Research

Methods and
Statistics

Muhammad Adeel Javaid

1
 Preface
I wrote this book because there is a large gap between the elementary statistics
course that most people take and the more advanced research methods courses
taken by graduate and upper-division students so they can carry out research
projects. These advanced courses include difficult topics such as regression,
forecasting, structural equations, survival analysis, and categorical data, often
analyzed using sophisticated likelihood-based and even Bayesian methods.
However, these advanced courses typically devote little time to helping students
understand the fundamental assumptions and machinery behind these methods.
Instead, they teach the material like witchcraft: Do this, do that, and voilà—
Statistics! Students thus have little idea what they are doing and why they are
doing it. Like trained parrots, they learn how to recite statistical jargon mindlessly.
The goal of this book is to make statistics less like witchcraft, to treat students
like intelligent humans and not like trained parrots—thus the title, Research
Methods and Statistics.

This book will cause students and researchers to think differently about things,
not only about math and statistics, but also about research, the scientific method,
and life in general. It will teach them how to do good modeling—and hence good
statistics—from a standpoint of deep knowledge rather than rote knowledge. It
will also provide them with tools to think critically about the claims they see in the
popular press, and to design their own studies to avoid common errors.

This book is not a ―cookbook.‖ Cookbooks tell you all about the what but nothing
about the why. With computers, software and the Internet readily available, it is
easier than ever for students to lose track of the why and focus on
the what instead. This book takes exactly the opposite approach. It will empower
students and researchers to use advanced statistical methods with confidence.

2
 Contents

THE NATURE OF DATA 5

ANALYSIS OF INDIVIDUAL OBSERVATIONS 12

DESCRIPTIVE STATISTICS 23

STANDARD PROBABILITY DISTRIBUTIONS 63

THE NORMAL DISTRIBUTION 71

STATISTICAL DECISION THEORY 74

SMALL SAMPLING THEORY 85

LINEAR REGRESSION AND CORRELATION 103

3
Research your idea. See if there's a demand. A lot of people
have great ideas, but they don't know if there's a need for it.
You also have to research your competition.

4
 I
THE NATURE OF DATA
Anything that can be counted or measured is called a variable. Knowledge of
the different types of variables, and the way they are measured, play a crucial
part in choice of coding and data collection. The measurement of variables can
be categorized as categorical (nominal or ordinal scales) or continuous (interval
or ratio scales).

Categorical measures can be used to identify change in a variable, however,

should you wish to measure the magnitude of the change you should use a
continuous measure.

A nominal scale allows for the classification of objects, individual and responses
based on a common characteristic or shared property. A variable measured on
the nominal scale may have one, two or more sub-categories depending on the
degree of variation in the coding. Any number attached to a nominal
classification is merely a label, and no ordering is implied: social worker, nurse,
electrician, physicist, politician, teacher, plumber, etc.

An ordinal scale not only categorizes objects, individuals and responses into sub-
categories on the basis of a common characteristic it also ranks them in
descending order of magnitude. Any number attached to an ordinal
classification is ordered, but the intervals between may not be constant: GCSE,
A-level, diploma, degree, postgraduate diploma, higher degree, and doctorate.

The interval scale has the properties of the ordinal scale and, in addition, has a
commencement and termination point, and uses a scale of equally spaced
intervals in relation to the range of the variable. The number of intervals
between the commencement and termination points is arbitrary and varies from
one scale to another. In measuring an attitude using the Likert scale, the
intervals may mean the same up and down the scale of 1 to 5 but multiplication is
not meaningful: a rating of ‗4‘ is not twice as ‗favourable‘ as a rating of ‗2‘.

In addition to having all the properties of the nominal, ordinal and interval scales,
the ratio scale has a zero point. The ratio scale is an absolute measure allowing
multiplication to be meaningful. The numerical values are ‗real numbers‘ with
which you can conduct mathematical procedures: a man aged 30 years is half
the age of a woman of 60 years.

5
 Categorical Continuous

Unitary Dichotomous Polytomous Interval or Ratio Scale

Name [1] . . . Yes Attitudes (Likert Scale): Income (£000s per annum)
[0] . . . No [5] . . . strongly agree
Occupation [4] . . . agree Age (in years)
[1] . . . Good [3] . . . uncertain
Location [0] . . . Bad [2] . . . disagree Reaction Time (in seconds)
[1] . . . strongly disagree
Site [1] . . . Female Absence (in days)
[0] . . . Male Age:
[4] . . . Old Distance (in kilometres)
[1] . . . Right [3] . . . Middle-aged
[0] . . . Wrong [2] . . . Young Length (metres)
[1] . . . Child
[1] . . . Extrovert Attitude (Thurstone & Cheve)
[0] . . . Introvert Income:
[3] . . . High
[1] . . . Psychotic [2] . . . Medium
[0] . . . Neurotic [1] . . . Low

[1] . . . Assertive Socio-Economic Status:

[0] . . . Passive [5] . . . A
[4] . . . B
[1] . . . Present [3] . . . C1
[0] . . . Absent [2] . . . C2
[1] . . . D
[0] . . . E

Qualitative Quantitative

Sex (Male/Female) Age (in years)

Age (Old/Young) Attitude (Guttman scale)
Attitude (Favourable/Unfavourable) Attitude (Thurstone & Cheve scale)
Attitude (Likert scale) Performance (errors or faults per minute)
Achieved Educational Level (High/Low) Achieved Educational Level (number of years
Style (Autocratic/Participative) post-secondary school education)
Location (Urban/Rural)
Performance (Good/Bad)

Table I
A Two-Way Classification of Variables

6
 Methods of Data Collection

The major approaches to gathering data about a phenomenon are from primary
sources: directly from subjects by means of experiment or observation, from
informants by means of interview, or from respondents by questionnaire and
survey instruments. Data may also be obtained from secondary sources:
information that is readily available but not necessarily directly related to the
phenomenon under study. Examples of secondary sources include published
academic articles, government statistics, an organization‘s archival records to
collect data on activities, personnel records to obtain data on age, sex,
qualification, length of service, and absence records of workers, etc. Data
collected and analyzed from published articles, research papers and journals
may be a primary source if the material is directly relevant to your study. For
instance, primary sources for a study conducted using the Job Descriptive Index
may be Hulin and Smith (1964-68) and Jackson (1986-90), whereas a study
using an idiosyncratic study population, technique and assumptions, such as
those published by Herzberg, et alia (1954-59), would be a secondary source.

Methods of Data Collection

Primary Sources Secondary Sources

Data primarily and directly gathered Data not primarily or directly gathered
for the purpose of the study for the purposes of the study

Interview Survey Census

Observation Experiment Instruments

Archives Other
records

Stuctured Participant True Captive

Previous
unrelated
studies
Unstructured Non-participant Quasi Mailed

7
 Figure 1
A Classification of Methods of Data Collection
Procedures for Coding Data

A coding frame is simply a set of instructions for transforming data into codes
and for identifying the location of all the variable measured by the test or
instrument. Primary data gathered from subjects and informants is amenable to
control during the data collection phase. The implication is that highly structured
data, usually derived from tests, questionnaires and interviews, is produced
directly by means of a calibrated instrument or is readily produced from raw
scores according to established rules and conventions. Generally, measures
such as physical characteristics such as height and weight are measured on the
ratio scale. Whereas psychological attributes such as measures of attitude and
standard dimensions of personality are often based on questions to which there
is no appropriate response. However, the sum of the responses is interpreted
according to a set of rules and provides a numerical score on the interval scale
but is often treated as though the measures relate the ratio scale. Norms are
available for standard tests of physical and psychological attributes to establish
the meaning of individual scores in terms of those derived from the general
population. A questionnaire aimed at determining scores as a measure of a
psychological attribute are said to be pre-coded; that is, the data reflects the
coder‘s prior structuring of the population. The advantages of pre-coding are
that it reduces time, cost and coding error in data handling. Ideally, the pre-
coding should be sufficiently robust and discriminating as to allow data
processing by computer.

A coding frame should include information for the variable to be measured:

 the source data (e.g. Question 7 or ‗achieved educational level‘);

 a list of the codes (e.g. number of years post-secondary school

education);

 column location of the variable on the coded matrix.

Once we have a coding frame, data relating to an individual or case can be read
off just as we would read the data in a sales catalogue or a coded matrix from a
holiday brochure:

Destination Number Class of Number Full/half Car hire Price

Paphos of days hotel of rooms board Included £
Athenaeum 10 *** 45 F N 450
Polis 7 ** 23 H N 299
Roman Court 5 ***** 225 F Y 999
Aphrodite 7 ** 30 F Y 520

8
 The following statements may or may not be true for your work organization. For each item
below, please answer by ticking (/) the appropriate box below the categories at the top of the
page. If you cannot decide then tick the box which is closest to your answer.

Definitely More True More False Definitely
True than False than True False
1. I feel that I am my own boss in most matters    

2. The company is generally quick to use
improved work methods    

3. The company has a real interest in the
welfare and happiness of those who work here     

4. The company tries to improve working conditions     
5. A person can makes his or her own decisions
around here without checking with anybody else    

6. Everyone has a specific job to do     
7. The company has clear-cut reasonable goals
and objectives      

8. Most people here make their own rules on the job    

9. Work activities are sensibly organized     

 
10. I feel that those above me are very receptive
to my ideas and suggestions    

Case # Age Sex Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10

101 18 1 1 3 1 2 2 2 3 1 1 3
102 27 1 5 3 5 1 2 1 2 1 5 3
103 54 1 2 3 1 1 2 1 2 1 4 3
104 23 1 2 4 1 1 2 1 2 1 4 3
105 19 0 3 3 4 1 2 2 3 2 2 4
106 34 1 4 2 3 3 3 1 2 2 3 5
107 51 1 4 3 3 1 3 3 3 1 3 3
108 25 0 5 5 1 3 3 3 2 3 3 2
109 25 0 5 3 2 3 2 3 3 3 4 3
110 31 0 2 3 2 1 2 3 2 1 4 4
111 21 1 1 4 5 1 2 2 3 2 4 3
112 44 1 1 4 3 2 2 2 3 3 4 2

Table 1

9
 Survey Instrument and Coding Frame Matrix for Research Data

If you choose to use an unstructured interview schedule or questionnaire,

containing open-ended questions, it is necessary to transform the responses
from these questions into a mode that allows for efficient analysis. It may be
possible to conduct a pilot study to establish the kind of responses expected from
the sample of respondents, thus deriving a code list that may allow for some
degree of pre-coding. For small samples, say less than thirty people, the
analysis may take the form of listening to the voice tape, or scanning the written
responses, to gain an overall impression of the categories of response.
Considering key statements relating to the phenomenon under study (see Table
2 below) can then reduce the data from the responses. This technique can be
used during the piloting the interview or questionnaire, and even if we decide not
to pre-code the responses, we have a useful set of anticipated responses that
can be used for training interviewers and be extended, if need be, by adding
further categories from the survey.

Key Responses to Statement:

What do you particularly like about 1 2 3 4 5 6 7 8
your job?
1. Convenience X X X X X
2. Responsibility X X X
3. Making decisions X X X
4. Friendships X X
5. Use of abilities and talents X X X
6. Working conditions X X X
7. Opportunities for promotion
8. Pay X X
9. Interesting work X X X
10. Job security X X X

Table 2
Response Matrix Representing Categories of Interest

The design of a coding frame is also determined by the approach we take in

respect of the data: what the data signifies, and useful ways of understanding the
data once collected. After Swift (1996), three approaches can be identified:

1. Representational Approach
The response of the informant is said to express the surface meaning of
what is ―out there‖ requiring the researcher to apply codes to reduce the
data, whilst at the same time, reflecting this meaning as faithfully as
possible. At this stage of the process, the data must be treated
independently from any views the researcher may hold about underlying
variables and meanings.

10
2. Anchored-in Approach
The researcher may view the responses as having additional and implicit
meanings that come from the fact that the responses are dependent on
the data-gathering context. For example, in investigating worker
involvement, we might want to conduct this with a framework comprising
of types of formal and informal worker/manager interactions. As a
consequence, the words given by informants can be interpreted to
produce codes on more than one dimension relating to the context: (a)
nature of the contact: formal versus informal, intermittent versus
continuous contact, etc. (b) initiator of contact: worker versus manager.
The coding frame using this approach takes into account ―facts‖ as being
anchored to the situation, rather than treating the data as though they are
context-free.

3. Hypothesis-Guided Approach
Although similar to the second approach, we may view the data as having
multiple meanings according the paradigm or theoretical perspective from
which they are approached (e.g. phenomenological or hermeneutic
approach to investigating a human or social phenomenon). The
hypothesis-guided approach recognizes that the data do not have just one
meaning which refers to some reality approachable by analysis for the
surface meaning of the words: words have multiple meanings, and ―out
there‖ is a multiverse rather than a universe. In the hypothesis-guided
approach, the researcher might use the data, and other materials, to
create or investigate variables that are defined in terms of the theoretical
perspective and construct propositions. For example, a data set might
contain data on illness and minor complaints that informants had
experienced over a period of say, one year. Taking the hypothesis-
guided approach, the illness data might be used as an indicator of
occupational stress or of a reaction to transformational change. Hence,
the coding frame is based on the researcher'‘ views and hypotheses rather
than on the surface meanings of the responses.1

1
In the case of anchored and hypothesis-guided coding frames, there may well be categories that
are not represented in small samples of data. Indeed, this is very likely for the hypothesis-guided
coded frame; that is, you may want the frame to be able to reflect important concept, if only by
showing that they did not occur among the responses.

11
 II
ANALYSIS OF INDIVIDUAL
OBSERVATIONS
In the analysis of individual observations, or ungrouped data, consideration will
be given to all levels of measurement to determine which descriptive measures
can be used, and under what conditions each is appropriate.

