1.

x O p E O E (O − E ) 2
(grouped) (grouped)
E
0 6 0⋅044337 3⋅059234
1 9 0⋅138151 9⋅532395 15 12⋅59 0⋅461326
2 11 0⋅215235 14⋅8512 11 14⋅85 0⋅998148
3 15 0⋅223553 15⋅42515 15 15⋅43 0⋅011983
4 13 0⋅174145 12⋅01597 13 12⋅02 0⋅079900
5 9 0⋅108525 7⋅488214 9 7⋅49 0⋅304419
≥6 6 0⋅096056 6⋅627836 6 6⋅63 0⋅059864
69 69 69.01 1.915641

The expected frequency for x = 0 is 3.06 < 5 so it has been grouped with x = 1.
Thus we have n = 6 classes (after grouping) and ν = n – 2 = 4
and χ 42 (5%) = 9.488 .
We have calculated χ2 = 1.92 < 9.488 which is not significant so we do not reject H0
and conclude that the Poisson distribution is a suitable model.

S3 JUNE 2016 SDB 25

The normal distribution

For H0 The Normal distribution is a good fit

we use the mean and variance of the Observed frequencies to calculate the Expected frequencies, and so
both Oi and Ei give the same mean, variance and total: thus there are 3 linear equations connecting the
frequencies and ν = n – 3
but For H0 The Normal distribution, N(14, 32), is a good fit
the means and variances using Oi and Ei will be different: thus there is only 1 linear equation, the total,
connecting the frequencies and so ν = n – 1.

Example: The sizes of men’s shoes purchased from a shoe shop in one week are recorded below.
size of shoe ≤6 7 8 9 10 11 ≥ 12
number of pairs 14 19 29 45 40 21 7
Is the manager’s assumption that the normal distribution is a suitable model justified at the 5%
level?

Solution: H0: The normal distribution is a suitable model

H1: The normal distribution is not a suitable model.

The total number of pairs, mean and standard deviation are calculated to be 175, 8.886 and
1.713 (taking ≤ 6 as 5 and ≥ 12 as 12)
Remembering that size 8 means from 7.5 to 8.5 we need to find the area between 7.5 and 8.5 and
multiply by 175 to find the expected frequency for size 8, and similarly for other sizes.

x x−m Φ(z) class area = p E = 175p O (O − E ) 2

z=
s E
6.5 -1.39 0.082 < 6.5 0.082 14.4 14 0.01
7.5 -0.81 0.209 6.5 to 7.5 0.209 – 0.082 = 0.127 22.2 19 0.46
8.5 -0.23 0.409 7.5 to 8.5 0.409 – 0.209 = 0.200 35.0 29 1.03
9.5 0.36 0.641 8.5 to 9.5 0.641 – 0.409 = 0.232 40.6 45 0.48
10.5 0.94 0.826 9.5 to 10.5 0.826 – 0.641 = 0.185 32.4 40 1.78
11.5 1.53 0.937 10.5 to 11.5 0.937 – 0.826 = 0.111 19.4 21 0.13
> 11.5 1 – 0.937 = 0.063 11.0 7 1.45
5.34

n = 7 classes & 3 linear equations connecting the frequencies (N, m ,s) ⇒ ν = n – 3 = 4.

χ 42 (5%) = 9.488 and we have calculated χ2 = 5.34 < 9.488 and so we do not reject H0 and
therefore conclude that the normal distribution is a suitable model.
26 S3 JUNE 2016 SDB
Contingency tables

For a 5 × 4 table in which the totals of each row and column are fixed the ‘?’ cells represent the degrees
of freedom since if we know the values of the ?s the frequencies in the other cells can now be calculated

A B C D E totals
W ? ? ? ? 
X ? ? ? ? 
Y ? ? ? ? 
Z 
totals      

Thus there are (5 –1) × (4 – 1) = 12.

Generalising we can see that for an m × n table the number of degrees of freedom is (m – 1)(n – 1).

