©UNZA-IDE

THE UNIVERSITY OF ZAMBIA

INSTITUTE OF DISTANCE EDUCATION

INTRODUCTION TO STATISTICS:
MAT 2602 MODULE

i
 ©UNZA-IDE

© Copyright

All rights reserved. This module is a copyright of the Institute of distance education of the
University of Zambia, and it cannot be reproduced anywhere without Permission from the
University of Zambia.

ii
 ©UNZA-IDE

Acknowledgements

The Institute of Distance Education of the University of Zambia wishes to thank Mr Urban N.
Haankuku for writing this module MAT 2602 Introduction to Statistics

iii
 ©UNZA-IDE

Table of Contents

© Copyright ............................................................................................................. ii
Acknowledgements ................................................................................................ iii
Table of Contents ................................................................................................... iv

iv
 ©UNZA-IDE

About this introduction to probability Module ix

How this statistics Module is structured ix

The course overview ix

The overview also provides guidance on: x

Module overview xii

Statistics Module —is this course for you? .................................................................... xii
Course outcomes ............................................................................................................. xii
Timeframe ...................................................................................................................... xiii
Study skills ..................................................................................................................... xiii
Need help? ..................................................................................................................... xiv

Assessments xiv

Getting around this Statistics Module. xv

Margin icons xv

UNIT 1 STATISTICS 1
1.5.1 Types of Data ........................................................................................................... 2
1.5.2 Frequency Distribution ............................................................................................ 3
1.1 UNIT ACTIVITY ....................................................................................................... 6
1.5.3 Descriptive Statistics................................................................................................ 9
1.5.4 Measure of dispersion ............................................................................................ 26
1.5.5 Presentation of Data ............................................................................................... 36
1.2 UNITY ACTIVITY .................................................................................................. 49
1.0 UNIT SUMARY ....................................................................................................... 52

UNIT 2 SAMPLING 53
2.5.1 The need for sampling ........................................................................................... 55
2.5.2 Sampling distribution ............................................................................................. 60
2.5.3 The sampling distribution of the mean .................................................................. 61
2.1 UNIT ACTIVITY ..................................................................................................... 65
2.5.4 An unbiased estimator of variance......................................................................... 66
2.5.5 Relative efficiency and consistency ....................................................................... 69
2.2 UNIT ACTIVITY ..................................................................................................... 72
2.5.6 The central limit theorem ....................................................................................... 74
2.3 UNIT ACTIVITY ..................................................................................................... 76
2.5.7 An unbiased estimator of population proportion ................................................... 77

v
 ©UNZA-IDE

2.4 UNIT ACTIVITY ..................................................................................................... 77

2.0 UNIT SUMMARY ................................................................................................... 79

UNIT 3 CONFIDENCE LIMITS 81

3.5.1 Confidence limits of the mean (  known) ........................................................... 82
3.1 UNIT ACTIVITY ..................................................................................................... 84
3.5.2 Confidence limits for a large sample (  unknown).............................................. 84
3.2 UNIT ACTIVITY ..................................................................................................... 85
3.5.3 Confidence limits for a small sample from a Normal population (  unknown) .. 87
3.3 UNIT ACTIVITY ..................................................................................................... 89
3.5.4 Confidence interval of a population from a large sample ...................................... 90
3.4 UNIT ACTIVITY ..................................................................................................... 91
3.5.5 Confidence Intervals for Variances ....................................................................... 95
3.5.6 Confidence interval of one Variance ..................................................................... 96
3.5 UNITY ACTIVITY .................................................................................................. 98
3.0 UNIT SUMMARY ................................................................................................... 99

UNIT 4 HYPOTHESES TESTING 100

4.5.1 Setting up a hypothesis ........................................................................................ 101
4.1 UNIT ACTIVIT ...................................................................................................... 105
4.5.2 Significance test on the mean of a Normal distribution (  known) ................... 106
4.2 UNIT ACTIVITY ................................................................................................... 109
4.5.3 Significance tests (  unknown): large samples .................................................. 110
4.3 UNIT ACTIVITY ................................................................................................... 111
4.5.4 Significance tests (  unknown): small samples ................................................. 112
4.4 UNIT ACTIVITY ................................................................................................... 113
4.5.5 Difference between two means for large samples ............................................... 114
4.5 UNIT ACTIVITY ................................................................................................... 116
4.5.6 Testing if two samples come from the same population ..................................... 117
4.6 UNIT ACTIVITY ................................................................................................... 119
4.5.7 Paired tests ........................................................................................................... 120
4.5.8 The sign test ......................................................................................................... 122
4.7 UNIT ACTIVITY ................................................................................................... 123
4.5.9 Significance of a proportion (large samples) ....................................................... 124
4.8 UNIT ACTIVITY ................................................................................................... 125
4.5.10 Difference between two proportions (large samples) ........................................ 126
4.9 UNITY ACTIVITY ................................................................................................ 127
4.5.11 Significance test using the Poisson distribution ................................................ 128
4.5.12 Type I and Type II errors ................................................................................... 129
4.10 UNIT ACTIVITY ................................................................................................. 132
4.0 UNIT SUMMARY ................................................................................................. 139

UNIT 5 THE  2 –TEST. GOODNESS OF FIT 141

5.5.1 An experiment...................................................................................................... 143
5.5.2 Degrees of freedom .............................................................................................. 144

vi
 ©UNZA-IDE

5.5.3 Significance testing using the X 2 –distribution................................................... 145

5.1 UNIT ACTIVITY ................................................................................................... 147
5.5.4 Test a distribution for normality .......................................................................... 148
5.5.5 Testing the fit of a Binomial distribution............................................................. 149
5.5.6 Testing the fit of a Poisson distribution ............................................................... 150
5.2 UNIT ACTIVITY ................................................................................................... 151
5.5.7 Contingency tables ............................................................................................... 153
5.5.8 Larger contingency tables .................................................................................... 155
5.3 UNIT ACTIVITY ................................................................................................... 158
5.5.9 Relationship between X2 and Normal distribution .............................................. 159
5.4 UNIT ACTIVITY ................................................................................................... 161
5.0 UNIT SUMMARY ................................................................................................. 164

UNIT 6 CORRELATION ANALYSIS 165

6.5.1 Scatter diagrams ................................................................................................... 166
6.5.2 Measurement of correlation ................................................................................. 170
6.1 UNIT ACTIVITY ................................................................................................... 176
6.5.3 Spearman’s rank correlation coefficient, rs ......................................................... 178
6.2 UNIT ACTIVITY ................................................................................................... 183
6.3 UNIT ACTIVITY ................................................................................................... 185
6.0 UNIT SUMMARY ................................................................................................. 188

UNIT 7 THE ANALYSIS OF VARIANCE 190

7.5.1 The Analysis of Variance .................................................................................... 192
7.5.2 A Comparison of More Than Two Means ........................................................... 197
7.5.3 Proof of Additivity of the Sums of Squares and E(MST) for a Completely
Randomized Design ...................................................................................................... 201
7.5.4 An Analysis-of-Variance Table for a Completely Randomized Design ............. 204
7.5.5 Estimation for the Completely Randomized Design ........................................... 205
7.5.6 The Analysis of Variance for a Randomized Block Design ................................ 207
7.5.7 Estimation for the Randomized Block Design .................................................... 211
7.5.8 The Analysis of Variance for a Latin-Square Design .......................................... 212
7.5.8 Estimation for the Latin-Square Design .............................................................. 218
7.5.9 Selecting the Sample Size .................................................................................... 219
7.0 UNIT ACTIVITY ................................................................................................... 222
7.0 UNIT SUMMARY ................................................................................................. 228

UNIT 8 LINEAR REGRESSION 229

8.5.1 Analysing the results of an experiment................................................................ 230
8.5.2 The method of least squares................................................................................. 233
8.5.3 Coded values ........................................................................................................ 236
8.1 UNIT ACTIVITY ................................................................................................... 238
8.5.4 Estimating a value of  Y2 ................................................................................... 240
X
8.5.5 Confidence limits for  ...................................................................................... 241

vii
 ©UNZA-IDE

8.5.6 Confidence limits for  ...................................................................................... 245

8.5.7 Confidence limits of predicted values ................................................................. 247
8.2 UNIT ACTIVITY ................................................................................................... 250
8.0 UNIT SUMMARY ................................................................................................. 253
9.0 MODULE SUMMARY .......................................................................................... 253
10.0 SYLLABUS: MAT 2602 - INTRODUCTION TO STATISTICS ...................... 254

2. Introduction to Statistics; (2016)., David Lane, Rice University, Publisher: Saylor

Foundation 255
11.0 Answers to Unit Activities .................................................................................... 255
12.1 Table A1 Z-Table.................................................................................................. 264
12.2 Table A3 Random Numbers ................................................................................ 265
12.3 Table A4 T-Distribution Table (One Tail and Two-Tails) ................................... 267
12.4 Table A6 Upper-tail critical values of chi-square distribution with ν degrees of
freedom ......................................................................................................................... 268
Table A7 ....................................................................................................................... 271

viii
About this introduction to probability Module

The introduction to statistics has been produced by The University of Zambia, Institute of
Distance education.

How this statistics Module is structured

The course overview

The course overview gives you a general introduction to the course. Information contained in
the course overview will help you determine:

 If the course is suitable for you.

 What you will already need to know.

 What you can expect from the course.

 How much time you will need to invest to complete the course.

ix
The overview also provides guidance on:

Study skills.

Where to get help.

Course assignments and assessments.

Activity icons.

Units.

We strongly recommend that you read the overview carefully before starting your study.

The course content

The course is broken down into units. Each unit comprises:

 An introduction to the unit content.

 Unit outcomes.

 New terminology.

 Core content of the unit with a variety of learning activities.

 A unit summary.

 Exercises and Answers to the exercises, as applicable.

Resources

For those interested in learning more on this subject, we provide you with a list of additional
resources at the end of this statistics Module; these may be books, calculator, articles or web
sites.

Your comments

After completing this module, we would appreciate it if you would take a few moments to
give us your feedback on any aspect of this course. Your feedback might include comments
on:

 Module content and structure.

 Module reading materials and resources.

 Module unit Activities

x
 Module Answers.

 Module duration.

 Module support (assigned tutors, technical help, etc.)

Your constructive feedback will help us to improve and enhance this course.

xi
Module overview

Welcome to Introduction to statistics Module.

This Module is designed to give students background knowledge/ information of statistics

concepts, knowledge and skills in statistics. The module gives a background of what is data
and how you would manipulate data into a more meaningful information that can be used for
decision making and planning. You will learn preliminary analysis under descriptive
statistics, data presentations for both discrete and continuous data, estimation theory,
statistical inference and linear regression analysis.

The module has a number of unit activities at the end of each unit for students to practice.
Answers to these activities are provided to give students feedback. Statistics is a practical
subject so you are advised to try all activities before you come for residential schools. During
the residential schools you are expected to come for consultations on areas you found
concepts difficult to understand.

Statistics Module —is this course for you?

This course is intended for people who are studying Mathematics for their Bachelor’s degree
in Mathematics Education, Demography and Eonomics.

Pre requisite: In order for you to study this module well you should have done Module MAT
1100 (Foundation Mathematics).

Course outcomes

By the end of this Module, you should be able to you will be able to:

 Make a distinction between discrete random variables and continuous random

variables.

 present data in graphical methods

 Carry out basic analysis using descriptive statistics

xii
  Carry out statistical inference using estimation theory, hypotheses testing methods
and linear regression analysis

Timeframe

It is recommended that you take 120 Hours to complete this module

Study skills

As an adult learner your approach to learning will be different from the way you used to learn
in your school days: you will choose what you want to study, you will have professional
and/or personal motivation for doing so and you will most likely be fitting your study
activities around other professional or domestic responsibilities.

Essentially you will be taking control of your learning environment. As a consequence, you
will need to consider performance issues related to time management, goal setting, stress
management, etc. Perhaps you will also need to reacquaint yourself with areas such as essay
planning, coping with exams and using the web as a learning resource.

Your most significant considerations will be time and space i.e. the time you dedicate to your
learning and the environment in which you engage in that learning.

We recommend that you take time now—before starting your self-study—to familiarize
yourself with these issues. There are a number of excellent resources on the web. A few
suggested links are:

 http://www.how-to-study.com/

The “How to study” web site is dedicated to study skills resources. You will find links to
study preparation (a list of nine essentials for a good study place), taking notes, strategies
for reading text books, using reference sources, test anxiety

 http://www.ucc.vt.edu/stdysk/stdyhlp.html

This is the web site of the Virginia Tech, Division of Student Affairs. You will find links
to time scheduling (including a “where does time go?” link), a study skill checklist, basic

xiii
 concentration techniques, control of the study environment, note taking, how to read
essays for analysis, and memory skills (“remembering”).

 http://www.howtostudy.org/resources.php

Another “How to study” web site with useful links to time management, efficient reading,
questioning/listening/observing skills, getting the most out of doing (“hands-on” learning),
memory building, tips for staying motivated and developing a learning plan.

The above links are our suggestions to start you on your way. At the time of writing these
web links were active. If you want to look for more go to www.google.com and type “self-
study basics”, “self-study tips”, “self-study skills” or similar search form.

Need help?

When you need help in this module you can contact the Director, UNZA IDE or
visit www.mathstutor.com

Assessments

Assessment

1. Continuous Assessment 30%

1.1 Assignments/Quizzes 10%

1.2 Tests 20%

2. Final Examination 70%

Total 100%

xiv
Getting around this Statistics Module.

Margin icons

While working through this module you will notice the frequent use of margin icons. These
icons serve to “signpost” a particular piece of text, a new task or change in activity. They
have been included to help you to find your way around.

A complete icon set is shown below. We suggest that you familiarize yourself with the icons
and their meaning before starting your study.

xv
Activity Assessment Assignment Case study

Group
Discussion Help Note it!
activity

Study
Outcomes Reading Reflection
skills

Terminolog
Summary Time Tip
y

xvi
UNIT 1 STATISTICS

1.1 Unit Introduction

Welcome to Unit 1 of this Module in which you will learn that Statistics is a science that deals
with the methods of data: Collections, Compilation, Presentation, analysis, and Interpretation of
results, Conclusion and above all making decisions based on the study of the data. Why
statistics? This field of study is used to study experiments and another natural phenomenon, to
solve a particular problem or find explanations of why specific events happen.

Figure 1.1 Pictorial picture of research

Figure 1.1 an illustration of inference from sample statistics to population parameters. Statistical
analysis is critical in Epidemiology, Agriculture, Engineering, Environment, Education, Finance,
economics etc. Figure1. 1 illustrates the schematic diagram of what is involved in statistics.

1.2 Unit Aim

The aim of this course is to give you concepts and skills of statistics.

1.3 Unit Objectives

By the end of the unit you should be able to:

 Calculate required statistics on any given data,

 Distinguish between discrete and continuous data

 Calculate statistics from any given data

1
  Calculate measures of central tendencies and measures of dispersion of data

Terminology

 x - The mean or average

  - Summation

  - Standard deviation

 f - Frequency

1.4 Unit Time required

You need 30 hours for this unit

1.5 Unit Topics

1.5.1 Types of Data

There are many examples of types of data that exist and table 1 gives examples of types of data
which gives basic distinction between quantitative variables (for which one asks "how much?") and
categorical variables (for which one asks "what type?"). The focus of this unit is on multivariate
analysis of discrete data. Here you will deal with data which are discretely measured responses such
as counts, proportions, nominal variables, ordinal variables, discrete interval variables with few
values, continuous variables grouped into a small number of categories, etc. You will learn basic
statistical methods and discuss issues relevant for the analysis of some discrete distribution.

Quantitative variables can be continuous or discrete. Examples of continuous variables, such as

height, weight, this is the type of data that can in theory take any value within a given range.
Examples of discrete variables are: number of children in a family, number of attacks of asthma per
week.

Categorical variables are either nominal (unordered) or ordinal (ordered). Examples of nominal
variables are male/female, alive/dead, blood group O, A, B, AB. For nominal variables with more

2
than two categories the order does not matter. For example, one cannot say that people in blood
group B lie between those in A and those in AB. Sometimes, however, people can provide ordered
responses, such as grade of breast cancer, or they can "agree", "neither agree nor disagree", or
"disagree" with some statement. In this case the order does matter and it is usually important to
account for it.

Table 1.1 Examples of types of data

Variables shown at the left of Table 1.1 can be converted to one’s further to the right by using "cut
off points". For example, blood pressure can be turned into a nominal variable by defining
"hypertension" as a diastolic blood pressure greater than 90 mmHg, and "norm tension" as blood
pressure less than or equal to 90 mmHg. Height (continuous) can be converted into "short", average"
or "tall" (ordinal). Data can be classified into two major types. These are discrete Data and
continuous Data. In the study of statistics, your interests are to study the several methods of statistics
of which the ultimate goal is to estimate the population parameters of interests.

1.5.2 Frequency Distribution

The frequency distribution is a statistical table which shows the value of a variable in order of
magnitude, either individually or in class interval, along with the corresponding frequencies side by
side. The data pertaining to a quantitative phenomenon can be classified in four ways:

3
  The set or series of individual observations- ungrouped (raw) or grouped (arrayed) data.
 Discrete frequency distribution.
 Continuous frequency distribution.

The data given in Example 1.1 are called the raw or ungrouped data which does not give us any
useful information. Your objective is to express the huge data in a suitable condensed form which
will highlight the significant facts and comparisons and furnish more useful information without
sacrificing any information of interest about the important characteristics of the distribution. A
better presentation of above raw data would be to arrange them in an ascending or descending order
of magnitude which is called arraying of data. However, this method is better than raw data but does
not reduce the volume of the data. This is the distribution of the ungrouped data.

Choice of class interval

It is important that the class interval is chosen so that the histogram gives a clear representation of
the data. Choice of too large a class interval will lead to loss of detail. On the other hand too small
a class interval would destroy the original point of grouping. A convenient rule of thumb is to choose
a class interval so that the average frequency is about 5. That is number of classes 
1
(total frequency) . If the values given are many and the difference between the smallest and
5
largest is greater than 10, you can group the data into suitable class intervals. Class intervals are
discussed in elementary mathematics as : (a, b) as open interval, [a, b] as closed interval, (a, b) as
open-closed interval and [a, b) as closed - open interval. Continuous data uses [a, b) closed-open
intervals.

Example 1.1: Consider the concentration of an antibody in blood serum from 50 donors, given
below:

4
 Table 1.2 Concentration of antibody in blood serum

11.6 15.8 17.0 14.2 16.2 17.3 15.7 17.3 12.0 11.5
8.4 13.2 12.5 15.7 14.0 10.8 9.9 14.0 15.0 10.2
19.2 12.0 8.2 13.8 12.6 9.6 16.0 13.7 11.5 14.8
17.5 15.2 7.2 17.0 5.2 9.4 7.5 15.5 10.7 14.2
15.3 7.8 11.2 10.0 12.1 10.5 12.2 8.6 9.9 16.1
Tally marks are used in order to track the individual values of the data set. Using the class interval
of 4.0 – 5.9, etc as discrete group class interval this is shown in column 1 and 3.95-5.95, 5.95-7.95
etc as grouped continuous class interval shown as true class limits.

Table 1.3 Grouped frequency table of concentration of an antibody in blood serum

The table above gives the first column as discrete grouped data, and this can be used for constructing
bar chart, while the second column is the continuous grouped data, and this can be used for
constructing a histogram, in which a frequency polygon can be constructed by joining the mid
points. The midpoint value calculated as (upper lines + low)/2 of column 1 and these mid points are
used to calculate the statistics from the frequency distributions. What this mean is that for Example
1.1 above is represented by 8.95. If you have 5.96 do not belong to the first interval or second
interval in the first column it is discrete. But using True class limit it belongs to the second interval.

While preparing the frequency distribution the following points must be kept in mind:

5
 1. The class interval should be uniform i.e. it should be of equal width. A comparison of different
frequency distributions is facilitated if the same class interval is used for all. The class interval
should be an integer as far as possible.
2. The class interval should be so chosen that all the observations should be reflected by the
frequency distribution.
3. The class interval should be continuous open-end. Classes less than a or greater than b should
be avoided. These classes create difficulty in analysis and interpretation.
5. There should not be too many or too small number of classes. The number of classes should
never be less than 6 and not more than 30 i.e. the number of class intervals should lie between
6 and 30. With less number of classes, the accuracy may be lost, and with more number of
classes the computations become tedious. The optimum number of classes is generally
considered as 10.

Advantages of grouping

(i) First advantage of grouping is that in subsequent calculations, much labour is saved
in numerical computation by treating all individuals in a class interval as having the
value at the centre of that interval.

(ii) The second advantage of grouping is where the observed sample is of moderate size
and from a large population. In such a case the frequency table is more likely to exhibit
a rise or fall of frequency against class interval.

1.1 UNIT ACTIVITY

1. Which of the following observations are discrete and which are continuous?

(a) The throws on a die (b) The height of a person

(c) The length of a train journey (d) the weight of an egg

(e) the age of a car

2. For the following classes, give the true class limits and mid-class values.

6
 (a) Length in mm (measured to nearest 0.1 mm) 20.0 - 20.4, 20.5 - 20.9, 21.0 - 31.4, 21.0 -
21.9

(b) Age in years 1, 2, 3, 4, 5

(c) Weight in Kg (to nearest 1Kg) 58, 59, 60, 61, 62, 63

(d) Height in CM (to nearest 1 cm) 140, 150, 160, 170 and less than 180

3. For the table 1.4 below: give the true class limits, mid-class values, illustrate by a Histogram
and a cumulative frequency Curve.

Table 1.4 Frequency distribution of heights

Height in cm(nearest cm) Frequency

120 – 129 2

130 – 139 12

140 – 149 17

150 – 159 18

160 – 160 7

170 - 179 1

4. For the table given below, gives the population of Zambia analysed by sex and age at 30
December 2017. Compare the figures for males and females graphically

5. The lives of 50 electric lamps in hours to the nearest hour, are given below: Form grouped
frequent table and illustrate by (a) a Histogram (b) a cumulative frequency curve.

724 695 716 730 689 700 689 726 662 681 676 732
676 697 710 694 715 738 696 696 682 699 714 707
697 710 660 703 717 692 698 684 695 682 721 708

7
 722 692 717 656 697 701 699 705 680 702 690 663
695 670

6. The numbers of pupils on the register of grant aided schools in Zambia are given below.
Show this information on a histogram.

Table 1.5 Frequency distribution of registered pupils of grant aided

school.

Age (years) Number (‘000)

2-4 197

5-10 3677

11 596

12 572

13 595

14 172

15-18 276

7. The following data pertain to first lactation milk yield (in kg) of 100 Daily cows

Table 1.6 Frequency distribution of first lactation milk yield (in kg) of 100 Daily cows

8
 The data given in example 1 are called the raw or ungrouped data which does not give us any
useful information. Our objective will be to express the huge data in a suitable condensed form
which will highlight the significant facts and comparisons and furnish more useful information
without sacrificing any information of interest about the important characteristics of the
distribution.

(a) Construct a frequency table

(b) Present the given data as a histogram

(c) Present the given data as a frequency polygon

1.5.3 Descriptive Statistics

Descriptive Statistics is the preliminary stage of analysis of statistical data. This involves the
calculations of values such as the median, mode, mean, variance, range etc. These values are
categorized as: measurers of central tendencies and measures of dispersion. In calculating these
measurers, we need to consider whether the given data is in discrete, ungrouped, discrete grouped
or continuous ungrouped or continuous grouped. Formulas adhere to the arrangement of the type of
data.

9
If you have a set of observations, it is useful to have a single value which represents all the
observations, and this value is called the central tendency and also you should have an idea of how
these values are spread called measures of dispersion.

Median: this is a middle value of the observations when the observations are arranged in order.

Note that if the number of observation is an odd number, then the median is one of the observations,
but if the number of observations is even, the median is not one of the observations.

Example 1.2 Find the median for the observations: 2, 2, 3, 3, 4, 4, 5, 6, 7, n = 9 observations

The median is one of the observations since n = 9 (odd).

Therefore, the median = 4 Median

The median is between 5 and 7 and it is not one of the observations since n = 10 (even)

57
Therefore; the median is 6
2

In general the median, M, of n observations is defined as follows: If the observations are arranged
in order according to size, then, for odd values of n, the median is the ½ (n+1)th observation; for
even values of n, the median is half the sum of the ½ nth and (½ n + 1)th observations.

Median from a frequency table for a discrete variable.

If the data are grouped the median can be easily found from the cumulative frequency table.
Table 1.4 gives the scores obtained for 50 throws of a die. The 25th value and the 26th value are
both 3, giving a median of 3.

Table 1.4 Cumulative frequency table for 30 throws of a die.

10
 Figure 1.1 Median of a continuous variety

To find the median divide total frequency by 2 and draw a straight line from the cumulative
frequency locating the position of the value obtained up to the curve and drop to the x-axis and
read the median as shown above in figure 1.1.

Example 1.3 Table 1.5 gives the data of civic of anti’s. Find the median civic antis using the
cumulative curve.

Table 1.5 gives the data of civic of anti's

11
 Using the upper true class limit you can construct a cumulative curve

Figure 1.2 Cumulative curve of civic anti’s

When using graphical methods answering statistics question you must construct your tables or
charts using graph papers.

The mode

12
Mode: is the most occurring observation in the set of values, for ungrouped data. If data are
grouped into classes, you call it the modal class, which is the class with the highest frequency. In
statistics we can have more than one mode in any set of values. If there is one mode, you call this
unimodal, those with two modes, you call it bimodal, and those with more than two modes is called
multimodal.

Example 1.4 Find the mode for each of the following observations:

(a) 3, 7, 4, 9, 6 . No Mode

(b) 2, 2, 2, 3, 3, 5, 7, 9, modes : 2

(c)

13
Mean

One of the important objectives of statistics is to find out various numerical values which explains
the inherent characteristics of a frequency distribution? The first of such measures is averages.
The averages are the measures which condense a huge unwieldy set of numerical data into single
numerical values which represent the entire distribution. The inherent inability of the human mind
to remember a large body of numerical data compels us to few constants that will describe the
data. Averages provide us the gist and give a bird’s eye view of the huge mass of unwieldy
numerical data. Averages are the typical values around which other items of the distribution
congregate. The mean at time called arithmetic mean is the most common measure of central
tendencies and used widely in everyday life. At times it is miscalculated, misused by many users.
The uses of the mean in statistical analysis is numerous. This is value in statistical analysis which
is used widely.

Definition The mean is defined as the central value of any given observations. This is calculated
by adding all the observations and divide by the number of observations. Since it is calculated by
using all observations it sits on the Centre of the observations that means some value are below
the meanwhile others are above the mean. The mean is not half of the observation are below and
half of the other observation are above, it depends on the magnitude of each value to the mean.
But if we sum the difference between those below and those above it is zero. This will be made
clear when we discuss the standard deviation.

14
Notation The following notation is being used here;

n
Mean, 
i 1
summing up the observations from i = 1,2, ….., n,

n  the number of observations. If the variates takes n values, they can be presented as
x1 , x2 ,..., x n , and x =mean is given by;

x i
x i 1
(1)
n

Equation (1) is the formula for calculating the mean from observations which is not grouped.

Example 1.5 Calculate the mean for the observations: 71, 58, 67, 82, 74, 66, 47, 90, 67, 82.

There are 10 observations so n = 10 the values are observations give.

71  58  67  82  74  66  47  90  67  82 704
The mean is =   70.4 .That is the mean is
10 10
70.4.

Equation (1) is used for calculating the mean from the ungrouped data.

Note that the observations given are integers but there is no reason why the mean should be an
integer, it can be any real number. The mean can be calculated from a set of ungrouped data as
the Example 1.5 given above or it can be calculated from grouped data (grouped as individual
observations or grouped in class intervals) as discussed in the next sections.

Calculation of the mean from a frequency table

The table below shows the scores observed for 50 tosses of a die. Instead of recording each
score, a column of frequency is added to make the tabulation easy and its frequency multiplies
each score. The total of this column is the total of the scores and its mean is calculated. This is
an example of grouped data.

15
 Table 1.6 The scores observed for 50 tosses of a die

160
The mean =  3 .8
50

If the data is grouped into classes, we can calculate the mean as follows:

Table 1.7

637.5
Mean  12.75
50

Note that data can be grouped individually just like data presented in table 1.8 below. The scores
represented as x in column 1 and their corresponding frequency under f column. The score of 0
has been observed 3 times while the score of 2 has been observed 5 times.

16
Table 1.8 Distribution of X values

From this table 1.8 the mean can be calculated using the formula below

fx i i
x i 1
n
(2)
f
i 1
i

This is the formula for calculating the mean from a grouped data

To calculate the mean from table 1.9 consider the column fx which is the total of the scores. If
you add 0 three times the answer is 0 (fx=0x3 = 0), if you add 2 five times the answer is 10 (
fx=2x5 = 10).

Table 1.9

Using the formula given in equation (2) the mean from table 1.9 is calculated as:

fx i i
15
x i 1
n
  1 .3
11
f
i 1
i

That is, the mean of the given observation is 1.3.

17
Data can be grouped using class intervals as illustrated in table 1.10 to read this table take an
interval 1 – 5 under the class interval column, it has a corresponding value of 1 under frequency
column. This means there is one value in this interval, interval 11 – 15 has the frequency of 5
which means there are 5 values in this interval but we do not know what these values are. So
these 5 values are approximated by the mid value of this interval called class mark (x). This is
upper lim it  lower lim it
calculated as: class mark (x) = . Then we can calculate the mean
2
n

fx i i
using the formula: x  i 1
n

f
i 1
i

Table 1.10 Distribution of grouped data

fx i i
151
Therefore, x  i 1
n
  2.32
65
f
i 1
i

That is the mean of the given table of values is 2.32

It should be observed that the class intervals given in table 1.10 are discrete class intervals. This
table can be used to calculate the mean as illustrated above and you can represent this data as a
bar chart. Table 1.10 can be made continuous by closing the gaps between intervals as follows:
Introduce an exact class limits by calculating

18
 upper lim it  lowe lim it 5.9  6.0 11.9
TCL = =   5.95 , 5.95 will be the upper limit for
2 2 2
interval one and lower limit for the next interval and this interval is closed-open interval
meaning if 5.95 appears it should be included in the interval because it is closed. Table 1.11
below illustrate this where the first column is discrete and the second column is continuous, the
third column is the class marks, the frequency column and the last column is the total of the
scores.

Table 1.11

The mean is calculated as;

fx i i
482.95
x i 1
n
  11.78
41
f
i 1
i

Calculation of the mean using the assumed mean.

At times calculating the mean may be extremely tedious even with a calculator. This can be
simplified by using assumed mean in which you take an arbitrary origin of the data. The formula
used is as follows:

19
 n n
A f i  B  f i u i
x i 1
n
i 1

f
i 1
i

x  A  Bu

This method can be generalized as follows:

If we have observations x1 , x 2 ,..., x n which are converted to u1 , u 2 ,..., u n , using A as arbitrary

origin and B as unit, then;

x1  A  Bu1
x 2  A  Bu 2
.....
x n  A  Bu n

And x  A  Bu
The u i s at times are called coded data.

Example 1.6 Using table 1.12 calculate the mean using the arbitrary origin of A = 12.95 and B =
2.

Table 1.12 Distribution of observations in class intervals

5
u  0.1
50

20
x  A  Bu
 12.95  2 x0.1
 12.75

The following table gives the distribution of 100 accidents during seven days of the week in a
given month. During a particular month there were 5 Fridays and Saturdays and only four each
of other days. Calculate the average number of accidents per day.

Table 1.13

Days: Sun. Mon. Tue. Wed. Thur. Fri. Sat. Total

Number of
accidents: 20 22 10 9 11 8 20 = 100

Solution: Calculation of Number of Accidents per Day

The formula for computation of arithmetic mean according to the short cut method is = where A
is assumed mean, dx = (X – A) and N = ∑f.

You can solve this example by short-cut method as given below:

Table 1.14

Calculation of Average Accidents per Day

21
f dx = X – A

128/30 = X - 10

X = 14 accidents per day

Skewness

The set of observations from a frequency distribution can either be symmetrical that is bell-
shaped in which case: mean = median = mode, Otherwise, the data is said to be skewed.

Figure 1.3 Normal distribution

If the observations are not symmetrical then they are skewed, either to the left or to the right. A
distribution is skewed if one of its tails is longer than the other. The first distribution shown has
a positive skew. That is it has a long tail in the positive direction. It is right for a distribution that

22
 is skewed to the right. For right-skewed distribution, the mean is higher than the median. Also
it be observed that the tail of the distribution on the right is longer than on the left-hand side.

Figure 1.4 Skewed data to the right

The distribution below it has a negative skew since it has a long tail in the negative direction. A
distribution which is skewed to the left has its mean smaller than its median

Figure 1.5 Skewed data to the left

In probability and statistics, skewness measures the symmetry of the distribution of a probability
of a real-valued function of a random variable about its mean which can be either positive or
negative.

Weighted mean

23
Suppose the journey from Lusaka to Livingstone which is 480km cost K160 and a journey from
Lusaka to Ndola a distance of 320km cost K120. If our interest is to find the average cost per km
overall, that is Ndola to Livingstone, it is wrong to add the cost per journey and divide by total
distance. Since the two journeys are of different distances and each distance has its own cost we
need to use weighted averages in calculating overall cost per km. To do this we use weighted
average to calculate overall cost per km weighted in the ratios of their distances. Therefore, the
example given above can be calculated as:

Table 1.15 Table for weighted data

320x0.38  480x0.33
The weight mean =  0.35
320  480

That means the overall cost per km is 0.35

In general if the values x1 , x2 ,..., x n are given and the weights are w1 , w2 ,..., wn then

w x  x 2 w2  ...  wn xn w x i i
Weighted mean = 1 1  i 1

w1  w2  ...  wn n

w
i 1
i

The weighted mean is used in financial calculations such as; knowing the average amount of
time left before mortgages in a mortgage-backed security expire

Moving Averages

In the study of times series data (when time effect is evident) data such as rainfall data, labour
statistics data (unemployment data), population data etc. may contain some seasonal variations,
trends and cyclic variations. Data given daily, monthly, yearly, a five or ten year moving average

24
or twelve months moving average can be calculated. A moving average is used to reveal the
general trends and this is generally done in economics, businesses and time series data analysis.

Example 1.7 The table below gives the quarterly unemployment figure for Zambia throughout
nearly three years. This is usually dealt in time series analysis, where a seasonal variation produces
a zig-zag effect which to some extent masks any general trend.

Table 1.16 Number of unemployment in Zambia

The annual variation can be removed by calculating the four quarterly moving averages. This is
done by averaging the values over four quarters and plotting the average obtained in the middle
of the range used. For example the average for the first four quarters is
1
(1471 1368  1464  1341)  1411 which plotted mid-way between Oct 2014 and Jan 2015.
4

The next moving average can be found most simply from 1535.25 + (value for July 2015 – value
for July 2014) and so on.

Calculation of M. average

Table 2.17 Calculations of moving averages

25
 Moving averages can be calculated from different types of data, but their underlying purpose
remains the same: that is to help technical traders track the trends of financial assets by smoothing
out the day-to-day price fluctuations, or noise. It is a tool used for smoothing out price data by
creating a continually updated average price. The average is calculated over a specific period, like
10 days, 20 minutes, 30 weeks or any period the trader chooses.

1.5.4 Measure of dispersion

When data is available, we are interested in the preliminary analysis the measures of central
tendencies of the observations which were discussed in the measures of central tendencies
discussed above. The next interest would be how spread is these observations from the measures
of central tendencies. These values are called measures of dispersion. Here we discuss measures
of dispersion and provides an interpretation of these values concerning measures of central
tendencies and their uses in everyday life.

The range

If we want to know how scattered the observations are; we can use them by finding the difference
between the smallest and largest values.

Range = highest value- smallest value

26
Interpretation: If the range is small then the values are clustered together. However, if the range
is large, it means the values are widely scattered around. If the range is zero, it means, we have
only one value.

Example 1.8 The numbers are 7, 4, 9, 8, 2. Find the range, IQR, variance, standard deviation
and coefficient of variation.

Solution: In order these numbers are 2, 4, 7, 8, 9. n = 5

Table 1.18

i) Range: range  9  2  7 .

Variance: First, find the sample mean. x   

x 30
ii)  6 . The preferred method for
n 5
finding the variance is the computational formula, which has the advantage that the
x  x  and x  x 2 columns are not needed.

1  34
iii) Var(x) =  2 
n
 (x  x) 2 
5
 6.8

IQR: To find the first quartile p  .25 so position  pn  1  .256  1.5  a.b . So a  1 and .b  .5 .
Thus 1Q  x.75  xa  .bxa1  xa   x1  .5x11  x1   x1  .5x2  x1 

27
 2  .54  2  2  1  3 . To find the third quartile p  .75 so position  pn  1  .756  4.5  a.b . So
a  4 and .b  .5 . Thus 3Q  x.25  xa  .bxa1  xa   x4  .5x41  x4   x4  .5x5  x4   8  .59  8

 8  0.5  8.5 . Finally, IQR  3Q  1Q  x.25  x.75  8.5  3  5.5 .

1  34
iv) Var(x) =  2 
n
 (x  x) 2 
5
 6.8 , standard deviation is square root of variance.

Standard deviation is  . Therefore   2.61

std.deviation  2.61
v) Coefficient of variation: C     0.435 or 43.5%.
mean x 6

Inter-quartile range
Interquartile range for ungrouped data
The other measure of dispersion in the inequality range which extends the idea of the median. The
median is just the middle value of the observations when the observations are ordered. The median
of the first set of the observation is called the first quarter, called the lower quartile denoted by Q1,
and the median of the second half is called the upper quartile denoted by Q3. In find Q1 and Q3,
you observe whether the number of observation n is either odd or even.
: - Interquartile range = Q3 – Q3
The IQR is an indicator of how observation is spread. When drawing box-whisker plots, the IQR
is used. This displays the variability of observation if IQR is small, it means the observations are
clustered around, but if the IQR is large then the observations are scattered around.
Example 1.9 Find the interquartile range for the observations given below

1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4,

Interquartile range for grouped data

The median for grouped data is found by using the cumulative curve. From the cumulative curve
which should be drawn on the graph paper, can be used to estimate the median, Q 1, and Q3. The
lower quartile (Q1) is found by reading off the income corresponding to a cumulative frequency
of one-quarter the total frequency and the upper quartile is found by reading off the corresponding
to a cumulative frequency which is three-quarters of the total frequency when finding Q1 and Q3,
it means the data set when ordered is partitioned into four equal parts as Q1, median and Q3.

28
You can partition Data into ten equal parts, and this gives deciles as Q1, median, and Q3.

Where D5 = Median = ½ (n+1)th value if n is odd and ½ (nth + (n + 1)th) value if n is even

2. 5
D2.5 is (n +1)th value if n is odd
10

7 .5
D7.5 is (n+1)th
10

If the data is partitioned into 100 equal parts, this is called percentiles, with

P50 = Md = D5, P75 = Q3, P25 = Q1

Example 1.10 Find the inter quartile range of the following data set given below:

Table 1.19 Frequency distribution of observations

29
The Standard Deviation
Dispersion is the extent to which the magnitudes or quantities of the items differ, the degree of
diversity. The word dispersion may also be used to indicate the spread of the data.
Measure of dispersion is a value that indicates the extent to which all other values are dispersed
about the central value in a particular distribution. The standard deviation can be calculated from
an ungrouped data or grouped data.