One of the most widely used descriptive measures is the ‗average‘. One speaks
of the ‗average age‘, average response time‘, or ‗average score‘ often without
being very specific as to precisely what this means. The use of the average is
an attempt to find a single figure to describe or represent a set of data. Since
there are several kinds of 'average', or measures of central tendency, used in
statistics, the use of precise terminology is important: each ‗average‘ must be
clearly defined and labelled to avoid confusion and ambiguity. At least three
kinds of common uses of the ‗average‘ can be described:

1. An average provides a summary of the data. It represents an

attempt to find one figure that tells more about the characteristics of
the distribution of data than any other. For example, in a survey of
several hundred undergraduates the average intelligence quotient
was 105: this one figure summarizes the characteristic of
intelligence.

2. The average provides a common denominator for comparing sets

of data. For example, the average score on the Job Descriptive
Index for British managers was found to be 144, this score provides
a quick and easy comparison of levels of felt job satisfaction with
other occupational groups.

3. The average can provide a measure of typical size. For example,

the scores derived for a range of dimensions of personality can be
compared to the norms for the group the sample was taken from;
thus, one can determine the extent to which the score for each
dimension is above, or below, that to be expected.

12
 The Mode

The mode can be defined as the most frequently occurring value in a set
of data; it may be viewed as a single value that is most representative of
all the values or observation in the distribution of the variable under study.
It is the only measure of central tendency that can be appropriately used
to describe nominal data. However, a mode may not exist, and even if it
does, it may not be unique:

1 2 3 4 5 6 7 8 9 10 ... ... ... ... No mode

YYNYNNNNY ... ... ... ... Unimodal (N)

1223444455 ... ... ... ... Unimodal (4)

1222345556 ... ... ... ... Bimodal (2, 5)

1223445667 ... ... ... ... Multimodal (2, 4, 6)

With relatively few observations, the mode can be determined by

assembling the set of data into an array. Large numbers of observations
can be arrayed by means of Microsoft EXCEL, or other statistical software
programs:

Subject Reaction Time Array

(in m/seconds)
000123 625 460
000125 500 480
000126 480 500
000128 500 500
000129 460 500
000131 500 500 Mode
000134 575 510
000137 530 525
000142 525 530
000144 500 575
000145 510 625

The Median

When a measurement of a set of observation is at least ordinal in nature, the

observations can be ranked, or sorted, into an array whereby the values are
arranged in order of magnitude with each value retaining its original identity.
The median can be defined as the value of the middle item of a set of items that
form an array in ascending or descending order of rank: the [N+1}/2 position. In

13
simple terms, the median splits the data into two equal parts, allowing us to state
that half of the subjects scored below the median value and half the subjects
scored above the median value. If an observed value occurs more than once, it
is listed separately each time it occurs:

Reaction Reaction Time Reaction Time

Subject Time in shown in array: shown in array:
m/secs shortest to longest longest to shortest
000123 625 460 625
000125 500 480 575
000126 480 500 530
000128 500 500 525
000129 460 500 510
000131 500 500 500 Median
000134 575 510 500
000137 530 525 500
000142 525 530 500
000144 500 575 480
000145 510 625 460

The Arithmetic Mean

Averages much more sophisticated than the mode or median can be used at the
interval and ratio level. The arithmetic mean is widely used because it is the
most commonly known, easily understood and, in statistics, the most useful
measure of central tendency. The arithmetic mean is usually given the notion 
and can be computed:

[x1 + x2 + x3 + x4 + x5 + x6 + x7 + x8 + . . . xn] / N

where, x1, x2, x3 . . . xn are the values attached to the observations; and, N is the
total number of observations:

000123 000125 000126 000128 000129 000131 000134 000137 000142

Subject
x1 x2 x3 x3 x4 x5 x6 x7 x8

Reaction Time 625 500 480 500 460 500 575 530 525
(m/secs)

Using the above formula, the arithmetic mean can be computed:

14
 =  (x)/N = 4695/8 = 586.875 m/secs.

The fact that the arithmetic mean can be readily computed does not mean that it
is meaningful or even useful. Furthermore, the arithmetic mean has the
weakness of being unduly influenced by small, or unusually large, values in a
data set. For example: five subjects are observed in an experiment and display
the following reaction times: 120, 57, 155, 210 and 2750 m/secs. The arithmetic
mean is 658.4 m/secs, a figure that is hardly typical of the distribution of reaction
times.

Measures of Dispersion

The degree to which numerical data tend to spread about an average value is
called the variation or dispersion of the data. Various measures of dispersion or
variation are available such as range, mean deviation, semi-interquartile range,
and the standard deviation.

The Mean Absolute Deviation

The mean absolute deviation of a data set of N numbers, x1, x2, x3 . . . xn is

defined by:
Mean Deviation = [x - ]/N

Where  is the arithmetic mean of the data set and [x-] is the absolute value of
the deviation of x from (note that in mathematics, the absolute value of a
number is the number without the associated sign and is indicated by two vertical
lines: |-4| = 4, |+4| = 4).

Compute, and make observation on, the mean absolute deviation for the two
sets of data derived from sample (a) and sample (b):

(a) 2, 3, 6, 8, 11 (b) 2, 4, 7, 8, 9

(a) Arithmetic Mean = [2+3+6+8+11]/5 = 6

Mean Absolute Deviation = [|2-6|+|3-6|+|6-6|+|8-6|+|11-6|]/5 = 2.8

(b) Arithmetic Mean = [2+4+7+8+9]/5 = 6

Mean Absolute Deviation = [|2-6|+|4-6|+|7-6|+|8-6|+|9-6|]/5 = 2.4

Conclusion

The mean absolute deviation indicates that sample (b) shows less dispersion
than sample (a).

15
 Variance and Standard Deviation

The most useful measures of dispersion, and those with the most desirable
mathematical properties, are variance and standard deviation. The standard
deviation is particularly useful when dealing with the Normal Distribution, since
any Normal Distribution can be completely determined when the arithmetic mean
and standard deviation are known. When it is necessary to distinguish the
standard deviation of a population from the standard deviation of a sample drawn
from this population, we often use the symbol  for the former and s for the
latter. For ungrouped data the variance and standard deviation from a sample
can be determined from the following derived formulae:

(1) s= [x -]2/N (2) s2 = [x2/N] - 2

where, s2 = variance; = the arithmetic mean; [x - ]2 = the squared difference
between the observed value and the arithmetic mean; and N = the number of
observations. To compute the variance or standard deviation of a sample from a
population, the value for N should be substituted by [N - 1]; however, with large
samples, say >50, the difference is so small as to be of no significance. The
standard deviation is the square root of the variance. Formula (2) above is often
easier to compute for manual operations:

Find the standard deviation of each set of data: (a) 2, 3. 6, 8, 11 and (b) 2, 4,
7, 8, 11.

(a) Arithmetic Mean = [2+3+6+8+11]/5 = 6

Variance = [(22+32+62+82+112)/5] - 62 = 10.8

hence, the standard deviation = 10.80.5 = 3.28634

(b) Arithmetic Mean = [2+4+7+8+9]/5 = 6

Variance = [(22+42+72+82+92)/5] – 62 = 6.8

hence, the standard deviation = 6.80.5 = 2.60768

The above results should be compared with those of the mean absolute
deviation computed previously. It will be noted that the standard deviation
does indicate that the data set (b) shows less dispersion than data set (a).
However, the effect is masked by the fact that extreme values affect the
standard deviation much more than the mean absolute deviation. This of
course is to be expected since the deviations are squared in computing the
standard deviation.

16
For grouped data the arithmetic mean is computed as: = fmf and the
variance: s2 = f[m - ]2/f where m is the class midpoint.

The properties of the standard deviation for approximately normal, or moderately

skewed distributions, are:

1. 68.27 percent of observations (x) are included between the value of the
arithmetic mean and +/- one standard deviation

2. 95.45 percent of observations (x) are included between the value of the
arithmetic mean and +/- two standard deviations; and,

3. 99.73 per cent of observations (x) are included between value of the
arithmetic mean and +/- three standard deviations.

For moderately skewed distribution the above percentages may hold

approximately. These powerful properties are of pivotal importance to
probability and hypothesis testing in applying the Normal deviate, or its
derivatives. Furthermore, such measures may be said to represent the
distribution for which they were calculated. For instance, in the validation of
learning gain acquired by two similar training programmes the respective
properties are:

Traditional Programme Computer Based Programme

1= 63.500 2 = 75.950

1 = 15.220 2 = 11.400
n1 = 60 n2 = 40

Whilst the traditional training programme shows a lower mean score than that of
the computer-based programme, how can we compare the relative variability
bearing in mind that one shows a lower mean than the other?

The Pearson Coefficient of Variability can assist in such instances, where s is the
standard deviation and is the arithmetic mean of the sample:

Variability () = 100 s/


Hence, the traditional training programme has a variabilityof 23.97 per cent
whereas the computer-based training programme has a variability of 15.01 per
cent. We can conclude that not only is the traditional training programme more

17
variable in an absolute sense, but the variability of the observed values
expressed as a percentage are also greater.

Empirical Relations between Measures of Dispersion

The consequences for the normal distribution are that the mean deviation and
the semi-interquartile range (or median) are equal respectively to 0.7979 and
0.6745 times the standard deviation. For moderately skewed distributions we
have the empirical formulae:

Mean Deviation = 0.7979(standard deviation)

Median (Q) = 0.6745(standard deviation)

Example for Computation of Mean and Standard Deviation for

Ungrouped Data

Scores of Ten Sampled Subjects as at 31 March 20XX

Name Score (x) [x - ] [x - ]2 x2

Abrams 22.5 16.00 256.00 506.25

Brown 56.0 17.50 306.25 3136.00
Curtis 42.3 3.80 14.44 1831.84
Cuthbert 20.8 17.70 313.29 432.64
Dunning 33.4 5.10 26.01 1115.56
Easton 48.2 9.70 94.09 2323.24
Fallowes 50.0 11.50 132.25 2500.00
Gorman 37.4 1.10 1.21 1398.76
Graham 41.0 2.50 6.25 168.10
Harris 33.4 5.10 26.01 1115.56

  = 385.0 1175.80 15998.30

Arithmetic Mean () = x/N = 385/10 = 38.50

Variance (s=[x - ]2/N-1 = 1175.80/9 = 130.6444

hence, s = 130.64440.5 = 11.43

Note: Since the set of data is a sample drawn from a population, the number of observation
(N) has been adjusted to N-1.

18
 Standardized Variables and Standard Scores

The variable z = [x - ]/s

Which measures the deviation from the mean in units of the standard deviation is
called a standardized variable and is a dimensionless quantity, that is,
independent of the units used. If deviations from the mean are given in units of
the standard deviation, they are said to be expressed in standard scores or
standard units. These are of great value in comparison of distributions and are
often used in aptitude and educational testing.

In a test of aptitude, a subject received a score of 84 in a numerical test for

which a mean score was 76 and the standard deviation was 10. In the test for
verbal reasoning for which the mean score was 82 and the standard deviation
was 16, the subject received a score of 90. In which test was the subject‘s
relative standing higher?

Computation

The standardized variable z = [x-]/s measures the deviation of x from the

mean  in terms of standard deviation s.

For the numerical test, z = [84-76]/10 = 0.8

For the verbal reasoning test, z = [90-82]/16 = 0.5

Thus, the subject had a score 0.8 of a standard deviation above the mean for
the numerical test, but only 0.5 of a standard deviation above the mean in
verbal reasoning. Thus, the subject‘s standing was higher in numerical
aptitude.

Miscellaneous Averages

19
 Weighted Arithmetic Mean

It is not always the case that values of a variable (x) are of equal importance, for
example in the assessment of a learning event. To compute the arithmetic
mean certain weights (w) may be attached to the variable:

= [w1x1 + w2x2 + w3x3 + . . . wnxn ] / [w1 + w2 + w3 + . . . wn ]

= wjxj[1/x]

Assessment Score (x) Weight (w) wx

Test of knowledge I 65 0.20 13

Practical assignment I 70 0.30 21
Test of knowledge II 60 0.30 18
Practical assignment II 60 0.20 12
Total: 1.00 64

Geometric Mean

It is possible to average observations of data on the ratio scale by means of the

geometric mean (G). This is particularly useful in the computation of the
average increase in index numbers or percentiles. The geometric mean can be
defined as the nth root of the product of n values of a variable (x):
1/n
G = [x1 * x 2 * x3 * . . . xn]

For large values of X, the geometric mean can be most easily computed by the
application of logarithms10

Log10G = log10x]1/n

Period end Number of Percentage increase over Log10x

year subjects previous period (x)
1950 138 - 0.000000
1960 292 211.594 2.324952
1970 385 131.849 2.120078
1980 596 154.805 2.189786
1990 938 157.383 2.196957
2000 1233 131.450 2.118760
Total 10.950533
hence,

20
 Log10G = 10.950533[1/5] = 2.1901066

G = 154.9196831 say, 155 percent increase over each ten year period.

Alternatively, a rough and ready estimate can be derived from the following:
1/n
G = [no. of subjects at end period /no. of subjects at beginning period]

1/5
= [1233/138]

= 154.9589975 percent

But this increase is for each period of ten years, to obtain the annual rate of
increase over the total of fifty years:
1/50
G = [1233/138] -1

= 1.044772373 – 1

= 4.47724 percent a year

The nature of the geometric mean is that if any value for x is zero, then the
product will be zero. It is also meaningless of any of the values are negative.
This can be overcome by expressing the values of x with care, and selecting the
values of x to be averaged.

Harmonic Mean

The harmonic mean (H) is useful for solving problems that involve variables
expressed in time, such as reaction times, number of errors per hour, etc. The
harmonic mean can be computed:

H = N/[1/x]

Subject Minutes to complete task Number of tasks completed per hour

(n) (x)
104/64 2.5 24
104/74 2.0 30
104/77 1.5 40
104/85 6.0 10
Total: 12.0 104

The arithmetic mean would be 3 minutes to complete the task (12 minutes
divided by 4 subjects), which on average is 20 tasks per hour/subject or 80 tasks

21
per hour for all four subjects. However, when computed separately for each
subject, the total number of tasks completed per hour is 104. The total number
of tasks completed per hour for all four subjects must be computed some other
way:

H = N/[1/x1 + 1/x2 + 1/x3 + 1/x4 . . . 1/xn] = 4/[0.4+0.5+0.67+0.17]

= 2.31 minutes

Since the four subjects participated in the experiment for one hour, representing
240 minutes of running time, the average number of tasks completed by the
group of four subjects is:

240 minutes/2.31 minutes per task = 104 tasks per hour

22
 III
Descriptive Statistics

PART A: DESCRIPTIVE STATISTICS

Data

Data are very important for scientific study, and statistics is a discipline that deals
with the collection, presentation and analysis of data. In this chapter we are going
to study how we can summarize and describe a set of data. When we study a set
of data we need to identify the following important characteristics of the dataset.