Example: Natives of England, Africa and China were classified according to blood group giving the
following table.
O A B AB
English 235 212 79 83
African 147 106 30 51
Chinese 162 135 52 43

Is there any evidence at the 5% level that there is a connection between blood group and
nationality?

Solution: H0: There is no connection between blood group and nationality.

H1: There is a connection between blood group and nationality.

First redraw the table showing totals of each row and column

O A B AB totals
English 235 212 79 83 609
African 147 106 30 51 334
Chinese 162 135 52 43 392
totals 544 453 161 177 1335

S3 JUNE 2016 SDB 27

 Now we need to calculate the expected frequency for English and group O. There are 609 English
and 1335 people altogether so 609/1335 of the people are English, and from H0 we know that there
is no connection between blood group and nationality, so there should be 609/1335 of those with
group O who are also English
609 609 × 544
⇒ expected frequency for English and group O is × 544 = = 248.2
1335 1335
this can become automatic if you notice that you just multiply the totals for the row and column
concerned and divide by the total number

O A B AB totals
English 609×544
1335
= 248.2 609× 453
1335
= 206.6 609×161
1335
= 73.4 609×177
1335
= 80.7 608.9

African 334× 544

1335
= 136.1 334× 453
1335
= 113.3 334×161
1335
= 40.3 334×177
1335
= 44.3 334

Chinese 392× 544

1335
= 159.7 392× 453
1335
= 133.0 392×161
1335
= 47.3 392×177
1335
= 52.0 392

totals 544 452.9 161 177 1335

The value of χ2 is calculated below

Observed Expected (O − E ) 2
frequency frequency
E
235 248.2 0.70
212 206.6 0.14
79 73.4 0.43
83 80.7 0.07
147 136.1 0.87
106 113.3 0.47
30 40.3 2.63
51 44.3 1.01
162 159.7 0.03
135 133.0 0.03
52 47.3 0.47
43 52.0 1.56
8.41
We have ν = (4 – 1)(3 – 1) = 6 degrees of freedom and χ 6 (5%) = 12.592 .
2

We have calculated χ2 = 8.41 < 12.592

⇒ do not reject H0 and therefore conclude that there is no connection between nationality and
blood group.
28 S3 JUNE 2016 SDB
6 Regression and correlation
Spearman’s rank correlation coefficient
Ranking and equal ranks
Ranking is putting a list of figures in order and giving each one its position or rank.
Equal numbers are given the average of the ranks they would have had if all had been different.

Example: Rank the following numbers: 45, 65, 76, 56, 34, 45, 23, 67, 65, 45, 81, 32.

Solution: First put in order and give ranks as if all were different: then give the average rank for
those which are equal.
Numbers: 81 76 67 65 65 56 45 45 45 34 32 23
Actual rank 1 2 3 4= 4= 6 7= 7= 7= 10 11 12
Rank (if all different) 1 2 3 4 5 6 7 8 9 10 11 12
average for equal ranks
4+5
2 = 4 12 7 +8+ 9
3 =8
Modified rank 1 2 3 4½ 4½ 6 8 8 8 10 11 12

You must now calculate the PMCC, not Spearman, using the modified ranks.

Spearman’s rank correlation coefficient

To compare two sets of rankings for the same n items, first find the difference, d, between each pair of
ranks and then calculate Spearman’s rank correlation coefficient
6∑ d 2
rs = 1 −
n(n 2 − 1)
This is the same as the product moment correlation coefficient of the two sets of ranks and so we know
that
rs = +1 means rankings are in perfect agreement,
rs = –1 means rankings are in exact reverse order,
rs = 0 means that there is no correlation between the rankings.

S3 JUNE 2016 SDB 29

Example: Ten varieties of coffee labelled A, B, C, ..., J were tasted by a man and a woman. Each
ranked the coffees from best to worst as shown.
Man: G H C D A E B J I F
Woman: C B H G J D I E F A
Find Spearman’s rank correlation coefficient.

Solution: Rank for each person, find d and then rs.