Ungrouped data

In calculating the standard deviation from an ungrouped data the following formulas are being
used:

 ( xi  x ) 2
1
a) Mean = x   x (b) standard deviation =   i
n i n

Example 1.11 Find the standard deviation for the ungrouped data 8, 9, 10, 10, 11

30
 5
 ( xi  x ) 2
1
x  xi =9.6  
5 i 1
i
n
=1.02

Grouped data: There are many ways of writing the formula for the standard deviation. The one
above is for a basic list of numbers. The formula for the variance when the data is grouped is as
follows. The standard deviation can be found by taking the square root of this value.

n n

 f i xi  f (x  x)
i
2

(a) Mean  x  i 1
n
, (b) Variance   2  i 1
n
or
f
i 1
i f
i 1
i

2
n
 n 
   f i xi
2
f i xi 
Variance    n
2 i 1
  i 1n  , These formulas can be used in calculating the mean
 

i 1
fi   fi
 i 1


and variance of any grouped data. Note that data can be grouped as individual values or grouped
in intervals.

31
Example 1.12 The table shows marks (out of 10) obtained by 20 people in a test. Calculate (i)
the mean marks, (ii) the standard deviation of the marks.

Table 1.20 Frequency distribution of marks obtained in a test

Solution

32
  fx 118
(i) x    5 .9
f 20

(ii)  
 f (x  x) 2


67.8
 3.39  1.84
f 20

Example 1.13 The frequency distribution of the lengths of 100 leaves from a certain species of
plant is given in below in table 1.21:

Table 1.21 Frequency distribution of the lengths of 100 leaves

Find the range and standard deviation of heights.

33
Solution The range is L = 52 and s =22 Range = 30

 fx 3740
(i) x    37.4
f 100

(ii)  
 f (x  x) 2


5734
 57.34  7.57
f 100

Finding the mean and standard deviation of a combined sample observations

Sometimes we have the mean and standard deviation of two or more sets of values, and we want
to find the mean & standard deviation of the combined set of values. This presents no problem if
the original data are available, it is possible to find the combined mean and standard deviation if
the sample means & standard deviation are available.

Example 1.14: Below is a table showing the mean and standard deviation of the marks obtained
by two classes in a test. Calculate

(a). the mean and b) the standard deviation for the two classes combined.

Table 1.22 Mean and standard deviation Combined tables

34
The method is based on finding  x and  x 2
for each sample and hence finding  x and
x 2
for the sample combined. Using the formula x 
x you have nx   x .You use
n
subscripts to distinguish between two samples  x A  n A x A =15x67=1005

x B  nB x B  18x58  1044

Combining the samples

 x  1005  1044  2049, n  15  18  33

2049
Giving a mean for the combined sample of x   62.1
33

Note that in effect you have calculated a weighted mean of .x

x 2

(a) From equation  2   x 2 this can be rearranged to give

n

x 2
 n( x   2 ) for the two samples separately you have

x  n( x A   A )  15(67 2  5 2 )  67710
2 2
A

x  n( x B   B )  18(582  8 2 )  61704 and the sample combined

2 2
B

35
 x 2
 67710  61704  129414

The variance for the combined sample is given by:

x 2

129414  2049 
2

 
2
x 
2
   66.355
n 33  33 

Interpretation of the measure of dispersion

In a statistical sense, dispersion has two meanings: first it measures the variation of the items
among themselves, and second, it measures the variation around the average. If the difference
between the value and average is high, then dispersion will be high, otherwise it will be low.
This displays how observations are spread about the mean.

1.5.5 Presentation of Data

Preliminary analysis of data involves data presentations in pictorial forms for non-mathematics
audience to understand what is being presented. Two types of presentation can be displayed.

Discrete data: This type of data can be presented using: scatter plots, bar chart, pie chart etc.
These charts can be drawn from frequency distributions.

Example 1.15 The frequency table of number of members per family. Construct a scatters
diagram.

Table 1.23 Frequency table of number of members per family

Solution. The graph below shows the scatter diagram where individual points are shown in x-y
plane.

36
 Figure 1.6 Number of family versus number of members

Bar graph

Example 1.16 The students in IDE math class have birthdays in each of the four seasons. Table
1.24 shows the four seasons, the number of students who have birthdays in each season, and the
percentage (%) of students in each group. Construct a bar graph showing the number of students.

Table 1.24 Seasons number of student’s proportions of population

Solution

Using the data from table 1.24 above, you can construct a bar graph showing the percentages.as
follows:

37
 Figure 1.7 Seasons number of student’s proportions of population

Example 1.17 The table below shows the number of members in each family

Table 1.25 number of members in each family

Construct a bar chart to illustrate the number of members in each family.

Solution. Figure 1.8 below illustrates on how to construct a bar graph which shows bars which
are the same and for family 3 there is zero member.

Figure 1.8

Example 1.18 The table below is a frequency distribution on types of jobs people are involved
in. Construct a bar chart for this distribution.

Table 1.26 Frequency table of part-time jobs

38
Solution You can read the values of the part time jobs from the Bar charts as:

Figure 1.9 Frequency distribution on types of jobs

Example 1.19 Suppose, of the 140 children, 20 lived in owner occupied houses, 70 lived in
council houses and 50 lived in private rented accommodation. Figures from the census suggest
that for this age group, throughout the county, 50% live in owner occupied houses, 30% in
council houses, and 20% in private rented accommodation. Type of accommodation is a

39
categorical variable, which can be displayed in a bar chart. We first express our data as
percentages: 14% owner occupied, 50% council house, 36% private rented. You then display the
data as a bar chart. The sample size should always be given.

Solution The bar chart of housing data for 140 children and comparable census data

Figure 1.10 The bar chart of housing data for 140 children and comparable census data

This example shows you that you can present two or more variables on the same bar chart.

Presentation of continuous data: For continuous data you can present data as; a line graph,
histogram, cumulative curve (Ogive curve), frequency polygon.

Line graph: In constructing line graph you start by plotting points and join these points with
straight lines. In joining straight line you can use a ruler and not by free hands.

Example 1.20 The graph in figure 1.11 represents the average temperatures during the months in
2018 at IDE offices. This is an example of a line graph which is continuous. You can easily tell
this by looking at the graph and seeing the data points connected together.

40
 Figure 1.11 Months-temperature graph for IDE offices

The graph above is continuous, hence you can read any temperature from the graph within the
given range. This is an example of a line graph.

Figure 1.12 Time – distance graph

The graph shown above is time - distance graph which is a line graph. As you can see from the
graph, there is no break in the line. In other words, you can choose any time between 8:45 am and
12:15 pm, even one involving a fraction of a second, and there will be a corresponding distance in
km. Therefore, this is an example of a line graph which represents continuous data.

Histogram: You can construct histogram for continuous data by making sure the bars are joined
and not separate bars as examples in bar graphs. These bars should be of the same width and the
values of the height should be the frequency values as shown below:

41
Example 1.21 Consider the concentration of an antibody in blood serum from 50 donor, given in
table 2.27(continued) below:

Table 1.27(cont’) Concentration of antibody in blood serum

11.6 15.8 17.0 14.2 16.2 17.3 15.7 17.3 12.0 11.5
8.4 13.2 12.5 15.7 14.0 10.8 9.9 14.0 15.0 10.2
19.2 12.0 8.2 13.8 12.6 9.6 16.0 13.7 11.5 14.8
17.5 15.2 7.2 17.0 5.2 9.4 7.5 15.5 10.7 14.2
15.3 7.8 11.2 10.0 12.1 10.5 12.2 8.6 9.9 16.1
Using the class interval of 4.0 – 5.9, etc as discrete ungroup class interval and 3.95-5.95, 5.95-
7.95 etc as grouped continuous class interval.

Table 1.28 Grouped frequency table of data

The table 1.28 above gives the first column as discrete grouped data, and this can be used for
constructing bar chart, while the second column is the continuous grouped data, and this can be
used for constructing a histogram, in which a frequency polygon can be constructed by joining

42
the mid points. The midpoint value calculated as (upper lines + low)/2 of column 1. What this
means is that for Example, 1.21 above is represented by 8.95.

Results for continuous data such as the data above can be represented by a histogram as shown
below. The divisions between the bars fall on the true class limits and the heights of the bars, and
consequently their areas, are proportional to frequency.

The frequency histogram would look like

Figure 1.13 Concentration of antibody in blood serum

Example 1.22: The table below is the frequency table on the nose length of IDE statistics
students.
Table 1.29 the nose length of IDE statistics

43
Construct a frequency table showing the relative frequency and density height
Construct a histogram using the density height using 5mm classes cantered at 30, 35, …,70
Solution
Table 1.30 Nose length

From table 1.30 you can construct a histogram using frequency column, relative frequency
column or density column see figure 1.14. You can construct a histogram using frequency values
or relative frequency from table 1.30. Note that a density histogram is just a modified relative
frequency histogram. A density histogram is defined so that:

 the area of each rectangle equals the relative frequency of the corresponding class, and
 the area of the entire histogram equals 1.

44
 Figure 1.14 Histogram of the length of Nose for IDE students
Frequency Polygon
A frequency polygon is a graphical form of representation of data. It is used to depict the shape
of the data and to depict trends. It is usually drawn with the help of a histogram but can be drawn
without it as well. A histogram is a series of rectangular bars with no space between them and is
used to represent frequency distributions.

Steps to Draw a Frequency Polygon

 Mark the class intervals for each class on the horizontal axis. We will plot the frequency
on the vertical axis.
 Calculate the class mark for each class interval. The formula for class mark is:

Class mark = (Upper limit + Lower limit) / 2

 Mark all the class marks on the horizontal axis. It is also known as the mid-value of every
class.
 Corresponding to each class mark, plot the frequency as given to you. The height always
depicts the frequency. Make sure that the frequency is plotted against the class mark and
not the upper or lower limit of any class.

45
  Join all the plotted points using a line segment. The curve obtained will be kinked.
 This resulting curve is called the frequency polygon.

Note that the above method is used to draw a frequency polygon without drawing a histogram.
You can also draw a histogram first by drawing rectangular bars against the given class intervals.
After this, you must join the midpoints of the bars to obtain the frequency polygon. Remember
that the bars will have no spaces between them in a histogram.

Example 1.23: Construct a frequency polygon using the data given below:

Table 1.31 Frequency distribution of observations

Solution: We first need to calculate the cumulate frequency from the frequency given.

Table 1.32 Continuous data Frequency distribution of observations

You now start by plotting the class marks such as 54.5, 64.5, 74.5 and so on till 94.5. Note that
you will also plot the previous and next class marks to start and end the polygon, i.e. you plot

46
44.5 and 104.5 as well. Then, the frequencies corresponding to the class marks are plotted
against each class mark. Like you can see below, this makes sense as the frequency for class
marks 44.5 and 104.5 are zero and touching the x-axis. These plot points are used only to give a
closed shape to the polygon. The polygon looks like this:

Figure 1.15 Frequency polygon

Note that the points plotted have been joined with straight lines. You need to complete by
joining the x-axis otherwise it will not be a polygon.

Cumulative curve

This is another way of representing data graphically. The cumulative frequency for a particular
value of the variety is frequency of observations, which accumulates, constructed using the
upper true class limits. The frequency is the number of times an event occurs within a given
scenario. Cumulative frequency is defined as the running total of frequencies. It is the sum of all
the previous frequencies up to the current point. It is easily understandable through a Cumulative
frequency table.

47
Example 1.24: The table below gives the frequency distribution number of students marks.
Construct a cumulative frequency curve.

Table 1.33: Frequency Table for the No. of students

Cumulative Frequency is an important tool in Statistics to tabulate data in an organized manner.

Whenever you wish to find out the popularity of a certain type of data, or the likelihood that a
given event will fall within certain frequency distribution, a cumulative frequency table can be
most useful. Say, for example, the Census department has collected data and wants to find out all
residents in the city aged below 45. In this given case, a cumulative frequency table will be
helpful.

Cumulative Frequency Curve

Figure 1.16: Cumulative curve of the marks of students

A curve that represents the cumulative frequency distribution of grouped data on a graph is
called a Cumulative Frequency Curve or an Ogive. Representing cumulative frequency data on a
graph is the most efficient way to understand the data and derive results.

48
Box and Whisker plots: This type of data presentation is commonly used in time series. This
shows the variability of data say in each month of the year. If the box is small that means the
variability of the inter quartile occurs within a short time. If the range is small then the values are
clustered together. But if the range is large it means the values are widely scattered around. If the
range is zero, it means, we have only one value. : - Interquartile range = Q3 – Q1

The IQR is an indicator how observation are spread. When drawing box – whisker plots, the IQR
is used. The box plots it looks like

Figure 1.17 Box and whisker plots

These box plots are used when comparing variability between data sets. The smaller the box the
more closely the data from the mean. This displays the variability of observation. If IQR is small,
it means the observations are clustered around the median, but if the IQR is large than the
observations are scattered away from the median.

1.2 UNITY ACTIVITY

1. Give the mean, median and mode(s) for the following values, which gives the best
measures of central tendency? 2, 4, 6, 3,1,2, 1,1, 5,4,4,2,5,6,8,15,7.

2. The mark distribution in the table below where obtained from two groups of children
taking the same school examination.

(a) Plot these distribution in the way that seems best to you, in order both to illustrate
them separately and to show up the difference between them.

(b) Estimate the median marks for the two groups

49
 Table 1.34 The marks of test results for two groups

3. The sunshine figures for Lusaka town for the year 2015-2017 are given in table2.35 below:
Each figure representing the mean hours of sunshine per day over the month.

(a) Draw a graph to illustrate this information

(b) Superimpose a second graph showing the twelve –month moving average

Table 1.35 The sunshine figures for Lusaka

Jan Feb Mar Apr May June July Aug Sep Oct Nov Dec

2015 1.2 2.8 5.0 5.6 6.6 5.4 8.5 6.4 5.2 3.8 1.8 1.4

2016 1.8 2.5 4.5 5.5 7.8 5.1 5.1 4.9 3.3 3.6 1.9 0.7

2017 1.5 2.7 3.1 5.2 6.7 9.6 4.5 4.7 3.7 2.7 2.1 1.7

4. The table below shows the cumulative distribution of gross weekly earnings of male full
time workers in Zambia Sugar company in April 2014. Estimate for the distribution

(a) The median earnings of a manual worker,

(b) The median earnings of a non- manual worker

(c) The proportion of manual workers earning less than the median earnings of non-manual
workers

Table 1.36 Distribution of gross weekly earnings of male full time workers in Zambia
Sugar Company

50
 Gross weekly 35 40 45 50 55 60 70 80 100 200
earning (K)

Manual workers 0.1 0.3 0.7 1.3 2.0 2.7 4.0 4.9 5.7 6.1
(‘000)

Non-manual 0.1 0.2 0.4 0.6 0.9 1.2 1.8 2.4 3.2 4.1
workers (‘000)

(5) Table 1.37 gives further unemployment figures for Zambia. Draw a graph to illustrate
this information and superimpose the four-quarterly moving average.
Table 1.37 Unemployed people in Zambia. Figures in thousands.
Year January April July October
1983 3021 3094
1984 3200 3108 3101 3225
1985 3341 3273 3235 3277
1986 3408

(6) Table 1.38 is a frequency table for the number of words in a sentence for a paragraph in a
book. Calculate the mean number of words in a sentence.
Table 1.38 Frequency table for the number of words in a sentence
Number of words Number of sentences
5-9 9
10-14 10
15-19 8
20-24 11
25-29 8
30-34 4
35-39 4
40-44 1
45-49 1
(7) For a project a student collected data on the hair-care habits of fellow students. She asked
50 boys and 50 girls how many times they had washed their hair in the previous week. The
results are shown in Table 1.39.
Table 1.39 Hair-care habits of fellow students

51
 Number of washes 1 2 3 4 5 6 7
Number of girls 2 22 22 4 0 0 0
Number of boys 0 10 19 13 2 2 4
Find the mean, median and mode for each group.
(8) A factory uses five raw materials A,B,C,D,E to manufacture a flash-gun. The masses of the
materials used in its production are in the ratios 1:1:4:3:1 respectively. The prices of the
materials, in kwacha per ton, in the years 1978 and 1980 are given in Table 1.40.
Table 1.40 The masses of the materials used
Raw material A B C D E
1978 4 3 2 5 3
1980 8 5 3 9 8
Taking 1978 as the base year, calculate an index number for the total cost of the raw
materials used for the manufacture of the flash-gun in 1980.

1.0 UNIT SUMARY

A frequency table gives the frequency with which each value of an observation occurs in a
group. For a discrete data it can be represented by a bar chart. Continuous data can be displayed
graphically by

(a) Histogram in which equal areas represent equal frequencies.

(b) A frequency polygon which is obtained by forming the mid points at the top of the
histogram blocks
(c) A cumulative frequency curve which gives the number of observations which are less than a
stated value of the observed value.
The Standard Deviation. This is the most used measure of spread of observations notation

The standard deviation is calculated as:

1. ungrouped data  
 (x  x) 2

2. grouped data  
 f (x  x) 2

f

52
 3. Variance is calculated as:  2

 (x  x) 2

for ungrouped data

n

and  2 
 f (x  x) 2

for grouped data.

f
Your interest here is to learn how to calculate the mean and standard deviation from grouped
data and ungrouped data. Their uses and applications are numerals in the next units.

UNIT 2 SAMPLING

2.1 Unit Introduction

Welcome to Unit 2 in which you will be introduced to procedures of sampling. Sampling is a
process used in statistical analysis in which a predetermined number of observations are taken

53
from a larger population. The methodology used to sample from a larger population depends on
the type of analysis being performed,

2.2 Unit Aim

The aim of this Unit is to introduce you procedures of Sampling in which you will learn how to
make inferences about the whole population.

2.3 Unit Objectives

By the end of the Unit you should be able to:

 Carry out sampling correctly

 Obtain the optimum results, i.e., the maximum information about the characteristics of the
population with the available sources at our disposal in terms of time, money and
manpower by studying the sample values only.

 Estimate means and proportions of the population

Terminology

S – Sample standard deviation biased

Ŝ - Sample standard deviation which is unbiased

n – Sample size

2.4 Unit Time required

You need 20 hours for this unit

2.5 Unit Topics

54
2.5.1 The need for sampling

In many cases it is not possible to obtain information about all Members of a population, for the
following reasons:

(1) The collecting of the information may destroy the sample, e.g. testing fireworks or electric
fuses.
(2) The population may be infinite, e.g. the measurements of a physical constant such as
gravity using a particular apparatus.
(3) It may be impracticable to make a measurement for every member of the population, e.g.
measuring the length of ants of a particular species.
(4) Even if a measurement could be made for each member of a population, considerations of
time and expense usually dictate otherwise.

Random sampling

In order for a sample to be representative of the whole population each member of the population
must have an equal chance of being chose. A sample chosen in this way is called a random sample.

The simplest method of selecting a random sample is by using a table of random numbers (see
Table A3). Such tables are now normally compiled electronically but could be made using any
device which gives the digits 0 to 9 with equal probability. It is a prism whose cross-section is a
decagon and whose faces are labelled 0 to 9.

Suppose we wish to select two days at random from the month of August using a random number
table. Each day is allocated a number 01, 02, 03, etc., up to 31. Note that each day must have the
same number of digits in its number so that each number has an equal probability of being chosen.
Any starting position can be chosen in Table A3. Suppose you obtain the pairs of digits: (46), (51),
06, (59), (60), 16. The numbers shown in brackets do not correspond to any members of our
population and are rejected. Pairs of digits are taken until we have sufficient to give a sample of
the required size. In this case the random sample will consist of August 6th and August 16th.

55
To reduce the amount of numbers which has to be rejected, you can allocate more than one
number to each member of the population, on a cyclic basis, thus;

August 1st 01, 32, 64

August 2nd 02, 33, 65 etc.
Each member of the population must have the same number of numbers allocated to it.
Using this method the first two random numbers we obtained before, i.e. 46 and 51, correspond to
August 15th and August 20th.
Periodic sampling
This is a method of sampling where every nth member of the population is chosen. It is quicker
and easier than using random numbers and might be appropriate for, say, selecting names from an
electoral register. In some situations it is not suitable, since, for example, choosing every tenth
item produced by machine might coincide with a periodicity of the machine.
Stratified random sampling
As its name implies, this method involves dividing the population into strata. A random sample is
then selected from each stratum. The size of each sample is in proportion to the size of the stratum
from which it is taken. The advantage of stratified random sampling is that the accuracy of the
mean is greater than for an unstratified sample of the same size. Sometimes, however the
differences between the strata may be so great that calculation of a mean for the whole population
may seem inappropriate.
Drawing a random sample from a discrete distribution
Random numbers can be used to simulate the drawing of a sample from a given distribution.
Suppose we wish to choose a random sample of five from a Binomial distribution for which

p  1 , n = 3.
3
The possible values of the random variable and the corresponding probabilities are given in Table
2.1. We can use random numbers to draw a sample if we assign numbers so that the probability of
selecting x = 0 is 8/27 etc. these are shown in the third column of Table 2.1, using a cyclic method
as before to use as many digits as possible. From the table of random numbers we obtain the pairs
of digits. (94), 68, 81, (97), 25, 39, 68

56
And the corresponding values of r are 1, 3, 2, 1, 1
Other discrete distributions can be treated similar way. Table 2.2 shows the method for a Poisson
distribution with = 0.3. In this case the probabilities have to be rounded off, here (arbitrarily) to
four decimal places, and so the values of the variable > 5 have been
Table 2.1 Random sample from a Binomial distribution
X P(X=x) Random digits
3
2
0    8 27 01-08, 28-35, 55-62
3
2
 2   1  12
1      27 09-20, 36-47, 63-74
 3  3
2
 2 1
2      6 27 21-26, 48-53, 75-80
 3  3
3
1
3    1 27 27, 53 , 81
 3

Table 2.2 Drawing a random sample from a Poisson distribution

Grouped together. The random numbers start from 0000 so that they all have four digits. If from
the table of random numbers we obtain, for example, the four digit number 7452, the
corresponding value of the variable is 1.

57
Drawing a random sample from a continuous distribution

Suppose we wish to simulate drawing a random sample from a Normal distribution. In this case
the variety is continuous and so can be selected with varying degrees of accuracy. The probability
of getting a value of the standard deviate below a certain value is found from Table A1. For
example the probability of a standard deviate z where z < 0.65 is 0.7422 and this probability could
be represented by assigning it the random numbers 0000 to 7421. Similarly, the probability that z
< 0.64 is 0.7389 and could be represented by the random numbers 0000 – 7388. Then the
probability 0.64 < z < 0.65 would be represented by the random numbers 7389 to 7421. This
suggests that to select a random sample from a Normal distribution we can first select four-figure
random numbers and then convert them to F(z), the area under the standardized Normal probability
distribution, by putting a decimal point in front. From Table A1 the corresponding value of z is
found and hence x, the value of the variable. This is indeed the method used apart from an
important provision. The four-figure random numbers are 0000 to 9999 and if we take the
corresponding value of F(z) as 0.0000 to 0.9999, the values of z will not be symmetrically
distributed about their mean, 0. To avoid this we add 0.00005 to each F(z) giving a range of
0.00005 to 0.999 95 which is symmetrically about the mean value of F(z), i.e. 0.5 and gives values
of z symmetrical about 0.

Example 2.1

Select a random sample of four values of the variable from a normal distribution mean 10, s.d. 2.
(The measurements should be correct to 1 decimal place.) The method is set out in Table 2.3.
Four 4-digit random numbers are taken from Table A3. They are converted to values of F(z)
adding a decimal point and 0.00005. From Table A1 the range in which Z lies is found and the
(x  )
values of z are converted to values of x using z  , with  = 10,  = 2. (If necessary the

range of Z can be reduced by using interpolation in Table A.1) Correct to one decimal place the
four randomly chosen values of X are 6.6, 12.8, 14.3 and 11.1.

Table 2.3 Drawing a random sample from a Normal distribution

58
A similar method can be used for other continuous distributions. Random numbers are taken and
converted to a value of the cumulative distribution function in the same way as they were for the
Normal distribution. Then the corresponding value of the variable is found using the c.d.f. For
example, a continuous Uniform distribution over the range 1 to 3 has probability density
1 1
function f ( x)  , 1< x < 3, and cumulative distribution function F ( x )  ( x  1) . Using the
2 2
random number 6432 gives 0.643 25 for the value of F(x) so that 25 = ½ (x – 1) giving x =
2,2865
Practical Sampling
In discussing and comparing sampling schemes the following criteria should be borne in mind:
(i) the randomness of the sample,
(ii) time,
(iii) cost,
(iv) Convenience to the person being questioned.
Consider, for example, some of the ways in which a survey might be made in a small town to find
out whether parents of children under five consider the nursery school facilities satisfactory. A
truly random sample is one in which all of the parents have an equal chance of being chosen. This
could be achieved by selecting names from the electoral register using random numbers and
interviewing those chosen. This has the disadvantages that (i) the time and expense involved in
travelling to peoples’ homes would be considerable, especially since more than one visit would be

59
required if they were out, and (ii) a large number of those chosen would not have children under
five and further names would have to be chosen to replace them.
The latter disadvantage could be overcome by selecting the sample from a list of parents with
children under five, possibly obtainable from the Local Health Authority. The sampling could be
stratified by dividing the parents into groups according to which district of the town they live in
so that different income (and possibly ethnic) groups are fairly represented.
An attractive way of overcoming the first disadvantage mentioned above, i.e. the time and expense
involved in traveling, would seem to be offered by selecting names at random from the telephone
directory. This method is not satisfactory since (i) many people, generally those less well off, do
not have a phone, and (ii) people are suspicious of being phoned by strangers.
One of the simplest ways of obtaining a sample would be to stop people with small children in the
town centre. This method is quick and cheap but has the disadvantage that the sample is not
necessarily random. In practice, however, a compromise may have to be made between obtaining
a random sample and considerations of time, cost and convenience.

2.5.2 Sampling distribution

Estimators
The main purpose of taking a sample is to obtain information about the parameters of the
population from which the sample is drawn. For example, the mean of a sample gives us an
estimate of the mean of the population.
If we took another sample from the population and calculated its mean we should be most unlikely
to obtain the same value as we did for the first sample. In fact, if we continued taking samples
and calculating their means, these means would have a frequency distribution of their own. If we
consider all possible values that the mean can take when all possible random samples (of a given
size) are drawn from the population then we can form the probability distribution of the sample
mean. The random variable defined by this probability distribution, in this case the sample mean,
is called an estimator. The probability distribution of an estimator is called its sampling
distribution. In this unit you look at sampling distributions, in particular that of the mean, and the
properties which an estimator needs to give a ‘good’ estimate of a population parameter.

60
An unbiased estimate of the mean
One of the first requirements of an estimator is that it should be unbiased. This means that the
expected value of the sampling distribution of the estimator should be equal to the parameter which
is being estimated. It seems intuitively obvious that the variable, X, the mean of a sample, should
give an unbiased estimate of u, the population mean. For a sample size n, X is calculated from n
independent observations of X, which you will call x1 , x2 ,..., x n you can show that X is an

unbiased estimate of  as follows:

x1  x2  ...  xn
x
n
E ( X )  E{( X 1  X 2  ...  X n ) / n}

x x x 
E  1  2  ...  n 
n n n
 E ( x1 ) E ( x 2 ) E ( xn ) 
   ...  
 n n n , (1)
  
   ... 
n n n
E( X )  

2.5.3 The sampling distribution of the mean

You know that each sample will give a different estimate of the population mean. In order to find
out how close an estimate will be to the population parameter you need to study the sampling
distribution of the mean in more detail and to find out how it is related to the population
distribution. You will start by considering a particular numerical example.
Consider a large population which consists of equal numbers of the digits 1, 2 and 3.
The mean of the population is given by
  xp( X  x)
1 1 1
 *1  * 2  * 3
3 3 3
2
and its variance by

61
 2   x 2 p( X  x)  xp( X  x)
2

all x

1 1 1
 * 12  * 2 2  * 3 2  2 2
3 3 3
14
 4
3
2

3
If you take samples, size two, from the population, the possible pairs of values are (1, 1), (1, 2),
(2, 1), (2, 2), (2, 3), (3, 2), (3, 3), (1, 3), (3, 1). (Since the population is large you need not
concern ourselves with whether or not sampling is with replacement but the theory of this unit is
concerned only with sampling with replacement.) Figure 2.1 gives the sample space for the
means of the samples. The sample means x with their corresponding probabilities are given in
Table 2.1 together with the calculation of  xp( X  x ) and  x
all x all x
2
p( X  x ) .

You have mean of sampling distribution =  xp( X  x )  2 , which, as you expected, is the
all x

mean of the original or parent population

Figure 2.1 Sample space for the example in Section 2.5.3

Table 2.4 Probability distribution

62
 =============================================
x p( X  x ) x p( X  x ) x 2 p( X  x )

1 ⅟9 1/9 ⅟9

3/2 2/9 ⅓ ⅟9

2 3/9 ⅔ 4/3
5/2 2/9 5/9 25/18
1 1
3 ⅟9 3 13
2 3
=============================================

Parent population sampling distribution of the mean

Figure 2.2 Probability distribution
Probability distribution for the parent population and the means of the sample of the example in
2
 
Section 2.5.3 variance of sampling distribution   x p( X  x )   x p( X  x )
2

all x  
13
  22
3
1

3
Figure 2.2 illustrates the probability distribution of the parent population and the sampling
distribution of the mean. The diagram shows that the values of the sample mean cluster more
closely about the mean that the values of the variable in the parent population. This is reflected

63
in the values for the variance: the variance of the sampling distribution of the mean is half that of
the parent population. This is a particular example of a theorem which is as follows:
Theorem 2.1
If all possible random samples, size n, are drawn (with replacement) from a population, mean 
and s.d.  , then the means of the samples have a probability distribution known as the sampling

distribution of the mean, with mean  and s.d.  . The standard deviation of the sampling
n
distribution of the mean is known as the standard error (s.e.) of the mean.
Proof of Theorem 2.1

You have already shown in Section 2.5.3 that E( X ) =  and E( X ) =  and E( X ) is the mean

of the sampling distribution of the mean. The variance of the sampling distribution is var ( X ).
You have
n X 
Var( X )  var  1 
 i 1 n 
1 n 
= 2
var  X i 
n 1 
1
= var( X 1 )  var( X 2 )  ...  var( X n )
n2
1
= * n 2
n2
2
= (2)
n

Thus the s.d. of the sampling distribution of the mean .
n
If the sampling is without replacement and the population size is N then equation (2) becomes
 2  N n
var( X )   
n  N 1 
it can be seen that if n is very much smaller than N then this formula gives approximately the same
result as equation (2). In this book it is assumed that the sample size is always much less than the

64
population size so that even if sampling is without replacement (the more usual situation), equation
(2) may still be used.
Whether the sampling is with or without replacement var ( X ) decreases as the sample size
increases. This confirms what we would expect intuitively; the larger the sample size, the closer
the sample mean is likely to be to the population mean.

2.1 UNIT ACTIVITY

(1) A large population consists of equal numbers of the digits 1 and 3. Find the mean and
variance of this population.
Find the probability distributed of the mean of samples size three taken from this
population and verify that its mean is equal to the population mean and its variance is
equal to one-third of the population variance.
(2) The discrete random variance J has the distribution given in Table 2.6
(a) Find the mean  and variance  2 of the distribution. Random samples size two
are taken from the distribution. By considering all possible samples, obtain the
probability distribution of the mean of such samples. Verify that this distribution
has mean  and variance ½  2 .
(b) What would be the mean and variance of the distribution of the mean of random
samples of size three from the original distribution?
Table 2.6 The discrete random variance
===================================================
J -2 -1 0 1 2
P(J = j) 0.1 0.2 0.4 0.2 0.1
===================================================

65
2.5.4 An unbiased estimator of variance

Just as the sample mean varies from sample to sample, so does the sample variance. Here you
are concerned with using a sample to obtain an unbiased estimate of population variance,  2 .
( X i  )2 n
You might expect the random variable  to be an unbiased estimator of  2 and this
i 1 n
can be proved as follows by finding n its expected value:
 n ( X i  )2  ( X1  )2 ( X 2  )2 ( X n  )2 
E    E   ...  
 i 1 n   n n n 
( X  )2  ( X 2  )2  ( X n  )2 
= E 1   E    ...  E  
 n   n   n 

=
1
n
 1
n
  1

E ( X 1   ) 2  E ( X 2   ) 2  ...  E ( X n   ) 2
n

1 2 1 2 1
=     ....   2
n n n
=2
Unfortunately we are not usually in the position of requiring an estimate of  when we know 
. In most cases we have only an estimate of  , i.e. x . Using x we can calculate the s.d., s, of
the sample, from the formula
n
( xi  x ) 2
s2  
i 1 n

( xi  x ) 2 n
However, we cannot use s   2
to give an unbiased estimate of variance, since the
i 1 n
sum of the squares of the deviation of the xi’s from x is less than the sum of the squares of the
deviations from  and consequently s2 under estimates  2 .

You can find an unbiased estimator of  2 as follows:

66
  n (X i  )2 
  E 
2

 i 1 n 
 ( X i  X )  (   X )2 
= E  
 n 

 
 ( X i  X ) 2  2( X i  X )(  X )  (   X ) 2 
= E  
 n 

 n ( X i  X )2 n
(X  X ) (  X ) 2 
= E   2(   X ) i  n*  The second term is zero since
 i 1 n i 1 n n 

(X i  X )   X i   X = nX  nX  0

 ( X  X )2 
So you have  2  E  i   E(   X ) 
2
(3)
 n 
( X i  X )2
The first term is E(S2) where S2 is the random variable  n . The second term,

E[(   X ) 2 ] or E ( X   )  , is var( X ) which was shown in the proof of Theorem 1 to be

2

2
. Thus
n
2
 2  E (S 2 ) 
n
Rearranging,
n
2  E (S 2 )
n 1
 n 
= E S2 (4)
n 1 
Substitute for S,

 n  n ( X i  X )
2

  E 
2
  

 n  1  i 1
 n 


 ( X i  X )2 
 2  E   (5)
 n 1 

67
Equation (5) gives us an unbiased estimator of variance, which you shall denote by Ŝ 2 ,
Where
n
(X  X ) 2
Sˆ 2   i (6)
i 1 n 1

 n 
and Ŝ and S are related by Sˆ   S (7)
 n 1
The term n – 1 in the denominator of equation (6) is referred to as the number of degrees of
freedom and is given the symbol  . The reason for this name is as follows. If you knew  then
n
( xi   ) 2
the variance could be calculated from an independent deviations, using  .
i 1 n
If instead you measure the deviation
from x , then you have only n – 1 independent deviations, as when (n – 1) deviations are given
n
the last deviation can be deduced using the fact that  ( x  x )  0 . One degree of freedom has
i 1

been ‘lost’ since only n – 1 of the deviations can be varied independently. Equation (7) shows

that, for large n, there is only a small error in takings S2 rather than Ŝ 2 as an estimate of o2.
Example 2.2
Five measurements of the volume of acid required in a titration are 25.1, 25.2, 25.2, 25.0, 25.5
cm. Use the results in table 2.7 to obtain estimate for the mean and s.d of the volume of acid
required.
First you need to calculate x and s are shown in Table 2.7 and the calculations that follows.

Table 2.7 Five measurements of the volume of acid

68
 Using A = 25.0, B = 0.1 (that is using thee assumed mean of 25.0)
n

fu i i
x  A B i 1
n

f
i 1
i

10
 25  0.1 *  25.2
5
  n 
 i i   f i ui
2
f u 
2  i 1 
s B  n
2
  i 1n
 
  fi   fi 
 i 1  i 1 

 34  10  2 
s  0.1       0.0280
2 2

 5  5  

The unbiased estimate of  is x = 25.2 cm.

n 2 5
The unbiased estimate of  2 is sˆ 2  s  * 0.0280  0.0350
n 1 4
So ŝ = 0.187 cm3

2.5.5 Relative efficiency and consistency

If you have two estimators which both give an unbiased estimate of a population parameter, you
would prefer to use the one for which the sampling distribution is more closely clustered about
the true value of the parameter, i.e. the sampling distribution with the smaller variance. This is

69
the more efficient estimator. The ratio of the variances of the two sampling distributions gives a
measure of relative efficiency on a scale between 0 and 1. For example, both the mean and the
median give unbiased estimates of the mean of a Normal distribution. However, it can be shown
that for large samples, the sampling distribution of the median is Normal with standard error
1.25
, and so the median is a less efficient estimator than the mean, which has a standard error
n

of .
n
A good estimator of variance should also be consistent. This means that the larger n the closer
the statistic is likely to be to the parameter it estimates. If G is an estimator of y then this is
achieved by having
E(G)  y (8)
var(G)  0 (9)
as n  
You have already shown that X as an estimator of  satisfies the first of these criteria

2
irrespective of the value of n, and it satisfies the second since var( X ) = (see equation (2)).
n
Example 2.3
1
A discrete random variable X can take values 0, 1 and 2 only, with respective probabilities  ,
2
1
1 –  and  , where  is an unknown number between 0 and 1. Let X1 and X2 denote two
2
randomly observed values of X. List the possible values of {X1, X2} that may arise and
calculate the probability of each possibility of each possibility; verify that your probabilities sum
to unity. By calculating the value of ( X 1  X 2 ) 2 for each possible {X1, X2} determine the
1
sampling distribution of ( X 1  X 2 ) 2 . Hence show that Y  ( X 1  X 2 ) 2 is an unbiased estimator
2
of  and express its sampling variance in terms of 0. Since  is the probability that X will not
take the value 1, another possible estimator of  is the proportion of sample values not equal to

70
 1
1; for a sample of two observations this estimator is given by Z  N , where N is the number
2
of observations (0, 1 and 2) not equal to 1. State, giving your reasons, which is Y and Z you
would prefer as the estimator of  . Table 2.8 sets out the possible values (x1, x2) together with
their probabilities. Then you should check that the sum of these probabilities is 1. The third
column gives the value of ( x1  x 2 ) 2 for each {x1, x2}. Table 2. 9 gives the sampling distribution
of (X1 – X2)2. The fourth and fifth columns of this table give yP(Y = y) and y2 P(Y = y) which
are required for
Table 2.8 discrete random variable X

Table 2.9 Probability discrete random variable

71
 1
Calculating the mean and variance of the sampling distribution of Y (where Y = X1  X 2 )2 .
2
You find E(Y) =  yP(Y  y)   , showing that Y is an unbiased estimator of  . The
all x

sampling variance of Y is given by

3 1
y 2
p (Y  Y )  [ yp( y  y )] 2   2     2
2 2
1 1
 2  
2 2
1
  (  1)
2
N, the number of observations in a sample of two which are not equal to 1, is Binomially
distributed with n = 2, p =  , q = 1 –  . Thus the mean and variance of the sampling
distribution of N are mean = np = 2  , variance= npq = 2  (1 -  )
Since Z = ½ N, the mean and variance of its sampling distribution are mean = ½ x 2  =  (see
equation (4)) variance = (½)2 x 2  (1 –  ) = ½  (1 –  ).
Since the sampling distribution of Z Has the smaller variance, Z is the preferred estimator.