 Primary and secondary data. When the data are collected by us it is called
primary data. We always have the individual values of the data. When the
dataset is collected by others, it is called secondary data. Sometimes the
data is grouped into a table, and is called grouped data.

 Population and sample data. Population refers to the totality of elements in

which we are interested. Suppose we want to study the salary of Hong
Kong people, our population includes all those persons who work in Hong
Kong. However as the population is so big that it is not practical and
economical to collect salary data of all the working people, we always
select randomly only a subset of the population and the data is sample.

 Discrete and continuous data. It is important to identify whether the data is

continuous or discrete. For example data on the number of persons in a
household is discrete, and data on salary is continuous. Different
statistical techniques are used for handling discrete or continuous data.

Frequency Distribution

Statistical data obtained by means of census, sample surveys or experiments

usually consist of raw, unorganized sets of numerical values. Before these data
can be used as a basis for inferences about the phenomenon under investigation
or as a basis for decision, they must be summarized and the pertinent
information must be extracted.

23
Example 1

A random sample of 100 households in a town was selected and their monthly
town gas consumption (in cubic metres) in last month were recorded as follows:

55 82 83 109 78 87 95 94 85 67
80 109 83 89 91 104 90 103 67 52
107 78 86 29 72 66 92 99 60 75
88 112 97 88 49 62 70 66 88 62
72 85 81 78 77 41 105 92 94 74
78 75 87 83 71 99 56 69 78 60
119 39 104 86 67 79 98 102 82 91
7
46 120 73 125 132 86 48 55 112 28
42 24 130 100 46 57 31 129 137 59
102 51 135 53 105 110 107 46 108 117

A useful method for summarizing a set of data is the construction of a frequency

table, or a frequency distribution. That is, we divide the overall range of values
into a number of classes and count the number of observations that fall into each
of these classes or intervals.

The general rules for constructing a frequency distribution are:

(i) There should not be too few or too many classes.

(ii) Insofar as possible, equal class intervals are preferred. But the first and
last classes can be open-ended to cater for extreme values.

In example 1, the sample size is 100 and the range for the data is 113 (137 - 24).
A frequency distribution with six classes is appropriate and it is shown below.

24
 Frequency distribution of household town gas consumption

Town gas monthly

consumption Number of
( in cubic metres) households
20 - 39 5
40 - 59 15
60 - 79 25
80 - 99 30
100 - 119 18
120 - 139 7
Total 100

Class limits: are the numbers that typically serve to identify the classes in a
listing of a frequency distribution. Thus, in the above frequency distribution, for
the class whose frequency is 30, its lower class limit is 80 and upper class limit is
99.

As contrasted to a class limit, a class boundary is the precise point that

separates one class from another, rather than being a value indicated in one of
the classes. A class boundary is typically located midway between the upper limit
of a class and the lower limit of the next higher class adjoining it. Therefore the
class boundary separating the class 60-79 and the class 80-99 is halfway
between 79 and 80, that is, at the point 79.5.

Class interval: is the width of a class. The class interval of a class is computed
by subtracting the lower limit (boundary) of the class from the lower limit
(boundary) of the next class.

Class midpoint or class mark: is the point dividing the class into equal halves
on the basis of class interval. This point can be obtained by adding the lower and
upper limits (boundaries) of a class and dividing by 2.

Relative frequency of a class: is the frequency of the class divided by the total
frequency of the distribution.

Cumulative frequency distribution: shows the number of items of a series that

are less than (or more than) certain specified values.

25
Measure of Central Tendency

A value that would describe the 'centre' of a distribution would be visually located
near the spot where most of the data seem to be concentrated. Consequently,
values that fulfil this role are called measures of central tendency.

The most common measures of the central tendency of a data set are arithmetic
mean or simply as mean, median and mode.

The mean of a set of numerical data is the sum of the set divided by the number
of observations, that is, their average.

The median of a distribution is the value which divides the distribution so that an
equal number of values lie on either side of it, i.e., half of the items have values
smaller or equal to it and half of the items have values larger or equal to it.

The mode of a set of numerical data is the value which occurs most frequently.

Example 1 (calculating mean, median and mode for individual data)

The following table shows the hourly wage rates of eight sampled construction
workers.

Worker i 1 2 3 4 5 6 7 8
Hourly wage
rate ( x i ) $35 38 46 60 65 69 72 78

x i
x1  x 2  x3  x 4  x5  x6  x7  x8
Mean ( x )  i 1
( )
8 8

463
  57.875 ($)
8

n 1 9
Location of the median:   4.5 th
2 2

x 4  x5 60  65
Median =   62.5 ($)
2 2

Mode: the sample size is too small, mode cannot be identified.

 Research Methods and Statistics

Example 2 (calculating mean, median and mode for grouped data)

The following table shows the daily wages of a random sample of construction
workers. Calculate its mean, median and mode.

Daily Wages ($) Number of Workers

200 - 399 5
400 - 599 15
600 - 799 25
800 - 999 30
1000 - 1199 18
1200 - 1399 7
Total 100

Solution

Number of
Daily Wages ($) Workers Class Mark f i xi
fi xi

200 - 399 5 299.5 1,497.5

400 - 599 15 499.5 7,492.5
600 - 799 25 699.5 17,489.5
800 - 999 30 899.5 26,985.5
1000 - 1199 18 1,099.5 19,791.0
1200 - 1399 7 1,299.5 9,096.5
Total 100 82,350.0

fx i i
82 ,350 .0
Mean ( x )  i 1
6
  823 .5 ($
f
100
i
i 1

2
 Research Methods and Statistics

Daily Wages ($) Number of Workers Cumulative

fi Frequency
Fi

200 - 399 5 5
400 - 599 15 20
600 - 799 25 45
800 - 999 30 75
1000 - 1199 18 93
1200 - 1399 7 100
Total 100

As 0.5n = 0.5(100) = 50, so the median lies in the 4th class.

0.5n  F3
Median = L4  (c 4 ) where L is the lower class boundary,
f4
c is the class interval.

0.5(100)  45
 799.5  (200)  832.8 ($)
30

Number of Class Relative Density

Daily Wages ($) Workers Interval fi
fi'  ( 200 )
fi ci ci

200 - 399 5 200 5

400 - 599 15 200 15
600 - 799 25 200 25
800 - 999 30 200 30
1000 - 1199 18 200 18
1200 - 1399 7 200 7
Total 100

As f 4'  30 is the largest relative density, so mode lies in the 4th class.

f 4'  f 3'
Mode  L4  (c4 )
( f 4'  f 5' )  ( f 4'  f 3' )

30  25
 799 .5  (200 )  858 .3 ($)
(30  18 )  (30  25 )

3
 Research Methods and Statistics

Advantages and disadvantages of each measure

Mean
Advantages: (i) All values in the distribution are used in its calculation, so
it can be regarded as more representative than the other
two measures.

(ii) Its method of calculation is simple and most people

understand the meaning of its result.

(iii) Its result can easily be used in further analysis.

Disadvantages: (i) Its result can be easily distorted by extreme values. As

such, its result may be rather lower or higher than the
bulk of the values and becomes unrepresentative.

(ii) In case of open end classes, mean can be calculated

only if their class marks are determined. If such classes
contain a large proportion of the values, then the mean
may be subjected to substantial error.

Median
Advantage: Its result will not be affected by extreme values and open
end classes.

Disadvantage: It has to be supplemented by other statistics because it

does not reflect the distribution in the way that the mean
does, that is, including all values.

Mode
Advantages: (i) Its result will not be affected by extreme values and open
end classes.

(ii) If data are not grouped, it can be determined easily.

Disadvantages: (i) It has to be supplemented by other statistics.

(ii) It is difficult to obtain an accurate estimate of the mode if

the values are classified into a frequency distribution.

4
 Research Methods and Statistics

How to select a suitable measure

(i) Always select the mean whenever there is no special reason for choosing
the other two measures.

(ii) Select the median is the distribution consists of substantial amount of

extreme large or small values.

(iii) Select the mode if integral result is preferred as in cases the data are in
ordinal scales.

5
 Research Methods and Statistics

Measure of data variation (variability)

A measure of central tendency is almost never, by itself, sufficient to provide an

adequate summary of the characteristics of a set of data. We will usually require,
in addition, a measure of the amount of variation in the data.

Example 1

Consider the following measurements, in grams, for two samples of strawberry

jam bottled by companies A and B:

Sample for 31 32 32 33 32
Company A
Sample for 28 29 32 35 36
Company B

Both samples have the same mean, 32 grams. It is obvious that company A, in
comparison with company B, bottles strawberry jam with a more consistent
content. We say that the variability of the observations is smaller for company A.
Therefore in buying strawberry jam we would feel more confident that the bottle
we select will be closer to the advertised average content if we buy from
company A.

The most important measures of variability or dispersion are the range, mean
deviation, standard deviation and variance.
(There are some other measures like quartile deviation and percentiles. We shall
not study these measures. Read our textbook if interested)

The range of a set of numbers is the difference between the largest and the
smallest number in the set.

Example 2 (For individual data)

The following table shows the hourly wage rates of eight sampled construction
workers.

Worker i 1 2 3 4 5 6 7 8
Hourly wage
rate ( x i ) $35 38 46 60 65 69 72 78

The range is $78 - $35 = $43.

6
 Research Methods and Statistics

Though range is simple and can be obtained easily, its result is unstable. This is
particularly true if the sample size is large. So whenever the sample size is over
10, we seldom choose to use range to indicate variability of the data.

Mean deviation is the average of the absolute deviation of the numerical data
from their mean.

Worker i 1 2 3 4 5 6 7 8
Hourly wage
rate ( x i ) $35 38 46 60 65 69 72 78
xi  x
22.87 19.87 11.87 2.125 7.125 11.12 14.12 20.12
 xi  57.875 5 5 5 5 5 5

x i  57.875
109.25
Mean deviation i 1
  13.656 ($)
8 8

The mean deviation is a good measure to show the extent of variation of the data
in a distribution. However, when this measurement is used in further analysis, it
would give rise to some unnecessary tedious mathematical problem as a result
of its absolute value term. To avoid this pitfall, we can use the standard deviation
instead.

Standard deviation of a population (  ) is the square root of the average of the

squared distances of the observations from the mean.

 (x i  )2
 i 1
, where  is the population mean
N

To compute the sample standard deviation (s) we use the above formula,
replacing  by x and N by n  1 .

 (x  x )
i
2

s i 1

n 1

7
 Research Methods and Statistics

Worker i 1 2 3 4 5 6 7 8 Total
Hourly wage
rate ( x i ) $35 38 46 60 65 69 72 78 463

( x  35.875) 2
s  16.226($)
7

Variance is the square of the standard deviation.

 (x  x )
i
2

s2  i 1

n 1

Example 3 (for grouped data)

The following table shows the daily wages of a random sample of construction
workers. Calculate its mean deviation, variance, and standard deviation.

Daily Wages ($) Number of Workers

200 - 399 5
400 - 599 15
600 - 799 25
800 - 999 30
1000 - 1199 18
1200 - 1399 7
Total 100

8
 Research Methods and Statistics

Solution

Number of
Daily Wages ($) Workers Class Mark f i xi  x
fi xi  f i xi  823.5

200 - 399 5 299.5 2,620

400 - 599 15 499.5 4,860
600 - 799 25 699.5 3,100
800 - 999 30 899.5 2,280
1000 - 1199 18 1,099.5 4,968
1200 - 1399 7 1,299.5 3,332
Total 100 21,160

f i xi  x
21,160
Mean deviation  i 1
6
  211 .60 ($)
f
100
i
i 1

Number of
Daily Wages Workers Class Mark fi ( xi  x)2
($) fi xi

200 - 399 5 299.5 1, 372,880

400 - 599 15 499.5 1,574,640
600 - 799 25 699.5 384,400
800 - 999 30 899.5 173,280
1000 - 1199 18 1,099.5 1,371,168
1200 - 1399 7 1,299.5 1,586,032
Total 100 6,462,400

6462400
Variance ( s 2 )   65, 276.77
99

Standard deviation =
65276.77  255.49

9
 Research Methods and Statistics

Comparison of the variation of two distributions

The values of the standard deviations cannot be used as the bases of the
comparison because:
(a) units of measurements of the two distributions may be different,
and
(b) average values of two distributions may be widely dissimilar.
The correct measure that should be used is the coefficient of variation (CV ) .
s
CV  100%
x
Example 4
The following table shows the summary statistics for the daily wages of two types
of workers.

Worker's Daily Wages

Type Mean Standard deviation
I $100 $20
II $150 $24

Compare these two daily wages distributions.

Solution
In comparison Distribution Reason
Average
magnitude II > I x II  150  x I  100

Variation I > II 20 24
CVI  100%  20%  CVII  100%  16%
100 150

10
 Research Methods and Statistics

PART B: Probability

Introduction and concepts

―Perhaps it was man‘s unquenchable thirst for gambling that led to the early
development of probability theory. In an effort to increase their winnings,
gamblers called upon the mathematicians to provide optimum strategies for
various games of chance.‖ ---- from Walpole R.E. Introduction to Statistics

Probability is the basis upon which the discipline of statistics has been developed
and applied in many fields associated with chance occurrences such as politics,
business, weather forecasting, and scientific research. Probability may be taken
as a tool with which we may solve problems involving uncertainties. In fact
uncertainty is a basic element of human experiences. To cite some examples:
travelling time, number of customers, rainfall, temperature, share price
movement, length of our life, etc.

There are three approaches to understand probability. In the empirical approach,

probability may be taken as a relative frequency. As such the probability of an
aeroplane arriving its destination on time may be taken as the proportion of times
the aeropline has been on time in the past, say, one thousand times.