Coffee Man Woman d d2
A 5 10 -5 25
B 7 2 5 25
C 3 1 2 4
D 4 6 -2 4
E 6 8 -2 4
F 10 9 1 1
G 1 4 -3 9
H 2 3 -1 1
I 9 7 2 4
J 8 5 3 9
86

6∑ d 2 6 × 86
rs = 1 − = 1− = 0.521212 = 0.521 to 3 S.F.
n(n − 1)
2
10 × 99

Spearman or PMCC

Use of Spearman’s rank correlation coefficient

(i) Use when one, or both, sets of data are not from a normal population.
(ii) Use when the data does not have to be measured on scales or in units (probably not normal).
(iii) Use when data is subjective − e.g. judgements in order of preference (not normal).
(iv) Can be used if the scatter graph indicates a non-linear relationship between the variables, since
the PMCC is used to indicate linear correlation.
(v) Do not use for tied ranks (Spearman formula depends on non-tied ranks).

Use of Product moment correlation coefficient

(i) Use when ranks are tied – see above: modify the ranks and then use PMCC on the modified
ranks.
(ii) Use when both sets of figures are normally distributed (this will not be the case when using
ranks).
(iii) Use when the scatter diagram indicates a linear relationship between the variables – i.e. when
the points lie close to a straight line.

30 S3 JUNE 2016 SDB

Testing for zero correlation
N.B. the tables give figures for a ONE-TAIL test

Product moment correlation coefficient

PMCC tests to see if there is a linear connection between the variables. For strong correlation, the points
on a scatter graph will lie close to a straight line.
S xy
Reminder: PMCC = ρ =
S xx S yy

(∑ x ) 2
(∑ y ) 2
(∑ x )(∑ y ) .
S xx = ∑ x i − S yy = ∑ y i − S xy = ∑ x i y i −
2 i 2 i i i
where , ,
n n n

Example: The product moment correlation coefficient between 40 pairs of values is +0.52. Is there
any evidence of correlation between the pairs at the 5% level?

Solution: H0: There is no correlation between the pairs, ρ = 0.

H1: There is correlation, positive or negative, between the pairs, ρ ≠ 0, two-tail test

From tables for n = 40 which give one-tail figures, we must look at the 2.5% column and the critical
values are ±0.3120
The calculated figure is 0.52 > 0.3120 and so is significant
⇒ we reject H0 and conclude that there is some correlation (positive or negative) between
the pairs.

Spearman’s rank correlation coefficient

Spearman tests to see if there is a connection (or correlation) between the ranks.
Example: It is believed that a person who absorbs a drug well on one occasion will also absorb a
drug well on another occasion. Tests on ten patients to find the percentage of drug absorbed gave
the following value for Spearman’ rank correlation coefficient, rs = 0.634. Is there any evidence at
the 5% level of a positive correlation between the two sets of results.
Solution: H0: There is no correlation between the two sets of results, ρs = 0,
H1: There is positive correlation between the two sets of results, ρs > 0, one-tail test.
From the tables for n = 10 and a one-tail test the critical value for 5% is 0.5364.
The calculated value is 0.634 > 0.5364 which is significant
⇒ reject H0; conclude that there is evidence of positive correlation between the two sets of
results.
Note that this shows correlation between the ranks of the two sets of results.

S3 JUNE 2016 SDB 31

 Comparison between PMCC and Spearman
Example: A random sample of 8 students sat examinations in Geography and Statistics. The product
moment correlation coefficient between their results was 0⋅572 and the Spearman rank correlation
coefficient was 0⋅655.
(a) Test both of these values for positive correlation. Use a 5% level of significance.
(b) Comment on your results.

Solution:
(a) H0 : ρ = 0 ; H1 : ρ > 0
For the PMCC
the 5% Critical Value is 0⋅6215
0⋅572 < 0⋅6215 ⇒ not significant at %5
⇒ there is evidence that there is no positive correlation.
For Spearman’s rank correlation coefficient
the 5% Critical Value is 0⋅6429
0⋅655 > 0⋅6429 ⇒ significant at 5%
⇒ there is evidence of positive correlation.