2.2 UNIT ACTIVITY

(1) Each trial of a random experiment has probability p (0 < p < 1) of yielding a success.
In n1 independent trials of the experiment the number of successes obtained was r1

72
 Write down an unbiased estimate, p1 of p and find its standard error in terms of n1 and

p. In a further n 2 independent trials of the same experiment the number of successes

obtained was r2 . Let p 2 denote the unbiased estimate of p from these n 2 trials. Verify
1
that ( p1  p 2 ) is an unbiased estimate of p, and find its standard effort in terms of n1
2
, n 2 and p.
1
Determine the range of values of the ratio n1/n2 for which the estimate ( p1  p 2 ) is to
2
be preferred to each of p1 and p 2
(2) Explain the terms (a) unbiased estimator, (b) consistent estimator, and indicate how
possession of these properties helps to ensure a ‘good’ estimator.
Let X 1 , X 2 ,..., X n be a random sample from some population with distribution f(x) and

let S 2 
(X i  X )2
, for n > 1
n 1
Show that S2 is an unbiased and a consistent estimator for  2 , the population variance.
(3) Explain what is meant by the sampling distribution of estimator.
Explain what you understand by (a) an unbiased estimator, (b) a consistent estimator,
(c) the relative efficiency of two estimators of the same parameter.
In order to estimate the mean  of a population, random observation x1 , x 2 , x3 are taken

of are taken of a random variable X which has variance  2 . Find the relative efficiency

of the two estimators 1 and  2 where

x1  x2  x3 x  x 2  x3
ˆ1  , ˆ 2  1
3 3
(4) A random sample of n1 observations is made from a population with unknown mean u
and variance  2 . For this sample the mean x1 and variance s12 are calculated. A second

sample, size n 2 , has mean x 2 and variance s 22 . Show that an unbiased estimate of the
population mean  is given by

73
 n1 x1  n 2 x 2
n1  n 2

and an unbiased estimate of the population variance  2 is given by

n1 s12  n2 s 22
n1  n2  2

(5) To estimate the mean of a population, two observations x1 and x 2 are made of the

random variable X, which has mean  and variance o2. Show that ˆ  kx1  (1  k ) x2 is
an unbiased estimator of  and find the value of k for which this estimator is most
efficient.

2.5.6 The central limit theorem

If a sample, size n, is taken from a normal population, the sampling distribution of the mean is
also normal. Even if the original population is not normal, it can be shown that, as n increases,
the sampling distribution of the mean approaches a normal distribution.
This result is called the central limit theorem and is of fundamental importance. It is usually
assumed that for values of n > 30 the sampling distribution of the mean is Normal.
Example 2.4
What is the probability that the mean of 100 digits taken from a random number table is greater
than 5.0? Random numbers form a discrete Uniform distribution. By symmetry mean = 4.5
The variance for a Uniform distribution where the variable takes the values 1, 2, 3, . . . n was
1 2
shown to be ( n  1) . The numbers 0, 1, 2, . . . , 9 will have the same
12

74
 Figure 2.3 Graph to illustrate Example 2.4
Variance as 1, 2, 3, . . . , 10, since a change of origin does not alter the variance. Thus variance
1
of random numbers = (10 2  1) = 8.25. Using equation (2), the variance of the sampling
12
2 8.25
distribution of the mean of 100 random digits is   0.0825, and the mean of the
n 100
sample distribution of the mean is 4.5. The central limit theorem tells us that the sampling
distribution of the mean is Normal as shown in Figure 2.3. The s.d. of this distribution is
0.0825  0.287 . The probability that the mean of 100 members is greater than 5 is given by the
5 . 0  4 .5
shaded area. z   1.74
0.287
Required probability = p ( z  1.74)
= 1 – P(Z < 1.74)
= 1 - 0.9591
= 0.0409

75
2.3 UNIT ACTIVITY

(1) The body length of a certain species of insect is normally distributed with mean 3.1
mm, s.d. 0.2 mm.
(a) What is the probability that n insect chosen at random has a body length greater
than 3.12 mm?
(b) What is the probability that the mean body length of a sample of (i) 9 insects,
and (ii) 100 insects is greater than 3.12 mm?
(2) Explain what is meant by the standard error of the mean.
For each of the following large populations (a) – (c), three possible standard
deviations (i)-(iii) are given. In each case, selected the standard deviation you believe
to be the most reasonable and use it to obtain a range of values within which the mean
of a sample of 100 values is like to lie with 95% probability.
(a) The weights in kg of adult males. Mean is 66 kg, s.d:
(i) 1, (ii) 8, (iii) 16.
(b) The salaries of typists. Mean is 795, s.d:
(i) 110, (ii) 425, (ii) 1410.
(c) The number of matches in a box of nominal contents 50. Mean is 51.6, s.d:
(i) 0.6, (ii) 5.3, (iii) 20.2.
(3) In a certain examination with a very large entry, the percentage marks obtained by the
male candidates were found to follow a Normal distribution with a mean of 54 and a
standard deviation of 16. Let X denote the mean of the percentage marks scored by a
random sample of four male candidates. What is the sampling distribution of X ?
Calculate the probability that X will exceed 70 and the value c such that there is a
probability of 0.95 that X will be within c marks of the mean mark of 54.
In the same examination the percentage marks obtained by the female candidates were
found to follow a Normal distribution with a mean of 59 and a standard deviation of 20.
Let Y denote the mean of the percentage marks scored by a random sample of five female

76
 candidates. What is the sampling distribution of Y  X ? Calculate the probability that the
value of Y will be greater than the value of X .

2.5.7 An unbiased estimator of population proportion

If the probability that a member of a population possesses an attribute is p then we know from
that X, the number of members of sample size n which possess this attribute, is Binomially
distributed, B(n, p). We might expect P s = X/n, the proportion in the sample which shows the
attribute, to be an unbiased estimator of p and this can be shown as follows:
X
E ( Ps )  E ( )
n
 E( X ) / n (10)
 np / n
p

You can also find the variance of the sampling distribution of Ps since

Var ( Ps )  Var ( X / n)

 Var ( X ) / n 2

 np(1  p) / n 2 (11)
 p (1  p) / n
When n is large the sampling distribution of X tends to a Normal distribution. Thus the sampling

distribution of Ps also tends to a Normal distribution with mean p/n and variance p(1  p) / n .
The standard deviation of this sampling distribution, p(1  p) / n , is called the standard error
of a proportion.

2.4 UNIT ACTIVITY

(1) The number of days that each of five employees (A, B , C, D, E) in an office was absent
from work during a year is shown in Table 2.10
Table 2.10

77
 =============================================
Employee A B C D E
Number of days absent 10 6 0 4 0
=============================================
Calculate the mean u and the variance  2 f the numbers of days these employees were
absent from work.
(a) Three of these employees are selected at random without replacement. Let X denote the
mean number of days absent for the chosen three employees.
(i) Determine the sampling distribution of X and display it in a table.
(ii) Determine whether or not X is an unbiased estimator of  .
(iii) Find the variance of X and verify that it is equal to one-half of the variance of the
sample mean if the three employees are chosen with replacement.
(2) A random sample x1 , x 2 is drawn from a distribution with mean  and standard

deviation  . state the mean and standard deviation of the distribution of (a) x1 + x 2 , (b)

x1 - x 2 , (c) x
A student’s performance is equally good in two subjects. The marks she might be
expected to score in each subject may be treated as independent observations drawn from
a Normal distribution with mean 45 and standard deviation 5. Two procedures might be
used to decide whether to give the student an overall pass. One is to demand that she
passes separately in each subject, the pass mark being 40; the other is to require that her
mean mark in the two subjects exceeds 40. Find the probability that the standard will
obtain an overall pass by each of these procedures.
(3) The random variable X 1 , X 2 ,..., X n are independent and each has a Normal distribution

with mean u and variance 1. The random variable X is defined to be ( X 1 , X 2 ,..., X n ) / n .

Determine, in terms of n, the value u which is such that, when u = 0, the probability of X
exceeding v is 0.05. For this value of v it is desired that the probability of X being less
than v when u = 0.2 should be at most 0.10. Calculate the smallest possible value of n
which satisfies this requirement.

78
 (4) Explain what is meant by the sampling distribution of the mean, and discuss briefly the
properties of the distribution.
By considering all possible outcomes, find the sampling distribution of the total score
obtained when two unbiased dice are thrown. Find the mean and variance of the
distribution. If 2n dice are thrown, and n is large, what distribution is a close approximation
to that of the total score?
(5) Explain what is meant by the expectation of a random variable.
A random sample is drawn from a large population in which a proportion p have a certain
rare disease. Sampling continues until a predetermined number a of the sample are found
to have the disease, and at this stage the sample size is r. find the probability distribution
a 1  a  1
of r, and show that is an unbiased estimate of p, i.e. that E  p
r 1  r  1 

2.0 UNIT SUMMARY

Unbiased estimate of population mean, 

Unbiased estimate of popular variance,  2 ,

Unbiased estimate of population proportion, p, PS  x

n
The sampling distribution of the mean has mean  and variance  2 where  and  2 are the
population mean and variance respectively. When n is large this sampling distribution
approximates to a normal distribution; this result is known as the central limit theorem.
The sampling distribution of the sample proportion has mean p and variance p(1  p) / n , where p
is the population proportion. When n is large this sampling distribution approximates to a Normal
distribution.
The sampling distribution of the mean
Using the table of random numbers, Table A3, as the population, take 200 samples of ten digits
each. Calculate the mean of each sample and present the results in a frequency table with classes
1.4- 2.0, 2.1, 2.7, 2.88 3.4, 3,5-4.1, 4.2- 4.8 etc. (These classes have been chosen because they
are symmetrically distributed about the expected mean.) Illustrate by a frequency polygon.

79
 Calculate:
(i) the mean variance for the original population, i.e. a discrete Rectangular distribution for
which P(r) = 1/10 for r = 0, 1, 2, . . . ,9;
(ii) the mean and variance for the frequency table of the sample means.
Verify that (a) the mean of the means of the samples is approximately equal to the mean of
the original population, (b) the variance of the distribution of the means is approximately equal
to one-tenth of the variance of the original population (since in this case n = 10). (The agreement
is only approximate since not all of the possible samples are considered.)
The population from which the sample were drawn has mean 4.5, variance 99/12 so that, using
the central limit theorem, we expect the sampling distribution of the mean to be approximately
99
Normal with mean 4.5, variance  0.825 . Using these parameters calculate the expected
12 *10
frequencies for classes 1.4- 2.0, 2.1- 2.7, 2.7 etc, for the means of 200 samples. Compare with
the observed frequencies by drawing a frequency polygon using the same axes as before.

80
UNIT 3 CONFIDENCE LIMITS

3.1 Unit Introduction:

Welcome to Unit 3 in which you will learn how to construct confidence interval (CI). You will
learn types of interval estimates, computed from the statistics of the observed data that might
contain the true value of an unknown population parameter. Most commonly, the 95% confidence
level is used. However, other confidence levels can be used, for example, 90% and 99%. An
interval estimate gives a range of values which has a certain probability of containing the
population parameter. This unit describes the way in which some interval estimates are calculated
from the appropriate sampling distribution.
In the previous unit you learnt that the mean of a sample gave an unbiased estimate of the
population mean. Such an estimate which is the form of a single value is sometimes called a point
estimate. This is to distinguish it from an alternative form of giving an estimate called an interval
estimate. An interval estimate gives a range of values which has a certain probability of containing
the population parameter. In this unit you will study methods of constructing confidence intervals
from the appropriate sampling distributions.
3.2 Unit Aim

The aim of this Unit is teach you on how to construct confidence interval (CI).

3.3 Unit Objectives

By the end of the unit you should be able to:

 Get good estimates of the unknown population parameter.

 Construct confidence intervals.

Terminology

ŝ : Sample standard deviation, this is unbiased estimator of the population standard deviation

s : Sample standard deviation biased

81
 2 : Population variance

 : Population standard deviation


x : Sample mean

 : Population mean

Z  : Degree of confidence, table value

2

3.4 Unit Time required

You need 20 hours for this unit

3.5 Unit Topics

3.5.1 Confidence limits of the mean (  known)

You will start by calculating an interval estimate for the population mean,  in the situation where
the population standard deviation,  , is known but  is not. If n is large, the central limit theorem
tells you that the sampling distribution of X will be normal, as shown in Figure 1, with mean 

and s.d.  . You can use this distribution to find arrange of values within which X will lie
n
with a certain probability. Suppose you wish to calculate a range of values for X so that X lies
within this range with 95% probability. Let x u and x l be the upper and lower values of this range
respectively. They are shown in Figure 1 placed symmetrically on either side of u. the shaded area
is 0.95, so that the area above x u is 0.025 and0 below it 0.975. Using Table A1, the standard

deviates corresponding to x l and x u are – 1.96 and 1.96 respectively, so that

xU   1.96 x  1.96
1.96  , giving xU    and  1.96  L , giving x L   
 n  n
n n

82
 1.96 1.96
Thus there is a 95% probability that X lies in the range   < X < 
n n
This inequality can be rearranged to give an interval which includes  with probability of 95%
1.96 1.96
X <  < X
n n
In this expressing X is a random variable. If, for a particular sample, X takes the value x , you
1.96 1.96
can calculate an interval x  <  < x . (1)
n n
This interval is called the 95% confidence interval of the mean and its end-points are called the
95% confidence limits of the mean.  is fixed, and so for a particular sample the interval
calculated in (1) either does or does not include  and we do not know which is the case. What
we do know is that if we calculate confidence intervals by this method then is a probability of 95%
that the interval does include  . Other confidence limits are sometimes used, e.g. 99 and can be
found by using the appropriate value of the standard deviate. The 99% confidence limits are
2.58 2.58
x <  < x (2)
n n
Notice that the confidence interval must increase if we are to become more confident that it
contains  . Equations (1) and (2) can also be used for a small sample from a Normal
population, since in this case the sampling distribution is Normal.

Figure 3.1 Graph to illustrate normal distribution

Example 3.1

83
The standard deviation for a method of measuring the concentration of nitrate ions in water is
known to be 0.05 ppm. If 100 measurements give a mean of 1.13 ppm, calculate the 95%
confidence limits for the true mean.
Solution
Using x = 1.13, n = 100 and  = 0.05,
1.96 * 0.05 1.96 * 0.05
1.13  <  < 1.13 
100 100
1.12 <  < 1.14

3.1 UNIT ACTIVITY

(1) A machine produces washers whose diameter is known to be normally distributed with a
s.d. of 0.04 mm. in order to find the mean diameter of the washers produced, a random
sample of nine washers is taken whose mean diameter is found to be 3.14 mm.
Calculate (a) 95%, (b) 98% confidence limits for the mean diameter of washers produced
by the machine.
(2) The masses of sweets produced by a machine are known to have a standard deviation of
0.5 g. A sample of 50 sweets has a mean mass of 15.21 g. Calculate a 99% confidence
interval for the mean mass of sweets produced by the machine. Why is it not necessary to
assume that the masses are normally distributed in this calculation?

3.5.2 Confidence limits for a large sample (  unknown)

In practice it is unusual to know  2 but not  . However, you can obtain an unbiased estimate

ŝ 2 of  2 from the sample. The estimate varies from sample to sample but for large samples the
variation is so small compared with ŝ 2 itself that we can replace  2 by ŝ 2 in equation (1). This
gives for the 95% confidence limits of the mean
1.96sˆ 1.96sˆ
x <  < x (3)
n n

84
 n
Since sˆ  s (see equation (7)), this expression can also be written as
n 1
1.96s 1.96s
x <  < x (4)
n 1 n 1
Example 3.2
Fifty measurements of the acceleration due to gravity, g, had a mean value of 9.8 ms -2 and a s.d.
of 0.75 s 2 . What are the 95% confidence limits for g?
Solution
Using equation (2) with x = 9.8, s = 0.75 and n = 50, the 95% confidence limits of the mean are
1.96 * 0.75 1.96 * 0.75
9.8  <  < 9.8 
49 49
Therefore, the 95% confidence interval is given by: 9.6 <  < 10.0. That is you are 95%
confidence that the true mean for the population is in the interval [9.6, 10]
Example 2.3
In the previous example how many measurements would be necessary to reduce the 95%
confidence limits to 9.7    9.9 ?
You have
1.96s
 0.1
n 1
Rearranging
n  1  (1.96 * 0.75 / 0.1) 2
n = 217

3.2 UNIT ACTIVITY

(1) Fifty children were selected at random from the pupils at a school and each was asked how
many hours a week he or she spent in watching television. The mean of the sample was
17.2 hours and the standard deviation was 5.3 hours. Calculate the (a) 95%, (b) 98%
confidence interval for the mean number of hours spent a week in watching television for
the population consisting of all the children in the school.

85
(2) Fifty boxes of matches were selected at random from a large carton of such boxes. The
numbers of matches in each of the 50 boxes were counted and the mean and standard
deviation of these numbers were found to be 48 and 0.5 respectively. Between what limits
would you expect the mean for all of the boxes in the carton to lie with 0.95 probability?
(3) On 1 January 100 new Eternity light bulbs were installed in a certain building, together
with a device which records how long each light bulb is in use. By 1 March all 100 bulbs
had failed, and the recorded lifetimes, t (in hours of use since 1 January) are summarised
in Table 3.1.
Obtain values for the sample mean and the sample standard deviation for this set of data.
(Assume measurements are to the nearest second.) Obtain values for the sample mean and
the sample standard deviation for this set of data. (Assume measurements are to the nearest
second.)
Assuming that the bulbs constituted a random sample of Eternity light bulbs, obtain a
symmetric 99% confidence interval for the mean lifetime of Eternity light bulbs.
Table 3.1 Random sample of Eternity light bulbs

Table 3.2 The diameters of the heads of 100 rivets

86
(4) The random variable X takes values xi with associated frequencies fi (i = 1, 2, . . ., n)
and the mean of these values is x , N   f . If c is a constant prove the formula

1 1
N
 f (x  x) 2

n
 f ( x  c) 2  ( x  c) 2

The diameters of the heads of 100 rivets from a production line were measured correct to
the nearest 0.1 mm and the results were as in Table 3.2.
Using an assumed mean of 11.4, calculate the sample mean and standard deviation of the
diameters.
Determine 95% confidence limits for the population mean diameter, stating any
assumptions that you make.

3.5.3 Confidence limits for a small sample from a Normal population (  unknown)

If small samples, size n, are drawn from a Normal population, their means are normally distributed

with mean  and s.d. . However you can no longer use the confidence limits derived for
n
(x  )
large samples because this derivation rested on the fact that was normally distributed. For

n

large samples, even when you did not know  , you could assume that ( x   ) /( sˆ ) was still
n
approximately Normal. This is no longer true for small samples because the variation of s from
(x  )
sample to sample is too large to be ignored. The statistic is called ‘t’ and its distribution
sˆ
n
is called the Student t-Distribution Its form depends on the ‘degrees of freedom’ , mentioned
in connection with the Estimation of  (see Section 2.5.4) and, as in that case, we have  = n – 1.
For large n, the t-distribution
Approximates to the Normal distribution, a property which was used in Section 3.5.2. As n
decreases the Distribution remains bell-shaped but becomes more spread out as shown in Figure
2.2. This spread Reflects the uncertainty introduced because of the variation in ŝ .
To denote the dependence of t on the degrees of freedom you write

87
 x
t n 1  (5)
sˆ
n

 n 
Since sˆ   s
 n 1
x
you can also write t n 1  (6)
s
n 1
Because n can take any value, tables for the t-distribution are not given in the form of cumulative
probability values. Instead they are given in the form shown in Table A4. This table gives what
are called the percentage points of the t-distribution for different values of n.

Figure 3.2 Graph comparing the standard Normal distribution and the t-distribution with
v = 4.
The percentage points give a value of t outside which t lies with a certain probability. For example,
if n = 8 (== v = 7) then, using Table A4, 5% of the values of t lie outside the range – 2.36 to +
2.36 and consequently 95% of the values lie within this range. In general if we require the  %
confidence limits for n- 1 degrees of freedom, the appropriate value of t will lie in the row for v
= n – 1 and the column for (100 –  )%. We can denote this value by t n1,(100 )% . Equation (2) for

the 95% confidence limits becomes

s s
x  t n 1,,5% *    x  t n 1,5% * (7)
n 1 n 1
and in general the  % confidence limits will be
s s
x  t n 1,,(100 )% *    x  t n 1,(100 )% * (8)
n 1 n 1

88
The last line of the t – table, v =  , gives the percentage points of the Normal distribution.
Example 2.4
Ten measurements of the zero error on an ammeter yield the results 0.13, - 0.09, 0.06, 0.15, -
0.02, + 0.03, + 0.01, - 0.02, - 0.07, + 0.05 A. Calculate the 95% confidence limits of the mean
zero error. (Assume the errors are normally distributed.)
1
x 0.13  0.09  0.06  0.15  0.02  0.03  0.01  0.02  0.07  0.05 x  0.023
10
You now calculate the sample standard deviation:

s2 
1
10
 
0.0132  0.09 2  0.06 2  0.152  0.02 2  0.032  0.012  0.02 2  0.07 2  0.052 -

2
 0.23 
 
 10 
s  0.0742
We have no = 10, v = 9. From Table A4, 9.5% = 2.26. Using equation (7), the 95% confidence
limits of the mean are
s 0.742 0.0742
0.023  2.26 *    0.023  2.26 * - 0.033 <  < 0.097
9 9

3.3 UNIT ACTIVITY

(1) The diameters of 25 steel rods are measured and found to have a mean of 0.980 cm and a
standard deviation of 0.015 cm. Find 99% confidence limits for the population mean
(2) The masses, in grams, of thirteen ball bearings taken at random from a batch are
21:4, 23.1, 25.9, 24.7, 23.4, 24.5, 25.0, 26.9, 26.4, 25.8, 23.2, 21.9
Calculate a 95% confidence interval for the mean mass of the population, supposed
Normal, from which these masses were drawn.
(3) A random sample of sic eggs taken from a day’s production of a poultry farm has the
following masses, measured in grams: 51.2, 52.6, 53.1, 53.2, 53.2, 54.7. Making suitable

89
 assumption, which should be stated, find 95% confidence limits for the mean mass of the
eggs produced that day.
(4) Twelve cotton threads are taken at random from a large batch, and the breaking strengths
are found to be 7.41, 7.01, 8.34, 8.29, 8.08, 6.60, 6.59, 7.39, 4.72, 8.65, 8.51 and 8.01
respectively. Assuming the breaking strengths of the threads form a Normal distribution,
find 95% confidence limits for the mean breaking strength of threads in the batch.

3.5.4 Confidence interval of a population from a large sample

In Section 2.5.7 you saw that the sampling distribution of Ps, the proportion of a sample possessing
a given attribute, was Normal with mean p/n and variance p(1 – p)/n, where p is the proportion of
the population possessing the attribute and n is the sample size. We can develop an argument
similar to that of Section 3.5.1 to give an equation corresponding to equation (1) for the 95%

confidence interval for p: p s  1.96  p(1  p) / n  P  ps  1.96  p(1  p) / n

Since we do not know the value of p in order to estimate the standard error of the proportion,
 p(1  p) / n , you must replace p by p s in this term to give as the approximate 95%

confidence interval of a proportion

p s  1.96  p s (1  p s ) / n  P  p s  1.96  p s (1  p s ) / n (9)

Example 3.5
An opinion poll is taken as to how an electorate will vote in a forthcoming referendum. Out of a
random sample of 100, 40 says ‘yes’ and 60 say ‘no’. What is the 95% confidence interval for
the proportion of the population who will vote ‘yes’?
We have n = 100 and p s = 40/100 = 0.4. Substituting these values in equation (1) gives the 95%
confidence interval of the population proportion voting ‘yes’ as
0.4  1.96 0.4  0.6 / 100  P  0.4  1.96 0.4 * 0.6 / 100 0.3 < p < 0.5
It should be noted that, provided the sample is sufficiently small compared with the population for
p to be regarded as constant, the accuracy with which we can estimate a proportion depends only
on the absolute size of the sample and not on its size relative to the whole population.

90
Example 3.6
How large should the sample have been in the previous example to reduce the confidence
interval to 1%?

 0.4 * 0.6 
The confidence limits required are 0.4 + 0.005 giving 1.96    0.005 , n = 36 880
 n 

3.4 UNIT ACTIVITY

(1) Out of a random sample of 50 children from a school, 24 were found to have been
vaccinated against whooping-cough. Calculate the 95% confidence limits for the
proportion of children at the school who have been vaccinated against whooping-cough.
(2) At a school of 1200 pupils it is found that a random sample of 40 pupils contains five who
are left-handed. Find the 95% confidence limits for
(a) the proportion of pupils in the school who are left-handed,
(b) the number of pupils in the school who are left-handed.
(3) In a sample of 400 shops taken in 1972, it was discovered that 136 of them sold carpets at
below the list prices which had been recommended by manufacturers.
(a) Estimate the percentage of all carpet-selling shops selling below list price.
(b) Calculate the 95% confidence limits for this estimate, and explain briefly what
this means.
(c) What size sample would have to be taken in order to estimate the percentage to
within + 2%?
(4) Table 3.3 shows the number of cases of and deaths from diphtheria at the City of Ndola
Hospital in the years 1900- 10 under antitoxin and ordinary treatments.
Obtain an estimate (with an estimate of the standard error) of the probability that a patient
treated with antitoxin died.

Table 3.3 Deaths from diphtheria at the City of Ndola

91
 Give an approximate 95% symmetric confidence interval for this probability. It was
required to estimate this probability with a standard error of 0.01. Estimate, to the
nearest 100, the number of patients that should have been treated.
Give an approximate 95% symmetric confidence interval for the probability that a
patient treated with the ordinary treatment died.
Comment on the relative effectiveness of the antitoxin treatment and the ordinary
treatment, stating, in coming to your conclusions, any assumption you have made about
the allocation of patients to the different treatments.
(5) (a) When an object is weighed on a chemical balance the readings obtained are
subject to random errors which are known to be independent and Normally
distributed with mean zero and standard deviation 1 mg. A certain object is to be
weighed nine times on such a balance and the mean of the nine readings is to be
calculated.
Find the probability that the mean of the nine readings will be within 0.5 mg of the
true weight of the object.
(b) Another weighing device is undergoing tests to determine its accuracy. A certain
object of known true weight 50 mg was weighed ten times on this device and the
readings in mg were 49, 51, 49, 52, 49, 50, 52, 51, 49, 48
(i) Calculate an unbiased estimate of the variance of the errors in readings
using this device.
(ii) Calculate 95% confidence limits for the mean error in readings using this
device.
(6) Discuss briefly the relative merits of estimating an unknown parameter by means of
either a single value or a confidence interval.

92
(a) Crates of bananas are packed in the West Indies with a nominal net mass of 55 kilograms.
However, on their arrival in Liverpool, this has usually decreased due to ripening and
shrinkage. A large batch of such crates has just arrived aboard SS Gauss. A random
sample of twelve crates is selected and their net masses are recorded in kilograms as
listed below. 56.4, 52.1, 49.5, 56.4, 48.1, 54.5, 47.8, 58.0, 48.4, 53.9, 46.7, 56.0
Assuming that this sample came from an underlying Normal population with variance 16,
calculate a 95% confidence interval for the mean mass of the population.
(b) The drained masses, in kilograms, of ten catering size tins of peaches taken at random
from a batch are 2.57, 2.05, 1.65, 2.62, 2.44, 1.48, 2.31, 1.58, 2.60, 1.85
Assuming that this sample came from an underlying Normal population, calculate a 95%
confidence interval for the mean of the population.
(7) (a) In order to calculate a confidence interval for the mean mass of packets of
butter produced by a machine, a random sample of ten packets is taken. These have
masses (measured in kg) of x1 , x 2 ,..., x10 such that
10 10

x
i 1
i  2.57 , 
i 1
xi2  06610

Calculate 95% confidence limits for the mean.

(b) If it is known that the standard deviation of the mass of a packet of butter is
0.008kg, what is the least number of packets that would need to be sampled in order to
give a 95% confidence interval for the mean mass whose width is less than 0.002kg?
(8) A random sample of 600 was chosen from the adults living in a town in order to
investigate the number x of days of work lost through illness. Before taking the sample it
was decided that certain categories of people would be excluded from the analysis of the
number of working days lost although they would not be excluded from the sample. In
the sample 180 were found to be from these categories. For the remaining 420 members
of the sample x = 1260 and x2 = 46000.
(a) Estimate the mean number of days lost through illness, for the restricted
population, and give a 95% confidence interval for the mean.
(b) Estimate the percentage of people in the town who fall into the excluded
categories, and give a 99% confidence interval for the this percentage.

93
 (c) Give two examples, with reasons, of people who might fall into the excluded
categories.
(9) (a) A random sample of n observations from a population distribution had the
values x1 , x 2 ,..., x10 , whose mean is x. Show that for any value of c,
n n

 ( xi  c) 2   ( xi  x ) 2  n(c  x ) 2
i 1 i 1

n
Hence find the value of c for which  (x
i 1
i  c) 2 is a minimum.

(b) A random sample of 12 values from a normal distribution, whose mean  and

variance  2 are unknown, were such that

12 12

x
i 1
i  5472,  (x
i 1
i  450) 2  1620

(i) Calculate unbiased estimate of  and  2 .

(ii) Determine a 95% confidence interval for  .
(iii) Given that (451, 463) was a 95% confidence interval for  based on
another random sample of 12 values from the same Normal
distribution, deduce the corresponding unbiased estimate of  and  2
from this sample
(10) A random sample of 500 fish is taken from a lake, marked, and returned to the lake. After
a suitable interval a second sample of 500 is taken and 25 of these are found to be
marked. By considering the number of marked fish in the second sample, estimate the
number of fish in the lake and, by considering a confidence interval for the proportion of
marked fish in the lake, obtain a 95% confidence interval for the number of fish.
(11) (a) A school dental service wishes to estimate the average number of teeth with fillings of
the 600 pupils at a Secondary School. Explain how you would select a sample of pupils
for a dentist to inspect.
(b) A survey of 3000 randomly chosen households in England revealed that 250 had
moved during the previous year. Estimate 95% confidence limits for the percentage of
households moving during the year considered.

94
(12) An estimate is required of the proportion of a large number of consumers who are likely
to purchase a particular brand of butter. Determine the smallest sample size that should
be taken in each of the following situations:
(a) The population proportion is known to be in the range from 0.1 and 0.2 and it is
required that there should be a probability of at least 0.99 that the difference
between the sample proportion and the population proportion is less than 0.02.
(b) Nothing is known about the value of the population proportion and it is required
that there should be a probability of at least 0.95 that the difference between the
sample proportion and the population proportion is less than 0.03.
(9) When s independent sets of n Binomial trials produce r1 , r2 ,..., rs
ri
successes the formula pˆ   is used to find an estimate p̂ of p, the
sn
probability of success in a single trial. Obtain the variance of p and show that
approximate 95% confidence limits for p are given by the roots of the equation
sn( sn  3.84) p 2  2( ri  1.92) snp  ( ri ) 2  0 if 1.962 is taken as 3.84.

In one set of observations n = 10, s = 5 and ri takes the values 1, 2, 1, 3 and 1. Determine
whether p = 0.1 lies within the confidence limits given by the equation.
(13) The continuous variable X has probability density function
 2(a  3  x)
 , for a  x  a  3
f ( x)   9
 0 , elsewhere

In this a is an unknown constant. In a single observation X is found to be 5. Find an unbiased

estimate for a and limits within which a lies with 95% confidence.
A sample of 50 values of X is found to have a mean of 5.1. Find an unbiased estimate for a
based on this sample and limits within which a will lie with approximately 95% confidence.

3.5.5 Confidence Intervals for Variances

You checked off the estimation of a number of population parameters already. Let's check off a
few more! In this unit, you will derive (1−α)100% confidence intervals for:

95
(1) a single population variance: 

 X2  Y2
(2) the ratio of two population variances: or
 Y2  X2

Along the way, we'll take a side path to explore the characteristics of the probability distribution
known as the F-distribution.

3.5.6 Confidence interval of one Variance

Let's start right out by stating the confidence interval for one population variance.

Theorem. If X 1 , X 2 ,...., X n are normally distributed and a  12 and b  12

2 n 1 2 n 1

, then a (1−α)% confidence interval for the population variance σ2 is:

 (n  1) s 2 (n  1) s 2 
 2  
 b a 

And a (1−α)% confidence interval for the population standard deviation σ is:

 n  1s n  1s 
   
 b a 


Proof. We learned previously that if X 1 , X 2 ,...., X n are normally distributed with mean μ
and population variance σ2, then:

(n  1) S 2
~  n21
 2

Then, using the following picture as a guide:

96
with a  12 and b  12 , we can write the following probability statement:
2 2

 (n  1) S 2 
P a   b  1  
  2


Now, as always it's just a matter of manipulating the quantity in the parentheses. That is:

(n  1) S 2
a b
2

Taking the reciprocal of all three terms, and thereby changing the direction of the inequalities,
we get:

1 2 1
 
a (n  1) S 2
b

Now, multiplying through by (n−1)S2, and rearranging the direction of the inequalities, we
get the confidence interval for σ2:

 (n  1) S 2 (n  1) S 2 
 2  
 b a 

as was to be proved. And, taking the square root, we get the confidence interval for σ:

 n  1s n  1s 
    , as was to be proved.
 b a 


Example 3.7

97
 A large candy manufacturer produces, packages and sells packs of candy targeted to weigh 52
grams. A quality control manager working for the company was concerned that the variation in
the actual weights of the targeted 52-gram packs was larger than acceptable. That is, he was
concerned that some packs weighed significantly less than 52-grams and some weighed
significantly more than 52 grams. In an attempt to estimate σ, the standard deviation of the weights
of all of the 52-gram packs the manufacturer makes, he took a random sample of n = 10 packs off
of the factory line. The random sample yielded a sample variance of 4.2 grams. Use the random
sample to derive a 95% confidence interval for σ.

Solution. First, we need to determine the two chi-square values with (n−1) = 9 degrees of
freedom. Using the table in the back of the text book, we see that they are:

a  12   02.975,9  2.7 and a  12   02.025,9  19.02

2 n 1 2 n 1

Now, it's just a matter of substituting in what we know into the formula for the confidence
interval for the population variance. Doing so, we get:

 9(4.2) 9(4.2) 
 2  
 19.02 2.7 

Simplifying, you get:

1.99   2
 14.0 

You can be 95% confident that the variance of the weights of all of the packs of candy coming off
of the factory line is between 1.99 and 14.0 grams-squared. Taking the square root of the
confidence limits, we get the 95% confidence interval for the population standard deviation σ:

1.41    3.74

3.5 UNITY ACTIVITY

1. A statistician chooses 27 randomly selected dates, and when examining the occupancy
records of a particular motel for those dates, finds a standard deviation of 5.86 rooms rented.

98
 If the number of rooms rented is normally distributed, find the 95% confidence interval for
the population standard deviation of the number of rooms rented.

3.0 UNIT SUMMARY

Confidence interval of the mean

(a)  known, large samples from any population and small samples from Normal
z z
populations: x  x where z takes the values shown in Table 4
n n
Table 3.4 Critical values of the Z- table
=======================================
Confidence
Interval 95% 98% 99%
z 1.96 2.33 2.58
=======================================
zs zs
(b)  not known, large samples from any population: x  x with
n 1 n 1
values of z as in Table 3.4
(c )  not known, small samples from a Normal population:
t n 1,(100 )% s t n 1,(100 )% s
x x . with values of t obtained from Table A4.
n 1 n 1
Confidence interval of a proportion
ps  z  p s (1  p s )  p  p s  z  p s (1  p s ) , with values of z as in Table 4

99
UNIT 4 HYPOTHESES TESTING

4.1 Unit Introduction

Welcome to Unit 4 in which you learn how to conduct Hypothesis Testing. Hypotheses testing is
a statistical procedure used to determine whether the assumed statement about a problem is true
or not. Simply.
4.2 Unit Aim

100
The aim of this unit is to determine whether there is enough statistical evidence in favour of a
certain belief, or hypothesis, about a parameter.

4.3 Unit Objectives

By the end of the unit you should be able to:

 Make an inference about the population of interest on the basis of a random sample taken
from that population.

 quantify evidence against a particular hypothesis being true

Terminology

H o : This is a null hypothesis normally to be proved. It is a positive statement about your claim.

This can be written in symbols or in a statement form

H 1 : This is an alternative hypothesis. This is a negative statement.

x : The sample mean

: The population standard deviation

n : Sample size

 : Population mean

4.4 Unit Time required

You need 20 hours for this unit

4.5 Unit Topics

4.5.1 Setting up a hypothesis

A person claims that she can tell whether the tea or the milk is put into a cup of tea first. To test
her claim we ask her to perform a series of trials. In each trial she tastes two cups of tea which are
identical in every respect except that one has the tea put in first and the other the milk put in first.