Suppose in a trial of an experiment, there are k possible outcomes which are

equally likely. The probability of the occurrence of an outcome is therefore 1/k.
Thus in throwing a coin, the probability of having a head is ½. In our course, we
shall adopt this approach but the empirical approach is always useful in giving us
some intuition to understand the problem.

The third approach is very mathematical. A number of axioms have been set up
and from these some theorems of probability have been developed. This
approach is too abstract and usually used by mathematicians.

Some Basic Concepts

Sample space: is a set of all possible outcomes of an experiment.

Event: is a subset of a sample space.

To find the probability of an event we need to count the number of outcomes of

the event and the number of all possible outcomes of the experiment, and then to
divide the former by the latter. Hence the following counting rules may be helpful.

11
 Research Methods and Statistics

Some counting rules

Example 1

Three items are selected at random from a manufacturing process. Each item is
inspected and classified defective (D) or non-defective (N).

DDD DDN DND NDD

Its sample space is =  
DNN NDN NND NNN 

Example 2

The event that the number of defectives in above example is greater than 1.

Its sample space is = {DDD DDN DND NDD}

The probability of the event is 4/8 or ½.

Example 3

Suppose a licence plate containing two letters following by three digits with the
first digit not zero. How many different licence plates can be printed?

1st 2nd 1st 2nd 3rd

Letter Letter Digit Digit Digit

Number of A-Z A-Z 1-9 0-9 0-9

Choices (26) (26) (9) (10) (10)

Number of different licence plates that can be printed is

(26)(26)(9)(10)(10) = 608,400

Example 4

Find the possible permutations (the number of ways where sequence of the
letters is counted) from 3 letters A, B, C.

 The number of permutations of n distinct objects is

12
 Research Methods and Statistics

n Pn  (n)( n  1)( n  2).....( 2)(1)  n!

 The number of permutations of n distinct objects taken r at a time is

n Pr  (n)( n  1)...( n  r  2)( n  r  1)


(n)( n  1)...( n  r  2)( n  r  1)(n  r )( n  r  1)...( 2)(1)   n!
(n  r )( n  r  1)...( 2)(1) (n  r )!

e.g. The number of 3-letter words formed from 5 letters is

5!
P3  60
(5  3)!
5

 The number of distinct permutations of n objects of which n1 are alike of

the first kind, n 2 are alike of the second kind, ....., n k are alike of the kth
n!
kind and n1  n2  ...  nk  n is
(n1!)( n 2 !)...( n k !)

Find the possible permutations of the following 5 letters: A A A B C

There are five objects of which three are alike.

P 5!
The answer  5 5 
3! 3!

Example 5

How many 7-letter words can be formed using the letters of the word
'BENZENE'?

(there are 1 B, 3 E, 2 N and 1 Z)

7!
The number of 7-letter words that can be formed is  420
(1!)(3!)( 2!)(1!)

 The number of combinations (number of ways where sequence is not

counted) of n distinct objects taken r at a time is

n!
Cr 
r!(n  r )!
n

13
 Research Methods and Statistics

Find the possible combinations of 5 distinct objects taken 3 at a time.

The answer
5!

3!(5  3)!

Example 6

The number of 3-person committees that can be formed from a group of 4

persons is

4!
C3  4
3!(4  3)!
4

Example 7

A box contains 8 eggs, 3 of which are rotten. Three eggs are picked at random.
Find the probabilities of the following events.
(a) Exactly two eggs are rotten.
(b) All eggs are rotten.
(c) No egg is rotten.

Solution:

(a) The 8 eggs can be divided into 2 groups, namely, 3 rotten eggs as the first
group and 5 good eggs as the second group.

Getting 2 rotten eggs in 3 randomly selected eggs can occurred if we

select randomly 2 eggs from the first group and 1 egg from the second
group.

The number of this outcome is  3 C 2  5 C1   15

Total number of possible outcomes of selecting 3 eggs randomly from the
total 8 eggs is 8 C 3  56 .

Thus the probability of having exactly two rotten among the 3 randomly
 C  C  15
selected eggs is 3 2 5 1 
8 C3 56

14
 Research Methods and Statistics

(b) Similarly, the probability of having all 3 rotten eggs is

 3 C 3  5 C 0  1

8 C3 56

(c) The probability of having no rotten egg is

 3 C 0  5 C 3  10 5
 
8 C3 56 28

Rules of probability

The following rules may help us to find the probability of an event.

Addition Rule: For any events that are not mutually exclusive

P( A  B)  P( A)  P( B)  P( A  B)

where A B is the union of two sets A and B, it is the set of elements that
belong to A or to B or to both.

A B is the intersection of two sets A and B, it is the set of

elements that are common to A and B.

Illustrative example

180 students took examinations in English and Mathematics. Their results were
as follows:

Number of students passing English = 80

Number of students passing Mathematics = 120
Number of students passing at least one subject = 144

Then we can rewrite the above results as:

80 4
Probability that a randomly selected student passed English = 
180 9
120 2
Probability that a randomly selected student passed Mathematics  
180 3

15
 Research Methods and Statistics

Probability that a randomly selected student passed at least one subject

144 4
 
180 5
Find the probability that a randomly selected student passed both subject.
Solution
Let E be the event of passing English, and M be the event of passing
Mathematics.
4 2 4
It is given that: P( E )  ; P( M )  ; P( M  E ) 
9 3 5

As P(M  E )  P( E )  P(M )  P(M  E )

4 2 4 14
 P( M  E )  P( E )  P( M )  P( M  E )      0.31
9 3 5 45

Example 8
A card is drawn from a complete deck of playing cards. What is the probability
that the card is a heart or an ace?

Solution
Let A be the event of getting a heart, and B be the event of getting an ace.
The probability that the card is a heart or an ace is P( A  B) .
P( A  B)  P( A)  P( B)  P( A  B)
13 4 1 16 4
    
52 52 52 52 13

For mutually exclusive events, P( A  B)  P( A)  P( B)

What is the probability of getting a total of '7' or '11' when a pair of dice are
tossed?
Solution
Total number of possible outcomes = (6)(6) = 36

16
 Research Methods and Statistics

Possible outcomes of getting a total of '7' :{1,6; 2,5; 3,4; 4,3; 5,2; 6,1}
Possible outcomes of getting a total of '11' : {5,6; 6,5}
Let A be the event of getting a total of '7', and B be the event of getting a total of
'11'.
The probability of getting a total of '7' or '11' is P( A  B) .
P( A  B)  P( A)  P( B)  P( A  B)  P( A)  P( B) ...A and B are mutually exclusive
6 2 2
  
36 36 9

If A and A' are complementary events then P( A)  1  P( A' )

Example 9
A coin is tossed six times in succession. What is the probability that at least one
head occurs?

Let A be the number of heads occurs in six successive tosses.

P( A  1)  1  P( A  0)

 1 1 21212121212  64
63

Conditional Probability
Let A and B be two events. The conditional probability of event A given that event
B has occurred, denoted by P( A / B) is defined as

P( A  B)
P( A / B)  provided that P(B) > 0.
P( B)
Similarly, the conditional probability of B given that event A has occurred is
P( A  B)
defined as P( B / A)  , provided P(A) > 0.
P( A)
Example 10

17
 Research Methods and Statistics

A hamburger chain found that 75% of all customers use mustard, 80% use
ketchup, and 65% use both, when ordering a hamburger. What are the
probabilities that:
(a) a ketchup-user uses mustard?
(b) a mustard-user uses ketchup?

Solution
Let A be the event of using mustard, and B be the event of using ketchup.
It is given that: P( A)  0.75 ; P( B)  0.80 ; P( A  B)  0.65
P( A  B) 0.65
(a) P(a ketchup-user uses mustard)  P( A / B)    0.8125
P( B) 0.80

P( A  B) 0.65
(b) P(a mustard-user uses ketchup)  P( B / A)    0.8667
P( A) 0.75

Multiplicative Rule
P( A  B)  P( A) P( B / A)
or = P( B) P( A / B)
Statistically Independence: the occurrence or non-occurrence of one event has
no effect on the probability of occurrence of the other event.

Two events A and B are independent if and only if P( A  B)  P( A) P( B)

Example 11
A pair of fair dice are thrown twice. What is the probability of getting totals of 7
and 11?

Solution
Let Ai be the event of getting '7' in the i-th throw and B j be the event of getting
'11' in the j-th throw.

P(Getting totals of 7 and ll)  P( A  B)  P( A1  B2 )  P( B1  A2 )

 P( A1 ) P( B2 / A1 )  P( B1 ) P( A2 / B1 )

 P( A1 ) P( B2 )  P( B1 ) P( A2 ) .... Ai , B j are independent

18
 Research Methods and Statistics

 6  2   2  6  1
        
 36  36   36  36  54

Theorem of Total Probability

If the events B1 , B2 ,..., Bk constitute a partition of the sample space S such that
P ( Bi  0) for i = 1, 2, ... , k, then for any event A of S
P( A)  P( A  B1 )  P( A  B2 )  ...  P( A  Bk )

 P( B1 ) P( A / B1 )  P( B2 ) P( A / B2 )  ...  P( Bk ) P( A / Bk )

Example 12
Suppose 50% of the cars are manufactured in the United States and 15% of
these are compact; 30% of the cars are manufactured in Europe and 40% of
these are compact; and finally, 20% are manufactured in Japan and 60% of
these are compact. If a car is picked at random from the lot, find the probability
that it is a compact.

Let A be the event that the car is compact,

B1 be the event that the car is manufactured is United States,
B 2 be the event that the car is manufactured in Europe, and
B3 be the event that the car is manufactured in Japan.

P( A)  P( A  B1 )  P( A  B2 )  P( A  B3 )

 P( B1 ) P( A / B1 )  P( B2 ) P( A / B2 )  P( B3 ) P( A / B3 )

 (0.50)(0.15)  (0.30)(0.40)  (0.20)(0.60)  0.315

Baye's Theorem
If E1, E 2 ,..., E k are mutually exclusive events such that E1  E 2  ...  E k contains

all sample points of S, then for any event D of S with P( D)  0 ,

P ( E i  D) P ( Ei  D)
P ( E i / D)   k

 P( E
P ( D)
j  D)
j 1

P( Ei ) P( D / Ei )

P ( E1 ) P ( D / E1 )  P ( E 2 ) P ( D / E 2 )  ... P ( E k ) P ( D / E k )

19
 Research Methods and Statistics

Example 13
Suppose a box contains 2 red balls and 1 white ball and a second box contains 2
red ball and 2 white balls. One of the boxes is selected by chance and a ball is
drawn from it. If the drawn ball is red, what is the probability that it came from the
1st box?

Solution

Let A be the event of drawing a red ball and B be the event of choosing the 1st
box.
1 2 2
Given: P( B)  P( B' )  ; P( A / B)  ; P( A / B' ) 
2 3 4
P(Coming from the 1st box/the drawn ball is red)  P( B / A)
P( A  B) P( A  B)
 
P( A) P( A  B)  P( A  B' )

P( B) P( A / B) ( 1 )(2 ) 4
  2 3 
P( B) P( A / B)  P( B' ) P( A / B' ) ( 1 )(2 )  ( 1 )(2 ) 7
2 3 2 4

20
 Research Methods and Statistics

PART C: Probability Distributions

To cope with uncertainties of outcome, a statistical model that describes the

behavior of the outcome is needed. These theoretical models which are very
similar to relative frequency distributions, are called probability distributions.

Random Variables - A random variable is a variable that takes on different

numerical values determined by the outcomes of a random experiment.

Example 1

An experiment of tossing a coin 3 times.

Let random variable, X be the number of heads achieved.

As S = {HHH HHT HTH THH TTH THT HTT TTT},

so X = {0, 1, 2, 3}

Discrete random variable - in a given interval, only a specified number of

values can occur.

Continuous random variable - in a given interval, any value can occur.

Probability Distribution of a random variable - is a representation of the

probabilities for all the possible outcomes.

Example 2

The probability distribution of the number of heads occurred when a coin is

tossed 4 times.

x 0 1 2 3 4
1 4 6 4 1
P(X=x) 16 16 16 16 16

Cx
That is, P( X  x)  4
, x  0, 1, 2, 4
16

21
 Research Methods and Statistics

Example 3

Consider an experiment of tossing two fair dice.

Let random variable, X be the sum of the two dice. Then the probability
distribution of X is:

x 2 3 4 5 6 7 8 9 10 11 12
1 2 3 4 5 6 5 4 3 2 1
P(X=x) 36 36 36 36 36 36 36 36 36 36 36

The probability function f (x) , of a discrete random variable X expresses the

probability that X takes the value x, as a function of x. That is

f ( x)  P( X  x)
where the function is evaluated at all possible values of x.

Properties of probability function P( X  x) :-

1. P( X  x)  0 for any value x.

2.  P ( X  x )  1.
x

Mathematical Expectations

The expected value, E(X), of a discrete random variable X is defined as

E ( X ) or  x   xP( X  x)
x

It is the mean of the probability distribution.

Let X be a random variable. The expectation of the squared discrepancy about

the mean, (X  x)2, is called the variance, denoted x2, and given by

Var ( X ) or  x  E [( X   x ) 2 ]
2

  ( x   x ) 2P( X  x)
x

  x 2 P( X  x)   x
2

22
 Research Methods and Statistics

Example 4
Calculate the mean and variance of the discrete probability distribution in
example 2 and 3.

The Normal Distribution

Normal distribution is probability distribution of a continuous random variable. It is

based on the Law of Errors which states that

1. Errors are inevitable.

2. Large errors are less likely than small errors.
3. Positive and negative errors are equally likely.

Definition :

A continuous random variable X is defined to be a normal random variable if its

probability function is given by

1 1 x 2
f (x )  exp[ ( ) ] for  < x < 
 (2 ) 2 

where  = the mean of X,  = the standard deviation of X,

 = 3.14154

Notation : X ~ N(, 2)

Properties of the normal distribution:-

1. It is a continuous distribution.

2. The curve is symmetric and bell-shaped about a vertical axis through the
mean .

3. The total area under the curve and above the horizontal axis is equal to 1.

4. Area under the normal curve:

 Approximately 68% of the values in a normally distributed population

are within 1 standard deviation from the mean.