(b) From the PMCC there is not enough evidence to conclude that as Statistics marks
increased Geography marks also increased
− i.e. conclude that the points on a scatter diagram do not lie close to a straight line.
From Spearman’s rank correlation coefficient there is evidence that students ranked
highly in Statistics were also ranked highly in Geography, or people with high scores in
Statistics also had high scores in Geography

32 S3 JUNE 2016 SDB

7 Appendix
Combining random variables

Let X and Y be random variables with probability distributions

X ∼ { X1, X2, X3, …, Xn} with probabilities (p1, p2, p3, …, pn), and
Y ∼ { Y1, Y2, Y3, …, Ym} with probabilities (q1, q2, q3, …, qm),
Then the random variable X + Y is all possible combinations xi + yj as i varies from 1 to n and j
varies from 1 to m.
Let P(xi + yj) = rij.
𝑚

Notice that � 𝑟𝑖𝑗 = 𝑟𝑖1 + 𝑟𝑖2 + 𝑟𝑖3 + … + 𝑟𝑖𝑚 = 𝑝𝑖

𝑗=1
𝑛

and, similarly, � 𝑟𝑖𝑗 = 𝑞𝑗

𝑖=1

E[X + Y] = E[X] + E[Y]

𝑛 𝑚

E[𝑋 + 𝑌] = � ��𝑥𝑖 + 𝑦𝑗 �𝑟𝑖𝑗 = � � 𝑥𝑖 𝑟𝑖𝑗 + � � 𝑦𝑗 𝑟𝑖𝑗

𝑖=1 𝑗=1 𝑖 𝑗 𝑗 𝑖

= � 𝑥𝑖 � 𝑟𝑖𝑗 + � 𝑦𝑗 � 𝑟𝑖𝑗
𝑖 𝑗 𝑗 𝑖
𝑛 𝑚

= � 𝑥𝑖 𝑝𝑖 + � 𝑦𝑗 𝑞𝑗
𝑖=1 𝑗=1

⟹ 𝐄[𝑿 + 𝒀] = 𝐄[𝑿] + 𝐄[𝒀]

S3 JUNE 2016 SDB 33

 Var[X + Y] = Var[X] + Var[Y]
In this case we take X and Y to be independent,
⇒ rij = P(xi and yj) = P(xi) × P(yj) = pi × qj.
Also notice that ∑ 𝑝𝑖 = ∑ 𝑞𝑗 = 1

Var[X + Y] = E[(X + Y)2] – (E[X + Y])2

2
= � ��𝑥𝑖 + 𝑦𝑗 � 𝑟𝑖𝑗 − (E[𝑋] + E[𝑌])2
𝑖 𝑗

= � � 𝑥𝑖 2 𝑝𝑖 𝑞𝑗 + 2 � � 𝑥𝑖 𝑦𝑗 𝑝𝑖 𝑞𝑗 + � � 𝑦𝑗 2 𝑝𝑖 𝑞𝑗 − ((E[𝑋])2 + 2E[𝑋]E[𝑌] + (E[𝑌])2 )

𝑖 𝑗 𝑖 𝑗 𝑖 𝑗

= � 𝑥𝑖 2 𝑝𝑖 � 𝑞𝑗 + 2 � 𝑥𝑖 𝑝𝑖 � 𝑦𝑗 𝑞𝑗 + � 𝑦𝑗 2 𝑞𝑗 � 𝑝𝑖 − (E[𝑋])2 − 2E[𝑋]E[𝑌] − (E[𝑌])2

𝑖 𝑗 𝑖 𝑗 𝑗 𝑖

= � 𝑥𝑖 2 𝑝𝑖 + 2E[𝑋]E[𝑌] + � 𝑦𝑗 2 𝑞𝑗 − (E[𝑋])2 − 2E[𝑋]E[𝑌] − (E[𝑌])2

𝑖 𝑗

= E[𝑋 2 ] − (E[𝑋])2 + E[𝑌 2 ] − (E[𝑌])2

⇒ Var[X + Y] = Var[X] + Var[Y], if X and Y are independent.