101
The two cups are presented in a random order and she has to identify the cup with the milk put in
first. If she is correct seven times out of eight, ought we to accept her claim?
You may feel that if she is right more than a certain number of times, we should accept her claim;
otherwise we should reject it. To decide what should be our boundary line we need to set up a
mathematical model. If we are sceptical of her claim and think she is guessing we would expect
the probability of success at a single trial to be ½. Consequently the probability of X successes in
eight trials would be binomially distribute with p = ½, n = 8. Using this hypothesis we can calculate
the probability of getting the observed result, i.e. not more than one failure. Such a hypothesis is
called a null hypothesis. We should also set up an alternative hypothesis which states, as its name
implies, an alternative to the null hypothesis. In this case it will be that the taster is not guessing,
i.e. p > ½. We do not need to specify an exact value of the parameter for the alternative hypothesis
since it is not used to calculate probabilities.
Denoting the null hypothesis by H o and the alternative hypothesis by H 1 we write
Model: Binomial
Ho : p=½

H1 : p>½
Using the null hypothesis we can calculate the probability of making no more than one mistake
by calculating the probability of seven or more correct guesses. The Binomial distribution gives
this probability as
P(X > 7) = P( X = 7) + P(X = 8)
= (8)(½)7(½)+ (½)8
= 0.035
This means that if the taster is guessing there is a probability of only 0.035 that she will be correct
seven or more times out of eight, and so such a result is unlikely. To decide whether to accept or
reject the null hypothesis we have to be more precise about what we mean by ‘unlikely’. The
usual convention is that events with a probability of 0.05 or less are ‘unlikely’. Using this
convention the above result is unlikely to arise by chance if the null hypothesis is true and so we
reject the null hypothesis.
If the taster had been correct only six times out of eight, the required probability is the

102
 Figure 4.1 Diagram showing the critical region for the example in Section 4.5.1

Probability that the taster makes no more than two mistakes, and, assuming the null hypothesis is
true, this probability is
6 2 7 8
 8  1   1   8  1   1   1 
p( X  6)               
 6  2   2   7  2   2   2 
p ( X  6) = 0.145

Using our convention, this would not be deemed unlikely and H o retained. Thus the possible
outcomes of the experiment can be divided into two sets: X < 7 and X > 7 according to whether
or not H o is retained. Those values of the variety for which H o rejected form the critical region.
This is shown diagrammatically in Figure 4.1. A result which falls in the critical region is said to
be significant. Further, if, on the null hypothesis, the probability of failing in the critical region is
< 0.05 of 5%, then the result is said to be ‘significant at the 5% level’, and 5% is called the
significance level of the test. Other levels of significance are sometimes used, notably 1% and
0.1%.
The significance test described above is called a one-sided or one-tailed test because the alternative
hypothesis specifies that the change in p, if any, occurs in one direction only, i.e. an increase in

103
this case, and so the critical region falls in one ‘tail’ of the distribution. The following example
uses a two-sided or two-tailed test.
Example 4.1
A coin is tested for bias by tossing it twelve times and counting the number of heads. If the result
is two heads, is there evidence, at the 2% significance level, that the coin is biased? What is the
critical region for this test?
Again we start by assuming that the coin is not biased, i.e. p = ½. Before we start the test we
have no idea in which direction, if any, the coin is biased and so the alternative hypothesis
specifies a change of p in either direction. We have:
Model: Binomial
Ho : p = ½

H1 p = ½
If the null hypothesis is true then the number of heads, X, is binomially distributed. B(12, ½). In
this case either a high or a low number of heads could lead us to reject the null hypothesis and so
you require;
P( X ,2 or X  10)  2 * P( X  2) (since the distribution is symmetrical)


 1  
12
 12  1  1 
11
12  1   1  
2 10

 2 *               

 2     1  2  2   2  2   2   
= 0.0386 = 3.86% > 2%
The probability of the observed (or a more extreme) result is greater than 2% and so the result is
not significant at the 2% level. The null hypothesis is retained: there is no evidence at this
significance level that the coin is biased.
The critical region gives those values of X for which the null hypothesis will be rejected. Table
4.1 shows the probability distribution for B(12, ½).
Table 4.1 Probability distribution for X where X is B(12, ½)

104
From this table you can see that
P(X < 1) + P( X > 11) = 0.006 34 = 0.6% < 2%
P(X < 2) + P( X > 10) = 0.038 56 = 3.9% > 2%
and so the critical region, which is shared between the two tails of the distribution, is X < 1 and
X > 11.
It is most important to realise that we can rarely prove or disprove a null hypothesis. For example,
when we perform a test at the  % significance level there is a probability of  % that the result
falls in the critical region when the null hypothesis is true. This means that there is a probability
of  % that a true null hypothesis will be rejected. In Example 4.1 this means that there is a 2%
risk of saying the coin is biased when it is not. If the test had been performed at the 5% significance
level the observed result would have been significant and the null hypothesis rejected but in this
case the risk of rejecting a true null hypothesis has risen to 5%. The risk of errors involved in
significance testing is discussed further in Section 4.5.2.

4.1 UNIT ACTIVIT

(1) The tea-tasting experiment is modified so that at each trial the taster is offered four cups of
tea. One of these is ‘milk-first’ and the others are ‘tea-first’. Again the cups are presented
in a random order. She has to identify the milk-first cup. She makes the correct

105
 identification three times out of ten. Test at the 5% level her claim that she can identify the
milk-first tea correctly, stating clearly your null and alternative hypothesis. How many
times would she have to be correct before you would accept her claim, using the same level
of significance?
(2) In a multiple choice test a student has to choose between three answers for each question
in a test consisting of ten questions. If he gets six questions right test, at the 5% level, the
null hypothesis that he is guessing.
(3) The national average pass rate for an exam is 70%. A teacher finds that six out of her
twelve pupils pass. Is there evidence that this group did significantly worse than average?
(4) It is suspected that a die is biased towards a 6. This is test by throwing the die eight
times. The result is three 6s. Is there evidence at the 2% significance level that the die is
biased towards a 6?

4.5.2 Significance test on the mean of a Normal distribution (  known)

You often wish to test whether a sample is drawn from population of specified mean,  . You
will start by considering the situation in which the population is normally distributed with known
standard deviation  .
Example 4.2
A machine should be set to produce bags of sugar whose weights are normally distributed with
 = 1000 g,  = 5 g. To check the setting, a sample of nine bags is taken and the mean weight
is found to be 1003 g. Is the machine correctly set? (Assume o cannot change.) You must choose
a null hypothesis that assumes the mean given for the machine is correct, so that you can
calculate the probability of observing a sample whose mean is 1003 g. you take
Model: Normal
Ho :  = 1000 g,  = 5 g
Since you are interested in differences either way from  = 1000 g, this is a two-tailed test and

H 1 :   1000 g

106
According to H o the sampling distribution for the mean, X , of a sample of nine bags is

Normally distributed with u = 1000 g, s.d. =  = 5 / 9 = 1.67 g. For the observed sample
n
mean, 1003, you have: Z = (1003 – 1000)/1.67 = 1.80
Since you are carrying out a two-sided test we require
P(z < 1.80) + P(z > 1.80 = 2 x P(z > 1.80) (since the distribution is symmetrical)
= 2 x [1 - P(z < 1.80)]
= 2 x (1 – 0.9641)
= 2 x 0.0359
= 0.0718 = 7.18%
If you adopt a 5% significance level for our test then this result is not significant: there is no
evidence that machine setting is incorrect.
A quicker method of arriving at the same result is to work with the critical region for the statistic
z rather than to calculate an exact value of the probability, as was done above. Figure 4.2 shows
the standard Normal distribution with the critical region for a test at the 5% significance level.
Since the test is two-sided the critical region is divided equally between the two tails of the
distribution. If a result falls in the critical region then the null hypothesis is rejected. For a two-
sided test, this happens when |z| > 1.96. In the calculation above you found that z = 1.80: this result
does not fall in the critical region and so there is no reason to reject the null hypothesis.

Figure 4.2 Graph showing the critical region for example 4.2

107
Suppose the previous example had been altered so that the question asked had been ‘Is the
machine set too high?’ We now need to carry out a one-tail test for which we have:
Model: Normal
Ho :  = 1000 g,  = 5 g

H1 :  > 1000 g,
The critical region is shown in figure 3 and is given by z > 1.64. The previous calculation gave
z = 1.80. This value is significant at the 5% level. you reject the null hypothesis and accept the
alternative hypothesis: the machine is set too high.

Figure 4.3 Graph showing the critical region for a one-tail test

The critical values of z can be conveniently found from the last row, v =  , of Table A4. For a
two-tail test at the (100   )% level the critical value is given in the --% column; for a one-tail
test in the 2 % column.
In practice you would not carry out a one-and a two-tail test on the same set of results. The null
and alternative hypotheses are defined before the data are collected and the choice between a
one-and a two-tailed test depends on prior knowledge. For example, if we wished to see whether
a catalyst increased the rate of a reaction then you know before you start that we are looking for
an increase. If the actual results showed a decrease then there would be no need to carry out the
test.

108
The method given in this section is applicable even if the population from which the sample is
drawn is not Normal provided that the sample is large. This is a consequence of the central limit
theorem.

4.2 UNIT ACTIVITY

(1) A machine designed to produce rope with a breaking strain of 1000 N. The breaking
strain is known to be normally distributed with s.d. 21 N. A new material is introduced
into the rope which is hoped will increase the breaking strain. A random sample of nine
pieces of rope had a mean breaking strain of 1012 N. Is there evidence at the 5%
significance level that the mean breaking strain has increased?
(2) A machine filling bottle of orange squash is adjusted to deliver 0.725 litre with a standard
deviation of 0.010 litre. A sample of 50 bottles is checked and the mean quantity is found
to be 0.721. Determine whether this should be taken to be significantly different from
0.725 at the 5% level.
In a later check the manufacturer decides on a more stringent test, using the 2% level.
If he now takes a sample of 40 bottles and the mean turns out to be 0.722, is this mean
significantly differently from 0.725?
(3) A factory produces washers of mean thickness 3 mm with a standard deviation of 0.2
mm. Each day, as a routine check, the thickness of a random sample of 100 washers is
measured and the mean calculated and recorded. Every 30 days, these means are
assembled into a frequency table. What would you expect the mean and standard
deviation of this table to be?
On one day, the mean of the sample of 100 washers is 3.036 mm. Is this significantly
different from the expected value?
(4) Explain the difference between a one-tail and a two-tail significance test.
A test of mental ability has been constructed such that, for adults in Great Britain, the test
score is normally distributed with mean 100 and standard deviation 15. A doctor wishes
to test whether sufferers from a particular disease differ from the general population in

109
 their performance on this test. He chooses a random sample of ten from his patients.
Their scores on the test are :119, 131, 95, 107, 125, 90, 123, 89, 103, 103.
What would you conclude?
(5) The mean number of letters per word in a dictionary of the English language is 6.8 and
the standard deviation of the number of letters per word is 2.5.
Work out the mean number of letters per word for the words in question 3 and test
whether it appears to come from this dictionary population. Is this a valid and test of the
hypothesis that the question is worded is typical English?
(6) Bill has used a particular type of razor for shaving for a long time. The length of time (in
seconds) that he takes to shave is normally distributed with mean 240 and s.d. 20. He
changes to a new razor and finds that his shaving times on nine consecutive days are 210,
220, 230, 220, 250, 230, 260, 210, 240. Assuming that he s.d. of his shaving time has not
changed, test whether his mean shaving time has changed with the new razor.
(7) The mass of packs of unwrapped chocolates is normally distributed with mean 508 g and
s.d. 4 g. You am told that a pack of chocolates weighs 520 g but you are not told whether
the chocolates are unwrapped or not. Carry out a significance test to show that there is
evidence that the chocolates are wrapped.

4.5.3 Significance tests (  unknown): large samples

If  is unknown but the sample is large, then the test statistic z can still be used, with 
replaced by ŝ (as in Section 3.5.2). This gives
x x
z 
sˆ s
n n 1
Example 4.3
100 100
One hundred measurements of a variety gave  xi = 151,
i
x
i 1
2
i = 390. Could these

measurements have come from a population, mean 1.5?

110
  n 2  n 
2

  xi   xi  
 
Sample s.d., s   i 1   i 1 

n  n 
   
   
 
2
 390  1.51 
s   
100  100 
 1.27

H o :   1.5

H1 :   1.5
x 1.51  1.5
z   0.078
s 1.27
n 1 99
This value of z is not significant at the 5% level since the critical region is |z| > 1.96. H o is
retained: the Measurements could have come from a population mean 1.5.

4.3 UNIT ACTIVITY

(1) A random sample of the mean marrying in a town had the age distribution given in Table
4.2.
Table 4.2 Distribution of marrying age
Age Under 18 18- 25- 32-38 39-45 46- Over Total
24 31 52 52
Coded 0 1 2 3 4
value
frequency 0 101 62 25 10 2 0 200

By using the suggested coding, or otherwise, estimate the mean and variance of the
population from which the sample is drawn. (Quote your results to two decimal places).
The mean age of men at marriage in the large country in which the town is situated is 25.90
years. Test whether the men in the town differ significantly from the men in the whole

111
 country in the age at which they marry. (State your null hypothesis and the significance
level you use.)
(2) The times taken by a salesman to travel between two shops on 50 occasions averaged 64
minutes with a s.d. of 6.3 minutes. He claims that this provided evidence that the journey
takes longer since he changed to a smaller car, as before he averaged 60 minutes. Do you
agree with the salesman?
(3) Explain briefly what is meant by a significance level of  %.
A student was asked to mark, by eye, the centre of a line drawn on a sheet of paper.
She repeated this 50 times, using a new test sheet each time. The signed deviations from
the centre, d, were measured in millimeters and the quantities d = - 60.0, d 2
=

197.44 calculated. Is there evidence that the student does not locate the mean position of
her bisection marks on the centre of the line?
(4) A pilot wishes to check whether the average depth of a section of river mouth is still 6.1
fathoms as his chart suggests. He takes 30 soundings at random points and finds that the
mean depth is 5.8 fathoms with a s.d. of 1.1 fathoms. Test at the 5% significance level
whether the mean depth has changed.
(5) A machine which fills orange squash bottles should be set to deliver 725 ml. A sample of
50 bottles is checked and the mean quantity is found to be 721 ml and the sample s.d. is 13
ml. Does this differ significantly from 725 ml at the 1% level?
(6) A new surgical technique has been developed in an attempt to reduce the time that patients
have to spend in hospital after a particular operation. In the past the mean time spent in
hospital was 5.3 days. For the first 40 patients on whom the new technique was used the
mean time spent in hospital was 5.0 days and the sample s.d. was 0.4 days. Is there evidence
that the new technique has decreased the time spent in hospital?

4.5.4 Significance tests (  unknown): small samples

For small samples the statistic

112
 x x
t n 1  
sˆ s
n n 1
must be used. It should only be used if the population from which the sample is drawn is normal.
Example 4.4
A manufacturer claims that his light bulbs have an average lifetime of 1500 hours. A purchaser
decides to check this claim and finds that for six bulbs the lifetimes are 1472, 1486, 1401, 1350,
1610, 1590 hours. Does this evidence support the manufacturer’s claim?
You have to assume the lifetimes of the light of the light bulbs are normally distributed.
Then you have ;
Model: Normal
H o :   1500h

H1 :   1500h
Two-tail test
Using Table A4 the critical region is |t5| > 2.57 for a test at the 5% significance level.
The reader should check that for the observed lifetimes
x = 1484.8 h, s = 93.2 h
x
t n 1 
s
n 1
1484.8  1500
t5   0.36
93.2
5
This result is not significant at the 5% level. The null hypothesis is retained and the
manufacturer’s claim is vindicated.

4.4 UNIT ACTIVITY

(1) Packets of breakfast cereal claim that the minimum net weight of the contents is 450 g.
The weights of the contents of seven packets are: 445, 453, 447, 451, 440, 460, 449
Is there any evidence at the 5% significance level that the packets are underweight?

113
(2) The lives of six candles are found to be 8.1, 8.7, 9.2, 7.8, 8.4, 9.4 hours. Estimate the
population mean and show that the estimate of the population variance is 0.388.
The manufacturer claims that the average life is 9½ hours. Making a suitable assumption
concerning the nature of the distribution of the life of a candle, carry out a statistical test
of the manufacturer’s claim. Give full details of your test.
(3) A student titrates 10 ml of 0.1 M acid against 0.1 M alkali five times and obtains the
following results for the volume of alkali: 9.88, 10.18, 10.23, 10.39, 10.25 ml
Is there any evidence that these results show a bias from the expected value of 10 ml?
(4) Eight volunteers tested a food which the manufacturer claimed would help people to
slim. At the end of the test they had lost 1, 2, 0, 1, - 2, 0, 3, 3 kg respectively. Assuming
that these losses are a random sample from a Normal distribution, carry out a t-test to
determine whether the mean loss differs significantly from zero, and comment on the
manufacturer’s claim.

4.5.5 Difference between two means for large samples

Suppose we take large samples from two populations as indicated below:

Population 1 Population 2
s.d. =  1 s.d. =  2

Sample size = n1 Sample size = n 2

Sample mean = x1 Sample mean = x 2

Sample s.d. = s1 Sample s.d. = s 2

X 1 will be normally distributed (since n1 is large) with mean 1 and s.d.

1 and X 2 will be
n1

normally distributed with mean  2 and s.d. 

2
. The difference between the means,
n2

 1 2    22 
X 1  X 2 , will also be normally distributed with mean 1   2 and s.d.  
n1   n2 
 

114
(see Section 4.5.4). You frequently with to test whether two samples come from populations
with the same mean in which case 1   2  0 .
Example 4.5
A firm employs 300 women and 100 men. The mean number of days absent last year for the
women was 5.3 with a s.d. of 2.2 and for the men the corresponding figures were 6.2 and 2.9. Is
the difference between the means significant?
Model: Normal
H o : 1   2 Therefore 1   2  0

H1 : 1   2
Two-tail test
The test statistic is
( x1  x 2 )  ( 1   2 )
z
 12  22 
  
 n1 n2 

x1  x 2
z
 12  22 
  
 n1 n2 

Since from . 1   2  0

If  1 and  2 are not known, then, because the samples are large,  1 and  2 can be replaced

 n 
by ŝ and ŝ . Since sˆ    s this gives
 n 1
x1  x 2
z
 s 2
s 22 

1
 
 n1  1 n2  1
5.3  6.2
z
 2.2 2 2.9 2 
  
 299 99 

z = - 2.83

115
The critical region for z is z  1.96 so that z is significant at the 5% (two-tail) level and we

reject the null hypothesis. There is a significant difference between the mean absences of men
and women .The theory for small samples is beyond the scope of this book.

4.5 UNIT ACTIVITY

(1) It is found that over a certain period at one telephone exchange 200 subscribers taken at
random made a total of 13 248 calls. During the same time, a random sample of 300
subscribers at another exchange made a total of 20 922 calls. The standard deviation of
the number of calls made by a subscriber at either exchange in the period is 8.
Is there any evidence of a difference between the subscribers at the two exchanges in
their average frequency of calls?
Find 95% confidence limits for the mean number of calls made in the period per
subscriber at each exchange.
(2) Samples of leaves were collected from two oak trees A and B. The number of galls was
counted on each leaf and the mean and standard deviation of the number of galls per leaf
was calculated with the results given in Table 4.3
Table 4.3 Samples of leaves

Assuming Normal distributions, do the data provide evidence at the 5% significance level
of different population means for the two trees?
(3) In an investigation into the effectiveness of a particular course in speed reading a group
of 500 students was split into two groups, A and B, of sizes 300 and 200 respectively,
thought to have been chosen at random.

116
 Those in group A were given no special instruction; those in group B were given a course
in speed reading. Each student was asked to read the same passage and the time taken
was measured. The results were
Group A: mean time 78.4 s, variance 14 s2
Group B: mean time 77.4 s, variance 15 s2
Carry out a significance test to see if there is evidence that the course has improved
reading speed. State carefully your null hypothesis, alternative hypothesis and final
conclusion.
You learn later that, of the original 500 students, 200 students had decided for themselves
that they wanted to take the course in speed reading and that these students became group
B. Discuss briefly how this might affect your previous conclusion.

4.5.6 Testing if two samples come from the same population

In this case we wish to test whether x1  x2 differs significantly from zero as in the previous
example.
However, since our null hypothesis is now that the two samples are from the same population,

we have  1   2   (say) and the s.d. of x1  x2 is therefore   1  1  . If we know o

 n1 n2 

we can use the test statistic

x1  x 2
z (1)
1 1
   
 n1 n2 
More usually  is not known. We need to make an unbiased estimate s of it from both the
samples and use the test statistic t. it can be shown that, if s1 and s2 are the standard deviations of
the two samples, the most efficient unbiased estimate of the s.d of the population is given by

 n s 2  n2 s 22 
ˆs   1 1 
 n1  n2  2 

117
The denominator represents the number of degrees of freedom of this estimate so that the test
statistic
x1  x 2
t n1  n2  2  (2)
 n1 s12  n 2 n22  1 1
   
 n1  n2  2  n1 n2 
can be used provided that the population are Normal and/or the samples are large
Example 4.6
A market gardener decides to test a new pesticide, which the manufacturer claims increases the
yield, by applying it to one of his two orchards. The treated orchard contains twenty trees and the
mean and s.d of the yield per tree are 98 kg and 10 kg respectively. The untreated orchard contains
fifteen trees and the corresponding values are 94 kg and 8 kg.
Test whether these results are consistent with the yields being drawn from the same population.
Model: Normal
H o : 1   2 ,  1   2  

H1 : Samples come from different populations

x1 = 98 kg, x 2 = 94 kg
s1 = 10 kg, s2 = 8 kg
n1 = 20, n2 = 15
The unbiased estimate of variance using equation (2) is

x1  x 2  20 *102  15 * 8 2 
t n1  n2 2      9.47
 n1 s12  n2 n22  1 1  20  15  2 
   
 n1  n2  2  n1 n2 
Using equation (3),
98  84
t 33   1.24
9.47
1 1
  
 20 15 
This is not significant at the 5% level (two-tail) and H o is retained. The results are consistent with
the samples being drawn from the same population. Strictly speaking, before such a test is made,

118
a test should first be made on the variances of the samples to see whether a pooled estimate of
variance is justified (see Section 4.5.4). If there is a significant difference between the variances,
then the analysis described is no longer valid.

4.6 UNIT ACTIVITY

(1) In a butter-packing plant the quantity of butter packed in a day using a certain type of
machine is a normal variable with a standard deviation of 39 kg. Two packers A and B
average 1518 kg and 1499 kg per day respectively over a 26-day month. Is the
performance of the two operatives significantly different?
A third packer C averages 1480 kg per day for her first 26-day month. Is this a
significantly worse performance than B’s?
(For each test you should state clearly your null and alternative hypothesis.)
Table 4.4 Measurements of butter-packing plant the quantity

(2) The lengths of the femur in samples of Mus homunculus from two sources (Zambia and
North Africa) are given in Table 4.4.
The mean length of the femur is known to be characteristic of each breed of Mus
homunculus. Test whether the data are consistent with the assumption that Mus
homunculus in Zambia and N. Africa are of the same breed.
(3) In an experiment 22 mice were divided into two groups, one of which was given a special
diet. After an interval the gains in weight of the mice were measured, with the results
given in Table 4.5.
The variance of weight increases may be assumed to be the same for both populations of
which these groups are samples, those being fed a special diet and those being ordinarily

119
 fed. Calculate an estimate of this common variance from the combined results of both
groups, and discuss whether the difference in the observed weight increases is significant.
Table 4.5 Experiment of 22 mice

4.5.7 Paired tests

The paired sample t-test, sometimes called the dependent sample t-test, is a statistical procedure
used to determine whether the mean difference between two sets of observations is zero. In a
paired sample t-test, each subject or entity is measured twice, resulting in pairs of observations.
Example 4.7
Table 4.6 gives the times taken (in minutes) by eight typists to type the same number of words
using two different typewrites. Do the data indicate any difference in speeds for the two
typewriters?
Table 4.6 Times taken (in minutes) by eight typists

In this case it would be incorrect to calculate the means for typewriter 1 and typewriter 2 as in
Section 4.5.6 since the variations between the typists could swamp any difference due to the
machines. Instead we compute the difference, D, for each typist, since if there is no difference
between the machines then we would expect the mean difference, D, to be zero. Provided the
original populations are Normal, D will also be approximately normally distributed.

120
 Table 4.7 shows the calculation of the mean d, and the s.d., sd, of the sample. (Note that var(D)
= var(X) + var(Y) since the values of X and Y used to calculate each di are not independent. In
fact var(D) < var(X) + var(Y).)
3.6
d =  0.45 min.
8

 5.34  3.6  2 
sd     
 8  8  

= 0.68 mim
If  D is the mean of the population from which D is drawn you have:
Model Normal.
Ho : D = 0
H1 :  D  0
Two-tail test
Using t as the test statistic, since the s.d sd is estimated from the sample, you have
Table 4.7 Calculation for paired t-test

Od
t n1 
sd
n 1

121
  0.45
=
0.67
7
= - 1.75
This value of t is not significant at the 5% (two-tail) level and H o retained. The data do not
indicate a difference in speed for the two typewriters.

4.5.8 The sign test

This test is sometimes used instead of the paired t-test. Consider again the data in example 4.1 (see
Table 4.8). A plus sign in the last row indicate that typewriter 1 was faster and a minus sign that
typewriter 2 was faster. If we adopt the null hypothesis that there is no difference between the two
typewriters we would expect an equal number of plus and minus signs. In fact we have two minus
signs out of eight. The number of minus signs will be binomially distributed with, on our null
hypothesis, p = ½, n = 8. The test is two-tailed since we are only concerned with whether the
typewriters differ in speed.
Model: Binomial
Ho : p=½

H1 : p ½
Two-tail test
The probability of observing two or less minus signs is
Table 4.8 Times taken (in minutes) by eight typists
Typist A B C D E F G H
Time using 6.3 4.5 7.1 8.4 3.7 3.9 4.7 5.2
Typewriter 1
Time using 5.1 + 4.4 + 6.2 + 7.3 + 4.5 - 4.0 - 3.6 + 5.1 +
Typewriter 1

8 7 6 2
 1   8  1   1   8  1   1 
               0.145
 2   1  2   2   2  2   2 

122
Thus the probability of observing < 2 or > 6 minus signs is 0.29 and the observed number of minus
signs is not significant at the 5% level. H o is retained (as it was when the paired t-test was
performed). The sign test can be used whether or not the populations are Normal.
The sign test is an example of a non-parametric test. Non-parametric tests make no assumptions
about the distribution from which the sample is drawn. In particular they do not assume that the
population distribution is Normal as many of the tests described in this unit do.

4.7 UNIT ACTIVITY

(1) Ten marksmen shot at targets with two types of rifle. Their scores out of 100 were as in
Table 4.9. Apply the sign test to the hypothesis that he rifles are equally good. Also use
the t-test for paired values to test the same hypothesis. What assumption is made in
applying this test, but not in applying the sign test?
Table 4.9 Ten marksmen
Marksman A B D E F G H I J
Rifle 1 93 99 90 87 85 94 88 91 96 79
Rifle 2 89 93 86 92 78 90 91 87 92 86

(2) During negotiations between the union and management of a large firm two alternative
offers are put to the union side:
(i) old rates plus 20% increase across the board,
(ii) new rates based on a production bonus scheme.
The union statistician, Percy Glum, considers these offers and calculates the pay that 25
typical employees would receive from each offer. These are summarised in Table 4.10.
Table 4.10 negotiations between the union and management
Employee 1 2 3 4 5 6 7 8 9 10 11 12 13
Old 56 43 59 62 38 49 53 37 71 53 47 39 37
rates+20%(K)
New rate(K) 67 58 58 75 47 51 52 49 75 59 56 41 42

123
 Employee 14 15 16 17 18 19 20 21 22 23 24 25
Old 68 27 68 75 42 53 61 56 58 35 46 37
rates+20%(K)
New rate(K) 65 31 72 84 45 54 65 61 57 39 49 39

By use of the sign test, or otherwise, determine whether the new rates will lead, on average, to
an increase of more than the 20% on the old rates.

4.5.9 Significance of a proportion (large samples)

In this unit you return to the situation considered at the beginning of the unit where you have a
variable which is binomially distributed, the difference being that you will consider large
samples so that the Normal approximation can be used.
Example 4.8
On a national basis the success rate for people taking their driving test for the first time is 40%.
A driving instructor claims that his record is superior because, of the 50 pupils of his who took
the test for the first time last year, 25 passed. Is his claim justified?
You have
Model: Binomial approximated by Normal
H o : p  0.4

H o : p  0.4

One-tail test
On the null hypothesis the number of people, X, who pass first time in a sample of 50, is a
variable which is binomially distributed with
  np  50 * 0.4  20

  npg  (50 * 0.4 * 0.6)  12

Since n is large, X is approximately normally distributed with mean 20, s.d. 3.46. The
Number in the sample who pass first time is 25. Since the variance of population is known we
can use the test statistic z. As we have a single measurement of the variable X,

124
 x
z

25  20
z  1.44
12
This value of z is not significant at the 5% level (one-tail). H o is retained and the instructor’s

1
claim is not justified. Strictly speaking a continuity correction of  should be applied. In this
2
24.5  20
case this would lead to a value of z   1.30 . However, this correction is frequently
12
omitted.
The solution to this problem can also be given in terms of proportions since the sampling
distribution of the proportion in the sample, p s , is N [ p, p (1  p ) / n] . We have

ps  p
z
 p(1  p) 
 n 

0.5  0.4
z  1.44
 0.4 * 0.6 
 50 

which, of course, gives the same result as before. In this case the continuity correction, if
1
applied, is  n
2

4.8 UNIT ACTIVITY

(1) After a survey a market research company asserted that 75% of TV viewers watched a
certain programme. Another company interviewed 75 viewers and found that 51 watched
the programme and 24 did not. Does this provide evidence at the 5% level of significance
that the first company’s figure of 75% was incorrect?
(2) In a multiple choice examination paper, a candidate has to select which of four possible
answers to a question is the correct one. On a paper with 100 questions he gets 34 correct

125
 answers. Explain carefully, with supporting calculations, whether you regard this result as
contradicting the supposition that his answers are obtained entirely by guesswork.
(3) Explain what is meant by a standard error.
A certain method of scaling examination marks is supposed to fix the quartiles at 25 and 75.
Out of 1000 candidates, 541 have marks between these limits. Is this evidence that the
method has failed?
(4) You are engaged as an expert witness for the prosecution in a Court case in which a gaming
club is accused of running an unfair roulette wheel. The evidence is that out of 3700 trial
spins, zero (on which the club wins) turned up 140 times. There are 37 possible scores on a
trial spin, labelled 0 t0 36, and these should have equal probability. Test whether there is
evidence that the wheel is biased. Explain briefly what this test of significance means,
bearing in mind you have to convince a non-mathematical jury.
(5) If births are equally likely on any day of the week then the proportion of babies born at the
weekend should be 2/7. Out of a random sample of 100 children it was found that 23 were
born at the weekend. Does this provide evidence that the proportion of babies born at the
weekend differs from 2/7?

4.5.10 Difference between two proportions (large samples)

Example 4.9
In a random sample of 500 people from a certain town there are 270 men, of whom 160 are
smokers, and 230 women, of whom 110 are smokers. Is there evidence that the men of the town
are more likely to smoke than the women? If p1 is the probability that a man from the town
smokes and p2 is the corresponding probability for women, then
Ho : p1 = p 2 = p (say)

H1 : p1 > p 2
One-tail test
The numbers of men and women who smoke are each binomially distributed but since the
number in the sample is large, the Binomial distribution can be approximated by the Normal
distribution. Our best estimate of p is found by combining all the data:

126
 number who smoke 160  110
p̂ = =  0.54
number in sample 500

According to H o , the observed proportion of men who smoke, P1, has s.d. pq / n1 , where n1
is the number of men in the sample (see Section 4.5.10). Similarly for women, the observed

proportion, P2, has s.d. pq / n2 where n2 is the number of women in the sample. Thus P1 - P2

has s.d.  1 2 –given by

 1 1 
 1 2   pq  
  n1 n2 

Using our estimate of p, p̂ = 0.54, this gives for an estimate sˆ12 of  12

 1 1 
sˆ12 0.54 * 0.46   0.0447
 270 230
According to H o the mean value of P1 – P2 is 0. The observed value is

160 110
  0.114
270 230
Using the test statistic z since n is large
0.114  0
z =  2.55
0.0447
This value is significant at the 5% level (one-tail). H o is rejected and there is evidence that men
are more likely to smoke than women.

4.9 UNITY ACTIVITY

(1) A television rental organisation supplies its service mechanics with small vans for use
when visiting customers. In 1975, the London branch used vans of type A and 15 out of a
total of 60 spent at least one day off the road being repaired. The Edinburgh branch used
vans of type B in the same year and the corresponding figures were 20 out of a total of 40.
Is there a significant difference between the two proportions?
A manager of the organisation wishes to use these figures to compare the reliability of the
vans of the two types to help him decide whether the organisation should use one type of

127
 van rather than the other in both cities. Comment on (a) what may be inferred about relative
reliability from the significance test, (b) what extra information, if any, would be useful
when comparing reliability.
(2) Test whether the proportion of patients who die when they receive antitoxin treatment is
significantly different from the proportion who die when they receive ordinary treatment
for the data given in unit activities 3.54, question (4).
(3) The average number of flaws per 100 metre length of a yarn produced by a machine has
been found to be seven. A new machine is installed and the first 100 metre length has
three flaws. Does this provide evidence that the new machine is better than the old?
(4) In an intensive survey of Nguzu land it was found that the average number of plants of a
particular species was 26 per square metre. After a hard winter the number of these plants
found growing in a randomly chosen area of one square metre was 15. Is there evidence
that the hard winter has tended to kill off this species of plant?

4.5.11 Significance test using the Poisson distribution

The ideas developed in this unit can also be applied when the Poisson distribution is a suitable
model for the population from which the sample is drawn.
Example 4.10
Over a number of years the average number of breakdowns of an office photocopies has been six
per month. A new photocopier is installed and the number of breakdowns in the first month is
reduced to one. Is this evidence that the new photocopier is more reliable than the old one?
Encouraged by this result the office decides to keep the photocopier and the total number of
breakdowns for the first year is 50. Does this confirm the idea that the new photocopier is more
reliable?
Assuming that breakdowns are events which are randomly distributed in time, we have
Model: Poisson
Ho :  =6
H1 :  <6
p( X  1)  e 6  6 * e 6

128
  0.0174=1.74%< 5%
The result is significant at the 5% level and there is evidence that the new photocopier is more
reliable than the old. Again adopting the null hypothesis that the new machine has the same
reliability as the old one, the mean number of breakdowns per year will be Poisson distributed
with mean = 12 x 6 = 72. Since this is large, the number of breakdowns per year will be
approximately Normally distributed with mean and variance both equal to 72. We have

Model: Poisson approximated by Normal

Ho :  = 72
H1 :  < 72
You calculate the test statistic
50  72
z  2.59  1.64
72
This result is also significant at the 5% level and suggests that the new photocopier is more
reliable than the old one. (Strictly speaking a continuity correction should be applied).

4.5.12 Type I and Type II errors

When we test a null hypothesis by applying a significance test to a sample we can never be certain
that our conclusion is correct. All we know is that our decision is probably correct. For example
if we reject a null hypothesis it is because we obtained a value of the test statistic in the critical
region and this event is unlikely if the null hypothesis is true. you can make two types of error:
Type I: the null hypothesis is rejected when it is in fact true.
Type II:the null hypothesis is retained when it is in fact false.

129
Figure 4.4 Diagram showing the probabilities of making Type I and Type II errors for the
example in Section 4.5.12
The probability of making a Type I error is easily calculated since it is the probability that the
test Statistic lies in the critical region and is thus equal to the significance level of the test.
The probability of a Type II error cannot be calculated unless we specify a particular value of the
Parameter for the alternative hypothesis. Suppose we are sampling from a Normal distribution
with known standard deviation, o equal to 5 and wish to test H o :  = 50 against H 1 :  = 65
by taking a single measurement. The solid curve in figure 4 show the distribution of a single
measurement, X, if the null hypothesis is true and the broken curve shows the distribution if the
alternative hypothesis is true.
At the 5% level of significance the probability of making a Type I error is given by the solid black
area corresponding to |z| > 1.96. The upper critical value of X, i.e. the value of X above which H o
is rejected, is given by
x
z

1.96 = (x – 50)/5
x = 59.8
You can now calculate the probability of a type II error which is given by the hatched area. You
have
z = (59.8 – 65)/5
= 1.04
P(Z < -1.04) = P(Z > 1.04)

130
 = 1 - P(Z < - 1.04
= 1 – 0.8508
= 0.1492
= 15%
You can reduce the first error by using a lower significance level, e.g. 1%. This will move the
upper critical value of X to the right in figure 15.4 which unfortunately increases the probability
of a Type II error.
Similarly reducing the Type II error will increase the Type I error. If, however, we are prepared to
take a larger sample and use it’s mean to calculate the test statistic then both errors can be reduced
because the standard deviation of the sampling distributions is reduced. If, for example, we take

samples size four then the standard error of the mean is 5 = 2.5. If the upper critical value of
4
X is kept as 59.8 you have
P(Type I error) = P[Z > (59.8 – 50)/2.5]
= P(Z > 3.92)
= 0.005%
P(Type II error) = P[Z > (59.8 – 65)/2.5]
= (Z < 2.08)
= 1.9%
In an experiment the critical value and the sample size can be chosen to give acceptable values
for both types of error. An example is given in question (3) of the next exercise.
Example 4.11
A box is known to contain either H o ten white counters or ninety black counters or ( H 1 ) fifty

white counters and fifty black counters. In order to test hypothesis H o against hypothesis H 1 ,
four counters are drawn at random from the box without replacement. If all four counters are
black H o is accepted. Otherwise it is rejected. Find the size of the Type 1 and Type II errors for
this test.

131
Type I error
H o is rejected when it is in fact true. Ho is rejected if less than four black counters are obtained.

The Probability of this event is most easily calculated by finding the probability that all four
counters are black.
90 89 88 87
P(four black) =  * * *  0.652
100 99 98 97
Therefore the probability of a Type I error is 1 - 0.652 = 0.348.
Type II error
H o is retained when it is false and H 1 is true. If H 1 is true the probability of selecting four

50 49 48 47
black counters is  * * *  0.059
100 99 98 97
This is the probability of a Type II error.

If the probability of a Type II error occurring for a specific alternative hypothesis is  , then 1 –
 gives the probability that we reject a false null hypothesis, i.e. we make the correct decision.
This probability is known as the power of a test. In the previous example, with the specified
alternative hypothesis, the Power was 1 – 0.059 = 0.941.

4.10 UNIT ACTIVITY

(1) You are provided with a coin which may be biased. In order to test this you are allowed
to toss it twelve times and count the number, r, of heads to decide. If the coin is really
fair you wish to have at least a 95% chance of saying so. For what value of r should you
decide the coin is fair?
If you adopt you procedure with a coin which is actually biased two to one in favour of
heads, what is the probability that you decide the coin is biased?
(2) A sample of twenty items is taken from what is believed to be a Normal population with
mean 24 and standard deviation 5. Obtain the range of values within which the mean of
the sample must lie in order that the sample may be accepted as coming from the

132
 population if the significance level of the test is 5%. What is the probability that we shall
reject the conclusion that the sample comes from this population when it in fact does so?
Assuming that he standard deviation is correct but the mean of the population is in fact
25, what is the probability that we shall accept the original hypothesis when it is untrue in
this particular way?
(3) (a) outcomes: success and failure. To test the null hypothesis that the probability, p, of a
success is 0.25, an experiment consisting of ten independent trials is performed. If six
or more successes are observed the null hypothesis is rejected. Find the significance
level of the test. If p is actually 0.5, find the power of the test.
(b) The life of a particular type of projector bulb has a mean value of 50 hours and standard
A machine produces nails whose lengths are normally distributed with mean u cm and
standard deviation 0.1 cm. When the machine is working correctly u = 3.0, but
occasionally the machine goes wrong, in which case u = 3.05, the standard deviation
remaining 0.1 cm. In order to decide whether the machine is working correctly, the
lengths of the nails in a sample of n nails are measured, and the sample mean x is found.
If the value of x exceeds a predetermined value v then it is concluded that the machine
has gone wrong; otherwise the machine is presumed to be working correctly. It is
required that there should be probability of no more than 5% of presuming that the
machine has gone wrong when in fact it is working correctly, and that there should also
be a probability of no more than 10% of presuming the machine to be working correctly
when it has gone wrong. Determine appropriate values of n and v, if n is to be as small
as possible.
(4) State what you understand by the critical region and the power of a test.
(a) A trial has two possible deviation of 10 hours. It is thought that a different design
of bulb will have a longer life, but it is expected that the standard deviation will not
change. A random sample of 25 bulbs of the new type is obtained. The sample mean
x, which may be assumed to be normally distributed, is to be used to test for an
increase in mean life. If the significance level of the test is 1%, find the critical
region for the sample mean.
(b) Find the power of the test when the mean life for the improved bulb is (i) 56 hours,

133
 (ii) 52 hours.
5. (a) In an industrial manufacturing process, when production is under control, one unit in fifty
is defective. A sample of 100 units is taken at random every half-hour. Sketch a control
chart for the number of defectives found in a sample, showing the 1 in 40 and 1 in 1000
lines.
(b) Something goes wrong with the process, so that one unit in twenty is defective.
Calculate the probability that the number in random samples of 100 will go outside (i)
the 1 in 40 line, (ii) the 1 in 1000 line.
(c) What is the probability that the process will continue out of control for two hours
without
any sample exceeding the 1 in 1000 limit?
(6) The mean of a random sample of n observations, x1, x2, . . . , xn from a Normal distribution
is x. It is proposed to test the null hypothesis that the population mean is uo, against some
alternative hypothesis. State how your decision as to whether to use the t-distribution for
such a test would be influenced by (i) the value of n, (ii) whether the population standard
deviation is known to you.
You are given that x = 12.9, n = 10,  i2 = 1683.
(a) Test the null hypothesis that the population mean is 12 against the alternative hypothesis
that the mean is not 12.
(b) Determine a two-sided, symmetric, 95% confidence interval for the mean and explain
carefully what this confidence interval means.
(7) Over a period of years the average number of road deaths during the New Year Festival
period in Ndola has been 27. This year there were 35, a rise of approximately 30% and the
country’s press was full of gloomy comment on declining driving standards. Make suitable
calculations to demonstrate that the probability of there being at least the observed number of
road deaths is about 7½%.
Write an account of your argument and conclusions as if addressing an intelligent reader but
one having no knowledge of statistics or probability theory.
(8) What are the conditions under which a variable might be distributed according to the Binomial
distribution?