23
 Research Methods and Statistics

 Approximately 95.5% of the values in a normally distributed

population are within 2 standard deviation from the mean.

 Approximately 99.7% of the values in a normally distributed

population are within 3 standard deviation from the mean.

The standard normal curve :

The distribution of a normal random variable with  = 0 and  =1 is called a

standard normal distribution. Usually a standard normal random variable is
denoted by Z.

Notation : Z ~ N(0, 1)

Remark : Usually a table of Z is set up to find the probability P(Z  z) for z  0.

Example 7

Given Z ~ N(0, 1)

(a) P(Z > 1.73) = 0.0418

(b) P(0 < Z < 1.73) = P(Z > 0) - P(Z > 1.73) = 0.5 - 0.0418 = 0.4582

(c) P(2.42 < Z < 0.8) = 1 - P(Z < -2.42) - P(Z > 0.8)
= 1 - 0.00776 - .2119 = 0.78034

(d) P(1.8 < Z < 2.8) = P(Z > 1.8) - P(Z > 2.8) = 0.0359 - 0.00256 = 0.03334

(e) the value z that has

(i) 5% of the area below it

Let the corresponding z value be z1, then we have P(Z < z1) = 0.05.

From the standard normal distribution table we have P(Z < -1.64) = 0.05.

So z1 = -1.64

(ii) 39.44% of the area between 0 and z.

Let the corresponding z value be z1, then we have P(0 < Z < z1) =
0.3944.

From the standard normal distribution table we have P(0 < Z < 1.25) =
0.3944.

24
 Research Methods and Statistics

So z1 = 1.25

Theorem :

If X is a normal random variable with mean  and standard deviation , then

X 
Z


is a standard normal random variable and hence

x1   x2  
P( x1  X  x2 )  P( Z )
 

Example 8

Given X ~ N(50, 102), find P(45 < X < 62).

 45   X   62   
Solution: P(45  X  62 )  P   
    

 45  50 62  50 
 P Z 
 10 10 

 P(0.5  Z  1.2)

 1  P(Z  0.5)  P(Z  1.2)

= 1 - 0.3085 - .1151 = 0.5764

Example 9

The charge account at a certain department store is approximately normally

distributed with an average balance of $80 and a standard deviation of $30.
What is the probability that a charge account randomly selected has a balance

(a) over $125;

(b) between $65 and $95.

Let X be the balance in the charge account. X  N(80, 30 2 )

25
 Research Methods and Statistics

X  125  
(a) P( X  125)  P(  )
 

125  80
 P( Z  )  P( Z  1.5)  0.0668
30

 65   X   95   
(b) P(65  X  95 )  P   
    
 65  80 95  80 
 P Z 
 30 30 

 P(0.5  Z  0.5)  1  P(Z  0.5)  P(Z  0.5)

 1  0.3085  0.3085  0.3830

Example 10

On an examination the average grade was 74 and the standard deviation was 7.
If 12% of the class are given A‘s, and the grades are curved to follow a normal
distribution, what is the lowest possible A and the highest possible B?

Let X be the examination grade and x1 be the lowest grade for A.

 x  74 
P( X  x1 )  0.12  P Z  1   0.12
 7 

From the standard normal distribution, we get

P(Z  1.17)  0.1210 , and P(Z  1.18)  0.1190

so P(Z  1.175)  0.12

x1  74
Thus  1.175
7

i.e. x1  74  (7)(1.175 )  82 .2  83

The highest possible B is 82.

26
 Research Methods and Statistics

The Binomial Distribution

A binomial experiment possesses the following properties :

1. There are n identical observations or trials.
2. Each trial has two possible outcomes, one called ―success‖ and the other
―failure‖. The outcomes are mutually exclusive and collectively
exhaustive.
3. The probabilities of success p and of failure 1  p remain the same for all
trials.
4. The outcomes of trials are independent of each other.

Example 11

1. In testing 10 items as they come off an assembly line, where each test or
trial may indicate a defective or a non-defective item.

2. Five cards are drawn with replacement from an ordinary deck and each
trial is labelled a success or failure depending on whether the card is red
or black.

Definition :

In a binomial experiment with a constant probability p of success at each trial, the

probability distribution of the binomial random variable X, the number of
successes in n independent trials, is called the binomial distribution.

Notation : X ~ b(n, p)

 n x n  x
P(X = x) =  p q x = 0, 1, , n
 x
p+q=1

Example 12

Of a large number of mass-produced articles, one-tenth is defective. Find the

probabilities that a random sample of 20 will obtain

(a) exactly two defective articles;

(b) at least two defective articles.

27
 Research Methods and Statistics

Let X be the number of defective articles in a random sample of 20. X  b(20,

1
)
10

 20  1 
2 18
9
(a) P( X  2)        0.28517
 2  10   10 

P( X  2)  1  P( X  0)  P( X  1)
 20  1   9   20  1  9 
0 20 19
(b)
 1             1  .12158  0.27017  0.60825
 0  10   10   1  10  10 

Example 13

A test consists of 6 questions, and to pass the test a student has to answer at
least 4 questions correctly. Each question has three possible answers, of which
only one is correct. If a student guesses on each question, what is the probability
that the student will pass the test?

1
Let X be the no. of correctly answered questions among 6 questions. X  b(6, )
3
 6
  2 3 
6 6 6 x
P( X  4)   P( X  x)     1
x

x4 x 4  x 
3

  2 3    65 13  2 3    66 13  2 3 

 6
   1
4 2 5 1 6 0
 0.10014
 4 3    

Theorem

The mean and variance of the binomial distribution with parameters of n and p
are
 = np and 2 = npq respectively where p + q = 1.

Example 14

A packaging machine produces 20 percent defective packages. A random

sample of ten packages is selected, what are the mean and standard deviation of
the binomial distribution of that process?

Let X be the no. of defective packages in a sample of 10 packages. X  b(10,

0.2)

Its mean is  = np = (10)(0.2) = 2

28
 Research Methods and Statistics

Its standard deviation is   npq  (10)(0.2)(0.8)  1.265

The Normal Approximation to the Binomial Distribution Theorem :

Given X is a random variable which follows the binomial distribution with

parameters n and p, then

(x  05
. )  np (x  05
. )  np
P( X  x)  P( Z )
(npq ) (npq )

if n is large and p is not close to 0 or 1.

Remark : If both np and nq are greater than 5, the approximation will be good.

Example 15

A process yields 10% defective items. If 100 items are randomly selected from
the process, what is the probability that the number of defective exceeds 13?

Let X be the no. of defective in a random sample of 100 items. X  b(100, 0.1)

  np  (100)(0.1)  10 ,   npq  (100)(0.1)(0.9)  3

P( X  13)  P( X '  13.5) by normal approximation

 X '  13 .5     13 .5  10 
 P    P Z    P( Z  1.167 )  0.121
     3 

Example 17

A multiple-choice quiz has 200 questions each with four possible answers of
which only one is the correct answer. What is the probability that sheer
guesswork yields from 25 to 30 correct answers for 80 of the 200 problems about
which the student has no knowledge?

Let X be the no. of correct answers for 80 with sheer guesswork. X  b(80, 0.25)

29
 Research Methods and Statistics

  np  (80)(0.25)  20 ,   npq  (80)(0.25)(0.75)  15

P(25  X  30)  P(24.5  X '  30.5) by normal approximation

 24.5  20 30.5  20 
 P Z   P(1.16  Z  2.71)  0.1230  0.00336  0.1196
 15 15 

The Poisson Distribution

Experiments yielding numerical values of a random variable X, the number of

successes (observations) occurring during a given time interval (or in a specified
region) are often called Poisson experiments.

A Poisson experiment has the following properties:

1. The number of successes in any interval is independent of the number of
successes in other interval.
2. The probability of a single success occurring during a short interval is
proportional to the length of the time interval and does not depend on the
number of successes occurring outside this time interval.

3. The probability of more than one success in a very small interval is

negligible.

Examples of random variables following Poisson Distribution

1. The number of customers arrived during a time period of length t.

2. The number of telephone calls per hour received by an office.

3. The number of typing errors per page.

4. The number of accidents occurred at a junction per day.

Definition :

The probability distribution of the Poisson random variable X is called the

Poisson distribution.

Notation : X ~ Po()

where  is the average number of successes occuring in the

given time interval.

30
 Research Methods and Statistics

 x
P(X = x) = e  x = 0, 1, 2, 
x!
e = 2.718283

Theorem: In a Poisson Distribution mean is equal to variance, i.e.,    2 .

Example 17

The average number of radioactive particles passing through a counter during 1

millisecond in a laboratory experiment is 4. What is the probability that 6
particles enter the counter in a given millisecond?

Let X be the no. of particles entering the counter in a given millisecond. X  Po(4)

e 4 4 6
P( X  6)   0.1042
6!

Example 19

Ships arrive in a harbour at a mean rate of two per hour. Suppose that this
situation can be described by a Poisson distribution. Find the probabilities for a
30-minute period that

(a) No ships arrive;

(b) Three ships arrive.

Let X be the no. of ship arriving in a harbour for a 30-minute period. X 

Po( 2  1 )
2

e 110
(a) P( X  0)   0.3679
0!

e 113
(b) P( X  3)   0.0613
3!

Theorem :

The mean and variance of the Poisson distribution both have mean .

31
 Research Methods and Statistics

Poisson approximation to the binomial distribution

If n is large and p is near 0 or near 1.00 in the binomial distribution, then the
binomial distribution can be approximated by the Poisson distribution with
parameter np.

Example 20

If the prob. that an individual suffers a bad reaction from a certain injection is
0.001, determine the prob. that out of 2000 individuals, more than 2 individuals
will suffer a bad reaction.

Soln : According to binomial :

The required probability

 2000 1998 
= 1     0.001  0.999
0 2000  2000
   0.001  0.999
1 1999  2000
   0.001  0.999 
2

 0   1   2  

Using Poisson distribution:

0 2
P(0 suffers) = 2 e  12   = np = 2
0! e

21 e 2 2
P(1 suffers) =  2
1! e
2 2
P(2 suffer) = 2 e  22
2! e
Then the required probability = 1  52  0.323
e
General speaking, the Poisson distribution will provide a good approximation to
binomial when
(i) n is at least 20 and p is at most 0.05; or
(ii) n is at least 100, the approximation will generally be excellent provided p<
0.1.

Example 21

Two percent of the output of a machine is defective. A lot of 300 pieces will be
produced. Determine the probability that exactly four pieces will be defective.

32
 Research Methods and Statistics

Let X be the no. of defective pieces among 300 pieces. X  b(300, 0.02)

P ( X  4) 300 C 4 (0.02 ) 4 (0.98 ) 296  0.1338

By Poisson Approximation:

  np  (300)(0.02)  6

e 6 6 4
P( X  4)   0.1338
4!

33
 Research Methods and Statistics

PART D: SAMPLING DISTRIBUTIONS AND ESTIMATION

Definition

1. A sample statistic is a characteristic of a sample.

A population parameter is a characteristic of a population.

2. A statistic is a random variable that depends only on the observed random

sample.

3. A sampling distribution is a probability distribution for a sample statistic. It

indicates the extent to which a sample statistic will tend to vary because of
chance variation in random sampling.

4. The standard deviation of the distribution of a sample statistic is known as

the standard error of the statistic.

An illustrating example

Suppose a population consists of four elements, {0,1,2,3}. A simple random

sample of two elements is to be drawn.

The population has two parameters: a mean  of 1.5 and a variance  2 of

1.6667.

Obviously there are six possible samples ( C 24  6 ). They are

Sample Sample mean error Probability

0,1 0.5 -1.0 1/6
0,2 1.0 -0.5 1/6
0,3 1.5 0 1/6
1,2 1.5 0 1/6
1,3 2.0 0.5 1/6
2,3 2.5 1.0 1/6

From the above table, we can see that if we draw a sample and use the sample
mean to estimate the population mean, the accuracy of our estimate depends on
which sample we have drawn, which in turn depends on chance.

The probability distribution of sample mean is known as a sampling distribution of

sample mean, as compiled in the following table:

34
 Research Methods and Statistics

Sample 0.5 1.0 1.5 2.0 2.5

mean
Probability 1/6 1/6 2/6 1/6 1/6

The expected value of sample mean is

E( y)   y * P( y)  0.5 *1/ 6  1.0 *1/ 6  1.5 * 2 / 6  2.0 *1 / 6  2.5 *1/ 6  1.5  Y .

Hence the average value of the sample mean is equal to the population mean.
We call the sample mean an unbiased estimator of the population mean.

The variance of the sample mean (i.e., the average square deviation of the
sample mean from the population mean) is: V ( y)   ( y  Y ) 2 P( y) =
1 1 2 1 1
(0.5 1.5) 2 *  (1.0 1.5) 2 *  (1.5 1.5) 2 *  (2.0 1.5) 2 *  (2.5  1.5)2 *
6 6 6 6 6
=0.4167

Sampling Distribution of Mean

The Central Limit Theorem

If repeated samples of size n are drawn from any infinite population with
mean  and variance 2, and n is large (n  30), the distribution of x , the
sample mean, is approximately normal, with mean  (i.e. E ( x )   ) and
2
variance 2/n (i.e. V ( x)  ), and this approximation becomes better as n
n
becomes larger.

Notes: As in the previous illustrating example, we can see the following

modifications:
n 2
(i) If the population is finite, V ( x)  (1  ) ; where (1-n/N) is known as
N n
the finite population correction factor. When N is very big, the factor is
equal to 1.

(ii) If n is small, say less than 30, the sampling distribution is not so
normal. A t-distribution will be used (discussed later).

35
 Research Methods and Statistics

n 2
In the above example, N=4, n=2, V ( x)  (1 
) =(1-2/4)(1.6667/2) = 0.4167. If
N n
the population is big (or the sample is drawn with replacement), then
2
V ( x)  =1.6667/2=0.8333.
n
In this course we assume a big population or sampling with replacement.