Unbiased & biased estimators

Unbiased estimators

An estimator 𝜆̂ for a parameter λ is said to be unbiased if E[𝜆̂] = λ.

Example:
A bag has 468 beads of two colours, white and green. 20 beads are taken at random and the number, i, of
green beads in the sample is counted.
To estimate the true number of green beads, g, in the bag, we calculate
𝑖
𝑔� = 20
× 468 .
If g is the true number of green beads in the bag
𝑔
then the probability of drawing a green bead in a single trial is 𝑝 = 468
,
and drawing n = 20 beads with replacement gives a Binomial distribution B (n, p).
𝑔
Thus µ = E[i] = np = 20 × 468
We do not actually know the number of green beads, and want to estimate this number after taking one
sample
𝑖
estimate 𝑔� = 20
× 468.

34 S3 JUNE 2016 SDB

We now find the expected value of this estimate
𝑖 468 468 𝑔
⇒ E[𝑔�] = E�20 × 468� = 20
× E[i] = 20
× 20 × 468 = g, the true number
⇔ the expected value of the estimator, 𝑔�, is equal to the true value, g
⇒ the estimator, 𝑔�, is unbiased.

Biased Estimators
An estimator 𝜆̂ for a parameter λ is said to be biased if E[𝜆̂] ≠ λ.
Example
A naturalist wishes to estimate the number of squirrels in a wood. He first catches 50 squirrels, marks
them and then releases them. Later he catches 30 squirrels and counts the number, i, which have been
marked.
The true number in the population, n, is then estimated as 𝑛� from the equation
50 𝑖 1500
𝑛�
= 30
⇒ 𝑛� = 𝑖
.
30
1500
Now E[𝑛�] = ∑ 0 i
× pi

i.e. it is possible that i = 0, in which case the estimate 𝑛� is infinite when i = 0,

⇒ E[𝑛�] is also infinite and so cannot be equal to its true value
⇒ in this case the estimator 𝑛� = 1500
i is biased.

Unbiased estimates of population mean and variance

Let X be a random variable drawn from a population with mean µ and variance σ 2, then
E[X] = µ , and Var[X] = σ 2.
A random sample, X1, X2, X3, …, Xn, of size n is taken from the population.
1
The sample mean is 𝑋� = 𝑛
(𝑋1 + 𝑋2 + 𝑋3 + … + 𝑋𝑛 ).

⇒ E[Xi] = µ , and Var[Xi] = σ 2 for i = 1, 2, 3, …, n.

Unbiased estimate of the mean

1 1
E[𝑋�] = E �𝑛 (𝑋1 + 𝑋2 + 𝑋3 + … + 𝑋𝑛 )� = 𝑛
(E[𝑋1 ] + E[𝑋2 ] + E[𝑋3 ] + ⋯ E[𝑋𝑛 ])
1
= 𝑛 (µ +µ +µ + … +µ) = µ
⇒ E[𝑋�] = µ, the true value of the mean
⇒ E[𝑋�] is an unbiased estimate of the mean of the population.
S3 JUNE 2016 SDB 35
 Unbiased estimate of the variance of the population

Preliminary results
(i) Var[X] = E[X 2] − (E[𝑋])2 = E[X 2] − µ 2
⇒ E[X 2] = Var[X] + µ 2 = σ 2 + µ 2 I