134
 An investigator suspects that there might be an unnecessary rounding off of weights recorded
by a spring balance. In a spot check, he finds that, out of ten recorded weights, six end in the
digits 0 or 5. How strong is the evidence that his suspicions are justified?
(9) Explain the meaning of the term “confidence interval’.
The expressions  xi2 (x1 – x)2/n and  xi2 (x1 – x)2/(n – 1) are both used in connection

with variance for a set of observations x1, x2, . . ., xn. Explain the circumstances in which each
is used and hence distinguish between them.
For such a set of 81 independent observations from a normal distribution,  xi = - 36 and

 xi2 = 736. Construct a 95% confidence interval for the mean of their probability

distribution, and use it to test whether this mean is likely to be zero.

Explain, either by carrying out the appropriate calculations, or otherwise, how you would
conduct this test without using a confidence interval.
(10) Explain how student’s t-distribution may be used to test the hypothesis that a random sample
of n observations is derived from a population whose mean is  o assuming the population to
be Normal.

Table 4.11 Test in geography before and after a series of tape-slide

========================================================
Pupil A B C D E F G H I J
Mark before
tape-sliders 3 15 14 18 10 3 5 8 9 11
Mark after
tape-slides 4 18 14 16 12 6 5 9 10 15
=========================================================
Ten pupils took a test in geography before and after a series of tape-slide sequences on that
subject. Their marks (out of 20) were in Table 4.11.
Test, at the 5% level of significance, the hypothesis that the pupils’ performance is not
affected by the series of tap-slide sequences.
State the assumptions made in applying the test.

135
(11) Explain what is meant by a 95% confidence interval for a population mean.
A factory manufacturing ammeters tests them for zero errors in their calibration.
From past routine tests, it is known that the standard deviation of these errors is 0.3.
A batch of nine ammeters, taken from one worker’s production, has zero errors of 1.0, -0.1, -
0.3, 1.6, 0.5, 0.4, 0.5, 0.2, - 0.2. Test whether there is evidence of a bias in the ammeters
produced by this worker and establish a 95% confidence interval for the mean zero of his
ammeters.
(12) Two alternative hypotheses concerning the probability density function of a random variable
are
 2 x, 0  x  1
H o : f ( x) 
0 elsewhere
2(1  x) x, 0  x  1
H 1 : f ( x) 
 0 elsewhere
Give a sketch of the probability density function for each case.
The following test procedure is decided upon. A single observation of X is made and if X
exceeds a particular value c, where 0 < c < 1, then H o is accepted, otherwise H1 is accepted.

Find the value of c if the probability of accepting H 1 given that H o is true is ⅟9. With this

value of c, find the probability of accepting H o given that H 1 is true.

(13) A blindfold subject was given a sample of butter and one of margarine and asked to state
which was butter. This was repeated to give six tests in all, the samples being presented in a
random order on each occasion. The subject correctly identified the butter five times out of
six. Find the probability of five or six correct identifications by a subject who is not able to
distinguish between butter and margarine. State, with reasons, whether the subject’s
performance would lead you to believe that she could identify the butter.
The experiment was carried out simultaneously with a total of 24 subjects. The total number
of correct identifications was 83 out of a total of 144 trials. Test if the group as a whole
could identify the butter significantly better than might be expected.
(14) Explain what is meant by the sampling distribution of the mean. What shape would you expect
this to have for large samples?

136
 A machine is designed to produce rods 2 cm long with a standard deviation of 0.02 cm.
The lengths may be taken as normally distributed. The machine is moved to a new position in
the factory, and, in order to check whether the setting for the mean length has altered, the
lengths of the first nine rods produced are measure. The standard deviation may be considered
to be unchanged. If these lengths, in cm, are as given below, test whether the setting has been
altered or not.
2.04, 1.97, 1.99, 2.03, 2.04, 2.10, 2.01, 1.98, 2.07
(15) An engineer wishes to compare the results obtained by two methods of measuring the breaking
strains of a certain type of wire rope. Describe carefully the design of an experiment for this
purpose which would require a two-sample t-test for its analysis.
In such an experiment the measurements of breaking strain (x) obtained from six observations
using the first method gave x = 492, x 2 = 43 850, whilst eight observations using the
second method gave x = 608, x2 = 50 766. Test if the measurements came from the same
population.
(16) In the course of a survey concerning the proportion of lest-handed children the figure in
Table 4.12 were obtained from two schools.
Show that an approximate 95% confidence interval for the population proportion, p, of left-
handed children derived from the data from school 1 is 0.25 < p < 0.32, and calculate a
corresponding interval for school 2.

Table 4.12 proportion of lest-handed children

==================================
Number of Proportion
Children left-handed
----------------------------------------------------------
School 1 620 0.284

137
 School 2 475 0.341
=================================
Explain briefly what is wrong with the following argument: ‘Since these two confidence
intervals overlap, we cannot reject, at the 5% significant level, the hypothesis that the
populations from which the children in the two schools are samples each have the same
proportion of left-handed children.’
Calculate the overall proportion of left-handed children in both schools, and show that the
observed difference in proportions is significant at the 5% level.
(17) Experimental data concerning a variable X, which measures the reliability of a certain
electronic component, are as follows: x = 1164.2, x2 = 13911.6, n = 100.
Calculate the sample mean and standard deviation from these figures. Explain whether, on the
evidence of this sample, you would reject the hypothesis that the mean value of X is 12.
Figures collected over a long period have established that the mean and standard deviation of
X are 12 and 2 respectively. After a change in the manufacturing process it is expected that
the mean will have been increased, but it may be assumed that the standard deviation remains
equal to 2. A sample of n values of X is taken, with sample mean m: if m is greater than some
critical value it will be accepted that the mean has in fact increased, but it m is less than the
critical value the increase is not established.
State carefully appropriate null and alternative hypotheses for the situation, and find, in terms
of n, the critical value for a 1% significance level.
(18) (a) In one county in England, a random sample of 225 twelve year old boys and 250
twelve year old girls was given an arithmetic test. The average mark for the boys was
57 with a standard deviation of 12, whilst the average for the girls was 60 with a
standard deviation of 15. Assuming that the distributions are Normal, does this provide
evidence at the 2% level that twelve year old girls are superior to twelve year old boys
at arithmetic?
(b) An IQ test which had been standardised giving a mean of 100 and a standard deviation
of 12 was given to a random sample of 50 children in one area. The average mark
obtained was 105.

138
 Does this provide evidence, at the 5% level, that children from this area are generally
more intelligent?
(19) Describe what is meant by a random sample of observations of a random variable. A battery
manufacturer claims that his batteries have a mean life of 8 hours when used in a particular
model of calculating machine. Describe how you would use a significance test to examine
this claim on the basis of a large random sample of observed lifetimes. You should explain
clearly the hypothesis you would consider, the choice of significance level, and the details
of the test. State what is meant by concluding that the null hypothesis is rejected at the 5%
significance level.
A random sample of 121 batteries has a mean life of 7.56 hours and, for this sample, s 2 =
5.30 hours2. Test whether these data would lead to rejection of the manufacturer’s claim, at
the 5% significant level.
(20) A chemical is delivered in batches to a factory for use in a production process. It is
important that the percentage of manganese in the chemical should not decrease
significantly in successive batches. The first batch delivered is to act as a control. Ten
determinations of percentage manganese are made, with results as follow: Control batch
(% manganese) 3.3, 3.7, 3.5, 4.1, 3.4, 3.5, 4.0, 3.2, 3.7.Show that the control mean is 3.62,
and estimate the standard deviation. The results of the ten determinations for a later batch
are: Batch X (% manganese) 3.2, 3.6, 3.1, 3.4, 3.0, 3.4, 2.8, 3.3, 3.6. Determine whether
the mean of this batch is significantly smaller than 3.62. Use a 5% level of significance,
and justify any assumptions you make.

4.0 UNIT SUMMARY

Test of Test statistic Value if

139
 x
Mean  known z  Normal population and/or large sample

n

x x
Mean  unknown z  large sample
sˆ s
n n 1

x x
Mean  unknown t n 1   Normal population
sˆ s
n n 1
x1  x 2
Equal population means, z  ,normal population and/or large sample  1 ,  2
 12  22 
  
 n1 n2 

known
x1  x 2
Equal population means,  1 ,  2 un known z  large samples
 s12 s 22 
  
 n1  1 n2  1
x1  x 2
Same population,  unknown z  normal population and/or large samples
1 1
   
 n1 n2 

x1  x 2
Same population,  unknown z  large samples
1 1
sˆ   
 n1 n2 

x1  x 2
Same population,  unknown t n1  n2  2  normal population
1 1
sˆ   
 n1 n2 

n1 s12  n2 s 22
Where sˆ 2 
n1  n2  2

140
UNIT 5 THE  2 –TEST. GOODNESS OF FIT

5.1 Unit Introduction

Welcome to Unit 5 in which you will learn Chi-Square goodness of fit test which is a non-
parametric test that is used to find out how the observed value of a given phenomenon is
significantly different from the expected value. In Chi-Square goodness of fit test, sample data is
divided into intervals. Chi-Square Test is used in Market Research when:

141
  They need to estimate how closely an observed distribution matches an expected
distribution. This is referred to as a “goodness-of-fit” test.
 They need to estimate whether two random variables are independent.

The goodness of fit test is a statistical hypothesis test to see how well sample data fit a
distribution from a population with a normal distribution. In other words, it tells you if your
sample data represents the data you would expect to find in the actual population.

5.2 Unit Aim

The aim of Chi square test for testing goodness of fit is to decide whether there is any difference
between the observed (experimental) value and the expected (theoretical) value.

5.3 Unit Objectives

By the end of the unit you should be able to:

 apply when you have one categorical variable from a single population

 Determine whether sample data are consistent with a hypothesized distribution.

 compare the observed sample distribution with the expected probability distribution

 Determine how well theoretical distribution (such as normal, binomial, or Poisson) fits the
empirical distribution.

Terminology

O i : Observed values

E i : Expected values

X 2 : Chi square distributions

5.4 Unit Time required

142
You need 20 hours for this unit

5.5 Unit Topics

5.5.1 An experiment

Table 5.1 shows the results of an experiment in which four coins were thrown 160 times. The
results are divided into classes according to the number of heads obtained each time. For
comparison the theoretical frequencies predicted by the Binomial distribution have been
calculated, assuming that the coins are unbiased and using the parameters p = ½, n = 4. These
frequencies are usually called the expected frequencies.
As we would expect there is not exact agreement between the observed and expected frequency
for each class, since the frequency for each class is a random variable which varies from sample
to sample. Our problem is to test whether the observed frequencies differ significantly from
those calculated using the null hypothesis that the distribution is Binomial with p = ½ , n = 4. Let
O i denote the observed and Ei the expected frequency in the ith class. As a first step we might

calculate the discrepancy between the two values and sum over the classes, as shown in the third
column of Table 5.2 This, of course, produces the result zero. We can avoid this by squaring the
discrepancies, as shown in the fourth column of Table 5.2. To add the values in this column as it
stands would mean that the difference between 10 and 15 would be given the same weight as the
difference between 35 and 40 although the percentage difference in the first case in much
greater. It can be shown (using mathematics beyond the scope of this book) that the appropriate
statistic to use is
Table 5.1 Frequency distribution of the number of heads when four coins are tossed
===============================================================
Number of heads Observed frequency Expected frequency
------------------------------------------------------------------------------------------------------------
0 15 160 x ( 4 )(½)4 = 10
(0)
1 46 160 x ( 4 )(½)3 (½) = 40

143
 (1)
2 54 160 x ( 4 )(½)2 (½)2 = 60
(2)
3 35 160 x ( 4 )(½)(½)3 = 40
(3)
4 10 160 x ( 4 )(½)4 = 10
(4)
----- ----
160 160
====================================================
Table 5.2 Calculation of X2 for data in Table 5.1
====================================================
Oi Ei Oi – E i ( O i – E i )2 ( O i – E i )2/ E i
---------------------------------------------------------------------------------------
15 10 +5 25 2.5
46 40 +6 36 0.9
54 60 -6 36 0.6
35 40 -5 25 0.625
10 10 0 0 0
0 4.625
 (Oi  Ei ) 2 
  E  , whose calculation is shown in the last column of Table 5.2. You shall call this
 i 
statistics X 2 . The square emphasizes that it is always a positive quantity.

5.5.2 Degrees of freedom

If the experiment described in Section 5.1 was repeated with the same coins, different values of
O i and consequently of O i would have been obtained. In other words, X 2 has a sampling

distribution. This sampling distribution is approximately the same as a theoretical distribution

144
known as the X 2 (or chi-squared) distribution (‘chi’ is pronounced as the ‘ki’ in ‘kite’). ( X 2 is
often used to denote X 2 as well as the sampling distribution of X 2 .) It can be shown that the
form of the X 2 -distribution depends on the ‘degrees of freedom’ (and consequently it is often
written X 2 ). You have already met the term ‘degrees of freedom’ in Section 2.5.2. In the present
instance v is the number of expected frequencies which can be varied independently. In the
example in Table 5.2 there are five expected frequencies but, since the total frequency is fixed at
160, which four frequencies are given, the fifth is known, and so we have v = 4. The fact that the
total frequency must equal 160 is known as a constraint. In general, v is given by degrees of
freedom = number of classes – number of constraints
Figure 5.1 shows the form of the X2 –distribution for some different values of v.

Figure 5.1 The x2 –distribution for various values of v

5.5.3 Significance testing using the X 2 –distribution

As the discrepancy between the observed and expected frequencies increases so does the value
of X 2 . To decide if a value of calculated using a given null hypothesis is significant we use
Table A6. This gives the percentage points of the X 2 – distribution for different values of v. it is
a one-tall table and gives the probability that X2v exceeds a certain value. This probability is
given by the shaded area in Figure 2. For example, the probability is 5% that X⅔ > 7.81, 99.5%
that 142 > 4.07, etc. Using this table we are in a position to calculate if the value of X 2 = 4.625

calculated in Table 2 is significant. The critical region is  42 > 9.49 at the 5% significance level,
so the value of X2 obtained is not significant and we retain the null hypothesis that the

145
distribution of throws is Binomial with p = ½ n = 4. Although in a X 2 –test we normally test if
X2 exceeds the critical value, very low values of X 2 should regarded with suspicion since they
also have a low probability. For example there is only a 5% probability that  42 < 0.711. Such a

low value of X 2 is almost too good to be true and may indicate that the sample is not random, or
the data are fictitious.

Figure 5.2 Schematic diagram showing the critical region for a X 2 –test
Conditions in which the x2 –test is applicable
The X2 – test should only be used when the sampling distribution of X2 approximates closely to
a X2 – distribution. The conditions for this are
(i) the total frequency is not less than 50,
(ii) the expected class frequencies are not less than 5.
If the second condition is not met, it can be complied with by combining one or more classes.
The agreement between the X2 and X2 distributions can be improved when v = 1 by making an
adjustment known as Yates’ correction, which involves reducing each value of |Oi – Ei| by ½.
Example 5.1
On a national basis the success rate for people taking their driving test for the first time is 40%.
A driving instructor’s records show that for the 50 pupils of his who took the test for the first
time last year, 25 passed. Does his success rate differ from the national success rate?
Table 5.3 Driving test results

146
You take the null hypothesis that the instructor’s results do not differ from the national success
rates, that is
H o : p = 0.4

H 1 : p = 0.4
This gives Table 5.3 for the observed and expected frequencies. Notice that to use all the
information we include the frequencies for pass and fail. In this case v = 1, since we have two
classes and one restriction (that the total frequency is 50). Since v = 1, Yates’ correction has been
applied. The critical region at the 5% level as 12 > 3.84. The value of X 2 obtained is not
significant at the 5% level and there is no evidence that the driving instructor’s success rate differs
from the national rate.
A problem similar to this was solved in Section 5.5.2 using z but the question asked was ‘Is the
instructor’s success rate better than the national one?’ so that a one-tail test was used. The X 2 –
test is basically a two-test since X 2 is always positive and takes no account of the sign of the
difference between the observed and expected frequencies. For a one-tail test the z-method is
preferable.

5.1 UNIT ACTIVITY

(1) The following table purports to be a sample of 100 random digits. Test whether the
frequencies of the digits significantly from expectation.
11584 42689 08394 57019 33922 22413 21138 83541 53216 74935
51186 49197 30157 28543 51328 49788 31489 18971 28719 97121

147
(2) Two fair dice are thrown 432 times. Find the expected frequencies of scores of 2, 3, 4, . . .,
12 . Two players A and B are each given two dice and told to throw them 432 times, recording
the results.
The frequencies reported are given in Table 4.
Table 5.4 Two fair dice are thrown 432 times

Is there any evidence that either pair of dice is biased? What can be said about B’s alleged
results?

5.5.4 Test a distribution for normality

In section 9.5 a normal distribution was fitted to data. You are now in a position to test how good
the fit is. Although the Normal distribution is continuous, the data were divided into classes. We
can use the observed and expected frequencies in each class to calculate X2 and test the null
hypothesis that the observed values come from a Normal distribution with a specified mean and
s.d. Table 5.5 gives the observed and expected frequencies taken from Table 5.4.
The first four and the last three classes have been combined so that all Ei are > 5. The table
shows the calculation of X 2 . In this example there are six classes and three constraints: the total
frequency is 50, the mean (698.3) and the s.d. (18.6) are calculated from the data. So we have v
= 6 –3 = 3
The critical region is 32 > 7.81. Thus the value of X2 is not significant and the null hypothesis is
retained. If the values for the mean and s.d. had been given rather than calculated from the data,
then there would have been only one constraint.
Table 5.5 Calculation of X2 for data in Table 5.4

148
5.5.5 Testing the fit of a Binomial distribution

The example given in Section 5.5.1 was one in which a distribution was tested to see if it was
Binomial with a given value of p. In a Binomial distribution was fitted to data and p was not given
but calculated from the observed frequencies. The observed and expected frequencies obtained
are reproduced in Table 6. You wish to test the null hypothesis that the observed frequencies are
binomially distributed with p = 0.49, n = 4.

Table 5.6 Calculation of X2 for data in Table 5.3

149
 In this case, besides the constraint that the total frequency is 100, there is the added constraint
that p was calculated from the data so that v = 5 – 2 = 3
The critical region is X2 > 7.81. Thus X2 is not significant at the 5% level , the null hypothesis
is retained and the observed distribution conforms to a Binomial distribution the p = 0.49, n = 4.

5.5.6 Testing the fit of a Poisson distribution

In a Poisson distribution was fitted to data: the observed and expected frequencies are reproduced
in Table 5.7. We wish to test the null hypothesis that the data are Poisson distributed with a mean
equal to the mean of the data, which was found to be 0.37.
The calculation of X2 is shown in Table 5.7. There are two constraints: first, that the total frequency
is 100 and, second, that the mean is calculated from the data. There are three classes since the last
three have been combined, giving. v = 3 – 2 = 1 since v = 1, Yates’ correction has been applied.
The critical region is x2 > 3.84. Thus X2 is not significant at the 5% level and we retain the null
hypothesis that the data are Poison distributed with a mean of 0.37. If a value for the mean and
been given rather than calculated from the data, then there would have been only one constraint.

150
Table 5.7 Calculation of X2 for the data in Table 5.2
=====================================================================
Numbers of girls O i Ei Oi – E i | O i – E i |- ½ {| O i – E E i |- ½ }2|Ei
--------------------------------------------------------------------------------------------------------------------
0 71 69.1 + 1.9 1.4 0.028
1 23 25.6 - 2.6 2.1 0.172
2 4 4.7 } 5.3
3 2 0.6 } 5.3 + 0.7 0.2 0.008
>4 0 0.0 } 5.3
_________
0 X2 = 0.208
====================================================================

5.2 UNIT ACTIVITY

(1) Two dice were thrown 216 times, and the number of 6s at each throw were counted.
The results were as in Table 5.8.
Test the hypothesis that the distribution is Binomial with the parameter p = ⅙.
Explain how the test would be modified if the hypothesis to be tested is that the
distribution is Binomial with the parameter p unknown. (Do not carry out the test.)
Table 5.8 Two dice were thrown 216 times
==================================================
Number of 6s 0 1 2 Total
Frequency 130 76 10 216
==================================================
(2) A man kept count of the number of letter he received each day over a period of 78 days
(excluding Sundays). The frequencies of the number of letters per day and of the Poisson
distribution with the same mean (0.88) and total frequency are given in Table 5.9
Do the observations show a significant departure from a Poisson distribution?

151
 Table 5.9 Number of letter he received each day over a period of 78 days
====================================================
Letter Poisson
Per day Number of days frequencies
------------------------------------------------------------------------------------
0 33 32.4
1 26 28.4
2 14 12.6
3 or more 5 4.6
====================================================
(3) For a period of six months 100 similar hamsters were given a new type of feedstuff.
The gains in mass are recorded in Table 5.10
Table 5.10 Types of feedstuff
=================================================
Gain in mass (g) Observed frequency
x f
-------------------------------------------------------------------------------
- oo < x < - 10 3
- 10 < x < - 5 6
- 5<x<0 9
0<x<5 15
5 < x < 10 24
10 < x < 15 16
15 < x < 20 14
20 < x < 25 8
25 < x < 30 3
30 < x <  2
==================================================
It is though that these data follow a Normal distribution, with mean 10 and variance 100.
Use the X2 –distribution at the 5% level of significance to test this hypothesis.

152
 Describe briefly how you would modify this test if the mean and variance were unknown.
Table 2.11 Types of feedstuff
====================================================
x 0 1 2 3 4 5 6 7 8 or more
f 3 7 12 10 8 5 3 2 0
===================================================
(4) A group of students are required to carry out an experiment in which the results are
expected to have a Poisson distribution. There results for 50 replications of the
experiment are as in Table 2.11.
Estimating the mean of the Poisson distribution by the mean of these observations, obtain
the appropriate estimated frequencies and test how well these data fit the expected form.
State what you might suspect about the students’ results.

5.5.7 Contingency tables

Sometimes we wish to test if two attributes of a population are associated, i.e. if they occur
together. A common example is whether inoculation is associated with the prevention of disease.
In this case one
Attribute is whether or not a person is inoculated and the other attribute is whether or not they are
attacked by the disease. Table 12 gives the (fictitious) results for a random sample of 100, where
the total frequency has been divided into rows and columns according to attribute. This is known
as a contingency table. On the null hypothesis that inoculation has no effect in preventing the
disease, the marginal totals can be used to calculate the expected frequencies as follows. First we
have
70 35
P(inoculated) = P (attacked) =
100 100
30 65
P (not inoculated) = P (not attacked) =
100 100
Since according to our null hypothesis the events attacked and not attacked are independent of
the events inoculated and not inoculated we have, using the product law.

153
 70 35
P (inoculated and attacked)) = *
100 100
Table 5.12 Contingency table for inoculation and disease

Table 5.13 Calculation of expected frequencies for Table 5.12

Table 5.14 Calculation of X2 for data in Table 5.12

70 35
P(inoculated and not attacked) = *
100 100
30 65
P (not inoculated and attacked) = *
100 100

154
 30 65
P (not inoculated and not attacked)= *
100 100
The expected frequencies are found by multiplying these probabilities by 100, and are shown in
Table 13. Notice that the row and column totals are the same as before. It can be seen that the
expected frequencies could be calculated more directly using the rule
expected frequency = row total x column total
grand total
You are now in a position to calculate X 2 . How many degrees of freedom will X 2 have? The
answer is v = 1 since in the calculation of the expected frequencies, when one frequency was
found, the other frequencies could be found by subtraction from the marginal totals. Table 14
shows the calculations of X 2 using Yates’ correction. This value of X 2 is not significant at the
5% level. The null hypothesis is retained and there is no evidence that the inoculation is effective
against the disease.

5.5.8 Larger contingency tables

Table 5.15 shows a random sample of houses classified by region and type. Does the type of
housing vary between regions? Taking as our null hypothesis that the type of housing and the
region are independent, the expected frequencies are calculated in Table 5.16.

Table 5.15 Contingency table for housing and region.

155
Table 5.16 Expected frequencies for data in Table 5.15

Table 5.17 Calculation of X2 for data in Table 5.15

156
There are four degrees of freedom, since, when the frequencies marked with an asterisk have
been calculated, the other expected frequencies can be found by subtracting from the marginal
totals. The calculation of X 2 is given in Table 5.17. The critical region is  42 > 9.49 so the value
of X2 is significant at the 5% level: in fact it is significant at the 0.1% level. The null hypothesis
is rejected and the result indicates that there is a strong association between type of housing and
region.

Figure 5.3. Diagram to illustrate the degree of freedom of an h x k contingency table

A table with h rows and k columns is called an h x k contingency table. Its degrees of freedom are
(k – 1)(h – 1), since, as indicated in Figure 3, the frequencies in the last row and column can be
calculated by subtraction of the other frequencies from the marginal totals.

157
5.3 UNIT ACTIVITY

(1) Explain why the X2 statistic calculated from a 2 x 2 contingency table only has one
degree of freedom. Out of 100 rats fed with diet A, 65 showed signs of vitamin
deficiency, while out of 100 rats fed with diet B, only 53 suffered from vitamin
deficiency. Do the proportions of vitamin-deficient rats on the two diets differ
significantly?
(2) Table 5.18 summarises the incidence of cerebral tumours in 141 neurosurgical patients.
the expected frequencies on the hypothesis that there is no association between the type
and site of a tumour. Use the X2-distribution to test this hypothesis.
Table 5.18 Summarises the incidence of cerebral tumours in 141 neurosurgical
patients
=========================================================
Type of tumour
--------------------------------------------------------------------------------------------------
Site of tumour Benign Malignant Others
-------------------------------------------------------------------------------------------------
Frontal lobes 23 9 6
Temporal lobes 21 4 3
Elsewhere 34 24 17
==========================================================

Table 5.19 Oral tests are conducted by three examiners separately

=================================================
A B C Total

158
 --------------------------------------------------------------------------------
Credit 10 5 13 28
Pass 31 38 28 97
Fail 29 20 26 75
---------------------------------------------------------------------------------
70 63 67 200
==================================================
(3) Oral tests are conducted by three examiners separately. The numbers of candidates in the
categories credit, pass, fail are as shown in Table 5.19.
Use a X2 –test to examine the hypothesis that the examiners do not differ in their
standards of awards. State the assumptions made.

5.5.9 Relationship between X2 and Normal distribution

If X1, X2, . . ., Xn are independent N (0, 1) distributions, then 12 is said to be a X 2 random

variable with n degrees of freedom. In Particular if X is N(0, 1) then X 2 is a X 2 random variable

with one degree of freedom. In Section 5.5.2 we stated that the statistic has a sampling distribution
which is X2. We will now prove that this is true for the particular case in which there are two
classes. Let the probability of being in the first class be p and the probability of being in the second
class be q where p + b = 1. If the observed frequencies are ni and n2, the total frequency is n = n1
+ n2. The expected frequencies will be np and n(1 – p). This information is summarized in Table
20.
Then
 (Oi  Ei ) 2 
 2   
 Ei 

159
 (n1  np) 2 (n2  np) 2
 
np np
(n1  np) 2 (n2  (n  np)) 2
 
np np
(n1  np) 2 (np  n1 ) 2
= 
np np
(np  n1 ) 2
 ( p  q)
npq
(np  n1 )
 , sin ce p  q  1
npq
Table 5.20 Observed and expected frequencies for two classes
=======================================
Observed frequency Expected frequency
------------------------------------------------------------------
Class 1 n1 np

Class 2 n2 nq
n1 is Binomially distributed since we have n trials and a probability p that the result of a trial is in
class 1. If n is large then the distribution of n1 can be approximated by a Normal distribution with
np  n1
mean np and s.d. npq Thus is a standardized Normal deviate z and X2 = z2. From the
npq

definition given at the beginning of this section 12  z 2 . Thus in this case X 2  12 . This is
consistent, with the rule for calculating the degrees of freedom since in this case there are two
classes and one constraint : n1 + n2 = n.

160
5.4 UNIT ACTIVITY

(1) In a large town, the number of road accidents reported daily over 300 working days gave
the results in Table 5.21
Table5.21 Number of accidents reported
==========================================================
Number of accidents
reported in a day (x) 0 1 2 3 4 5 6 7 8
Number of days (f) 17 43 69 50 28 13 8 4
==========================================================
Compare these frequencies with those of a Poisson distribution having the same mean and
total frequency. Comment on the agreement.
(2) (a) A large haulage company employers drivers to work a ‘round-the-clock’ system
.The drivers’ union is concern that some periods are more dangerous than others. The
management contests this, claiming that no one period is more dangerous than any other. The
following data on the incidence of traffic accidents by time of days, for this group over the
previous two years is presented (Table 5.22). Are these results consistent with the
management’s claim?
Table 5.22 Haulage company employers’ drivers work hours
=====================================================================
Time of day
(24 hr clock) 00.01–04.00 04.01–08.00 08.00–12 12.01–16.00 16.01–20.00 20.01–24.00 Total

Number of
accidents 14 16 24 22 24 20

=====================================================================
(b) The same company wishes to examine whether there is an association between accident
proneness and coloured blindness. The results for a group of 80 drivers (with a minimum
of five years’ employment) are as in Table 5.23.

161
 Table 5.23 Records of accidents due to Coloured blind
==================================================
Coloured blind
No Yes
----------------------------------------------------------------------------------
Accidents during None 22 5
Last five years One or more 38 15
===================================================
Is there any evidence of an association between colour blindness and accident proneness?
(3) Explain how to calculate the degrees of freedom for the X2-statistic in
(a) a goodness-of-fit test,
(b) a test of no association of the two factors in an h x k contingency table.
An ecologist collected organisms of a particular species from three beaches and
counted the number of females in each sample (the remainder were males).
Table 5.24 Ecologist counts of organisms
Beach 1 2 3
Number of females 44 86 110
Total number in sample 100 200 200
Test if the proportion of females differed significantly between the beaches.
Find the percentage of females at each beach and comment on the results
(4) In routine test of seeds of a certain flower, batches of hundreds are allowed to germinate
and the colour of the resulting blooms is noted. 25% are expected to be red. It is suspected
that a certain laboratory assistant may have been recording data carelessly. In eight
successive batches, he returns the number of red blooms (out of 100 in each case) as 18,
20, 26, 21, 37, 16, 30, 32. Does this evidence support the suspicion?
(5) Table 5.25 gives the distribution for the number of heavy rainstorms reported by 330
weather stations in the United States of America over a one-year period.
(a) Find the expected frequencies of rainstorms given by the Poisson distribution
having the same mean and total as the observed distribution.

162
 (b) Use the X2-distribution to test the adequacy of the Poisson distribution as a model
for these data.
Table 5.25 Distribution for the number of heavy rainstorms
==========================================================
Number of rainstorms (x) 0 1 2 3 4 5 More than 5
Number of stations (f)
reporting x rainstorm 101 114 74 28 10 2 0
==========================================================
(5) A doctor made a comparison of ointment A and ointment B for the cure of a skin disease.
She chose at random which ointment to use for each patient and assessed the result after
two weeks. Of the 95 patients receiving ointment A, 85 were cured, while of the 105
patients receiving B, 78 were cured. Is there evidence that one ointment is butter than the
other? Suppose ointment A had been used by doctor C on her patients and ointment B
had been used by doctor D on his patients and the same numbers of patients and cures
resulted. Would this affect your interpretation of the data?
(6) Given that a random variable which has a X2 –distribution with n degrees of freedom can
be represented as the sum of the squares of n independent standardised normal variables,
show that the sum of two independent X2 variables is itself a X2 variable and explain, with
justification, how the degrees of freedom of the three variables are related.
In making a scientific measurement a technician has to estimate the position of a
Table 5.26 Measurements of readings
===============================================================
Final digit 0 1 2 3 4 5 6 7 8 9
Frequency 11 27 16 23 13 31 14 18 24 23
needle between two successive divisions of a scale in order to obtain the final digit of his
reading. Table 5.26 shows the final digit recorded in 200 such measurements. State what
frequencies you would have expected, and justify your choice. Test whether the data are
consistent with your expectation, and comment as fully as possible on the result of your
analysis.

163
(7) A chain store offers rain hats for sale in three colours: red, blue and green. A sales manager
wonders if the colour preference of customers differs in London and a county town. In the
London store 48 hats were sold in a week and of these 19 were red, 14 blue and 15 green. In the
county two branch 32 hats were sold and of these 6 were red, 16 blue and 10 green. Is there
evidence of differential colour preference between London and the county town? Are any
colours significantly preferred to the others?

5.0 UNIT SUMMARY

Chi-squared (X2) test for frequencies, test statistic X2 where

 (Oi  Ei ) 2 
 2   
 Ei 
X2 = (Oi – Ei)2
Ei
If v = 1, use Yates’ correction:

 (  i  Ei  0.5) 
 2

 2   

 Ei 

X 2 has a 2 distribution with v degrees of freedom:
===============================================================
Type of test Conditions v
----------------------------------------------------------------------------------------------------------
Binomial fit, n classes (a) p know (a) n - 1
(b) p unknown (b) n - 2
Poisson fit, n classes (a) known (a) n - 1
(b) unknown (b) n - 2
Normality, n classes (a) u. o known (a) n -1
(b) u. o unknown (b) n - 3
Contingency table, hxk (h – 1)(k – 1)
===============================================================

164
UNIT 6 CORRELATION ANALYSIS

6.1 Unit Introduction

Welcome to Unit 6 in which you learn how to carry out the correlation analysis using statistical
methods. You will learn what correlation is and how to calculate and interpret the correlation
coefficients.

6.2 Unit Aim

The aim of this Unit is to teach you on how to carry out correlation analysis..

6.3 Unit Objectives

By the end of the Unit you should be able to:

 measure the degree of association between numerical variables

 Establish if there are possible connections between variables.

 investigating the relationship between two quantitative, continuous variables

Terminology

 : Summation

r : Correlation coefficient

6.4 Unit Time required

You need 20 hours for this unit

6.5 Unit Topic

165
6.5.1 Scatter diagrams

In the unit you will develop a theory to test whether two attributes of a population are associated
with each other: the theory of this unit is concerned with the relationship between two variates for
a population. Table 6.1 gives the marks, X and Y, obtained by a group of students in each of two
mathematics examinations.
As we would expect, each student tends to obtain similar marks in both papers. This is
illustrated graphically in Figure 6.1. Such a graph is known as a scatter diagram. Each
Table 6. 1 Marks obtained in a Mathematics examination

166
Figure 6.1 Scatter diagram for the data in Table 6.1
Figure 6.1 shows a scatter diagram of X and Y. You just plot as coordinates of (x,y) and the
relationship between X and Y and compare your graph with standard graphs in figure 6.4. The
data shown in Table 6.1 are shown on a scatter diagram in figure 6.2. In figure 6.2 the graph
shows an inverse correlation.
The data in Table 6.3 gives the scores obtained on each of two dice when they were thrown
together 15 times, and the corresponding scatter diagram is shown in figure 6.3.
In this case there is no obvious correction.
Table 6.2 Number of working horses and number of tractors in Zambia (in ten
thousands) (Data from An Agricultural Statistics,)

167
Figure 6.2 Scatter diagram for the data in Table 6.2
Table 6.3 Scores of two dies

168
 Figure 6.3. Scatter diagram for the data in Table 6.3

Figure 6.4 Standard scatter diagrams

169
Figure 6.5. Diagram showing why cov(X, Y) is positive for positive correlation

6.5.2 Measurement of correlation

The scatter diagrams in Figure 6.4 show a difference in the degree of correlation which can be
readily appreciated by eye. How can we measure the degree of correlation quantitatively? You
know an expression E ( X   X )(Y  Y ) which is termed as covariance, cov(X, Y), of two random
variables X and Y, and it was proved there that cov(X, Y) is zero if X and Y are independent. If
X and Y are not independent cov(X, Y) differs from zero. This can be seen by reference to Figure
6.5. For the point ( x1 , y1 ) , x1  x and y1  y are both positive and therefore their product is

positive also. For the point ( x2 , y 2 ) , x2  x and y 2  y are both negative so that their product is

positive. In fact for the majority of points ( xi  x )( y i  y ) is positive leading to a positive estimate
for cov(X, Y). Similarly for an inverse correlation such as that shown in Figure 6 the product (xi
– x)(yi – y) is generally negative and the estimated value of cov(X , Y) is also negative. Only in
cases where there is no correlation will cov(X, Y) be zero since then positive and negative values
of ( xi  x )( y i  y ) are equally likely.
It can be shown that the unbiased estimate of cov(X, Y) is
1 n
 ( xi  x )( yi  y )
n  1 i 1

170
This depends not only on the degree of correlation but also on the spread of values of X and of Y.
To form a basis of comparison the covariance must be standardised so that it does not depend on
the scales on the X and Y axes. This can be done by making xi  x into a standardised deviate by

dividing by ŝ x , the unbiased estimate of  x , and correspondingly for Y. This gives as an

expression for the degree of correlation
1 n ( xi  x ) ( y i  y )

n  1 i 1 sˆ X
*
sˆY
This statistic is called r and is known as the (product moment) correlation coefficient.
Substituting

 1 n 2  1 n 
sˆ X   
 n  1 i 1
( x i  x ) 

and ˆ
s Y   
 n  1 i 1
( yi  y ) 2 


Figure 6.6. Diagram showing why cov(X, Y) is negative for negative correlation.
This becomes
n

 (x i  x )( y i  y )
r i 1
(1)
n n

 
 i 1
( x i  x ) 2

i 1
( yi  y ) 2 

To simplify calculation this formula can be rewritten using the following relationships
(see equation (3.7.1)):

171
 2
n n
 n 

i 1
( x i  2
x )  
i 1
x2
i   xi  / n
 i 1 

2
n
 n  n

 ( yi  y )   y    yi  / n
 i 1 
2 2
i n
i 1 i 1
 (x
i 1
i  x )( yi  y ) can also be

rewritten in an alternative form, since

n n n n n

 ( xi  x )( yi  y)   xi yi   xyi   xi y   xy
i 1 i 1 i 1 i 1 i 1

n n n
  x i y i  x  y i  y  x i  nx y
i 1 i 1 i 1
n
  x i y i  nx y  n y x  n x y
i 1
n
  x i y i  nx y
i 1
n
 n n

  xi y i    xi  y i  / n
i 1  i 1 i 1 
Using these alternative forms, equation (1) becomes
n n n
n  xi y i   xi  y i
r i 1 i 1 i 1
(2)
 n  n 
2
 n 2  n  2 
n xi2    xi  n  y i    y i  
 i 1  i 1   i 1  i 1  

Table 6.4 Life expectancy and number of people per physician.