Example 1

An electrical firm manufactures light bulbs that have a length of life that is
approximately normal distributed with mean equal to 800 hours and a standard
deviation of 40 hours. Find the probability that a random sample of 16 bulbs will
have an average life of less than 775 hours.

2 40 2
Let X be the average life of the 16 bulbs. X  N(  x    800,  x2   )
n 16

 
 
 X   x 775   x  775  800 
P( X  775)  P    P Z 
  x   40 
 
x

 16 

 P(Z  2.5)  0.00621

Example 2

The mean IQ scores of all students attending a college is 110 with a standard
deviation of 10.

(a) If the IQ scores are normally distribution, what is the probability that the
score of any one student is greater than 112?
(b) What is the probability that the mean score in a random sample of 36
students is greater than 112?
(c) What is the probability that the mean score in a random sample of 100
students is greater than 112?

Solution

(a) Let X be the student's IQ score. X  N(110, 102 )

36
 Research Methods and Statistics

 112  110 
P( X  112 )  P Z    P( Z  0.2)  0.4207
 10 
(b) Let X 1 be the mean score of a sample of 36 students.

2 10 2
X 1  N(  x    110 ,  x2   )
n 36

 
 
 112  110 
P( X 1  112)  P Z   P( Z  1.2)  0.1151
 10 
 
 36 

(c) Let X 2 be the mean score of a sample of 100 students.

2 10 2
X 2  N(   110,  )
n 100

 
 
 112  110 
P( X 2  112)  P Z   P( Z  2)  0.0.228
 10 
 
 100 

Estimation

Estimation is the process of using statistics from sample data to estimate the
parameters of the population. A statistic is a random variable which depends on
which sample is drawn from a population.

The followings are some examples

Estimator Population parameter

1. x 

2. s2 2

3. P P

37
 Research Methods and Statistics

There are two important properties for an estimator, namely, unbiasedness and
efficiency.

Unbiased estimator: An estimator, for example, x , is unbiased if and only if

E ( x)   .

Efficiency: The efficiency of an estimator, for example, x , is given by V ( x) . The

smaller the V ( x) , the more accurate will be the x as an estimator.

There are two types of estimate

1. A point estimate is a single-value estimate of a population parameter, for

^ ^
example,   x; P  p .

2. An interval estimate of a population parameter gives an interval that may

contain the true value of the parameter with a certain probability (i.e.
confidence); for example, Pr(a    b)  0.99.

For a point estimate, both the accuracy and reliability of the estimation are
unknown. For an interval estimate, the width of the interval gives the accuracy
and the probability gives the reliability of the estimation.

Examples 3

(a) The mean and standard deviation for the quality point averages of a
random sample of 36 college seniors are calculated to be 2.6 and 0.3,
respectively. Find a 95% confidence interval for the mean of the entire
senior class.

(b) How large a sample is required in (a) if we want to be 95% confident of 

is off by less than 0.05?

Solution

Let  be the mean of the entire senior class.

38
 Research Methods and Statistics

Given: n = 36, x  2.6 , s = 0.3, (1 -  ) = 0.95    0.05

(a) A 95% confidence interval estimate for the  is

ˆ ˆ  0.3   0.3 
x  z 0.025    x  z 0.025  2.6  1.96     2.6  1.96 
n n  36   36 

 2.502    2.698
(b) Let n1 be the required sample size.

To be 95% confident that  if off by less than 0.05 would imply

 ˆ   
z 0.025    0.05  1.96  0.3   0.05
 n   n 
 1  1 

2
 (1.96)(0.3) 
 n1     138.30  139
 0.05 

A summary table for constructing (1  )% confidence interval for mean and
proportion

Estimating Conditions Formula

Mean Large samples (n  30)  
X  Z    X  Z
OR  is known 2 n 2 n
Mean * Small samples and  s s
X  t ,    X  t ,
unknown 2 n 2 n
v=n-1
Proportion Large sample pˆ qˆ
pˆ  Z 
2 n
Difference of means Large sample OR  1  12 ˆ 2
and  2 are known ( X 1  X 2 )  Z 
2 n1 n2
Difference of means Small sample &  1 and
 2 are unknown, 1 1
( X 1  X 2 )  t , ( s p ) 
assume  1   2 2 n1 n2

  n1  n2  2 ,

pooled estimate of sample

standard deviation:

39
 Research Methods and Statistics

(n1  1) s12  (n2  1) s 22

sp 
n1  n2  2

Difference of means Small sample &  1 and s12 s22

 2 are unknown, ( X 1  X 2 )  t , 
2 n1 n2
assume  1   2
( s12 / n1  s 22 / n2 ) 2

( s12 / n1 ) 2 ( s 22 / n2 ) 2

n1  1 n2  1

Difference of means Paired observations sd

d  tv , / 2 ; d  x1  x 2
n
and v=n-1
Difference of Large samples pˆ 1qˆ1 pˆ 2 qˆ 2
proportions ( pˆ 1  pˆ 2 )  z 
2 n1 n2
Example 4

The contents of seven similar containers of sulfuric acid are 9.8, 10.2, 10.4, 9.8,
10.0, 10.2 and 9.6 liters. Find a 95% confidence interval for the mean of all such
containers, assuming an approximate normal distribution.

Solution

Let  be the mean of all such containers.

Given: n = 7  x  70  x  700.48
2

x  
2
( x)
x 
2
70
2
700.48  70
x   10 s 2
 n  7  0.08
n 7 n 1 6

s= 0.08  0.2828 ; (1 -  ) = 0.95    0.05 , t 6, 0.025  2.447

A 95% confidence interval estimate for the  is

s s  0.2828   0.2828 
x  t 6,0.025    x  t 6,0.025  10  2.447     10  2.447 
n n  7   7 

 9.738    10.262

Example 5

40
 Research Methods and Statistics

In a random sample of n = 500 families owning television sets in the city of

Hamilton, Canada, it was found that x = 340 owned color sets. Find a 95%
confidence interval for the actual proportion of families in this city with colour
sets.

Let P be the actual proportion of families in this city with colour sets.

Given: n = 500, pˆ  x  340  0.68 , (1   )  0.95    0.05

n 500

A 95% confidence interval for P is

 ˆˆ   ˆˆ 
pˆ  z 0.025  pq   P  pˆ  z 0.025  pq 
 n   n

(. 68 )(. 32 ) (. 68 )(. 32 )
 0.68  1.96  P  0.68  1.96  0.64  P  0.72
500 500

Examples 6

A standardized chemistry test was given to 50 girls and 75 boys. The girls made
an average grade of 76 with a standard deviation of 6, while the boys made an
average grade of 82 with a standard deviation of 8. Find a 96% confidence
interval for the difference 1 and 2, where 1 is the mean score of all boys and 2
is the mean score of all girls who might take this test.

Given: n1  75 , n2  50 , x1  82 , s1  8 , x 2  76 , s2  6 ,

(1   )  .96    0.04

A 96% confidence interval for 1   2 is:

ˆ 12 ˆ 22 ˆ 12 ˆ 22
( x1  x 2 )  z 0.02   1   2  ( x1  x 2 )  z 0.02 
n1 n2 n1 n2

82 6 2 82 6 2
 (82  76)  2.05   1   2  (82  76)  2.05  ,
75 50 75 50

n1  30 & n2  30 , so ˆ 1  s1 & ˆ 2  s 2

41
 Research Methods and Statistics

 3.43  1   2  8.57

Example 7

In a batch chemical process, two catalysts are being compared for their effect on
the output of the process reaction. A sample of 12 batches is prepared using
catalyst 1 and a sample of 10 batches was obtained using catalyst 2. The 12
batches for which catalyst 1 was used gave an average yield of 85 with a sample
standard deviation of 4, while the average for the second sample gave an
average of 81 and a sample standard deviation of 5. Find a 90% confidence
interval for the difference between the population means, assuming the
populations are approximately normally distributed with equal variances.

Solution

Let  1 and  2 be the mean population yield using catalyst 1 and catalyst 2,
respectively.

Given: n1  12 , n2  10 , x1  85 , s1  4 , x 2  81 , s2  5 ,

(1   )  .90    0.10 ,   n1  n2  2  12  10  2  20 , t 20, 0.05  1.725

(n1  1) s12  (n2  1) s 22

pooled estimate of sample standard deviation s p 
n1  n2  2

(12  1)4 2  (10  1)5 2

  4.478
12  10  2

A 90% confidence interval for 1   2 is:

1 1 1 1
( x1  x2 )  t 20,0.05 ( s p )   1   2  ( x1  x2 )  t 20,0.05 ( s p ) 
n1 n2 n1 n2

1 1 1 1
 (85  81)  (1.725 )( 4.478 )   1   2  (85  81)  (1.725 )( 4.478 ) 
12 10 12 10

 0.69  1   2  7.31

42
 Research Methods and Statistics

Example 8

The weight of 10 adults selected randomly before and after a certain new diet
was introduced was recorded as follows:

Adult Before ( x1 ) After ( x2 ) Difference

1 76 81 -5
2 60 52 8
3 85 87 -2
4 58 70 -12
5 91 86 5
6 75 77 -2
7 82 90 -8
8 64 63 1
9 79 85 -6
10 88 83 5
Find a 98% confidence interval for the mean difference in weight.

Solution

d
d i
= -1.6 s2

 (d i  (1.6)) 2
 40.7
n 1
d
n

For v = n-1 = 9; t0.01  2.821 .

 6.38 
A 98% confidence interval is 1.6  (2,821)  
 10 
That is 7.29  d  4.09

Example 9

A certain change in a manufacturing procedure for component parts is being

considered. Samples are taken using both the existing and the new procedure in
order to determine if the new procedure results in an improvement. If 75 of 1500
items from the existing procedure were found to be defective and 80 of 2000
items from the new procedure were found to be defective, find a 90% confidence
interval for the true difference in the fraction of defectives between the existing
and the new process.

Solution

43
 Research Methods and Statistics

Let P1 and P2 be the true fraction of defectives of the existing and the new
processes, respectively.

Given: n1  1500 , x1  75 , pˆ1  75  0.05

1500

n2  2000 , x 2  80 , pˆ 2  80  0.04
2000

(1   )  .90    0.10

A 90% confidence interval for P1 - P2 is:

pˆ 1qˆ1 pˆ 2 qˆ 2 pˆ 1qˆ1 pˆ 2 qˆ 2
( pˆ 1  pˆ 2 )  z 0.05   P1  P2  ( pˆ 1  pˆ 2 )  z 0.05 
n1 n2 n1 n2

(. 05 )(. 95 ) (. 04 )(. 96 ) (. 05 )(. 95 ) (. 04 )(. 96 )

 (0.01)  1.64   P1  P2  (0.01)  1.64 
1500 2000 1500 2000

 0.001697  P1  P2  0.021697

44
 Research Methods and Statistics

PART E: Introduction to Test of Hypothesis

Statistical Hypothesis

Consider the following example:

A manufacturer of sports equipment has developed a new synthetic fishing

line that he claims has a mean breaking strength of 8 kilograms with a
standard deviation of 0.5 kilogram. Test the hypothesis that  = 8 kilograms
against the alternative that   8 kilograms if a random sample of 50 lines
is tested and found to have a mean breaking strength of 7.8 kilograms. Use
a 0.01 level of significance.

When a random sample is drawn from a population (the 50 lines randomly

selected), the sample information can be used assess the validity of some
conjecture, or hypothesis. Here  = 8 kilograms is known as the null hypotheis
and   8 kilograms is the alternative hypothesis. They are complementary to
each other, and we need to decide which one to accept on the basis of the
sample result of 50 lines.

Now let us make a 95% confidence interval about the mean breaking strength of
the population as below:

 0.5 0.5 
P  7.8  1.96*    7.8  1.96*   0.95 ;
 50 50 

i.e., P(7.6614    7.9386)  0.95 .

As there is a probability of 0.95 that the mean breaking strength is between 7.66
kg and 7.94 kg, it is highly unlikely that the null hypothesis  = 8 kg is true and
hence should be rejected.

There are four possible situations for the above decision making exercise:

H 0 is correct H 0 is wrong
Accept H 0 Correct Type 2 error
decision
Reject H 0 Type 1 error Correct
decision

45
 Research Methods and Statistics

We still have a probability of 1-0.95, or 0.05 to reject a true H 0 . We call this

probability ‗level of significance‘ or  , which is the probability of committing a
type 1 error.

The rationale of hypothesis testing is simply outlined as above. There are

however some formal concepts and procedures to conduct the test. The details
are put down below.

Some Hypothesis Testing Terminology

1. Null hypothesis, H0

A hypothesis that is held to be true until very strong evidence to the

contrary is obtained.
H0 :   0

2. Alternative hypothesis, H1

It is a hypothesis that is complement to the null hypothesis. Hence it will

be accepted if the null hypothesis is rejected.

H1 :   0 (two-tail test)
H1 :   0 (One-tail test)
H1 :   0 (One-tail test)
In the one-tail test we have some expectation about the direction of the
error when the null hypothesis is wrong, while in the two-tail test we don‘t
have such expectation.

3. Test statistics

is the value, based on the sample, used to determine whether the null
hypothesis should be rejected or accepted.

4. Critical region

is a region in which if the test statistic falls the null hypothesis will be
rejected.

46
 Research Methods and Statistics

5. Types of error

(a) Type I error: Reject H0 when H0 is true

(b) Type II error: Accept H0 when Ha is true

6. The significance level, 

is the probability of committing a type 1 error, i.e., P(Type I error) = .

The probability of committing a type 2 error is ; i.e., P(Type II error) = .

Basic Steps in Testing Hypothesis

1. Formulate the null hypothesis.

2. Formulate the alternative hypothesis.

3. Specify the level of significance to be used.

4. Select the appropriate test statistic and establish the critical region.

5. Compute the value of the test statistic.

6. Conclusion: Reject H0 if the statistic has a value in the critical region,

otherwise accept H0.