(ii) Var[ 𝑋�] = E[𝑋� 2 ] − (E[𝑋�])2 = E[𝑋� 2 ] − 𝜇 2

⇒ E[𝑋� 2 ] = Var[ 𝑋�] + µ 2
1
= Var �𝑛 (𝑋1 + 𝑋2 + 𝑋3 + … + 𝑋𝑛 )� + µ 2
1
= 𝑛2 Var[𝑋1 + 𝑋2 + 𝑋3 + … + 𝑋𝑛 ] + µ 2
1
= 𝑛2 (Var[𝑋1 ] + Var[𝑋2 ] + Var[𝑋3 ] + ⋯ + Var[𝑋𝑛 ]) + µ 2
1
= (𝜎 2 + 𝜎 2 + 𝜎 2 + … + 𝜎 2 ) + µ 2
𝑛2
1
⇒ E[𝑋� 2 ] = 𝑛
σ2 + µ2 II

Proof
The variance of X1, X2, X3, … , Xn is defined to be
1
Variance = (s.d.)2 = 𝑛 ∑ 𝑋𝑖 2 − 𝑋� 2
1
⇒ E[(s.d.)2] = E �𝑛 ∑ 𝑋𝑖 2 − 𝑋� 2 �
1
= E � 𝑛 ∑ 𝑋𝑖 2 � − E[𝑋� 2 ]
1
= 𝑛
E�∑ 𝑋𝑖 2 � − E[𝑋� 2 ]
1
= ∑ E�𝑋𝑖 2 � − E[𝑋� 2 ]
𝑛
1
= ∑(𝜎2 + 𝜇2 ) − E[𝑋� 2 ] since E�𝑋𝑖 2 � = (𝜎 2 + 𝜇 2 ) from I
𝑛
1 1
= 𝑛
�𝑛(𝜎 2 + 𝜇 2 ) − �𝑛 𝜎 2 + µ2 �� 𝑛
1
since E[𝑋� 2 ] = � 𝜎 2 + 𝜇 2 � from II

1 𝑛−1
⇒ E[(s.d.)2] = (σ 2 + µ 2) − �𝑛 𝜎 2 + µ2 � = 𝑛
𝜎2

Thus E[(s.d.)2] is not equal to the true value, and so (s.d.)2 is a biased estimator of σ 2,
𝑛
but multiplying both sides by 𝑛−1
, we can see that
𝑛
𝑛−1
(s.d.)2 is an unbiased estimator of σ 2.

36 S3 JUNE 2016 SDB

 Bias
Example: A large bag contains counters: 60% have the number 0, and 40% have the number 1.
(a) Find the mean, µ, and variance, σ 2.
A simple random sample of size 3 is drawn.
(b) List all possible samples.
𝑋1 + 𝑋2 +𝑋3
(c) Find the sampling distribution for the mean 𝑋� = 3

(d) Use your answers to part (c) to find E[𝑋�], and Var [𝑋�].
(e) Find the sampling distribution for the mode M.
(f ) Use your answers to part (e) to find E[𝑀], and Var [𝑀].

Solution:
(a) µ = ∑ 𝑥𝑖 𝑝𝑖 = 0 × 0⋅6 + 1 × 0⋅4 = 0⋅4
σ 2 = ∑ 𝑥𝑖 2 𝑝𝑖 – µ 2 = (02 × 0⋅6 + 12 × 0⋅4) – 0⋅42 = 0⋅24
(b) Possible samples are
(0, 0, 0) (1, 0, 0) (1, 1, 0) (1, 1, 1)
(0, 1, 0) (1, 0, 1)
(0, 0, 1) (0, 1, 1)
(c) From (c) we can find the sampling distribution of the mean
1 2
𝑋� 0 3 3
1

p 0⋅63 3 × 0⋅62 × 0⋅4 3 × 0⋅6 × 0⋅42 0⋅43

0⋅216 0⋅432 0⋅288 0⋅064
1 2
(d) E[𝑋�] = 0 × 0⋅216 + 3
× 0⋅432 + 3
× 0⋅288 + 1 × 0⋅064 = 0⋅4
1 2 2 2
Var[𝑋�] = (02 × 0⋅216 + �3� × 0⋅432 + �3� × 0⋅288 + 12 × 0⋅064) – 0⋅42