172
It can be shown that r can only take values between + 1 and – 1. For exact direct linear correlation,
i.e. when the points on the scatter diagram lie exactly on a straight line of positive slope, r = + 1:
for exact inverse linear correlation, r = - 1. If there is no correlation r is zero. It is important to
remember that r only measures linear correlation. For the scatter diagram shown in Figure 7 r
would be close to zero: there is obviously correlation between X and Y but the relationship is not
linear. Conversely for the data shown in Table 6.4 and illustrated in figure 8 r takes the value –
0.83 even though the relationship is evidently not linear. In fact we should not expect a linear
relationship since, however few physicians there are, the life expectancy can never be negative!
Example 6.1
Calculate r for the data in Table 6.4.

173
 Figure 6.8 Scatter diagram for the data in Table 6.4
Denoting the marks in paper 1 and paper 2 by X and Y respectively, the calculation is shown in
Table 6.5. Substitution in equation (2) gives = 0.825
This rather tedious calculation can sometimes be simplified by using coded values. If we use
coded values u i and v i for x i and y i such that

xi  A  Bu i  x  A  Bu
y i  C  Dvi  y  C  Dv

Table 6.5 Calculation of r for data in Table 6.1

174
( then from equation (1)
n

 (x i  x )( y i  y )
r i 1

n n

 i   ( yi  y ) 2 
2
( x x )
 i 1 i 1 
n

 ( Bu i  bu )(Dvi  Dv 0
r i 1

n n

  ( Dvi  Dv ) 2 
2
( Bu i Bu )
 i 1 i 1 
n

 (u i  u )(vi  v )
r i 1

n n

 
 i 1
(u i  u ) 2

i 1
(v i  v ) 

which using the identity proved above (see equation(3) becomes
n n n
n u i vi   u i  vi
r i 1 i 1 i 1
(3)
 n  n   n 2  n  
2 2

n u i    u i  n vi    vi  
2

 i 1  i 1   i 1  i 1  

175
Modern calculators simplify the computation of r, many of them being able to calculate it directly
when the values of X and Y are fed in. However, it is wise to draw a scatter diagram first to whether
or not the calculation of r is appropriate. (What r is calculated the values of xi2 , yi2 and xiyi should
not be rounded off too much as this may result in a value of |r| greater than one!)

6.1 UNIT ACTIVITY

(1) Calculate the correlation coefficient for the data in (a) Table 2, (b) Table 3.
(2) The following experiment was performed at the Much more Crops Institute. Carnations
were grown on standard plots inside and outside glasshouses. For each plot the amount of
a certain vital, but toxic, chemical was measured together with the average number of
blooms per plant.
The results were as in table 6.6 and 6.7.
Calculate the product moment correction coefficient for each set of data.
Table 6.6 Glasshouse crop
=====================================================
Amount of chemical
in standard plot x 3 4 6 7 8 10 12 15 16
(micrograms)
Average number of y
blooms per plant for this plot 3.2 2.9 3.7 2.2 1.8 2.3 1.7 0.8 0.3
===========================================================
(You are given that ∑v = 18.9. ∑ v2 = 49.33: other sums should be calculated.)

Table 6.7 Outdoor crop

================================================
Amount of chemical
in standard plot x
(micrograms)
Average number of 3 5 6 10 11 12 14 15

176
 blooms per plant y
for this plot 4.0 4.2 3.6 2.3 2.5 1.9 1.3. 1.1
======================================================
(You are given that ∑y = 20.9, ∑y2 = 64.65: other sums should be calculated.)
(3) (a) Pupils transferring from primary to secondary education are given two verbal
reasoning tests. The score of twelve pupils (quoted relative to the mean score for each
test) are in Table 6.8. Show that the product moment correlation coefficient of the test
scores is approximately 0.871.
Table 6.8 Two verbal reasoning test
===============================================================
Pupil A B C D E F G H J K L M
Test 1 - 15 - 14 -10 -5 -4 -1 2 5 10 13 17 20
Test 2 - 10 -7 -4 -6 -1 -6 -2 5 9 0 10 12
===============================================================
(b) Corresponding values of the variable X and Y are given in Table 6.9.
These values give a product moment correlation coefficient of 0.845
Table 6.9 Corresponding values of the variable X and Y
===============================================================
X -15 - 14 -10 -5 -1 4 10 17 20
Y - 10 -7 0 5 -8 10 11 10 9
Compare and contrast, without further calculation, the relationship between the test
scores in (a) and the relationship between X and Y.

Some fallacies in the interpretation of r.

(i) Correlation should not be confused with causation. Figure for the years 1955-65 show high
correlation between the number of television licences and the number of road accidents in
this country but no one would suppose that one of these factors causes the other. Such a
spurious relationship can result when each of the two variables depends on a third variable,
in this case the ‘standard of living’, which produces a simultaneous change in both of them.
It is particularly common if the variables are measured over a period of time.

177
(ii) Even if it appears that there is a causal relationship between two variables, the correlation
coefficient does not help us to establish which variable depends on which.
(iii) As already mentioned a non-significant value of r only implies an absence of linear
correlation.
(iv) A linear relationship established over a particular range of values of X and Y should not be
assumed to hold outside that range by extrapolation. For example, there is a strong direct
correlation between the age and weight of a baby, but to extend this linear relationship
would imply that older children and adults continued to grow at this same rate!

6.5.3 Spearman’s rank correlation coefficient, rs

A manufacturer of margarine has been experimenting with different recipes and wishes to test
public Reaction to them. Table 6.12 shows how two tasters rank eight varieties in order of
preference. Hardly surprisingly the testers do not show the same order of preference but the
manufacturer would like to know whether there is any degree of correlation between their
rankings. In this case you do not have two continuous variates which are normally distributed
but

Table 6.10 Public Reaction to taste

178
two discrete varieties in the form of the rankings. Using these rankings we can calculate the
correlation coefficient using formula (2). However, since the values of X (and Y) are the first n
integers, this formula can be considerably simplified. If there are n ranks then
n n n
1
 x   y   i  2 n(n  1)
i 1
i
i 1
i
i 1

n n n
1
x
i 1
2
i   yi2   i 2 
i 1 i 1 6
n(n  1)(2n  1)

Using these expressions the denominator of formula (2) becomes

2
1 1 
n * n(n  1)(2n  1)   n(n  1)
6 2 
1 2 2
 n (n  1) (after some manipulation)
12
The numerator of ( equation 1 ) can also be simplified by letting di be the difference between
ranks for the ith item. Then d i  xi  yi
n n n n n

d
i 1
i
2
  ( xi  yi ) 2   xi2  2 xi yi   yi2
i 1 i 1 i 1 i 1

Rearranging
n
1  n 2 n 2 n 2
 xi y i 
i 1
 xi  
2  i 1 i 1
yi   d i 
i 1 
1 1 n

  2
2 6
* n( n  1)(2n  1)  
i 1
d i2 


179
Using this expression for ∑ xiyi and substituting for ∑ xi and ∑ yi as before, the numerator
of equation (3.) becomes
2
1  1 n
 1 
 n 2 * n(n  1)(2n  1)   d i2    n(n  1)
2  6 i 1  2 
1 2 2 1 n
 n (n  1)  n d i2
12 2 i 1
Combining the expression obtained for the numerator and denominator we obtain for the rank
correlation coefficient, r s,
n
6 d i2
rs  1  i 1
(4)
n(n  1)
2

This correlation coefficient is due to Spearing (1904) and is usually called after him. Obviously
when the rankings are identical the di s are all zero and rs = 1.
Example 6.2 Calculate the Spearman rank correlation coefficient for the data in Table 6.10
(reproduced in Table 6.11).
Table 6.11 Public Reaction to taste
Variety Taster 1, xi Taster 2, yi di d i2
------------------------------------------------------------------------
A 5 7 -2 4
B 4 3 +1 1
C 2 1 +1 1
D 1 2 -1 1
E 6 6 0 0
F 7 6 -1 1
G 3 8 -1 1
H 8 5 +3 9
0 18
==========================================

180
 n
6 d i2
6 *18
rs  1  i 1
 1  0.79
n(n  1)
2
8(8 2  1)
As for r, rs can vary between – 1 and +. The critical values of rs at the 5% level, taking the
null hypothesis that there is no correlation, are given in Table 6.12. As the number of
Table 12 Public Reaction to taste
===============================================
Number of ranks Critical value of |rs|
-----------------------------------------------------------------------------
5 1.00
6 0.89
7 0.79
8 0.74
9 0.70
10 0.65
20 0.45
25 0.40
50 0.28
ranks increase the cortical values appropriate for r can be used. Using this table the value of rs =
0.79 obtained in the preceding example is significant at the 5% level since the critical region is |
rs | > 0.74 indicating there is a degree of correlation between the rankings of the two tasters.

In the example given here the variate measure by the tasters, i.e. how much they liked the
margarine, was qualitative and the correlation could only be measured by computing rs. Rank
correlation is also used for quantitative variables when neither of them is normally distributed.
One problem that may arise in this case is that two more values may be equal. What ranking value
should they be given? Consider the following arranged in increasing order:
4.3, 5.1, 5.1, 6.3, 7.5, 8.6, 8.6, 8.6
The second and third values are ‘tied’ and are both given an average ranking of 2½. Similarly the
6th, 7th and 8th values are given the average ranking of 7. The complete sequence of rankings is

181
therefore: 1, 2½, 2½, 4, 5, 7, 7, 7. Note that the sequence has the same total as 1, 2, 3, 4, 5,
6, 7, 8.
Example 6.3
Two judges of an international beauty competition award the marks given in Table 6.13.
Calculate the Spearman rank correlation coefficient.
Do the judges agree on the order in which they place the candidates?
Table 6.13 international beauty competition
=====================================================================
Competitor
Zambia France Malawi Germany Botswana USA Canada
--------------------------------------------------------------------------------------------------------------------
-
Judge A 5.8 5.5 5.9 4.9 5.6 5.6 5.0
Judge B 5.5 5.4 5.6 5.3 5.7 5.7 5.7
=====================================================================
Judges A’s marks, placed in descending order, are
Mark 5.9 5.9 5.8 5.6 5.5 5.0 4.9
Rank 1½ 1½ 3 4 5 6 7
with the rankings below.
For judge B, marks and their ranks are
Mark 5.8 5.7 5.7 5.7 5.5 5.4 5.3
Rank 1 3 3 3 5 6 7
The calculation of rs is shown in Table 6.14.
Our null hypothesis is that there is no correlation between the judges’ rankings. Using Table 6.13
the critical region for n = 7 is | rs | > 0.79. The value obtained is not significant at the 5% level.

We retain H o :

Table 6.14 Results of a beauty competition

===============================================================

182
 Competitor Judges A’s rank Judge B’s rank di d i2
------------------------------------------------------------------------------------------------------------
Zambia 3 5 -2 4
France 5 6 -1 1
USSR 1½ 1 ½ ¼
Zimbabwe 7 7 0 0
Germany 1½ 3 -1½ 2¼
RSA 4 3 +1 1
Canada 6 3 +3 9
0 17½
============================================================
Using equation (1),
rs = 1 - 6 x 17 ½ = 0.6875

7(72 – 1)
there is no evidence that the judges agree on the order in which they place the competitors.
(Strictly speaking, the formula for rs should be modified for tied ranks since its derivation used
the fact that the ranks were the first n natural numbers. However, the correction involved is
small.)

6.2 UNIT ACTIVITY

(1) The Organisers of a flower completion intend to base their order of merit on the opinions
of judge X and Judge Y. The marks given by the two judges are set out in Table 6.15.
(a) Carry out a rank correlation test to determine whether the two judges’ opinions are
consistent.
(b) Determine an order of merit of the competitors.
(2) In a village flower show the six competitors in the finals had their entries assessed by
three judges. The placing’s were as in Table 6.16
Calculate a rank correlation coefficient between (a) the first and second judges’ placing’s,
and (b) the first and third judges’ placing’s. Comment on your results.

183
 Table 6.15 Village flower competition
=============================================================
Competitor A B C D E F G H I J
Judge X marks
(max. 100) 48 50 55 51 51 47 48 46 52 50
Judge Y marks
(max. 100) 18 19 29 22 26 14 22 11 24 17
=============================================================
Table 6.16
===========================================================
Competitor P Q R S T U
-------------------------------------------------------------------------------------------------
1st Judge 3 5 6 1 4 2
2nd Judge 4 3 5 2 6 1
3rd Judge 2 6 4 1 3 5
=============================================================
Table 17
===========================================================
x 29 81 60 88 91 91 86 99 73 42
y 74 86 63 74 70 63 41 81 29 56
=============================================================

Table 18 Results of X and Y for judges at a beauty contest

===============================================================
Competitor A B C D E F G H I
------------------------------------------------------------------------------------------------------------
Judge X 4 9 2 5= 3 10 5= 7 8

184
 Judge Y 6 10 2 8 1 9 7 4 5
(3) Explain briefly how a scatter diagram can show whether there is direct or indirect
correlation between two quantities. Give one example each of data where (a) direct, (b)
indirect correlation might be expected. Table 6.17 shows some values of a variable x and
the corresponding values of a second variable y. Calculate a coefficient of rank
correlation between x and y and comment briefly on the result you obtain.
(4) (a) X and Y were judges at a beauty contest in which there were ten competitors.
Their rankings are shown in Table 6.18.
Calculate a coefficient of rank correlation between these two sets of ranks and
comment briefly on your result.
(b) Illustrate by means of two scatter diagrams rank correlation coefficients of 0 and - 1
between two variable X and Y.

6.3 UNIT ACTIVITY

(5) In an experiment the values of two variables, X and Y, were measured. Ten such experiments
were performed with the following result for (X, Y):
(6, 13), (12, 2,) (9, 12), (5, 15), (2, 17), (12, 5), (8, 10), (3, 13), (11, 12), (7, 11).
Illustrate these data by means of a scatter diagram and comment briefly on the type of
correlation that this shows. Calculate a coefficient of rank correlation between the ten values
of X and Y.
(6) In each of the following section, which purport to be extracts from reports, the second
sentence is an inference from the statement made in the first sentence. State whether the
inferences are valid or invalid, and give the reasons for your decisions.
(a) The amount of fertiliser applied was varied from plot to plot and the correlation
coefficient between the yield of corn from a plot and the amount of fertiliser
applied was found to be 0.02. This shows that there was no relation between yield
and amount of fertiliser applied.
(b) Inspection of the police reports of car accidents in the town during 1975 revealed
that, when the number of accidents involving drivers of a particular age was

185
 correlated with that age, the correlation coefficient was – 0.72. We conclude that
older drivers are less likely to have an accident than younger drivers.
(c ) The correlation coefficient between the sugar content s of the peas and the length
of time t they have been in the greengrocer’s shop was negative. It follows that s
decreases with increasing t.
(d) The correlation coefficient of percentage of children over sixteen at school and
the Gross National Product, over the years for which we have records, is 0.91,
which is a high correlation. It is obvious that if many more children can be
persuaded to stay on at school after sixteen then the Gross National Product will
increase substantially.
(7) (a) Explain the use of (i) the product moment correlation coefficient, (ii) the
coefficient of correlation by ranks.
Give two examples of the appropriate application of each.
(b) At the final judging of a ‘Cow of the Year Show’, two judges gave the descending
orders of merit of ten cows as EAHJBIFCGD and EHJAFICBDG. Find the rank
correlation coefficient between these two orders. Discuss the significance of the result.
(8) Define the product moment correlation coefficient and explain how you would interpret
values of 0 and 1. Guess the value of the correlation coefficient between the variables in the
following situation and interpret your estimate.
(a) The height of water and the volume of road traffic at London Bridge, if high tide is at
7 a.m. The interval between successive high tides is about 12 hours.
(b) The marks in paper I and total marks in a two-paper examination.
(9) Explain clearly what is meant by the statistical term ‘correlation’.
Vegboost Industries, a small chemical firm specialising in garden fertilisers, set up an
experiment to study the relationship between a new fertiliser compound and the yield from
tomato plants. Eight similar plants were selected and treated regularly throughout their life
with x grams of fertilizer diluted in a standard volume of water. The yield y, in kilograms,
of good tomatoes was measure for each plant. Table 6.19 summarises the results.
(a) Calculate the product moment correlation coefficient for these data.
(b) Calculate Spearman’s rank correlation coefficient for these data.

186
 (c) Is there any evidence of a relationship between these variables? Justify your answer.
(No formal test is required.)
Table 6.19 Fertiliser compound and the yield from tomato plants
===============================================================
Plant A B C D E F G H
Amount of
fertilizer x (g) 1.2 1.8 3.1 4.9 5.7 7.1 8.6 9.8
Yield y (kg) 4.5 5.9 7.0 7.8 7.2 6.8 4.5 2.7
===============================================================
(10) Explain how you would calculate the product moment correlation coefficient from a
sample of n pairs of values xi, yi (i = 1, 2, . . . , n).
The number of eggs laid, x, by a certain species of bird is either 1 or 2, and this is related to
the age of the bird, y, which is also 1 or 2. The frequencies of the four possible
combinations of values x, y are recorded for a total of n birds, as shown in Table 17.28.
Show that the product moment correlation coefficient between x and y is (ad – bc)/ {(a +
b)(c + d)(a + c) (b + d)}1/2.

Table 6.20 Number of eggs laid, x, by a certain species of bird

============================================
Numbers of eggs laid (x)
1 2 Total
---------------------------------------------------------------------------------------------
Age y 1 a b a+b
(years) 2 c d c+d
---------------------------------------------------------------------------------------------

187
 Total a+c b+d n
===========================================
(11) What is meant by the statistical terms ‘independent’ and uncorrelated’, and how are they
related?
At the University of Zambia, students for the degree in chemical engineering take two
examinations. They have a part I examination at the end of the second year of their studies,
and a part II examination at the end of the third year. Degree classification
Table 6.21 Marks of a Chemical Engineering Examination
==========================================================
Part I marks
40-49 50-59 60-69 70-79 80-89 90-99
------------------------------------------------------------------------------------------------
40-49 3 5 4 0 0 0
50-59 3 6 6 2 0 0
Part II marks 60-69 0 5 9 5 2 0
70-79 0 0 5 10 8 1
80-89 0 0 0 5 6 5
90-99 0 0 0 2 4 4
==========================================================
depends upon an average of those marks. Table 21 gives the part I and part II marks for a
particular set of 100 students. Use the coding method to calculate the product moment
correlation coefficient for these data. (Assumed means of 64.5 for part I marks and 74.5 for
part II marks should be used.)

6.0 UNIT SUMMARY

The following formulas are very useful in the calculation of correlation coefficient r.

188
 n

 (x i  x )( y i  y )
1. r  i 1

n n

 i   ( yi  y ) 2 
2
( x x )
 i 1 i 1 

n n n
n  xi y i   xi  y i
2. r  i 1 i 1 i 1

 n  n 
2
 n 2  n  2 
n xi2    xi  n  y i    y i  
 i 1  i 1   i 1  i 1  

This rather tedious calculation can sometimes be simplified by using coded values. If we use
coded values u i and v i for x i and y i such that

 (u i  u )(vi  v )
r i 1

3.  n n

 i   (vi  v )
2
(u u )
 i 1 i 1 

n n n
n u i vi   u i  vi
4. r  i 1 i 1 i 1

 n  n 
2
 n 2  n  2 
n u i2    u i  n vi    vi  
 i 1  i 1   i 1  i 1  

5. Spearman’s rank correlation coefficient, rs

n
6 d i2
rs  1  i 1

n(n  1) 2

189
 Correlation coefficient is used to find the relationship between to variable X and Y.

UNIT 7 THE ANALYSIS OF VARIANCE

7.1 Unit Introduction

Welcome to Unit 6 in which you will learn how to conduct analysis of variance. In this Unit you
are going to learn how to design experiments. You will learn how to allocate units in block
designing and carry out the analysis for each design.

7.2 Unit Aim

The aim of this Unit is teach you on how to design an experiment and analysis.

7.3 Unit Objectives

190
By the end of the unit you should be able to:

 Perform analysis of variance by hand

 Appropriately interpret results of analysis of variance tests.

 Distinguish between one and two factor analysis of variance tests.

 Identify the appropriate hypothesis testing procedure based on type of outcome variable
and number of samples

Terminology

Total SS Total sum of squares

SST Sum of squares of Treatments

SSC – Sum of squares of columns

SSR – Sum of squares of rows

SSE – sum of squares of errors

MSC- Mean squares of columns

MSR – means squares of errors

MST-Mean squares of Treatments

F- Distribution

7.4 Unit Time required

You need 20 hours for this unit

7.5 Unit Topics

191
7.5.1 The Analysis of Variance

The methodology for the analysis of experiments involving several independent variables can best
be explained in terms of the linear probabilistic models. Although elementary and unified, this
approach is not susceptible to the condensation necessary for inclusion in an elementary text.
Instead, we shall attempt an intuitive discussion using a procedure known as the analysis of
variance. Actually, the two approaches are connected and the analysis of variance can easily be
explained in a general way in terms of the linear model.

As the name implies, the analysis-of-variance procedure attempts to analyze the variation of a
response and to assign portions of this variation to each of a set of independent variables. The
reasoning is that response variables vary only because of variation in a set of unknown independent
variables. Since the experimenter will rarely, if ever, include all the variables affecting the
response in his experiment, random variation in the response is observed even though all
independent variables considered are held constant. The objective of the analysis of variance is to
locate important independent variables in a study and to determine how they interact and affect
the response.

The rationale underlying the analysis of variance can be indicated best with a symbolic discussion.
The actual analysis of variance-that is, “how to do it” it can be illustrated with an example.

You will recall that the variability of a set of n measurements is proportional to the sum of squares
n
of deviations,  ( y  y ) 2 , and that this quantity is used to calculate the sample variance. The
i 1

analysis of variance partitions the sum of squares of deviations, called the total sum of squares of
deviations, into parts, each of which is attributed to one of the independent variables in the
experiment, plus a remainder that is associated in Figure 1 for three independent variables. If a
multivariate linear model were written for the response, as suggested

192
Figure 7. 1 The Partitioning of the total Sum of Squares of Deviations

The portion of the total sum of squares of deviations assigned to error would be labelled SSE.

For the cases that we consider, and when the independent variables are unrelated to the
response, it can be shown that each of the pieces of the total sum of squares of deviations,
divided by an appropriate constant, provides an independent and unbiased estimator of  2 , the
variance of the experimental error. When a variable is highly related to the response, its portion
(called the “sum of squares” for that variable) will be inflated. This condition can be detected by
comparing the estimate of  2 for a particular independent variable with that obtained from SSE
using a F test. If the estimate for the independent variable is significantly larger, the F test will
reject a hypothesis of “no effect for the independent variable: and produce evidence to indicate a
relation to the response.

The mechanism involve in an analysis of variance can best be illustrated by considering a

familiar example, say the comparison of means for the unpaired experiment for the special case
where n1 = n2. This experiment, formerly analysed by the use of Student’s t, will now be
approached from another point of view. The total variation of the response measurements about
their mean for the two samples is

n n
Total SS   ( yi , j  y ) 2
i 1 j 1

193
Where y i , j denotes the jth observation in the ith sample and y is where yij denotes the jth

observation in the ith sample and y is the mean of all 2n1  n observations. This quantity can
be partitioned into two parts. That is,

2 n1
Total SS   ( yi , j  y ) 2
i 1 j 1

2 2 n1
 n1  ( yi  y )   ( yij  yi ) 2
2

i 1 i 1 j 1

(proof deferred ), where y1 is the average of the observations in the ith sample i = 1, 2, and the
second quantity on the right-hand side of the equality is simply the pooled sum of squares of
deviations used to calculate s2 . (Recall that we assume equal population variances for this
statistical method). We will denote this quantity as SSE.

Note that we have partitioned the “total sum of squares” of deviations into two parts. One part,
SSE, can be used to obtain a pooled estimator of  2 . the other part,

2
n1
n1  ( yi  y ) 2  ( y1  y 2 ) 2  SST
i 1 2

which we will call the sum of squares for treatments, will increase as ( y1  y 2 ) increases.

Hence, the larger SST, the weight of evidence to indicate a difference in (1   2 )

How large is “large?” When will SST be large enough to indicate a real difference between 1

and  2 ?

SSE SSE
s 2  MSE   , (since n1  n2 )
n1  n2  2 2n1  2

provides an unbiased estimator of  2 . (MSE denotes mean square for error.) Also, when the null
hypothesis is true (that is, µ1 = µ2), SST divided by an appropriate number of degrees of freedom
yields a second unbiased estimator for …2, which we will denote as MST. For this example, the
number of degrees of freedom for MST is equal to 1. When the null hypothesis is true, MSE and

194
MST (the mean square fir treatments) will estimate the same quantity and should be “roughly” for
the same magnitude. When the null hypothesis is false and µ1 ….. µ2, MST will tend to be larger
than MSE. You can show that E(MST) =  2 . when µ1 = µ2.

MST
The preceding discussion, together with a review of the variance ration, suggests the use of
MSE
as a test statistic to test the hypothesis, µ1 = µ2, against the alternative, µ1 ≠ µ2. Indeed, when both
populations are normally distributed, it can be shown that MST and MSE are independent and that
MST
F follows the F probability distribution. Disagreement with the null hypothesis is
MSE
indicated by a large value of F, and hence the rejection region for a given  will be F  F

The analysis-of-variance test results in a one-tailed F test, the degrees of freedom for F will be
those associated with MST and MSE, which we will denote as v1 and v2, respectively. Although
we have not indicated, in general, how one determines v1 and v2, v1 = 1 and v2 = (2n1 – 2) for the
two sample experiment described.

Proof of the independence of MST and MST will not be considered here because it is a special
case of a; more general and more complicated theorem that applies to many different analyses of
variance.

Example 7.1: The coded values for the measure of elasticity in plastic, prepared by two
different processes, for samples of six drawn randomly from each of the two processes, are as
follows:

Table 7.1 Measure of elasticity in plastic

195
Do the data present sufficient evidence to indicate a difference in mean elasticity for the two
processes?

Solution: Although the student’s t could be used as the test statistic for this example, we shall
use our analysis-of-variance F test since it is more general and can be used to compare more than
two means.

The three desired sums of squares of deviations are

2 6

2 6 2 6
( y ij ) 2
 ( yij  y ) 2   yij2 
i 1 j 1
Total SS =
i 1 j 1 i 1 j 1 12

(91.9) 2
 711.35   7.5492
12

2
SST  n1  ( y1  y ) 2  6 ( y1  y ) 2  1.6875
i 1

2 6
SSE   ( yij  yi ) 2  5.8617
i 1 j 1

(You may verify that SSE is the pooled sum of squares of the deviations for the two samples.
Also, note that Total SS = SST + SSE.) The mean squares for treatment and error are,
respectively,

196
 SST
MST =  1.6875
1

SSE 5.8617
MSE =   0.58617
2n1  2 10

To test the hypothesis, µ1 = µ2, we compute the test statistic

MST 1.6875
F= 
MSE 0.58617

The critical value of the F statistic for α = .05 is 4.96. Although the mean square for treatments is
almost three times as large as the mean square for error, it is not large enough to reject the null
hypothesis. Consequently, there is not sufficient evidence to indicate a difference between µ1 and
µ2.

As noted, the purpose of the preceding example was to illustrate the computations involved in a
simple analysis of variance. The F test for comparing two means is equivalent to a student’s test
because an F statistic with one degree of freedom in the numerator is equal to t2. You can easily
verify that the square of t.025 = 2.228 (used for the two-tailed test with α = .05 and v = 10 degrees
of freedom) is equal to F.05 = 4.96. Similarly, the value of the t statistic for Example7.1 would
equal the square root of the computed F = 2.88.

7.5.2 A Comparison of More Than Two Means

An analysis of variance to detect a difference in a set of more than two population means is a
simple generalization of the analysis of variance of Section 2.5.2. The random selection of
independent samples from p populations is known as a completely randomized experimental
design.

Assume that independent random samples have been drawn from p normal populations with means
µ1, µ2, … µp, respectively, and variance  2 . All populations are assumed to possess equal
variances. To be completely general, we shall allow the sample size to be unequal and let ni, i = 1,

197
2, …, p, be the number in the sample drawn from the ith population. The total number of
observations in the experiment will be n = n1 +…+ np.

Let yij denote the measured response on the jth experimental unit in the ith sample and let Ti and
Ti represent the total and mean, respectively for the observations in the ith sample. (The
modification in the symbols for sample totals and averages will simplify the computing formulae
for the sums of squares.) Then, as in the analysis of variance involving two means,

Total SS = SST + SSE

n n1 p n1
Total SS   ( yij  y )   yij2  CM
2

i 1 j 1 i 1 i 1

(total of all observations) 2

CM 
n

2
 p ni 
  yij 
 
   ny 2
i 1 j 1

(the symbol CM denotes “correction for the mean”),

p p
Ti 2
SST   ni (Ti  y ) 2    CM
i 1 i 1 ni

SSE = Total SS - SST

Although the easy way to compute SSE is by subtraction as shown above, it is interesting to note
that SSE is the pooled sum of squares for all p samples and is equal to

p ni
SSE   ( yij  Ti ) 2
i 1 j 1

The unbiased estimator of  2 based on (n1 + n2 + …+ np – p) degrees of freedom is

MSE
s 2  MSE 
n1  n2  ...  n p  p

198
The mean square for treatments will possess (p-1) degrees of freedom, that is, one less than the
SST
number of means, and MST 
p 1

To test the null hypothesis, H o : 1   2  ...   p against the alternative that at least one of the

equalities does not hold, MST is compared with MSE using the F statistic based upon v1 = (p – 1)
 p 
and v2 =   ni  p (n  p ) degrees of freedom. The null hypothesis will be rejected if
 i 1 

MST
F  F
MSE

where F is the critical value of F for probability of a type I error,  .

Intuitively, the greater the difference between the observed treatment means, T1 , T2 ,..., T p , the

greater the will be the evidence to indicate a difference between their corresponding

Population means. It can be seen from the above expression that SST = 0 when all the observed
treatment means are identical, because then T1  T2  ...  T p  y and the deviations appearing in

SST, (Ti  y ) , i= 1, 2, ……p, will equal zero. As the treatment means get farther apart, the

deviations, (Ti  y ) , will increase in absolute value and SST will increase in magnitude.
Consequently, the larger the value of SST will increase in magnitude. Consequently, the larger the
value of SST, the greater will be the weight of evidence favouring a rejection of the null
hypothesis. This same line of reasoning will apply to the F tests employed in the analyses of
variances for all designed experiments.

The assumptions underlying the analysis-of-variance F tests should receive particular attention.
The samples are assumed to have been randomly selected from the p populations in an independent
manner. The populations are assumed to be normally distributed with equal variances,  2 , and
means, 1 ,  2 ,...,  p . Moderate departures from these assumptions will not seriously affect the

properties of the test. This is particularly true of the normality assumption.

199
 Example 7.2: Four groups of students were subjected to different teaching techniques tested at
the end of a specified period of time. As a result of dropouts from the experimental groups (due
to sickness, transfer, and so on), the number of students varied from group to group. ; do the data
shown present sufficient evidence to indicate a difference in the mean achievement for the four
teaching techniques?

Table 7.2 Different teaching techniques

Solution:

2
 4 ni 
  yij 
  (1779) 2
CM   i 1 j 1    137,601.8
n 23

4 ni
Total SS   yij2  CM
i 1 j 1

 139,511  137,601.8  1909.2

4
Ti 2
SST    CM
i 1 ni

 138,214.4  137,601.8  712.6

SSE = Total SS –SST =1196.6

200
The mean squares for treatment and error are

SST 712.6
MST =   237.5
p 1 3

SSE SSE 1196.6

MSE =    63.0
n1  n2  ...  n p  p n  p 19

The test statistic for testing the hypothesis, 1   2   3   4 , is

MST 237.5
F   3.77
MSE 63.0

p
Where 1  p  1  3 ,  2   ni  4  19
i 1

The critical value of F for  = .05 is F.05 = 3.13. Since the computed value of F exceeds F.05, you
reject the null hypothesis and conclude that the evidence is sufficient to indicate a difference in
mean achievement for the four teaching procedures.

You may feel that the above conclusion could have been made on the basis of visual observation
of the treatment means. It is not difficult to construct a set of data that will lead the “visual”
decision maker to erroneous results.

7.5.3 Proof of Additivity of the Sums of Squares and E(MST) for a Completely
Randomized Design

The proof that Total SS = SST + SSE

For the completely randomized design is presented in this section for the benefit of the interested

reader. It may be omitted without loss of continuity.

The proof utilizes elementary results on summations that appear in the exercises for unit 1, and the
device of adding and subtracting TI within the expression for the Total SS. Thus

201
 p ni
Total SS =  ( y
i 1 j 1
ij  y) 2

p ni
=  ( y
i 1 j 1
ij  Ti T i  y ) 2

p ni
= [( yij  Ti )  (T i  y )]2
i 1 j 1

p ni
= [( y
i 1 j 1
ij  Ti ) 2  2( yij  Ti )(Ti  y )  (Ti  y ) 2 ]

Summing first over j, we obtain

 ni p ni

Total SS =   ( yij  Ti )  2(Ti  y ) ( y ij  Ti )  ni (Ti  y ) 2 
2

i 1  j 1 i 1 

where

ni

(y
i 1
ij  Tij )  Ti  ni Ti  Ti  Ti  0

Consequently, the middle term in the expression for the Total SS is equal to zero.

Then, summing over i, we obtain

p ni p
Total SS =  ( yij  Ti ) 2   ni (Ti  y ) 2 ,
i 1 j 1 i 1

= SSE + SST.

The first expression is SSE, the pooled sum of squares of deviations of the sample measurements
about their respective means. The second is the formula for SST.

Proof of the additive of the analysis-of-variance sums of squares for other experimental designs
can be obtained in a similar manner. The procedure is tedious.

202
You will now derive the expected value of MST for a completely randomized design. Assume that
Yij, the jth sampled value from the ith population, has E(Yij) = 1, …, p, j = 1, …,ni. since Ti is the
average of ni independent random variables, Yij, j = 1, …,ni, it follows that E(Ti) = , µi and V(Ti)
2
= . Similarly, Y is given by
ni

1 p ni
Y   Yij ,
n i 1 j 1

and hence

p ni
1 1 p
E (Y ) 
n
  i 
i 1 j 1
 ni  i
n i 1

We will denote E (Y ) by 

1 p ni
2
V (Y )  2
n
V (Y ) 
i 1 j 1
ij
n

Then

1  p 
E ( MST )  E  ni (Ti  Y ) 2 
p  1  i 1 

1  p 
 E  ni (T 2 i  2Ti Y  Y 2 )
p  1  i 1 
1  p 
 E  ni Ti 2  nY 2 
p  1  i 1 

1  p 
  
p  1  i 1
ni E (Ti 2 )  nE(Y 2 )


1  n  2 2  2 
  i n    
i   n   2 
p  1  i 1  ni   n 

203
On noting that, for random variable U, E(U 2 )  V (U )  E(U )2 . It then follows that

1  p 2 p 
E ( MST )       ni  i2   2  n i2 
p  1  i 1 i 1 

1  2 p

 
p 1 
 ( p  1)   ni  i2  n i2 
i 1 

1
2  
p  1 i 1
ni (  i   ) 2

Under H o : 1   2  ...   p ,  i   , i  1,2,.., p , and hence E(MST) =  2 Thus MST/MSE is

a ratio of unbiased estimators of  2 when H o is true.

7.5.4 An Analysis-of-Variance Table for a Completely Randomized Design

The calculations of the analysis of variance are usually displayed in an analysis-of-variance

(ANOVA or AOV) table. The table shown in Table 7.1.

Table 7.3 ANOVA table for a Complete Randomized Design

204
The first column shows the source of each sum of squares of deviations; the second column gives
the respective degrees of freedom; the third and fourth columns give the corresponding sums of
squares and mean squares, respectively. A calculated value of F, comparing MST and MSE, is
usually shown in the fifth column. Note that the degrees of freedom and sums of squares add to
their respective totals.

The ANOVA table for Example 7.2, shown in Table 7.3, gives a compact presentation of the
appropriate computed quantities for the analysis of variance.

Table 7.4 ANOVA Table for Example 7.2

7.5.5 Estimation for the Completely Randomized Design

Confidence intervals for a single treatment mean and the difference between a pair of treatment
means. The confidence interval for the mean of treatment i or the difference between treatments i
and j are, respectively,

205
 t S
Ti  2

ni

1 1
And (Ti  T j )  t s  ,
2 n1 n2

SSE
Where s  s 2  MSE 
n1  n2  ...  n p  p

And t  is based upon (n – p) degrees of freedom

2

Note that the confidence intervals stated above are appropriate for single treatment means or a
comparison of a pair of means selected prior to observation of the data. The stated confidence
coefficients are based on random sampling. If one were to look at the data and always compare
the largest and smallest sample means, the assumption of randomness would be disturbed.
Certainly the difference between the largest and smallest sample means is expected to be larger
than for a pair selected at random.

Example 7.3: Find a 95 percent confidence interval for the mean score for teaching technique 1,

Solution: The 95 percent confidence interval for the mean score is

t 0.025 s
T1 
6

(2.093)(7.94)
75.67 
or 6
75.67  6.78

Example 7.4: Find a 95 percent confidence interval for the difference in mean score for teaching
techniques 1 and 4, Example 2

Solution: The 95 percent confidence interval is

(T1  T4 )  (2.093)(7.94) 1 / 6  1 / 4

206
 12.08  10.74

Hence the 95 percent confidence interval for (1   4 ) is -22.82 to - 1.34. This suggests that

 4  1

7.5.6 The Analysis of Variance for a Randomized Block Design

The method for constructing a randomized block design was presented in Section 7.5.5.

The randomized block design implies the presence of two qualitative independent variables,
“blocks” and treatments.” Consequently, the total sum of squares of deviations of the response
measurements about their mean may be partitioned into three parts, the sums of squares for
blocks, treatments, and error.

Denote the total and average of all observations in block I as Bi as and Bi , respectively.