47
 Research Methods and Statistics

Tests concerning means

H0 Conditions Test statistic

Large samples (n  30) OR  is x  0
  0 known z
 n
  0 Small samples and  unknown x  0
t with   n  1
s n

1   2  d 0 Large samples OR  1 and  2 ( x1  x 2 )  d 0

z
are known  12  22

n1 n2

1   2  d 0 Small sample &  1 and  2 are ( x1  x 2 )  d 0

t
unknown, assume  1   2 1 1 
s p   
 n1 n 2 
with   n1  n2  2 and
(n1  1) s12  (n2  1) s22
s 2p 
n1  n2  2
if  1 =  2 but unknown
1   2  d 0 Small sample &  1 and  2 are ( x  x2 )  d 0
t 1
unknown, s12 s 22
assume  1   2 
n1 n2
( s12 / n1  s 22 / n2 ) 2
with  
( s12 / n1 ) 2 ( s 22 / n2 ) 2

n1  1 n2  1

1   2  d 0 Paired observations d  d0
t with   n  1
sd n
p  p0 Large sample pˆ  p0
z
p0 (1  p0 )
n
p1  p2  0 Large samples ( p1  pˆ 2 )
ˆ
z
1 1
pˆ (1  pˆ )  
 n1 n2 
n pˆ  n pˆ
pˆ  1 1 2 2
n1  n2

48
 Research Methods and Statistics

Example 1

A manufacturer of sports equipment has developed a new synthetic fishing

line that he claims has a mean breaking strength of 8 kilograms with a
standard deviation of 0.5 kilogram. Test the hypothesis that  = 8 kilograms
against the alternative that   8 kilograms if a random sample of 50 lines
is tested and found to have a mean breaking strength of 7.8 kilograms. Use
a 0.01 level of significance.

Null hypothesis:   8 kilograms

Alternative hypothesis:   8 kilograms

Level of significance: 0.01

Critical region: Z > z 0.005  2.58 or Z <  z 0.005  2.58

Computation:

n = 50 x  7.8   0.5

x 7.8  8
z   2.828
 0.5
n 50

Conclusion: As the sample z (= -2.828) falls inside the critical region, so reject
the null hypothesis at 0.01 level of significance and conclude that  is
significantly smaller than 8 kilograms.

Example 2

The average length of time for students to register for fall classes at a certain
college has been 50 minutes with a standard deviation of 10 minutes. A new
registration procedure using modern computing machines is being tried. If a
random sample of 12 students had an average registration time of 42 minutes
with a standard deviation of 11.9 minutes under the new system, test the
hypothesis that the population mean is now less than 50, using a level of
significance of (1) 0.05, and (2) 0.01. Assume the population of times to be
normal.

Let  be the population mean time for students to register in the new registration
procedure.

49
 Research Methods and Statistics

(1) Null hypothesis:   50 minutes

Alternative hypothesis:   50 minutes

Level of significance: 0.05

Critical region: (n = 12 < 30; and the new  is unknown, so t-test should
be used ) degree of freedom ( ) = n -1 = 12 -1 =11

 t < t11, 0.05  1.796

Computation:
n = 12 x  42 s = 11.9

x 42  50
 t   2.329
s 11.9
n 12

Conclusion: As sample t (= -2.329) falls inside the critical region, so reject

the null hypothesis at 0.05 level of significance and conclude that  is
significantly smaller than 50 minutes.

(2) Identical with those of (1) except the critical region would be replaced by:

t  t11, 0.01  2.718

and the corresponding conclusion would be changed as follows:

As sample t (= -2.329) falls outside the critical region, so reject the

alternative hypothesis at 0.01 level of significance and conclude that  is
not highly significantly smaller than 50 minutes.

Example 3

An experiment was performed to compare the abrasive wear of two different

laminated materials. Twelve pieces of material 1 were tested, by exposing each
piece to a machine measuring wear. Ten pieces of material 2 were similarly
tested. In each case, the depth of wear was observed. The samples of material
1 gave an average (coded) wear of 85 units with a standard deviation of 4, while
the samples of material 2 gave an average of 81 and a standard deviation of 5.
Test the hypothesis that the two types of material exhibit the same mean
abrasive wear at the 0.10 level of significance. Assume the populations to be
approximately normal with equal variances.

50
 Research Methods and Statistics

Let  1 and  2 be the mean abrasive wear of material 1 and 2 respectively.

Null hypothesis: 1   2 , i.e. 1   2  0

Alternative hypothesis: 1   2 , i.e. 1   2  0

Level of significance: 0.10

Critical region: (As both n1 and n 2 are smaller than 30 and their standard
deviations are unknown, so t-test has to be used.)

  n1  n2  2  12  10  2  20 ,  t  t 20, 0.05  1.725 or

t  t 20, 0.05  1.725

Computation:

n1  12 x1  85 s1  4

n2  10 x 2  81 s2  5

(n1  1)s12  (n2  1)s 22 (11)4 2  (9)5 2

s 2p    20.05
n1  n2  2 12  10  2

( x1  x 2 )  ( 1   2 ) (85  81)  0
t   2.086
1 1 1 1
sp  20.05 
n1 n2 12 10

Conclusion: As the sample t (=2.086) falls inside the critical region, so reject the
null hypothesis at 0.10 level of significance and conclude that the mean abrasive
wear of material 1 is significantly higher than that of the material 2.

Example 4

Five samples of a ferrous-type substance are to be used to determine if there is a

difference between a laboratory chemical analysis and an X-ray fluorescence
analysis of the iron content. Each sample was split into two sub-samples and the
two types of analysis were applied. Following are the coded data showing the
iron content analysis:

51
 Research Methods and Statistics

Sample
Analysis 1 2 3 4 5
x-ray 2.0 2.0 2.3 2.1 2.4
Chemical 2.2 1.9 2.5 2.3 2.4

Assuming the populations normal, test at the 0.05 level of significance whether
the two methods of analysis give, on the average, the same result.

Let  1 and  2 be the mean iron content determined by the laboratory chemical
analysis and X-ray fluorescence analysis respectively; and

 D be the mean of the population of differences of paired measurements.

Null hypothesis: 1   2 or D  0

Alternative hypothesis: 1   2 or D  0

Level of significance: 0.05

Critical region: (As n = 5 < 30, so t-test should be used.)

 t  t 4, 0.025  2.776 or t  t 4, 0.025  2.776

Computation:

Sample
Analysis 1 2 3 4 5
x-ray 2.0 2.0 2.3 2.1 2.4
Chemical 2.2 1.9 2.5 2.3 2.4
di -0.2 0.1 -0.2 -0.2 0

d i2 0.04 0.01 0.04 0.04 0

5 5

 d i   0 .5
i 1
d
i 1
i
2
 0.13

d 
 d   0.5  0.1
5 5

52
 Research Methods and Statistics

n d 2   d 
2
(5)(0.13)  (0.5) 2
s 
2
  0.02
n(n  1)
d
(5)(4)

d   D (0.1)  0
t   1.5811
sd 0.02
n 5

Conclusion: As the sample t (=-1.5811) falls outside the critical region, so reject
the alternative hypothesis at 0.05 level of significance and conclude that there is
no significant difference in the mean iron content determined by the above two
analyses.

Tests Concerning Proportions

Example 5

A manufacturing company has submitted a claim that 90% of items produced by

a certain process are non-defective. An improvement in the process is being
considered that they feel will lower the proportion of defective below the current
10%. In an experiment 100 items are produced with the new process and 5 are
defective. Is this evidence sufficient to conclude that the method has been
improved? Use a 0.05 level of significance.

Let P be the proportion of defective product in the new production process.

Null hypothesis: P = 0.1

Alternative hypothesis: P < 0.1

Level of significance: 0.05

Critical region: Z   z 0.05  1.64

Computation:

x 5
n = 100 x=5 pˆ    0.05
n 100

pˆ  P 0.05  0.1
z   1.667
P (1  P ) (0.1)( 0.9)
n 100

53
 Research Methods and Statistics

Conclusion: As the sample z (=-1.667) falls inside the critical region, so reject the
null hypothesis at 0.05 level of significance and conclude that P is significantly
smaller than 0.1. That is, the production method has been improved in lowering
the proportion of defective below the current 10%.

Example 6

A vote is to be taken among the residents of a town and the surrounding country
to determine whether a proposed chemical plant should be constructed. The
construction site is within the town limits and for this reason many voters in the
country feel that the proposal will pass because of the large proportion of town
voters who favor the construction. To determine if there is a significant difference
in the proportion of town voters and county voters favoring the proposal, a poll is
taken. If 120 of 200 town voters favor the proposal and 240 of 500 county
residents favor it, would you agree that the proportion of town voters favoring the
proposal is higher than the proportion of county voters? Use a 0.025 level of
significance.

Let P1 and P2 be the proportions of town voters and country voters, respectively,
favouring the proposal.

Null hypothesis: P1  P2 or P1  P2  0

Alternative hypothesis: P1  P2 or P1  P2  0

Level of significance: 0.025

Critical region: Z  z 0.025  1.96

Computation:

x1 120
n1  200 x1  120 pˆ 1    0 .6
n1 200

x 2 240
n2  500 x 2  240 pˆ 2    0.48
n 2 500

n1 pˆ 1  n 2 pˆ 2 (200 )( 0.6)  (500 )( 0.48 )

pˆ    0.514
n1  n 2 200  500

( pˆ 1  pˆ 2 )  ( P1  P2 ) (0.6  0.48 )  0
z   2.870
1 1  1 1 
pˆ (1  pˆ )    (0.514 )( 0.486 )  
 n1 n 2   200 500 

54
 Research Methods and Statistics

Conclusion: As sample z (=2.870) falls inside the critical region, so reject the null
hypothesis at 0.025 level of significance and conclude that the proportion of town
voters favouring the proposal is significantly larger than that of the country voters.

55
 Research Methods and Statistics

PART F: Chi-square Tests

There are two types of chi-square tests: goodness-of-fit test and tests for
independence.

Goodness-of-fit Test

A test to determine if a population has a specified theoretical distribution. The

test is based on how good a fit we have between the frequency of occurrence of
observations in an observed sample and the expected frequencies obtained from
the hypothesized distribution.

Theorem: A goodness-of-fit test between observed and expected frequencies

is based on the quantity

(Oi  Ei ) 2
 2
test 
Ei

where  2test is a value of the random variable whose

sampling distribution is approximated very closed by the Chi-square
distribution,

Oi is the observed frequency of cell i, and Ei is the expected

frequency of cell i.

The number of degrees of freedom in a Chi-square goodness-of-fit

test is equal to the number of cells minus the number of quantities
obtained from the observed data that are used in the calculations of
the expected frequencies.

For a level of significance equal to  ,  2test   2 constitutes the critical region.

The decision criterion described here should not be used unless each of the
expected frequencies is at least equal to 5.

Example 1

Consider the tossing of a die 120 times.

Faces
1 2 3 4 5 6
Observed 20 22 17 18 19 24
Expected

56
 Research Methods and Statistics

By comparing the observed frequencies with the expected frequencies, one has
to decide whether the die is fair die or not.

Null hypothesis: the die is a fair die, i.e. P( X  i)  1 for i = 1, 2, 3, 4, 5, and 6

6

Alternative hypothesis: the die is not a fair die

Level of significance: 0.05

Critical region:   n  1  6  1  5 ;   2   52,0.05  11.07

Computation:

1
Expected value = nP( X  i)  120( )  20
6

i 1 2 3 4 5 6
Observed (Oi ) 20 22 17 18 19 24
Expected ( E i ) 20 20 20 20 20 20
Oi  E i 0 2 -3 -2 -1 4

6
(Oi  Ei ) 2
2  
i 1 Ei

0 2 2 2 (3) 2 (2) 2 (1) 2 4 2

       1.7
20 20 20 20 20 20

Conclusion: As the sample  2 (= 1.7) falls outside the critical region, so reject the
alternative hypothesis and conclude that the die is a fair die.

Example 2

The following distribution of battery lives may be approximated by the normal

distribution.

Class Oi z-value p-value Ei

boundaries

57
 Research Methods and Statistics

1.45 - 1.95 2
1.95 - 2.45 1
2.45 - 2.95 4
2.95 - 3.45 15
3.45 - 3.95 10
3.95 - 4.45 5
4.45 - 4.95 3

Chi-squared test can be applied to test whether the above frequency distribution
can be approximated by a normal distribution or not.

Null hypothesis: the distribution can be approximated by a normal distribution

Alternative hypothesis: the distribution cannot be approximated by a normal

distribution

Level of significance: 0.05

Critical region:  2  2,0.05 where   n  3 , and n is the number of cells.

Computation:

For finding the expected values, the mean and standard deviation of the
frequency distribution have to be found first.

( fi ) Class mark
Class Oi ( xi )
boundaries
1.45 - 1.95 2 1.7
1.95 - 2.45 1 2.2
2.45 - 2.95 4 2.7
2.95 - 3.45 15 3.2
3.45 - 3.95 10 3.7
3.95 - 4.45 5 4.2
4.45 - 4.95 3 4.7

n = 40  fx  136.5  fx 2
 484.75

x
 fx  136.5  3.4125
 f 40

58
 Research Methods and Statistics

s
 fx   fx
2 2
/n


484.75  136.5
2

40  0.6969
n 1 39

 L  3.4125 U  3.4125
z  value   i Z i ;
 0.6969 0.6969 

where L i and U i are the Lower and Upper Boundaries of the ith class.

 L  3.4125 U  3.4125
p  value   i Z i
 0.6969 0.6969 

 L  3.4125 U  3.4125
Ei  (40) P  i Z i
 0.6969 0.6969 

Class Oi z-value p-value Ei

boundaries
1.45 - 1.95 2 Z<-2.10 .0179 0.716
1.95 - 2.45 1 -2.10<Z<-1.38 .0659 2.636
2.45 - 2.95 4 -1.38<Z<-0.66 .1708 6.832
2.95 - 3.45 15 -0.66<Z<0.05 .2653 10.612
3.45 - 3.95 10 0.05<Z<0.77 .2595 10.38
3.95 - 4.45 5 0.77<Z<1.49 .1525 6.1
4.45 - 4.95 3 1.49<Z .0681 2.724

In order to satisfy the rule that the expected value in each cell is larger than or
equal to 5, we have to combine the first three classes in to one cell and the last
two classes into another cell. As such, the number of cells (n) is 4.