⇒ Var[𝑋�] = 0⋅08
(e) From (c) we can find the sampling distribution of the mode
M 0 1
p 0⋅63 + 3 × 0⋅62 × 0⋅4 3 × 0⋅6 × 0⋅42 + 0⋅43
0⋅648 0⋅352
(f ) E[𝑀] = 0 × 0⋅648 + 1 × 0⋅352 = 0⋅352

Var[𝑀] = (02 × 0⋅648 + 12 × 0⋅352) – 0⋅3522

⇒ Var[𝑀] = 0⋅228096
Thus the sample mean is an unbiased estimator of the mean of the population
since E[𝑋�] = 0⋅4 = µ , the true value
S3 JUNE 2016 SDB 37
but the sample mode is a biased estimator of the mode of the population
E[𝑀] = 0⋅352, but the true value of the mode of the population is 0.
We say that the bias is E[𝑀] – (the true value) = 0⋅352 – 0 = 0⋅352

In general, if 𝜆̂ is a biased estimator of the parameter λ then the bias is defined to be

bias = E[𝜆̂] – λ
In the above example, the bias in estimating the mode from the sample is
bias = E[𝑀] – true value
= 0⋅352 – 0 = 0⋅352

38 S3 JUNE 2016 SDB

Probability generating functions
Probability functions are a neat idea, and are useful for finding the expected mean and variance for
distributions which have a probability generating function which is easy to differentiate.
If X is a random variable on the set [1, n], then G(t) = 𝑝0 + 𝑝1 𝑡 + 𝑝2 𝑡 2 + … + 𝑝𝑛 𝑡 𝑛 is a probability
generating function, p.g.f.,
𝑛

if (i) � 𝑝𝑖 = 1, and (ii) 𝑝𝑖 ≥ 0 ∀ 𝑖

1
P(X = i) = the coefficient of ti. The probability generating function can be thought of as a probability
labelling function, where ti acts as a label for the probability that X = i.

Expected mean and variance for a p.g.f.

We know that E[X] = ∑ 𝑥𝑖 𝑝𝑖 = 0 × 𝑝0 + 1 × 𝑝1 + 2 × 𝑝2 + … + 𝑛 × 𝑝𝑛 = ∑ 𝑖𝑝𝑖
and that Var[X] = E[𝑋 2 ] − (E[𝑋])2
= 02 × 𝑝0 + 12 × 𝑝1 + 22 × 𝑝2 + … + 𝑛2 × 𝑝𝑛 − (∑ 𝑖𝑝𝑖 )2

Notice that G′(t) = 0 × 𝑝0 + 1 × 𝑝1 + 2 × 𝑝2 𝑡 + 3 × 𝑝3 𝑡 2 + … + 𝑛 × 𝑝𝑛 𝑡 𝑛−1

⇒ G′(1) = 0 × 𝑝0 + 1 × 𝑝1 + 2 × 𝑝2 + 3 × 𝑝3 + … + 𝑛 × 𝑝𝑛
⇒ Expected mean = E[X] = G′(1)

and G′′(t) = 0 × (−1) × 𝑝0 + 1 × 0 × 𝑝1 + 2 × 1 × 𝑝2 + 3 × 2 × 𝑝3 𝑡 + … + 𝑛(𝑛 − 1) × 𝑝𝑛 𝑡 𝑛−2

𝑛 𝑛 𝑛

⟹ G"(1) = � 𝑖(𝑖 − 1)𝑝𝑖 = � 𝑖 2 𝑝𝑖 − � 𝑖𝑝𝑖

1 1 1
𝑛 𝑛

⟹ � 𝑖 2 𝑝𝑖 = G"(1) + � 𝑖𝑝𝑖
1 1
𝑛 𝑛 2

⟹ Var[𝑋] = � 𝑖 2 𝑝𝑖 − �� 𝑖𝑝𝑖 �
1 1

⇒ Var[X] = G′′(1) + G′(1) − (G′ (1))2

Thus for a probability generating function G(t) = 𝑝0 + 𝑝1 𝑡 + 𝑝2 𝑡 2 + … + 𝑝𝑛 𝑡 𝑛 ,

E[X] = G′(1) and Var[X] = G′′(1) + G′(1) − (G′(1))2.