Similarly, let T j and T j represent the total and the average for all observations receiving

treatment j. Then, for a randomized block design involving b blocks and p treatments,

Total SS = SSB + SST + SSE

b p b p
  ( yij  y )   yij2  CM
2

i 1 j 1 i 1 j 1

b B i
2

SSB  p  ( Bi  y ) 2  i 1
 CM
i 1 p

p T j
2

SST  b (T j  y ) 2 
j 1
 CM
j 1 b

SSE = Total SS –SSB -SST

In the formulas, y  (averageof all n  bp observations)

207
 b p

 y
i 1 j 1
ij


n

(total of all observations) 2

CM  and
n

2
 b p 
  y ij 
 
 
i 1 j 1

The analysis of variance for the randomized block design is presented in table 7.4. The degrees
of freedom associated with each sum of squares is

Table 7.5 ANOVA Table for a Randomized Block Design

shown in the second column. Mean squares are calculated by dividing the sums of squares by
their respective degrees of freedom.

To test the null hypothesis “there is no difference in treatment means,” you use the F statistic,

MST
F
MSE

and reject if F  F based on  1  b  1 and  2  n  b  p  1 degrees of freedom.

208
Blocking not only reduces the experimental error, it also provides an opportunity to see whether
evidence exists to indicate a difference in the mean response for blocks. Under the null hypothesis
that there is no difference in mean response for blocks, MSB provides an unbiased estimator for
 2 based on (b-1) degrees of freedom. Where real differences exist among block means, MSB
will tend to be inflated in comparison with MSE and

MSB
F
MSE

provides a test statistic. As in the test for treatments, the rejection region for the rest will be

F  F

Based on  1  b  1 and  2  n  b  p  1 degrees of freedom.

Example 7.5 A stimulus-response experiment involving three treatments was laid out in a
randomized block design using four subjects. The response was the

Table 7.6 Allocating of units in blocks

length of time to reaction measured in seconds. The data, arranged in blocks, are shown in Table
7.6. The treatment number is circled and shown above each observation. Do the data present
sufficient evidence to indicate a difference in the mean response for stimuli (treatments)?
Subjects?

209
Solution: The sums of squares for the analysis of variance are shown individually below and
jointly in Table 7.5 Thus

(total) 2 (21.2) 2
CM    37.45
n 12

4 3 3 3
Total SS =  ( y ij  y ) 2  i   yij2  CM
i 1 j 1 i 1 j 1

 46.86  37.45  9.41 ,

B i
2

SSB  i 1
 CM  40.93  37.45  3.48
3

T
j 1
j
2

SSE   CM  42.93  37.45  5.48

4

SSE = Total SS – SSB – SST

= 9.41 – 3.48 – 5.48 = 0.45

Table 7.7 ANOVA Table

You use the ratio of mean-square treatment to mean-square error to test a hypothesis of no
difference in the expected response for treatments. Thus

MST 2.74
F   36.53
MSE 0.075

210
The critical value of the F statistic (a = .05) for v1 = 2 and v2 = 6 degree of freedom is F.05 = 5.14.
Since the computed value F exceeds the critical value, there is sufficient evidence to reject the null
hypothesis and conclude that real difference do exist among the expected responses for the three
stimuli. Table 7.7 shows the ANOVA table for RBD.

A similar test may be conducted for the null hypothesis that no difference exists in the mean
response for subjects. Rejection of this hypothesis would imply that subject-to-subject variability
does exist, and that blocking is desirable. The computed value of F based on v1 = 3 and v2 = 6
degrees of freedom is

MST 1.16
F   15.47
MSE 0.075

Since this value of F exceeds the corresponding tabulated critical value, F.05 = 4.76, we reject the
null hypothesis and conclude that a real difference exists in the expected response for the group
of subjects.

7.5.7 Estimation for the Randomized Block Design

The confidence interval for the difference between a pair of means is exactly the same as for the
completely randomized design, Section 7.5.6. It is

2
(Ti  T j )  t s
2 b

where ni = nj = b, the number of observations contained in a treatment mean, and s = MSE. The
difference between the confidence intervals for the completely randomised block design and the
randomized block designs is that s, appearing in the expression above, will tend to be smaller
than for the completely randomized design.

Similarly, one may construct a (1-  ) confidence interval for the difference between a pair of
block means.; each block contains p observations corresponding to the p treatments. Therefore,
the confidence interval is

211
 2
( Bi  B j )  t s
2 p

Example 7.6 Construct a 95 percent confidence interval for the difference between treatments 1
and 2, Example 7.5.

Solution: The confidence interval for the difference in mean response for a pair of treatments is

2
(Ti  T j )  t 0.025 S
b

where for our example t.025 is based upon six degrees of freedom. For treatments 1 and 2 we have

2
(0.98  2.63)  (2.447)(0.27)
4

 1.65  0.47

7.5.8 The Analysis of Variance for a Latin-Square Design

The method for constructing a Latin-square design for comparing p treatments is presented in
Section 7.5.5. The purpose of the design is to remove unwanted variation as might occur in the
mechanized application of icing to cakes on a conveyor belt. Variation in the thickness of icing
could occur across the belt due to the variation in pressure at the applicator nozzles. Similarly the
thickness of icing could vary somewhat along the length of the belt due to variations in the
consistency of the icing supplied to the machine. Now suppose that we wish to compare three
different types of cake mixes, A, B, and C, that result in different porosities which affect absorption
of the icing into the cakes. Then the thickness of the resulting icing, y, could be compared for the
three treatments (mixes) by employing a 3 x 3 Latin-square design. Each mix would appear in
each column (across the conveyor belt) and in each row as one proceeds down the belt. The design
configuration is shown in Figure 7.2.

Table 7.8 Latin square Design

212
Figure 7.2 A 3 x 3 Latin-Square Design

The three independent variables in a Latin-square design are “rows”, “columns,” and treatments.
All are qualitative variables, although the treatments could be levels of a single quantitative factor
or combinations of levels for two or more factors. Thus the total variation in an analysis of variance
can be partitioned into four parts, one each corresponding to the variation in rows, columns,
treatments, and experimental error. The analysis-of –variance table for a p x p Latin-square design
is shown in Table 7.6. As for previous designs, the four sums of squares add to the total sum of
squares of deviations. Table 7.9 gives an ANOVA table for a Latin square design.

Table 7.9 ANOVA Table for a Latin-Square Design

The formulas for computing the total sums of squares, SSR, SSC, and SST, are identical to the
corresponding formulas given for the randomized block design. Let yij denote an observation in
row i and column j. Then

213
 p p p p
Total SS   ( yij  y ) 2   yij2  CM
i 1 j 1 i 1 j 1

(total of all observations) 2

Where CM 
n
2
 p p 
  y ij 
 
 
i 1 j 1

n
Similarly, let Ri and Ri , C i Cj, and Tk and Ti represent the total and average for all observations
in row i, column j, and treatment k, respectively. Then the sums of squares for rows, columns,
and treatments are
p

p R i
2

SSR  P ( Ri  y ) 2  i 1
 CM
i 1 p
p

p C 2
j
SSC  P (C j  y ) 2  i 1
 CM and
i 1 p
p

p T 2
ik
SST  P (Tk  y ) 2  i 1
 CM
i 1 p
The sums of squares for error, SSE, can be obtained by subtraction. Thus
Total SS = SSR + SST + SSE
Hence
SSE = Total SS – SSR – SSC - SST

The mean squares corresponding to rows, columns, and treatments can be obtained by dividing
the respective sum of squares by (p – 1). Thus

SST
F
P 1
The hypothesis “no difference in mean response for treatments” is tested using the F statistic,

214
 SST
F
MSE

Similarly, the F statistic can be used to test a hypothesis of no difference between rows (or
columns) by using the ratio of MSR (or MSC) to MSE. We will illustrate with an example.

Example 7.7 An experiment was conducted to investigate the difference in mean time to assemble
four different electronic devices, 1, 2, 3, and 4. Two sources of unwanted variation affect the
response – the variation between people and the effect of fatigue if a person assembles a series of
the devices over time. Consequently, four assemblers were selected and each assembled all four
of the devices in the Latin-square design of Table 7. (The observed responses, in

Table 7. 10 Latin Square Design

minutes, are shown in the cells. Circled numbers in the design above indicates the treatments
employed.) Do the data provide sufficient evidence to indicate a difference in mean time to
assemble the four devices? A difference in the meantime to assemble for people? Is there evidence
of a fatigue factor (a difference in mean response for positions in the assembly sequence).

Solution: The totals for rows, columns, and treatments are as shown in Table 8

215
 Table 7.11 Latin Square Design revised

And

(736) 2 541,696
CM    33,856.0
16 16

Then

4 4
Total SS =  y
i 1 j 1
2
ij  CM

 35,186.0  33,856.0  1330.0

R i
2

SSR  i 1
 CM
4

(155) 2  (181) 2  (193) 2  (207) 2

  CM
4

 34,221.0  33,856.0  365.0

C 2
j
SSC  i 1
 CM
4
(202) 2  (161) 2  (191) 2  (182) 2
  CM
4
= 34,082.5 – 33,856.0 = 226.5,

216
 4

T j
2

SST  i 1
 CM
4
(202) 2  (142) 2  (204) 2  (188) 2
SST =  CM
4
= 34,482.0 – 33,856.0 = 626.0.
Finally,
SSE = Total SS – SSR – SSC – SST
= 1330.0 – 365.0 – 226.5 – 626.0
= 112.5.
The analysis-of variance table for this example is shown as Table 9. Note that the mean squares
were obtained by dividing the sums of squares by their, respective degrees of freedom.
Table 7.12 ANOVA table

All the computed F statistics are based on v1 = 3 and v2 = 6 degrees of freedom. The corresponding
tabulated critical value is F3,6 = 4.76 (a = .05). A comparison of the computed F statistics with the
tabulated value indicates that the computed F’s for both rows and treatments exceed the critical
value. Thus the data provide sufficient evidence to indicate a difference in the meantime to
assemble the four devices. The data also show a difference in mean time to assemble for rows or,
equivalently, positions in the sequences of assembly. It appears that fatigue, boredom, or some
other factor increases the mean time to assembly as the length of employment increases.

217
7.5.8 Estimation for the Latin-Square Design

The (1 –  ) confidence interval for the difference between a pair of treatment means is obtained
in the same manner as for the completely randomized and the randomized block designs. Since
each treatment mean will contain p observations, n1 = n2 = p and the standard deviation of the
difference between a pair of means is

1 1
 (T T )   
i j
p p

The (1 –  ) confidence interval is

2
(Ti  T j )  t S
2 p

Corresponding confidence intervals for the difference between a pair of row or column means
are respectively,

2
( Ri  R j )  t S
2 p

2
(Ci  C j )  t S
2 p

Example 7.8: Refer to Example 7.7. Estimate the difference in mean response between treatments
1 and 2 using a 95 percent confidence interval.

Solution: From Table 7.12 observe that s2 = MSE = 18.75 is based on six degrees of freedom.
Consequently, the tabulated value, ta/2, with six degrees of freedom is t.025 = 2.447. Then the 95
percent confidence interval for the difference between the treatment means, (µ1 - µ2), is

2
(Ti  T j )  t S
2 p

218
 2
(50.5  35.5)  (2.447)(4.33)
4

15 7.49

Thus we estimate the difference in mean time to assemble the two devices to be between 7.51 and
22.49 minutes.

7.5.9 Selecting the Sample Size

Selecting the sample size for the completely randomized or the randomized block design and the
Latin-square design is an extension of the procedures of Section 7.5.7. We confine our attention
to the case of equal sample sizes, n1 = n2 = … = np, for the treatments of the completely randomized
design. The number of observations per treatment is equal to b for the randomized block design
and for a b x b Latin-square design. Thus the problem is to select n1 or b for these three designs
so as to purchase a specifies quantity of information.

The selection of sample size follows a similar procedure for all three designs; we will outline a
general method. First the experimenter must decide on the parameter (or parameters) of major
interest. Usually, he will wish to compare a pair of treatment means. Second, he must specify a
bound on the error of estimation that he is willing to tolerate. Once determined, he need only select
ni (the number of observations in a treatment mean for a randomized block or Latin-square design)
that will reduce the half-width of the confidence interval for the parameter so that it is less than or
equal to the specified bound on the error of estimation. It should be emphasized that the sample-
size solution will always be an approximation, since α is unknown until the sample is acquired.
The best available value will be used for s in order to produce an approximate solution. We will
illustrate the procedure with an example.

Example 7.9: A completely randomized design is to be conducted to compare teaching techniques

in classes of equal size. Estimation of the difference in mean response on an achievement test is
desired correct to within 30 test-score points, with probability equal to .95. It is expected that the
test scores for a given teaching technique will possess a range approximately equal to 240. Find

219
the approximate member of observations required for each sample in order to acquire the specified
information.

Solution: The confidence interval for the difference between a pair of treatment means is

1 1
(Ti  T j )  t S 
2 ni n j

Therefore, we will wish to select ni and nj so that

1 1
t s   30
2 ni n j

value of  is unknown and s is a random variable. However, an approximate solution for ni = nj

can be obtained by guessing s to be roughly equal to one-fourth of the range. Thus s = 240/4 =
60. The value of ta/2 will be based upon (n1 + n2 + … + n5 – 5) degrees of freedom, and for even
moderate values of ni, t.025 will approximately equal 2. Then

1 1 2
t 0.025 s   (2)(60)  30
ni n j ni

ni  32 , i= 1, 2,…,5.

Example 7.10: An experiment is to be conducted to compare the toxic effect of three chemicals
on the skin of rats. The resistance to the chemicals was expected to vary substantially from rat to
rat. Therefore, all three chemicals were to be tested on each rat, thereby blocking out rat-to-rat
variability.

The standard deviation of the experimental error was unknown, but prior experimentation
involving several applications of a given chemical on the same rat suggested a range of response
measurements equal to 5 units.

220
Find a value for b such that the error of estimating the difference between a pair of treatment means
is less than one unit, with probability equal to .95.

Solution: A very approximate value for s would be one-fourth the range, or s = 1.25. Then we
wish to select b so that

1 1 2
t 0.025 s   t 0.025 s 1
b b b

Since t 0.025 will depend upon the degrees of freedom associated with s 2 , which will be (n – b – p

+ 1), we will guess t 0.025  2 . Then

2
(2)(1.25) 1
b

Or b  13, i=1,2,3.

Approximately 13 rats will be required to obtain the desired information.

The degrees of freedom associated with s2 will be 24, based on this solution. Therefore, the guessed
value of t would seem to be adequate for this approximate solution.

The sample-size solutions for Examples 7.9 and 7.10 are very approximate, and are intended to
provide only a rough estimate of approximate size and consequent cost of the experiment. The
experimenter will obtain information on α as the data are being collected and can recalculate a
better approximation to n as he proceeds.

Selecting the sample size for Latin-square designs follows essentially the same procedure as for
the completely randomized and the randomized block designs. The only difference is that one must
decide on the number of p x p Latin squares that will be needed to acquire the desired information.
We omit discussion of this topic because we have not shown how to conduct an analysis of
variance for more than one p x p Latin square. The reader interested in this topic should consult
the references at the end of the unit.

221
7.0 UNIT ACTIVITY

1. State the assumptions underlying the analysis of variance of a completely randomized

design.

2. Refer to Example 7.2. Calculate SSE by pooling the sums of squares of deviations within
each of the four samples, and compare with the value obtained by subtraction. Note that
this is an extension of the pooling procedure used in the two-sample case discussed in
Section 7.5.2.

3 To compare the strengths of concrete produced by four experimental mixes, three

specimens were prepared from each type of mix. Each of the 12 specimens was subjected
to increasing compressive loads until breakdown. The following compressive loads in
tons per square inch were attained at breakdown. Specimen numbers 1-12 are indicated in
parentheses for identification purposes.

Table 7.13 Completely Randomized Design

Mix A Mix B Mix C Mix D

(1) 2.30 (2) 2.20 (3) 2.15 (4) 2.25

(5) 2.20 (6) 2.10 (7) 2.15 (8) 2.15

(9) 2.25 (10) 2.20 (11) 2.20 (12) 2.25

Assuming that the requirements for a completely randomized design are met, analyze the
data. State whether there is statistical support at the α = .05 level of significance for the
conclusion that the four types of concrete differ in average strength.

4 A clinical psychologist wished to compare three methods for reducing hostility levels in
university students. A certain psychological test (HLT) was used to measure the degree
of hostility. High scores on this test were taken to indicate great hostility. Eleven students
obtaining high and nearly equal scores were used in the experiment. Five were selected at
random from among the 11 problem cases and treated by method A. three were taken at

222
 random from the remaining 6 students and treated by method B. the other 3 students were
treated by method C. All treatments continued throughout a semester. Each student was
given the HLT test again at the end of the semester, with the following results:

Table 7.14 Records of hostility levels

Method A Method B Method C

73 54 79
83 74 95
76 71 87
68
80
(a) Perform an analysis of variance for this experiment.
(b) Do the data provide sufficient evidence to indicate a difference in mean student
response for the three methods after treatment?

5. A study was initiated to investigate the effect of two drugs, administered simultaneously,
in reducing human blood pressure. It was decided to utilize three levels of each drug and
to include all nine combinations in the experiment. Nine high-blood-pressure patients were
selected for the experiment, and one was randomly assigned to each of the drug
combinations. The response observed was a drop in blood pressure over a fixed interval of
time.

(a) Is this a randomized block design?

(b) Suppose that two patients were assigned to each of the nine drug combinations.
What type of experimental design is this?

6. A dealer has in stock three cars (car A, car B, and car C) of the same make and model.
Wishing to compare these cars in gas consumption, a customer arranged to test each car
with each of three brands of gasoline (brand A, brand B, brand C). In each trial, a gallon
of gasoline was added to an empty tank and the car was driven without stopping until it ran

223
 out of gasoline. The following table shows the number of miles covered in each of the nine
trials.

7.15 Distance (miles) Brand of

Gasoline Car A Car B Car C

A 22.4 17.0 19.2
B 20.8 19.4 20.2
C 21.5 18.7 21.2
(a) Should the customer conclude that the three cars differ in gas mileage? Test at the
α = .05 level.
(b) Do the data indicate that the brand of gasoline affects gas mileage?
7. A portion of a questionnaire was constructed to enable judges to evaluate a certain aspect
of observed classroom teaching. Four films portraying teaching performances, and
differing markedly in the teaching characteristic under study, were viewed by each of eight
judges. The order of viewing the four films was assigned in a random manner to each judge.
The data obtained are as follows:
Judges.
Table 7.16 Evaluation of classroom teaching
Films 1 2 3 4 5 6 7 8
1 9 10 7 5 12 7 8 6
2 4 9 3 0 6 8 2 4
3 12 16 10 9 11 10 10 14
4 9 11 7 8 12 7 7 8
(a) Give the type of design employed for this experiment and justify your diagnosis.
(b) How many degrees of freedom are available for estimating α2? Perform an
analysis of variance on the data.
(c) Do the data provide sufficient evidence to indicate that the mean questionnaire
score varies from film to film? Test using α = .05.
(d) Suppose that the data did provide sufficient evidence to indicate differences
among the mean questionnaire scores for the four films. Would this imply that the

224
 questionnaire was able to detect a difference in the teaching characteristic
exhibited in the four films?
8. A completely randomized design was conducted to compare the effect of five stimuli on
reaction time. Twenty-seven people were employed in the experiment, which was
conducted using a completely randomized design. Regardless of the results of the analysis
of variance, it is desired to compare stimuli A and D. the reaction times (in seconds) were
as follows:
Table 7.17 Effect of five stimuli on reaction time
Stimulus
A B C D E
.8 .7 1.2 1.0 .6
.6 .8 1.0 .9 .4
.6 .5 .9 .9 .4
.5 .5 1.2 1.1 .7
.6 1.3 .7 .3
.9 .8
.7
Total 2.5 4.7 6.4 4.6 2.4
Mean .625 .643 1.067 .920 .48
(a) Conduct an analysis of variance and test for a difference in mean reaction time
due to the five stimuli.
(b) Compare stimuli A and D to see if there is a difference in mean reaction time.

9. An experiment was conducted to determine the effect of three methods of soil preparation
on the first-year growth of slash pine seedlings. Four locations (state forest lands) were
selected and each location was divided into three plots. Since it was felt that soil fertility
within a location was more homogeneous than between locations, a randomized block
design was employed using locations as blocks. The methods of soil preparation were A
(no preparation), B (light fertilization), and C (burning). Each soil preparation was
randomly applied to a plot within each location. On each plot the same number of

225
 seedlings were planted, and the observation recorded was the average first-year growth
(in centimeters) of the seedlings on each plot.

(a) Conduct an analysis of variance. Do the data provide sufficient evidence to indicate
a difference in the mean growth for the three soil preparations?

(b) Is there evidence to indicate a difference in mean growth for the four locations?

(c) Use a 90 percent confidence interval to estimate the difference in mean growth for
methods A and B.

Table 7.19 Effect of three methods of soil preparation

Location

Soil
Preparation 1 2 3 4
A 11 13 16 10
B 15 17 20 12
C 10 15 13 10
10. Give the analysis of variance for the following 3 x 3 Latin-square design.

Table 7.20 Latin Square Design

Columns

Rows 1 2 3
1 B A C
12 7 17

2 C B A
10 7 4
3 A C B
2 8 12
Find a 95 percent confidence interval for the difference in mean response for
treatments A and C assuming that this is a preplanned comparison.

226
11. The following measurements are of the thickness of cake icing for the Latin-square design
discussed in Section 7.5.2. The only difference between this exercise and the text discussion is
that five (not three) mixes, A, B, C, D, and E, were employed in a 5 x 5 Latin-square design.
Thickness measurements are given in hundredths of an inch.
Table 21 The thickness of cake icing
Columns

Rows 1 2 3 4 5
1 D E C A B
9 18 6 8 11

2 B D E C A
12 17 10 4 5
3 A B D E C
6 16 10 9 4
4 C A B D E
4 13 11 8 13
5 E C A B D
14 11 7 10 15
Do the data provide sufficient evidence to indicate a difference in mean thickness of icing for the
cake mixtures? Estimate the difference in mean thickness for mixtures A and B.
12. An experiment was conducted to investigate the toxic effect of three chemicals, A, B, and
C, on the skin of rats. One-inch squares of skin were treated with the chemicals and then
scored from 0 to 10, depending on the degree of irritation. Three adjacent 1-inch squares
were marked on the backs of eight rats, and each of the three chemicals was applied to
each rat. The experiment was blocked on rats to eliminate the variation in skin sensitivity
from rat to rat. The data are as follows:
Table 7.22 Investigate the toxic effect of three chemicals, A, B, and C
1 2 3 4 5 6 7 8
B A A C B C C B
5 9 6 6 8 5 5 7

227
 A C B B C A B A
6 4 9 8 8 5 7 6
C B C A A B A C
3 9 3 5 7 7 6 7
(a) Do the data provide sufficient evidence to indicate a difference in the toxic
effect of the three chemicals?
(b) Estimate the difference in mean score for chemicals A and B using a 95
percent confidence interval.
13. Consider the following one-way classification consisting of three treatments, A, B, and C,
where the number of observations per treatment varies from treatment to treatment.
Table 7.23 classification of three treatments, A, B, and C,
A B C
24.2 24.5 26.0
27.5 22.7
25.9
24.7
Do the data present sufficient evidence to indicate a difference between treatments?

7.0 UNIT SUMMARY

The completely randomized, the randomized block, and the Latin-square designs are illustrations
of experiments involving one, two and three qualitative independent variables, respectively. The
analysis of variance partitions the total sum of squares of deviations of the response measurements
about their mean into portions associated with each independent variable and the experimental
error. The former may be compared with the sum of squares for error, using mean squares and the
F statistics, to see whether the mean squares for the independent variable are unusually large and
thereby indicative of an effect on the response.

228
 In this unit you have presented a very brief introduction to the analysis of variance and its
associated subject, the design of experiments. Experiments can be designed to investigate the
effect of many quantitative and qualitative variables on a response. These may be variables of
primary interest to the experimenter as well as nuisance variables, such as blocks, which we
attempt to separate from the experimental error. These experiments are subject to an analysis of
variance when properly designed. A more extensive coverage of the basic concepts of
experimental design and the analysis of experiments will be found in the references.

UNIT 8 LINEAR REGRESSION

8.1 Unit Introduction

Welcome to Unit 8 in which you will learn methods of linear regression analysis. In this unit you
will learn methods Statistical Data Analysis technique. You will further learn that linear
regression is used to determine the extent to which there is a linear relationship between a
dependent variable and one or more independent variables. You will also learn that you can use
this analysis to make predictions of future values.

8.2 Unit Aim

The aim of this Unit is to teach you how to conduct linear regression analysis.

8.3 Unit Objectives

By the end of the unit you should be able to:

 Estimate the relationship between a set of independent variables (regressors) and some
dependent variable (outcome).

 Predict" the value of the dependent variable based upon the values of one or more
independent variables.

 Predict a continuous dependent variable from a number of independent variables. If the

dependent variable is dichotomous, then logistic regression should be used

229
Terminology

â : Intercept estimate

b̂ : Slope estimates

8.4 Unit Time required

You need 20 hours for this unit

8.5 Unit Topics

8.5.1 Analysing the results of an experiment

Figure 8.1 illustrates the apparatus used in a simple experiment which may people will have
performed for themselves. When weights are added to the scale pan, the spring stretches. Table
8.1 shows the result obtained when different loads were applied in a random order, and the
corresponding scatter diagram is shown in Figure 2. In this experiment the values of Y, the
length of the spring in cm, were measured for preselected values of X, the load in newtons, and
for this reason X is called the independent variable and Y the dependent variable.
Figure 8.2 suggests that there is a linear relationship between X and Y: how should we draw a
straight line to represent this relationship? Drawing a line by eye is obviously not satisfactory,
since the result will be subjective, but as a method it can lead us to a mathematical method.
Figure 3 shows schematically the points on a scatter diagram and a possible straight line which
relates them. In this experiment the values of X are extremely accurately known. The
measurement of the length of the spring is less accurate. If the experiment were repeated (using
the same loads) the values of Y for a given value of X would show random variation. It is this
variation which we may reasonably assume causes the points on the scatter diagram to deviate
from a straight line. For example, in Figure 3 (xi, y i ) deviates by e1 and (x2, y2) by e2 (in the
opposite direction) from the straight line which has been drawn. We can represent the
relationship between any pair of values (xi y i ) by

y i    xi  ei (1)

230
Where   xi represents the linear relationship between y i and xi and ei is a random error.  is
called the coefficient of regression of Y on X. We shall assume that the ei’s are independent of xi
and normally distributed with mean 0 and s.d.  Y
X

Table 8.1 Results for extension of a spring

231
 Figure 8.1. Scatter diagram for the data in Table 8.1

Figure 8.2. Schematic diagram showing the deviation of the measurements from a possible
straight line relating X and Y.

232
Figure 8.3. Diagram showing the distribution of repeated values of Y for a given value of X
Distribution of repeated measurements of y i for each xi and the appropriate line relating xi and y i

which Passes through the mean values of the y i distributions.

8.5.2 The method of least squares

The parameters  and  are estimated by the method of least squares, so called because
n
estimates, â and b̂ , of  and  respectively are chosen so as to minimise e
i 1
2
i (n is the

number of points on the scatter diagram.) The squares of the deviates are used for the same
reasons as the squares of the deviations are used in the calculation of standard deviation.
n
Rearranging equation (1) ei  y i    xi e
i 1
2
i   ( y i     xi ) 2 .

This expression can be varied by varying  and  , â and b̂ , the estimates of  and  , are

chosen so that it is minimised. The approximate values of â and b̂ are found by putting the

partial derivatives of (y i  aˆ  bˆxi ) 2 with respect to â and to b̂ equal to zero:

233
  n n

 i
aˆ i 1
( y  ˆ
a  b i 
ˆx ) 2   2( y  aˆ  bˆx )  0
i 1
i i

 n n

 i
bˆ i 1
( y  ˆ
a  b i 
ˆx ) 2   2 x ( y  aˆ  bˆx )  0
i 1
i i i

This gives two simultaneous equation in â and b̂ . These equations can be rewritten as follows:
n

  2( y
i
i  aˆ  bˆxi )  0

Gives
n n

y
i 1
i  naˆ  bˆ xi  0
i 1
(2)

and
n

  2x ( y
i 1
i i  aˆ  bˆxi )  0

gives
n n n

 xi yi  aˆ  xi  bˆ xi2  0
i 1 i 1 i 1
(3)

(2) and (3) are known as the normal equations. They can be solved simultaneously.
n
Subtracting (3) multiplied by n from (2) multiplied by x
i 1
i gives:

2
n n
 n  n n


i 1
xi  y i  bˆ  xi   n xi y i  nbˆ xi2  0
i 1  i 1  i 1 i 1

Rearranging
n n n
n xi y i   xi  y i
bˆ  i 1 i 1 i 1
2
(4)
 n
 n
n x    xi  2
i
i 1  i 1 
From (2) a, the estimated value of a, is given by
 n n

aˆ    yi  bˆ xi  / n
 i 1 i 1 

234
giving aˆ  y  bˆx
The relationship between x and y is given by

y  aˆ  bˆxi (5)
and this is known as the regression line of Y on X, An alternative form is given by substituting
the expression (5) for a in equation (5): y  y  bˆx  bˆx .

Giving y  y  bˆ( x  x ) (6)

From (6) you can see that the regression line passes through (x, y), a fact which is useful when
drawing it on a scatter diagram.

Table 8.2 Strength of the string

======================
xi yi xi y i x i2
--------------------------------------
0.1 10.7 1.07 0.01
0.2 11.3 2.26 0.04
0.3 12.0 3.60 0.09
0.4 12.4 4.96 0.16
0.5 13.0 6.50 0.25
0.6 13.7 8.22 0.36
0.7 14.5 10.15 0.49
0.8 15.1 12.15 0.64
0.9 15.6 14.04 0.81
1.0 16.0 14.04 1.00
5.5 134.3 78.88 3.85

235
Example 8.1
Calculate the regression line of Y on X for the data given in Table 8.1 (reproduced in Table 8.2),
and use it to predict the length for a load of 0.65 N. n = 10
n

x i
5 .5
x i 1
  0.55
n 10
n

y i
134.3
y i 1
  13.43
n 10
n n n
n xi y i   xi  y i
bˆ  i 1 i 1 i 1
2

n
 n
n x    xi 2
i
i 1  i 1 
10 * 78.88  5.5 * 134.3
bˆ   6.08
10 * 3.85  5.5 2
The regression line of Y on X is
y  y  bˆ( x  x ) (6)
y  13.43  6.08( x  0.55)
y  6.08x  10.1
To predict the length for a load of 0.65 N we substitute this value of x into the equation of the
regression line.
y  6.08 * 0.65  10.1
 14.1

8.5.3 Coded values

The calculation in the previous section can be greatly simplified using coded values as in Section
7.5.2. With xi  A  Bu i y i  C  Dvi

236
  n n n 
 n  u i vi   u i  vi 
D 
bˆ   i 1 n i 1 i 1
 (7)
B  n
 
 
n ui    ui 
2

i 1  i 1  
n
(Calculation may often be further simplified by choosing a value of A which makes u
i 1
i 0

Example 8.2
Calculate the regression line of Y on X for the data in Table 1 using coded values.
Table 8.3 Strength of the string

Solution
Taking A = 0.6, C = 13.0, B = 0.1 , D = 1 you have table 8.3. Using equation (7)

237
 1 10 * 48.0  (5) * 4.3
bˆ  *  6.08
0.1 10 * 85  (5) 2
x  A  Bu
 5
x  0.6  0.1 *  
 10 
 0.55
y  C  Dv

4.3
y  13  1 *  13.43
10
And the regression line is:
y  6.08x  10.1

8.1 UNIT ACTIVITY

(1) A scientist, working in an agricultural research, believes there is a relationship between

the hardness of the shells of eggs laid by chickens and the amount of a certain food
supplement put into the diet of the chickens. He selects ten chickens of the same breed
and collects the data of Table 4.
(Hardness is measured on a 0 – 10 scale, 10 being the hardest. There are no units
attached).
(a) Calculate the equation of the regression line of y and x.
(b) Calculate the product moment correlation coefficient.
(c) Do you b8elieve that his linear model will continue to be appropriate no matter how
large or small x becomes?
Just your reply
Table 8.4 Measurements of hardness of egg shells
========================================================
Chicken A B C D F G H I J
Amount of food
supplement x(g) 7.0 9.8 11.6 17.5 7.6 8.2 12.4 15.5 9.5 19.5

238
Hardness of shells 1.2 2.1 3.4 6.1 1.3 1.7 3.4 6.2 2.1 7.1
Table 8.5 Consumptions of petrol in Lusaka
===============================================================
Year (x) 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972
Consumption (y)
(millions of gallons)32.5 37.1 35.5 37.7 41.5 46.4 44.8 45.8 53.9 62.0

===============================================================
(2) The figures in Table 8.5 give the wine consumption in Lusaka in millions of gallons (y)
for the years 1963 to 1972 (x).
Draw a scatter diagram to show these data.
Determine the last squares estimate of the regression line of y and x, showing all your
working. Draw this line on your scatter diagram and use it to estimate the consumption
for 1973.
Comment on the appropriateness of a linear regression model in this case, given also that
the actual wine consumption in 1973 was 78.3 million gallons.
(3) The body and heart masses of fourteen ten-month-old male mice are given in Table 8.6.
(a) Draw a scatter diagram of these data.
(b) Calculate the equation of the regression line of y and x and draw this line on
the scatter diagram.
(c) Calculate the product moment coefficient of correction
(4) A regression line, y = a + bx, is to be fitted to a set of data points (x,y). The data are
coded to X = (x – c1)d1, Y = (y – x2)/d2, and in terms of these the regression line is Y
= a T BX. Find a and b in term s of A and B and the coding constants.
In such a problem the values of the independent variable are equally paced. Explain,
with reference to the appropriate least squares equations, how this simplifies the fitting
of the line.
Table 8.6 Measurements of body mass
===============================================================
Body mass (x)

239
(grams)27 30 37 38 32 36 32 32 38 42 36 44 33 38
Heart mass(y) (milligrams)
118 136 156 150 140 155 157 114 144 156 149 170 131 160
===============================================================
Table 8.7 shows the mass of a certain animal at weekly intervals.
=============================================================
Age (weeks) 11 12 13 14 15 16
Mass (g) 357 382 404 423 440 451
=============================================================
State whether it would be more appropriate to fit a regression line of age on mass or mass on
age to these data, and justify your choice. What value does the line of your choice give for the
mass at 17 Weeks? Comment on the validity of your estimate. (You are advised to plot rough
graph.)

8.5.4 Estimating a value of  Y2

X

For a value of xi the value of Y predicated by the regression line is given by y i  aˆ  bˆxi

The difference between y i and y i known as a residual, gives the values of ei for that point.

Each value of y i  y i gives a value of ei from a distribution which is N (0,  Y2 X ) . The unbiased

estimate ŝY of  Y is given by

X X

 n 
  ( y i  y i ) 2 
sˆY   i 1  (8)
X  n2 
 
 
The divisor is (n – 2) since two degrees of freedom are lost because there are two constraints:

the vlues of â and b̂ are calculated from the values of X and Y.

Example 8.3
Obtain an estimate of  Y2 / X for the data in Table 1.

240
  n 
  ( y i  y i ) 2 
sˆY   i 1 
X  n2 
 
 
0.11784
sˆY   0.121
X 8
In section 8.5.2 the regression line of Y and X was calculated as y = 6.08x + 10.1. Using this
equation, the predicted values for Y are given in Table 8.8.
From equation (8)
Table 8.8 Strength of string

8.5.5 Confidence limits for 

How accurate is the estimate, b, which we have made of  ? This section will show how the

confidence limits for  can be found in terms of  Y the s.d. of the random errors, e y . To
X

recapitulate, we assume the ei ’s are randomly distributed with a Normal distribution mean 0.s.d.

Y thus by definition, E (ei )  0 , (9)

X

var( ei ) =  2 Y X (10)

241
We will now calculate the expected value and variance of b̂ in terms of the ei ’s
From equation
n n n
n xi y i   xi  y i
bˆ  i 1 i 1 i 1
2
(11)
n
  n
n xi2    xi 
i 1  i 1 
Substituting yi  x    ei (equation (1) the numerator becomes
n n n
n xi (  xi    ei )   xi  ( xi    ei )
i 1 i 1 i 1

2
n
 n  n n n n n
 n  x  n  xi  n xi ei     xi   n  xi   xi  ei
2
i
i 1 i 1 i 1  i 1  i 1 i 1 i 1

 n  n  
2 n n
  n xi2    xi    n xi ei  n x ei
 i 1  i 1   i 1 i 1

 n 2  n 2  n
  n xi    xi    n ( xi  x )ei
 i 1  i 1   i 1

2
 n  n
Dividing by the denominator n x    xi  in the expression for b̂ , we have 2
i
i 1  i 1 
n
n ( xi  x )ei
bˆ    i 1
2
n
 n 
n x    xi  2
i
i 1  i 1 
n
n  ( x i  x ) ei
 i 1
n
n ( xi  x ) 2
i 1
n

 (x i  x ) ei
 i 1
n

 (x
i 1
i  x)2

242
In calculating the expected value and variance of b̂ , the term in x i is constant for each ei since

the values of x i are predetermined.

This gives
 n

  ( x i  x ) ei 
E (bˆ)  E    i n1 
 2 
 i 1
( xi  x ) 

n

 (x i  x ) E ( ei )
 i 1
n

 (x
i 1
i  x)2


Since E (ei )  0

Thus b̂ gives an unbiased estimate of  . Also

 n 
  ( x i  x ) ei 
Var( b̂ )=var  i 1 n 
i  1 (x  x)2 
 
i 1
i 

Since  is constant

Using the relationship var(cy)= c 2 var( y ) where c is a constant

n

 (x i  x ) 2 var( ei )
var( bˆ)  i 1
2
n 2
 ( x i  x ) 
 i 1 
n

 (x i  x ) 2  Y2
X
var( bˆ)  i 1
2
n 2
 ( x i  x ) 
 i 1 

243
  Y2
var( bˆ)  X
(12)
 n
2
 ( xi  x ) 
 i 1 

Assuming the sampling distribution of b̂ is Normal, the 95% confidence limits of  are given
by
1.96 Y
  bˆ  X
(13)
 
 ( xi  x ) 
2

 
If  Y X is estimated from the samples as then the 95% confidence limits of  are

t n  2,5% sˆY
  bˆ  X
(14)
 
 ( xi  x ) 
2

 
Where the t-distribution is used because the variance is estimated from the sample with n – 2
degrees of freedom.
Example 8.4
Calculate the 95% confidence limits for the estimate of  for the data in Table 8.1.

In equation (12) we cannot substitute  Y X but only the estimate of it, since ŝY , which are
X

n
found in Section 8.5.4 to be 0.2.1  (x
i 1
i  x ) 2 . Can be found most simply using the identity

2
n n
 n 

i 1
( x i  x ) 2
 
i 1
x 2
i    xi  / n
 i 1 
5.5 2
 3.85   0.825 (Using the values obtained in Example 8.2)
10

Which gives our estimate of the variance of b̂ as

sˆY 0.121
var( bˆ)  X

 n 2  n  2  0.825
 xi    xi  / n 
 i 1  i 1  

244
The 95% confidence limits of  are, from equation (14),


  bˆ  t n 2,5% var( bˆ) 
 6.08  2.31* 0.133
 6.08  0.31

8.5.6 Confidence limits for 

From equation (5) aˆ  y  bˆx

1 n
  yi  bˆx
n i 1
Thus
1 n
E (aˆ )   E( yi )  E(bˆx ) Since
n i 1
y i    xi  ei

E ( yi )    xi  E (ei )

    xi

Giving
1 n
E (aˆ )   (  xi )  xE(bˆ)
n i 1
   x  x

Showing that â is an unbiased estimate of  . Also
n 
var( aˆ )  var  yi / n  bˆx 
 i 1 
It can be shown that the covariance of y and b̂ is zero so that

245
  n 
  yi 
var( aˆ )  var  i 1   var( bˆx )
 n 
 
 
n
1

n2
 var( y )  x
i 1
i
2
var( bˆ)

Since y i    xi  ei

Var( y i ) = var( ei )   Y2 and equation(12) gives a value for var(bˆ) substituting, you have
X

1 x 2 Y2
var( aˆ )  2 * n Y2  n
X

n
 (x
X
i  x)2
i 1

 
1 x 2 
var( aˆ )   Y2   n  (15)
X n 2 
 
i 1
( xi  x ) 

Example 8.5
Find the s,d, and 95% confidence limits for  for the data in table 8.1
Solution
You do not have an exact value for  Y , but only its estimate, found in section 8.5.4 to be
X

n
0.121. The summation  (xi 1
i  x ) 2 was found in example 8.5 to be 0.825 and x = 0.55

substituting these values in (equation 15) gives

 1 0.552 
var( aˆ )  0.121  2

10 0.825

var( aˆ )  0.0827
The 95% confidence limits of  are;
  aˆ  t n 2,5% var( aˆ )
Where the t distribution is use because var( â ) is estimated from the sample. Substituting

246
  10.1  2.31* 0.0827
 10.1  0.2

8.5.7 Confidence limits of predicted values

In example 8.5 the regression line was used to predict the true value of Y, y o , for a value of X,

x o . To do this you can use equation (6). yo  y  bˆ( xo  x )

Which gives yo  bˆ( xo  x )  y

Since the covariance of y and b̂ is zero, gives

Var( y o ) = ( xo  x ) 2 var( bˆ)  var( y )

From equation ( 15)
 Y2
Var( bˆ)  n
X

 (x
i 1
i  x)2

And
 Y2
Var( y )  X

n
Giving
 
1 (x  x) 2 
var( y o )   Y2   n o  (16)
X n 2 
 
i 1
( xi  x ) 

Returning to example in section 8.5.2, the regression line of Y on X was
y  6.08x  10.1

When x = 0.65, y  6.08 * 0.65  10.1  14.10

In this example 8.5 you found sˆY  0.121 and in example 8.5 using equation (16) gives your
X

estimate of the variance of y o as

247
  1 (0.65  0.55) 2 
var( y o )  0.121   2

n 0.825 
s.d ( y o )  0.0405

And 95% confidence limits of y o are y o  14.10  2.31* 0.0405  14.10  0.09

Whereas before t n  2 has been used since the s.d. was estimated from the sample.

Study of equation (16) shows that the confidence limits for y o will have the smallest range when

xo  x , and the range increases towards the ends of the regression line.

Example 8.6
It is known that the true response Y in a certain chemical experiment is a linear function of the
operating temperature X. However, the experimental determinations of Y are subject
Table 8.9 Measurements of temperature
Temperature(X) 30 40 50
Observed response(Y) 14 10 7
12 11 6
To random errors, so that when an experiment is performed at temperature x i the observe

response y i in such that y i    xi  ei

Where   xi is the true response and ei is the error. Table 8.9 gives the observed responses in
six experiments two at each of the three temperature.
Use the data to obtain the least estimate of the linear relationship connecting X and Y (You are
n
given that x y
i 1
i i  2270)

The errors ei are independent and normally distributed with zero mean and unit standard
deviation. Calculated 90% confidence limits for
(a) The value of  (b) the value of  (c) the true value of Y, when X is 50.
Solution
Using the six pairs of values in the table you have
n n n n

 xi yi  2270 ,
i 1
 xi2  10000,
i 1
 xi  240 ,
i 1
y
i 1
i  60

248
From equation (4)
n n n
n xi y i   xi  y i
bˆ  i 1 i 1 i 1
2

n
 n
n x    xi  2
i
i 1  i 1 
6 * 2270  240 * 60

6 *10000  2402
 0.325
From equation (6)
aˆ  y  bˆx
60 240
  (0.325) *
6 6
 23
From equation (6),
 Y2
var( bˆ)  n
X

 (x
i 1
i  x)2

 Y2
var( bˆ)  X

 n
 
2
 n

 xi    xi  / n
2

 i 1  i 1  

You are given  Y2 = 1 so

X

12 1
var( bˆ)  
 240  2
400
10000  
 6 

s.d .(bˆ)  0.05

From equation (15)
 
1 x 2 
var( aˆ )   Y2   n 
X n 2 
 
i 1
( xi  x ) 


249
  1 40 2 
1  
2

 6 400
s.d. ( aˆ )  2.04
(a) Using the value of â and its variance obtained above the 90% confidence limits of  are
  23  1.64 * 2.04  23  3.35
(Where z is used since  Y is known).
X

i. Using the value of b̂ and its variance obtained above , the 90% confidence limits of  are
  0.325  1.64 * 0.05  0.325  0.082
ii. Using the regression line of Y on X, the value of Y, y o , when xo  50 is given by

 
1 (x  x) 2 
var( y o )   Y2   n o 
X n 2 
 
i 1
( xi  x ) 

 1 (50  40) 2 
 12   
6 400 

var( y o )  0.645
The 90% confidence limits of the predicted value are
y o  6.75  1.64 * 0.645

 6.8  1.1

8.2 UNIT ACTIVITY

1. Two variable X and Y of interest in an experiment are known to be linearly related, but the
coefficients in this relationship are not known. A series of 15 experiments was conducted in
which three determinations of y were made for each of the five values of x. The x- values are
accurate but the determinations of the y- values are subject to independent random errors that are
normally distributed with mean zero and standard deviation 1.25.The observed values in the

250
series of experiments are shown in table 8.10. The following equations were calculated from the
table

Table 8.10 Two variable X and Y

 x  45 ,  y  180,  x 2
 165 ,  xy  420
(a) Calculate the equation of the least square estimate of the linear relationship between x and y.

(b) Estimate the true value of y when x – 4. Determine the standard error of this estimate.
Explain why this estimate of y when x = 4 is preferable to that obtained from averaging the
three observed values of y when x = 4.

(c)Calculate a 90% confidence interval for the true value of y when x = 4.

2. In an experiment to find the young modulus for a brass wire the eleven pairs of values x

Table 8.11 Measurements of young modulus for a brass wire

X 1 1.5 2 2.5 3 3.5 3 2.5 2 1.5

y -1.1 -0.6 0 0.4 0.9 1.5 1.0 0.6 0.1 -0.5

(suspended mass kg) and y (length of wire in mm-7000mm) given in table 11 are obtained.

The equation connecting x and y is assumed to take the form y = c + kx

Obtain the least squares estimate values for c and k and 95% confidence limits for k.

x 2
 57.25 ,  y 2  7.22 ,  xy  10.0

251
3. The size (z) of an organism was measured at different times 9x giving the data shown in the
first two rows of table 8.12. The third row of the table gives the values of y  log10
z
to two
decimal places.

Table 8.12 The size of an organism

Time(x) 1 3 5 7 9

Size(z) 1.48 3.02 5.37 9.55 17.38

y  log10
z 0.17 0.48 0.73 0.98 1.24

Without carrying out any nontrivial calculations show that the assumption of a linear relationship
between y and x is more realistic than between z and x.

The following quantities were calculated from the data:

 x  25 ,  x 2
 165 ,  y  3.6 ,  xy  23.28
Suppose that y    x and that for given values of x the observed values of y are subject to
independent errors which are normally distributed with mean zero and standard deviation 0.02.

(a) Calculate the least squares estimates of  and 

(b) Obtain 95% confidence limits for the value of y when x = 10. Hence find 95%
confidence limits for the size of the organism at time x = 10.

4. In an investigation of the effect of duration of training (x) on performance time (y) for a
certain repetitive job, the following observations were obtained from 26 trainees;

 x  104 ,  y  208 ,  ( x  x ) 2
 56 , ( x  x )( y  y )  56 ,  ( y  y) 2
 62

Plot the linear relationship of y to x which best fits the observation. Calculate 95% confidence
limits for the value predicted for y from this relationship when x = 6. Sketch the form of these
confidence limits for varying values of x.

252
 Calculate the value predicted for y when x = 12: What does this value suggest to you concerning
the form of the relationship between y and x?

8.0 UNIT SUMMARY

Regression line of Y on X, y    x

Model: y i    xi  ei where ei is N (0,  Y2 )

X

n n n
n  xi y i   xi  y i
bˆ(least square estimate of  )  i 1 i 1 i 1
2
n
  n
n xi2    xi 
i 1  i 1 

aˆ (least squaresestimateof a)  y  bˆx

 n 
  ( y i  y i ) 2 
sˆY (unbiased estimate of  Y )   i 1 
X X  n2 
 
 

Where yi   yi (the residuals) = aˆ  bˆxi  yi

9.0 MODULE SUMMARY

The module MAT 2602 is an introduction to statistics but this will give you a very good start in the field of
statistics. As defined in unit 1, were statistics is a science that deals with the methods of data: Collections,
Compilation, Presentation, analysis, and Interpretation of results, Conclusion and above all making
decisions based on the study of the data. What you will learning in this module will be up to
conclusion about a particular problem. Unit 1 gives you basic ideas about data in general in which
you will learn that data can be discrete or continuous and how you can calculate the statistics. In
this unit you will learn how to present different types of data in a pictorial why to make non-
mathematician audience understand what the data all about. In unit 2 you will learn sampling
procedures and how to estimate parameters from samples. In this unit you will learn how to estimate
into the population and make a distinction between biased and unbiased estimates which will lead

253
you to: Confidence interval and significance test in unit 3 and 4 respectively. Unit 5 will give you
methods of test of goodness which is similar to the previous unit. Unit 6 will give you procedures
of how to conduct a correlation analysis in which you will learn how to calculate the correlation
coefficient between two variables, say X and Y. You will learn that a relationship can either be
shown as a scatter diagram or calculate the correlation coefficient. In the previous units under tests
of hypotheses you learnt how to conduct tests between a value or two values. In unit 7 under the
analysis of variance you will learn how to find differences between means of more than two
variables. Lastly, in unit 8 you will learn how to conduct a linear regression analysis. In this unit
you will learn how to estimate a least regression line and how to predict the future. This is a basis
for advanced statistical analysis.

10.0 SYLLABUS: MAT 2602 - INTRODUCTION TO STATISTICS

Pre-requisites: MAT 1100-Foundation Mathematics

1. Introduction
What is statistics, elements of statistics; Sampling Methods
(Data collection); Data representation: graphical techniques, frequency distribution.
2. Estimation and sampling distributions
Point estimates for mean, difference of means, proportions, difference of proportions and their
sampling distributions; Confidence intervals for mean, difference of means, proportion,
difference of proportions, variance and ratio of variances.
3. Statistical hypothesis testing
Definition of statistical hypothesis, type I and type II errors; Tests of hypothesis about the mean,
proportion, the difference of means and proportions for large and small samples; Tests of
hypothesis about variance and ratio of variances; Tests of independence and goodness of fit
test.
4. Analysis of variance
Introduction to linear model and experimental design; Completely randomized design (CRD),
balanced and unbalanced; Randomized block design (RBD), unreplicated and replicated.
5. Linear regression and correlation analysis
Introduction; Simple linear regression model, least squares, ANOVA table, t-test, and F-tests,
confidence intervals for the intercept and slope; Correlation analysis, coefficient of correlation.

254
Prescribed textbook
1. Statistics. Mclave, J.T.and Dietrich, F.H. (1979). Delloen Publishing Company.

2. Introduction to Statistics; (2016)., David Lane, Rice University, Publisher: Saylor

Foundation

3. Introductory Statistics: (2017); Douglas S. Shafer, University of North Carolina. ISBN

13: 9781453344873

Recommended text book

1. Introduction to statistics. Walpole, R.E. 1984, Macmillan.

11.0 Answers to Unit Activities

1.1 Unit Activities

1. (a) discrete (b) cont. (c) discrete (d) cont. (d) cont.

1.2 Unit Activities

1. 4.47; 4; 1;2; 4; median

2. (b) P 65.9 , q 71.3

3. (b)4.475, 4.525,4.500,4.458,,4.450, 4.550, 4.525, 4.242, 4.117,, 3.958, 3.942, 3.950,3.892,

3.867, 3.883, 3.767, 3.742, 3.650,4.025, 3.975, 3.958, 3.992,3.917,3.933,4.017 (h)

4. (a)k62, (b) k 74 72%(taking upper limits as; k35.885, k39.995, et) true class limits; 0,
34.995, 35.995, 39.995, etc . Corresponding frequencies; 0.1,0.2,0.4, 0.6, 0.7, 0.7, 1.3, 0.9, 0.8,
0.4. mean = 68

2.1 Unit Activities

1. 2, 1

X 1 5/3 7/3 3

p( X ) 1/8 3/8 3/8 1/8

2.

255
2. (a)   0 (b)  2  1.2

2.2 Unit Activities

1
r  p(1  p)  1  1 1  2 1 n1
1. 1 ,   ;  p (1  p )
 n n  ; 3  n  3

n1  n1  2   1 2  2

3. ̂1 is more efficient; var( ˆ 1 ) ; var( ˆ 2 )  3 ; 5

5. k = 0.5
2.3 Unit Activities
1. (a) 0.46.2 (b) (i) 0. 3821 (ii) 0.1587
2. 9a) (ii) (64.4 – 67.6)kg (b) (i) K773 – 817 (c) 51.48 – 51.72
3. N(54,64), 0.0228, c = 16 ; N(5,144), 0.662
2.4 Unit Activities
1. (a) 4, 14.4
2 1 2 1 1
(b) (i) x 6 ,5 , 4 ,3 , 2, 1
3 3 3 2 3
P( x ) 0.1 0.2 0.2 0.3 0.1 0.1
(iii) 2.4

2. (a) 2  ,  2 (b) 0 ,  2 (c)  ,  ; 0.71 , 0.92

2

3. v  1.645 / n , n  215

4. (a) E ( X )  2 , var(x) = 
2
(b) N (  , 1 ) ; n  97
n n
5. 7, 5.83; N (7 n,5.83n)

n1
6.
n2

 r  1 a
7. P( R  r )    p (1  p) r a , r  a
 a  1
3.1 Unit Activities

256
1. (a) (3.114 -3.166)m (b) (3.109 – 3.171)m
2. (15.03 – 15.39)g
3.2 Unit Activities
1. (a) (15.7 – 18.68)h (b) (15.44 – 18.96)h
2. 47.86 – 48.14
3. 105.5h, 78.1h ; (87.7 – 125.3)h
4. 11.41mm, 0.120mm; (11.39 – 11.43)mm
3.3 Unit Activities
1. (0.9714 -0.9886)cm
2. (23.1 – 25.3)g
3. N(51.81, 54.18)g
4. (6.754 – 8.182)m
3.4 Unit Activities
1. 0.34 – 0.62
2. (a) 0.023 – 0.227, (b) 27 – 27.3
3. (a) 34% (b) 29.4% - 38.6% (c) 2155
4. 0.162 s.e = 0.024; 0.114 – 0.210; 1400; 0.054 – 0.112
3.5 Unit Activities
1. (a) 0.866 (b) (i) 2.00mg (ii) -0.01 to 1.01mg
2. (a) (50.0 – 54.6)kg (b) (1.80 – 2. 44)kg
3. (a) (0.252 – 0.262)kg (b) 246
4. (a) 2.04 – 3.96 (b) 25.2% - 34.8% (c) retired unemployed
5. (a) c  x (b) (i) 456, 108 (ii) 449.4-462.6 (iii) 457, 89
6. 10000, 7240 – 16200
7. (b) (7.3 – 9.3)%
8. (a) 26,500 (b) 1068
9. Yes

257
10. 4, 2.47 – 4.96; 4.1, 3.90 – 4.10
4.1 Unit Activities
1 1
1. H o : p  , H1 : p 
4 4
2. 0.0766
3. 0.1178
4. 0.1348
4.2 Unit Activities
1. z = 1.71
2. z = 2.83
3. 3mm, 0.02m; z =1.8
4. z= 1.79
5. mean = 4.6, z = 7.7
6. z=-1.50
7. z = 3
4.3 Unit Activities
1. 26.75yr 42.22yr2; z = 1.85
2. z = 4.44
3. z = 5.3
4. z = -1.47
5. z = -2.2
6. z= -4.68
4.4 Unit Activities
1. t 6  0.298

2. 8.6, t 5  3.54

3. t 4  2.21

258
4. t 7  1.67

4.5 Unit Activities

1. z = 4.8; 65.13 – 67.33 68.83 – 70.65
2. z = 1.44
3. z = 2.86
4. z = 4.6
5. z = -2.15
4.6 Unit Activities
1. 0.34 ; t 9  1.16

2. 0.0005; yes; t 24  5.17

4.7 Unit Activities

1. no, z = 1.40
2. yes, z = 2.08
3. yes ,z = 2.59
4. yes , z = 4.06
4.8 Unit Activities
1. yes ; z = 2.57
2. yes ; z = 2.99
4.9 Unit Activities
1. 0.082
2. z = -2.16
4.10 Unit Activities
1. r = 3 to 9 ; 0.182
2. 21.81 – 26.19 ; 0.05 ;0.855
3. n = 35 , v = 3.028
4. (a) 2%, 0.377

259
 (b) x  54.66 , (i) 0.749 (ii) 0.0918
(a) 1 in 40 line between 5 and 6 defectives, 1 in 100 line between 8 and 9 defectives
(b) (i) 0.364 , (ii) 0.0681 (iii) 0.754

3. (a) t 9  1.96 , (b) 12.9  13.04

3. 0.0064
4. -1.01 -0.21

5. t 9  2.33

6. yes, z = 4; 0.2 – 0.6

1 4
7 a ;
3 9
8. 0.109, ; yes, z = 1.83
9. z = 4

10. yes; t12  0.429

11. 0.30 < p< 0.38; 0.309; z = =3.03

12. 11.642 , 1.892; no z  t 99  1.88 ; 12  4.66

n
13. (a) yes, z= 2.4 (b) yes ; z = 2.95

14. yes, t120  2.1

15. 3.62, 0.29; t 9  4.5

5.1 Unit Activities

1. yes, X 2  20.2,9df

2. 12,24,36,48,60,72,60,48,36,24,12; A, X 2  14.5 , B , X 2  2.12

5.2 Unit Activities

1. reject null hypothesis, X 2  9.6, 2df

2. no, X 2  0.23, 1df

3 retain null hypothesis , X 2  3.2, 7df

260
4. 3; 2.5, 7.5, 11.2, 11.2, 8.4, 5.0, 2.5, 1.1, 0.6,; X 2  0.28, 3df

5.3 Unit Activities

1. no , X 2  2.8, 1df

2. 21.0 10.0 7.0

15.5 7.4 5.2
41.5 19.7 13.8

No evidence, X 2  7.9, 4df

6.1 Unit Activities

1. (a) -0.98 (b) -0.03
2. -0.91, -0.98
6.2 Unit Activities
1. (a) r = 0.88 (b) C,I,E,D,B,G,J,A,F,H
2. (a) r = 0.66 , (b) r = 0.54
3. r = 0.28
4. (a) r = 0.75
6.3 Unit Activities
1. r = -0.86
2. (a) invalid (b) invalid (c) valid (d) invalid
3. (b) r = 0.87
5. (a) -0.37 (b) -0.27 (c) no evidence
6. r = 0.97
7. r = 0.78
7.0 Unit Activities

3 SSE =0.020; F = 2.0; do not reject the null hypothesis

4 (a) 2.20   A  2.30

261
 (b)  0.01  ( A   B  0.18

5 (c) SSE = 498.67 , yes

9 (a) no, it is completely RD (b) CRD with n treatment and two treatment

10 (c) SSE = 4.99 (d) no, no

11 (e) RBD, (f) 21 (g) SSE = 58.91; F = 3.07 ; YES; (h) no

12 (a) SSE = 0.571; SST = 1.212;F = 11.7; YES; (b) t =--2.73 ; v = 7; yes

13

source df SS MS F

BLOCKS 3 0.140 0.047 6.62

TREATMENT 4 0.787 0.197 27.7

ERRORS 12 0.085 0.0071

TOTAL 19 1.012

F = 27.7 , REJECT THE NULL HYPOTHESIS

14

source df SS MS F

TREATMENT 2 38.00 19.00 10.05

BLOCKS 3 61.67 20.56 10.88

ERRORS 6 11.33 1.89

TOTAL 11

i. yes; F = 10.05

262
 ii. yes ; F = 10.88
iii. 3.50  1.89
15

source df SS MS F

ROWS 2 46.89 23.45 11.11

COLUMNS 2 22.89 11.45 5.43

TREATMENTS 2 91.56 45.78 21.70

ERRORS 2 4.22 2.11

TOTAL 8 165.56

7.33  5.11
18. F= 1.535 no evidence of a difference between treatment
8.1 Unit Activities
1. (a) y  0.486x  2.397 (b) 0.997 (c) no

2. y = 2.79x – 5451 ; 53.7 x106

3. (b) y = 2.75x +48.35 (c) 0.79
4. y = 18.9x + 154 ; 476g
8.2 Unit Activities
1. (a) y = -4x +24 (b) 8, 0.395 (c) 7.4 – 8.6
2. k = 0.9948, c = -1.998; 0.92 – 1.07mm

3. (a) aˆ  0.06 , bˆ  0.132 (b) 1.344 – 1.416; 22.1 – 26.1

263
4. y = -x + 12 ; 5.67 – 6.33; 0 ,no linear
5. (a) 18.8 – 27.2
(b) (-0.428 ) – (-0.222)

12.1 Table A1 Z-Table

1.1 – Negative Z Table

264
Use the positive Z score table below to find values on the right of the mean as can be seen in the
graph alongside. Corresponding values which are greater than the mean are marked with a
positive score in the z-table and represent the area under the bell curve to the left of z.

1.2 – Positive Z Table

12.2 Table A3 Random Numbers

1000 Random Numbers

265
0980 19931 69587 34846 65137 89808 87910 21015 08217 80808 79191 79323 61832 59344 50445 69982
1598 29815 94718 95097 77449 73675 73906 50231 58945 13486 26566 00967 05591 01990 00727 84468
2144 36166 11700 36044 78926 96293 24899 40709 57141 10013 63645 09751 19181 41396 41981 82285
3946 04999 22488 51825 87148 04928 99092 17839 26631 98905 04880 62315 32680 87327 44912 51325
3332 27328 54034 61582 44692 71845 79687 32010 16245 33677 89725 40301 01165 74548 05222 04476
1630 19171 58291 09456 19137 94311 74888 58615 99810 92844 82632 50348 04714 62825 43615 79631
5169 21542 12212 64587 03061 94590 17898 98043 76155 79923 73055 24097 48407 36923 05793 39680
5277 00458 17751 21116 14042 19949 66919 75685 69107 23429 75448 18190 62320 06608 62459 72513
8281 46933 01860 98182 94658 42653 05156 41932 27496 76878 29262 14266 01533 06647 09317 91022
5510 76984 16493 39928 12487 59967 40811 72174 25148 18785 98952 18967 46122 96803 67069 33096
3293 31626 38602 15494 60107 94157 07087 66803 60892 52665 73970 63932 01168 61844 81420 61626
2956 76874 70142 19019 55096 92490 62188 51207 58556 07036 68769 05803 96746 11534 23072 64503
0828 03745 38065 69646 40033 27229 89316 75780 66969 35411 22804 26221 87936 80640 87728 54618
8238 28254 94324 11514 82496 95642 96675 15909 59958 16307 00525 31824 54816 28926 96091 69860
1514 91416 15872 13764 75062 19994 17468 71988 36352 53240 91388 54256 02463 41159 93864 73122
1982 59248 04662 70536 35310 47550 09898 69405 19978 74724 09845 50155 60957 27105 51751 52293
5973 22764 87649 00973 05819 50576 57168 58553 59665 34276 95349 01163 85374 21528 65455 20264
4814 97981 16783 49213 54165 26337 67481 14820 29241 10412 12353 13920 55794 84194 55678 14926
4832 25914 69020 84397 72928 97100 74040 38599 27534 06646 50310 20067 77294 99870 68124 99169
7357 72578 46312 00667 79685 10130 35049 55279 78372 51494 23674 76245 79953 78320 73497 60884
8715 28990 75611 17162 17014 37692 33711 15252 30584 98473 71898 64094 18848 54310 67498 12962
0237 11165 08872 78025 52495 77052 78228 93907 57997 26098 59434 90254 74058 87606 70832 93293
8248 60118 23047 52394 78704 50476 01804 91636 59439 37884 75736 86367 50661 62974 29201 64767
4079 41866 20090 34411 13220 05229 62642 33829 58739 62347 35258 09073 78734 90179 13309 15276
4056 47423 26953 43216 98620 13577 43945 70877 04725 53003 77451 40013 13357 03987 18858 57710
9507 92806 95590 22950 02484 72786 53917 78170 71225 88366 66770 85826 34607 64163 66552 29960
6762 19030 02953 42754 17390 05121 25759 56099 57067 71954 92666 78155 47562 55473 02094 94916
7148 92110 50997 96768 60914 57848 17676 30054 86079 32281 69595 73639 82301 87683 96187 40469
6993 03063 66673 04132 75636 03232 12871 50427 07089 22864 37008 53920 93901 95558 44534 79509
7626 67904 01936 08988 29134 38539 19547 92608 93926 22217 59359 25637 62441 38864 67839 49928
5374 27648 33738 80477 13986 86035 70906 56987 45417 33856 26100 20432 92998 02367 11804 14615
1565 48312 64033 12738 47769 92986 25152 68955 85391 86298 51686 94944 15652 95832 47132 61462
2529 20910 19773 24912 73063 31656 65488 78903 62676 33057 75569 30815 85480 54024 69527 03359
9592 61459 52912 61140 57444 99312 00118 62222 51099 52649 62202 34686 82560 36710 99692 78197
9083 10974 40402 90563 93492 18220 62244 19301 49490 48591 01038 38196 11970 28655 83330 48866
2330 35470 84198 07321 98784 35644 10069 75532 81676 49396 06531 97987 03789 39625 91143 03256
5662 25964 11251 30555 15899 23850 26797 81306 96485 08582 18134 95699 13094 04385 18836 54260
0405 98471 73516 04405 12414 80105 20631 12012 67745 04664 00299 37973 19964 82701 87112 39766
2890 94006 47945 98013 80909 26673 24773 68041 94286 41157 42480 56449 29477 59338 41672 22667
4443 78586 36491 05518 38960 10821 97428 74785 99933 46577 59130 07768 71711 44063 53221 90851
1392 43225 66108 46061 29146 04814 21797 82633 33517 74440 89702 73571 01743 71640 17462 57211
5067 03940 88749 63784 19045 05917 06085 66977 68864 03260 98650 60498 99653 01836 55097 59637
1794 01746 81367 27826 01235 36397 83494 28207 04804 36555 75426 00190 70519 19415 11831 61018
2903 02659 45914 32328 93295 43518 12227 91713 65163 02840 34483 49288 44418 34145 58365 49247

266
4721 97694 09345 96431 37802 99547 75489 07887 98771 55629 15275 89432 76428 80789 90404 37213
8576 02866 23611 64276 02335 01749 81480 52119 79456 73439 82255 27262 46628 19115 28650 62808
0796 59193 65667 05002 79472 72390 63406 24799 06045 59786 24875 94420 60096 33887 18806 88921
2711 32688 37466 12751 86791 58735 79469 51318 02366 43184 57590 47546 46910 34871 31576 03444
0647 14369 65960 06664 66869 33050 12871 82067 17249 77150 19851 68280 62148 56833 62662 19266
5932 20250 03102 30881 87286 92600 71817 92436 24520 40200 76113 63264 85888 59539 32768 74280
4896 91346 25847 54542 40619 59652 89510 82816 23008 35848 15829 64138 23208 27753 87598 31374
8866 93675 36146 71788 44446 81471 24151 78598 36542 92611 53450 27542 04187 71243 97853 78441
4170 18251 60954 01760 41346 30322 58946 30172 13949 64004 28432 35117 83902 24772 59237 28370
2137 81090 52222 88599 98897 03471 02096 12929 29057 28542 05226 93504 42433 48107 92454 19052
9094 23207 86905 01445 09676 23184 86805 27259 80695 89742 73286 34344 84856 26014 43206 13379
6369 10619 66956 34438 91262 36726 71276 98554 92690 76119 47065 14276 41249 68258 82196 99239
0291 34340 19327 06720 86233 16363 38711 77203 13871 41189 66016 14942 66692 09605 05128 89525
2189 10387 95971 62430 62168 59361 97316 38597 18066 63467 94592 42983 41263 09820 32705 22126
8389 79775 73843 70711 11643 87985 80291 83128 04103 46469 22247 30038 70038 53670 10630 59934
0220 08520 54894 25701 52238 49501 45841 36185 95811 18710 63433 32749 83097 50919 09606 90838
4010 87636 66078 46543 26531 16174 77169 48160 42016 33707 57604 32756 35518 36333 25336 27730
1088 33141 59349 77375 02099 85271 82477 94442 16548 69974 69213 36700 07378 69678 78523 08500
7782 86050 11659 87340 71521 79056 94969 75423

12.3 Table A4 T-Distribution Table (One Tail and Two-Tails)

df a = 0.1 0.05 0.025 0.01 0.005 0.001 0.0005

267
 ∞ ta = 1.282 1.645 1.960 2.326 2.576 3.091 3.291
1 3.078 6.314 12.706 31.821 63.656 318.289 636.578
2 1.886 2.920 4.303 6.965 9.925 22.328 31.600
3 1.638 2.353 3.182 4.541 5.841 10.214 12.924
4 1.533 2.132 2.776 3.747 4.604 7.173 8.610
5 1.476 2.015 2.571 3.365 4.032 5.894 6.869
6 1.440 1.943 2.447 3.143 3.707 5.208 5.959
7 1.415 1.895 2.365 2.998 3.499 4.785 5.408
8 1.397 1.860 2.306 2.896 3.355 4.501 5.041
9 1.383 1.833 2.262 2.821 3.250 4.297 4.781
10 1.372 1.812 2.228 2.764 3.169 4.144 4.587
11 1.363 1.796 2.201 2.718 3.106 4.025 4.437
12 1.356 1.782 2.179 2.681 3.055 3.930 4.318
13 1.350 1.771 2.160 2.650 3.012 3.852 4.221
14 1.345 1.761 2.145 2.624 2.977 3.787 4.140
15 1.341 1.753 2.131 2.602 2.947 3.733 4.073
16 1.337 1.746 2.120 2.583 2.921 3.686 4.015
17 1.333 1.740 2.110 2.567 2.898 3.646 3.965
18 1.330 1.734 2.101 2.552 2.878 3.610 3.922
19 1.328 1.729 2.093 2.539 2.861 3.579 3.883

12.4 Table A6 Upper-tail critical values of chi-square distribution with ν degrees of

freedom

Probability less than the critical value

ν 0.90 0.95 0.975 0.99 0.999

1 2.706 3.841 5.024 6.635 10.828

2 4.605 5.991 7.378 9.210 13.816
3 6.251 7.815 9.348 11.345 16.266
4 7.779 9.488 11.143 13.277 18.467
5 9.236 11.070 12.833 15.086 20.515
6 10.645 12.592 14.449 16.812 22.458
7 12.017 14.067 16.013 18.475 24.322
8 13.362 15.507 17.535 20.090 26.125
9 14.684 16.919 19.023 21.666 27.877
10 15.987 18.307 20.483 23.209 29.588
11 17.275 19.675 21.920 24.725 31.264

268
12 18.549 21.026 23.337 26.217 32.910
13 19.812 22.362 24.736 27.688 34.528
14 21.064 23.685 26.119 29.141 36.123
15 22.307 24.996 27.488 30.578 37.697
16 23.542 26.296 28.845 32.000 39.252
17 24.769 27.587 30.191 33.409 40.790
18 25.989 28.869 31.526 34.805 42.312
19 27.204 30.144 32.852 36.191 43.820
20 28.412 31.410 34.170 37.566 45.315
21 29.615 32.671 35.479 38.932 46.797
22 30.813 33.924 36.781 40.289 48.268
23 32.007 35.172 38.076 41.638 49.728
24 33.196 36.415 39.364 42.980 51.179
25 34.382 37.652 40.646 44.314 52.620
26 35.563 38.885 41.923 45.642 54.052
27 36.741 40.113 43.195 46.963 55.476
28 37.916 41.337 44.461 48.278 56.892
29 39.087 42.557 45.722 49.588 58.301
30 40.256 43.773 46.979 50.892 59.703
31 41.422 44.985 48.232 52.191 61.098
32 42.585 46.194 49.480 53.486 62.487
33 43.745 47.400 50.725 54.776 63.870
34 44.903 48.602 51.966 56.061 65.247
35 46.059 49.802 53.203 57.342 66.619
36 47.212 50.998 54.437 58.619 67.985
37 48.363 52.192 55.668 59.893 69.347
38 49.513 53.384 56.896 61.162 70.703
39 50.660 54.572 58.120 62.428 72.055
40 51.805 55.758 59.342 63.691 73.402
41 52.949 56.942 60.561 64.950 74.745
42 54.090 58.124 61.777 66.206 76.084
43 55.230 59.304 62.990 67.459 77.419
44 56.369 60.481 64.201 68.710 78.750
45 57.505 61.656 65.410 69.957 80.077
46 58.641 62.830 66.617 71.201 81.400
47 59.774 64.001 67.821 72.443 82.720
48 60.907 65.171 69.023 73.683 84.037
49 62.038 66.339 70.222 74.919 85.351
50 63.167 67.505 71.420 76.154 86.661
51 64.295 68.669 72.616 77.386 87.968
52 65.422 69.832 73.810 78.616 89.272
53 66.548 70.993 75.002 79.843 90.573
54 67.673 72.153 76.192 81.069 91.872
55 68.796 73.311 77.380 82.292 93.168

269
56 69.919 74.468 78.567 83.513 94.461
57 71.040 75.624 79.752 84.733 95.751
58 72.160 76.778 80.936 85.950 97.039
59 73.279 77.931 82.117 87.166 98.324
60 74.397 79.082 83.298 88.379 99.607
61 75.514 80.232 84.476 89.591 100.888
62 76.630 81.381 85.654 90.802 102.166
63 77.745 82.529 86.830 92.010 103.442
64 78.860 83.675 88.004 93.217 104.716
65 79.973 84.821 89.177 94.422 105.988
66 81.085 85.965 90.349 95.626 107.258
67 82.197 87.108 91.519 96.828 108.526
68 83.308 88.250 92.689 98.028 109.791
69 84.418 89.391 93.856 99.228 111.055
70 85.527 90.531 95.023 100.425 112.317
71 86.635 91.670 96.189 101.621 113.577
72 87.743 92.808 97.353 102.816 114.835
73 88.850 93.945 98.516 104.010 116.092
74 89.956 95.081 99.678 105.202 117.346
75 91.061 96.217 100.839 106.393 118.599
76 92.166 97.351 101.999 107.583 119.850
77 93.270 98.484 103.158 108.771 121.100
78 94.374 99.617 104.316 109.958 122.348
79 95.476 100.749 105.473 111.144 123.594
80 96.578 101.879 106.629 112.329 124.839
81 97.680 103.010 107.783 113.512 126.083
82 98.780 104.139 108.937 114.695 127.324
83 99.880 105.267 110.090 115.876 128.565
84 100.980 106.395 111.242 117.057 129.804
85 102.079 107.522 112.393 118.236 131.041
86 103.177 108.648 113.544 119.414 132.277
87 104.275 109.773 114.693 120.591 133.512
88 105.372 110.898 115.841 121.767 134.746
89 106.469 112.022 116.989 122.942 135.978
90 107.565 113.145 118.136 124.116 137.208

270
 Table A7 F Distribution Tables

F Table for
α = 0.05

/ df1=1 2 3 4 5 6 7 8 9 10
df2=
161.4476 199.5000 215.7073 224.5832 230.1619 233.9860 236.7684 238.8827 240.5433 241.8817 24
1
2 18.5128 19.0000 19.1643 19.2468 19.2964 19.3295 19.3532 19.3710 19.3848 19.3959 1
3 10.1280 9.5521 9.2766 9.1172 9.0135 8.9406 8.8867 8.8452 8.8123 8.7855
4 7.7086 6.9443 6.5914 6.3882 6.2561 6.1631 6.0942 6.0410 5.9988 5.9644
5 6.6079 5.7861 5.4095 5.1922 5.0503 4.9503 4.8759 4.8183 4.7725 4.7351

6 5.9874 5.1433 4.7571 4.5337 4.3874 4.2839 4.2067 4.1468 4.0990 4.0600
7 5.5914 4.7374 4.3468 4.1203 3.9715 3.8660 3.7870 3.7257 3.6767 3.6365
8 5.3177 4.4590 4.0662 3.8379 3.6875 3.5806 3.5005 3.4381 3.3881 3.3472
9 5.1174 4.2565 3.8625 3.6331 3.4817 3.3738 3.2927 3.2296 3.1789 3.1373
10 4.9646 4.1028 3.7083 3.4780 3.3258 3.2172 3.1355 3.0717 3.0204 2.9782

11 4.8443 3.9823 3.5874 3.3567 3.2039 3.0946 3.0123 2.9480 2.8962 2.8536
12 4.7472 3.8853 3.4903 3.2592 3.1059 2.9961 2.9134 2.8486 2.7964 2.7534
13 4.6672 3.8056 3.4105 3.1791 3.0254 2.9153 2.8321 2.7669 2.7144 2.6710
14 4.6001 3.7389 3.3439 3.1122 2.9582 2.8477 2.7642 2.6987 2.6458 2.6022

271
15 4.5431 3.6823 3.2874 3.0556 2.9013 2.7905 2.7066 2.6408 2.5876 2.5437

The F distribution is a right-skewed distribution used most commonly in Analysis of Variance.

When referencing the F distribution, the numerator degrees of freedom are always given first,
as switching the order of degrees of freedom changes the distribution (e.g., F(10,12) does not equal
F(12,10) ). For the four F tables below, the rows represent denominator degrees of freedom and the
columns represent numerator degrees of freedom. The right tail area is given in the name of the
table. For example, to determine the .05 critical value for an F distribution with 10 and 12 degrees
of freedom, look in the 10 column (numerator) and 12 row (denominator) of the F Table for
alpha=.05. F(.05, 10, 12) = 2.7534. You can use the Java Applet or the HTML5/JavaScript Webapp
interactive F-Distribution calculators to obtain more accurate measures of probability or critical
values.

272