S3 JUNE 2016 SDB 39

 Mean and variance of a Binomial distribution

If X ∼ B(n, p) then P(X = i) = nCi piqn−i, where p + q = 1.

These probabilities are the coefficients of ti in the expansion of (q + pt)n
⇒ the p.g.f. for the binomial distribution B(n, p) is G(t) = (q + pt)n.
⇒ G′(t) = np(q + pt)n−1, and G′′(t) = n(n − 1)p2 (q + pt)n−2
⇒ µ = E[X] = G′(1) = np since p + q = 1

and σ 2 = Var[X] = G′′(1) + G′(1) − (G′(1))2

= n(n − 1)p2 + np − (np)2 = n2p2 − np2 + np − n2p2
⇒ σ 2 = Var[X] = np(1 − p) or npq.

Mean and variance of a Poisson distribution

𝜆𝑖 𝑒 −𝜆
If X ∼ PO(λ) then, in a given interval, P(X = i) = , where λ is the mean number of occurrences
𝑖!
in an interval of the same length, i = 0, 1, 2, 3, …

∞ ∞
𝜆𝑖 𝑒 −𝜆 𝑖 𝜆𝑖
⟹ G(𝑡) = � 𝑡 = 𝑒 −𝜆 � 𝑡 𝑖 = 𝑒 −𝜆 𝑒 𝜆𝑡
𝑖! 𝑖!
𝑖=0 𝑖=0

⇒ G′(t) = 𝜆𝑒 −𝜆 𝑒 𝜆𝑡 and G′′(t) = 𝜆2 𝑒 −𝜆 𝑒 𝜆𝑡

⇒ µ = E[X] = G′(1) = λ 𝑒 −𝜆 𝑒 𝜆 = λ
⇒ µ = E[X] = λ

and σ 2 = Var[X] = G′′(1) + G′(1) − (G′(1))2

= λ 2 + λ − λ2 since 𝑒 −𝜆 𝑒 𝜆 = 1

⇒ σ 2 = Var[X] = λ

40 S3 JUNE 2016 SDB

 Index

χ2 test Poisson distribution p.g.f.

binomial dist, 24 expected mean and variance, 40

continuous uniform dist., 24 Probability generating functions, 39

expected mean and variance, 39
degrees of freedom, 23
discrete uniform dist., 23 Random number tables, 5

general points, 23 Ranks

equal ranks, 29
normal dist, 26
Poisson dist., 24
Sample variance
estimator of population variance, 11
Bias, 9, 37
Sampling, 4
Biased estimators, 8 quota sampling, 7
bias, 38
sample means, 14
examples, 35
simple random sampling, 5
Binomial distribution p.g.f.
stratified sampling, 6
expected mean and variance, 40
systematic sampling, 5
Census, 4
with and without replacement, 6
Central limit theorem, 15
Significance test
Combinations of random variables zero correlation, 31
expected mean of X ± Y, 3
Significance test – variance of population known
expected variance of X ± Y, 3 difference between means, 19
independent normal variables, 3 mean of normal distribution, 18
Var[X+Y], 34
Significance test – variance of population NOT
𝐄[𝑿 + 𝒀], 33 known
Confidence intervals, 16 difference between means, 22

Contingency tables mean of normal distribution, 21

χ2 test, 27 Spearman
degrees of freedom, 27 comparison with PMCC, 32

Data Spearman’s rank correlation coefficient, 29

primary data, 7 when to use, 30

secondary data, 7 Standard error, 15

Estimators Unbiased estimators, 8
population mean, 11 examples, 34

population variance, 11 of the population mean, 35

Lottery sampling, 5 of the population variance, 36

PMCC variance of (X + Y), 33

comparison with Spearman, 32
when to use, 30

S3 JUNE 2016 SDB 41