(O  E ) 2 (7  10 .184 ) 2 (15  10 .612 ) 2 (10  10 .38 ) 2 (8  8.824 ) 2

2      
E 10 .184 10 .612 10 .38 8.824
 2.901

Conclusion: Since 12,0.05  3.841, so the sample  2 (= 2.901) falls outside the
critical region. As such, reject the alternative hypothesis and conclude that the
distribution of battery lives can be approximated by a normal distribution.

59
 Research Methods and Statistics

Test for Independence

The Chi-square test procedure can also be used to test the hypothesis of
independence of two variables/attributes. The observed frequencies of two
variables are entered in a two-way classification table, or contingency table.

Remark: The expected frequency of the cell in the ith row and jth column in
the contingency table

(total of row i) * (total of column j)

E ij 
grand total

The degrees of freedom for the contingency table is equal to (r  1) (c 1) where
r is the number of rows and c is the number of columns in the table.

Example 3

Suppose that we wish to study the relationship between grade point average and
appearance.

Grade Point Average

Appearance 1 2 3 4 Totals
attractive 14 ( ) 11 ( ) 10 ( ) 5( ) 40
ordinary 10 ( ) 16 ( ) 16 ( ) 14 ( ) 56
unattractive 3( ) 4( ) 7( ) 10 ( ) 24
Totals 27 31 33 29 120

Null hypothesis: There is no relationship between grade point average and

appearance. That is, the two characteristics are independent.

Alternative hypothesis: There is a relationship between grade point average and

appearance. That is, the two characteristics are not independent.

Level of significance: 0.05

Critical region:  2  2,0.05 , where  = (r -1)(c - 1)

Computation:

Grade Point Average

60
 Research Methods and Statistics

Appearance 1 2 3 4 Totals
attractive 14 11 10 5 40
(9) (10.33) (11) (9.67)
ordinary 10 16 16 14 56
( 12.6) ( 14.47) (15.4) (13.53)
unattractive 3 4 7 10 24
(5.4) (6.2) (6.6) (5.8)
Totals 27 31 33 29 120

(O  E ) 2 (14  9) 2 (11  10 .33) 2 (10  11) 2 (5  9.67 ) 2

2      
E 9 10 .33 11 9.67

(10  12 .6) 2 (16  14 .47 ) 2 (16  15 .4) 2 (14  13 .53) 2

   
12 .6 14 .47 15 .4 13 .53

(3  5.4) 2 (4  6.2) 2 (7  6.6) 2 (10  5.8) 2

     10 .818
5.4 6.2 6.6 5.8

Conclusion: Since  62,0.05  12.596 , so sample  2 (=10.818) falls outside the

critical region. So reject the alternative hypothesis and conclude that there is no
evidence to support there is relationship between grade point average and
appearance.

Test for Homogeneity

To test the hypothesis that several population proportions are equal.

Remark: The approach for the test of homogeneity is the same as for the
test of independence of variables/attributes.

Example 4

A study of the purchase decisions for 3 stock portfolio managers, A, B, and C

was conducted to compare the rates of stock purchases that resulted in profits
over a time period that was less than or equal to one year. One hundred
randomly selected purchases obtained for each of the managers showed the
results given in the table. Do these data provide evidence of differences among
the rates of successful purchases for the three portfolio managers? Test with
  0.05.

61
 Research Methods and Statistics

Manager
Result A B C
Purchases show profit 63 71 55
Purchase do no show profit 37 29 45
Total 100 100 100

Null hypothesis: the rates of stock purchases that resulted in profit were the
same for the three stock portfolio managers

Alternative: their rates were not all the same

Level of significance: 0.05

Critical region:   (2  1)(3  1)  2 ;   2   22,0.05  5.991

Computation:

Manager
Result A B C Total
Purchases show profit 63 71 55 189
(63) (63) (63)
Purchase do no show profit 37 29 45 111
(37) (37) (37)
Total 100 100 100 300

(63  63) 2 (71  63) 2 (55  63) 2 (37  37 ) 2 (29  37 ) 2 (45  37 ) 2

2        5.49
63 63 63 37 37 37

Conclusion: As the sample  2 (= 5.49) falls outside the critical region so reject
the alternative hypothesis and conclude that there is no sufficient evidence to
support the rates of purchases resulted in profit of the three portfolio managers
were different.

62
 Research Methods and Statistics

IV
Standard Probability
Distributions
Suppose an event E can happen in h ways out of a total of n possibly equally
likely ways. Then the probability of occurrence of the event is denoted by

p = h/n

and, the probability of non-occurrence of the event is denoted by

q = 1–p

The probability scale runs from impossibility on the one hand to absolute
certainty on the other. The two poles of this continuum are given the numerical
values of 0 and 1 respectively. Therefore, the probability of any event occurring
must lie somewhere between these two extremes. The probability of my death
is absolutely certain indeed, the probability of my death next year can be
determined by a study of death rates in the past:

Age Group Deaths per 1000 per Year Probability of Dying in Year+1

0-4 3.4 0.0034

5-9 0.3 0.0003
10 -14 0.3 0.0003
15 -19 0.9 0.0009
20 - 24 0.9 0.0009
25 - 34 0.9 0.0009
35 - 44 2.0 0.0020
45 - 54 6.6 0.0066
55 - 64 18.7 0.0187
65 - 74 47.6 0.0476
75 - 84 110.3 0.1103
> 85 234.4 0.2344

63
 Research Methods and Statistics

However, the above definition of probability is circular because we are essentially

defining probability in term of itself: ―equally likely‖ appears to be synonymous
with ―equally probable‖. For this reason a statistical definition of probability is
advocated: the estimated probability, or empirical probability, of an event is the
relative frequency of occurrence of the event when the number of observations is
very large. The probability itself is the limit of the relative frequency as the
number of observations increase indefinitely.

Population

In collecting data concerning characteristics of a group of individuals, such as

attitudes, aptitudes and personality dimensions of university students, it is often
impossible or impractical to observe the entire group, especially if it is large.
Instead of examining the entire group, called the population or universe, one
examines a small part of the group called a sample. A population may be finite
or infinite. A finite population is defined as an entity with a boundary around it,
this boundary may be one of sex, time, location or other attribute determined by
the researcher; for example, the population of students may be defined as: all
students registered with the University of Newcastle upon Tyne as at 31 March
2001. Whereas an infinite population is one consisting of all possible outcomes
of an event (successive tosses of a coin or guessing the colour of a card is
infinite).

Sampling Theory and Sampling

Sampling theory is the study of relationships existing between a population and

samples drawn from the population. It is of value in estimating of unknown
population parameters (such as population arithmetic mean, variance, etc.) from
knowledge of corresponding sample statistics. Sampling theory is also useful in
determining whether observed differences between two samples are actually due
to chance variation or whether they are really significant. The answers involve
use of so-called tests of significance and hypotheses. In general, a study of
inferences made concerning a population by use of samples drawn from it,
together with indications of the accuracy of such inferences using probability
theory, is called statistical inference.

A sample is a set of individuals or objects selected from a population. The

purpose of sampling is to infer a characteristic or characteristics of a given
population from a subset of measurements obtaining from the sample. In a
quantitative research design, the sampling methods and the sample design are
crucial to obtain unbiased and consistent estimates of the characteristic
population inferred from the sample; and, enable one to statistically validate any
conclusions based on the observations and findings. In order that conclusions
of sampling theory and statistical inference be valid, samples must be chosen so
as to be representative of the population. A study of methods of sampling and
the related problems, which arise, is called the design of the experiment. One

64
 Research Methods and Statistics

way in which a representative sample may be obtained is by a process of

random sampling, according to which each member of the population has an
equal chance of being included in the sample. One technique for obtaining a
random sample is to assign a number to each member of the population, write
each number on a separate piece of paper, place them in a container and then
draw numbers from the container, being careful to mix thoroughly before each
drawing. Alternatively, this can be replaced by using a table of random numbers
specially constructed for such purposes.2

How well a sample represents a given population depends on the sample frame,
the sample size and the specific sample design:

 Sample Frame: the set of subjects who have a chance of being

selected from the study population, given the sampling approach
chosen

 Sample Design: the specific procedures to be used for selecting the

subjects in the sample

 Sample Size: the planning of, and reasons for choosing, the
number of subjects in the sample.

Sample Frame

A sample frame is a complete list of individuals or units in the population to be

studied. An initial step in sampling is to provide a clear and accurate definition
of the population exposed to your study. This study population may comprise of
a group of individuals who go somewhere or do something that enables them to
be sampled. Sampling is carried out in two stages; the first involves sampling
something other than the individuals to be finally selected, the second entails
creating a list of sampling units (individuals) from which a final selection is made.
Whilst the sample frame should be representative of the study population, it may
be necessary to control mediating or intervening variables. For instance,
suppose I wish to ascertain the attitude of workers, employed by the Kenkei
Electronics Company, toward some object or subject. A number of workers in
the study population will have been recently employed by the company, and may
retain beliefs, opinions and attitudes from previous employers. To eliminate
these from the study, the sampling frame may be constructed to include workers
with at least one year‘s service with the Kenkei Electronics Company.

2
Random numbers can be generated from scientific calculators, or be extracted from statistical
tables.

65
 Research Methods and Statistics

Sample Design

The sample design influences the precision of the subset of measurements

obtained from the sample. Sampling strategies can be categorized as random
probability, non-random probability and mixed sampling. A random probability
sample design provides for an equal and independent chance of selection, i.e.
each individual has a known probability of selection set by the sampling
procedure. Non-random probability sample designs are used when the
population size is either unknown or cannot be discretely identified. A mixed
sampling design has characteristics of both random and non-random probability
sampling. Systematic sampling is the most common strategy used, based on a
known population with the sample size determined a priori. The advantages of
systematic sampling over simple random sampling is that it is more efficient, in
terms of information per respondent cost, and easier to perform thus reducing
sampling error. Systematic sampling will generally be an adequate form of
random sampling to the extent that the placing of any sampling subject or unit is
independent of the placing of other sampling individuals or units, i.e. there is no
systematic bias introduced into the listing of sampling units. Should such a risk
of sampling bias be known, it may be avoided by choosing a suitable sampling
interval; or, after a predetermined number of units have been drawn, a fresh
random start can be made.3

Sample Size

How big should my sample be? It all depends on what you want to do with your
findings and the type of relationship you want to establish in your study. The
sample size is crucially important in a correlational research design, i.e. tests of
hypotheses and significance, or establishing an association or relationship
between two of more variables. However, there is no relationship between the
sample size and the size of the study population, sample size is determined by
the variability of the factor, element or characteristic prevalent in the study
population. Other things being equal, precision increases steadily up to sample
sizes of 50-200; after that, there is only a modest gain in precision to increasing
sample size. Whilst increasing sample size reduces errors attributable to
sampling, managing large amounts of data may increase non-sampling errors
(e.g. field-work problems, interviewer-induced bias, clerical errors in transcribing
data, etc.). In determining the size of a sample, consideration should be given
to:

 the degree of accuracy required in the estimation of the variables in

the chosen study population

 the level of confidence demanded from the sample to test the

significance of the findings or hypotheses

3
See Jakobowitz, K., (1999), How to Develop and Write a Business Research Proposal, pp.58-
74, ISBN 0-9538664-0-8

66
 Research Methods and Statistics

 the extent to which the variability of the factor in the chosen study
population is known, or can be estimated.4

You have developed a survey instrument and, after a pilot-study of ten

subjects, have determined a sample estimate of mean to be 52 with a
standard deviation of 15. Having decided a priori that the sample of
the population needs to be sufficient to allow for 95 percent confidence
in your findings and conclusions, it is necessary to compute the
required sample size:

Computation

The standard deviation of the study population is unknown, hence the normal
distribution cannot be used but a derivative, Student‘s t-distribution, can be
applied:

ns = [t0.025s/L]2

where:
ns = sample size
t = confidence limit value derived from statistical tables for v = n – 2
degrees of freedom, and t/2 (t-0.025 – t+0.025)
s = sample standard deviation obtained from the pilot-study
L = level of tolerance or error (say, 5 points on the instrument scale of
100)

Hence:

ns = [2.306(15)/5]2 = 47.858724 (say, 48)

2
Note The normal distribution provides a sample size of [1.96(15)/5] = 35; and, using an
estimate of the population standard deviation [(n/n-1)/s] provides a sample size
2
of [1.96(16.667)/5] = 43

4
It is rare indeed to have a known population variance and so be able to determine the
population standard deviation. Hence, deviates of the normal distribution cannot be applied, and
a derivative such as Student‘s t-test should always be used for an unknown standard deviation of
the study population.

67
 Research Methods and Statistics

Sampling Error

One purpose of using random probability sampling methods is to apply a variety

of statistical techniques to estimate the confidence one can have that the
characteristics of a sample accurately reflect the study population. Sampling
error is a random product of sampling. However, when random sampling
methods are used it is possible to compute how much the sample-based
estimate of the characteristic will vary from the study population by chance
because of sampling. The larger the sample size and the less variability of the
characteristic being measured, the more accurate a sample-based estimate will
be. Sampling error can be defined as the variation around the true population
value that results from random sample differences drawn from the population.

The standard error of mean is the most commonly used statistic to describe
sampling error5:
SE = [s2/n]0.5
Where: s2 is the variance derived from the sample
n is the sample size
and [sum]0.5 is the square root of the product

Alternatively, the standard error of mean is more easily computed from a

proportion statement, since the variance of a proportion is expressed as p[1-p]:
the standard error of mean of a proportion is computed from: [p(1-p)/n]0.5.

In a survey of job satisfaction, a pilot study of 21 subjects provided a

sample mean of 136.81 with a sample standard deviation of 27.42.
Estimate the true population mean from the pilot-study.

Computation
The standard error of mean = [s2/n]0.5 = [(27.42)2/21]0.5 = 5.98

Conclusion
We can be 67 percent certain that the true population mean lies + one standard
error from the sample mean, i.e. 136.81 + 5.98, or between 130.83 and 142.79
on the survey instrument scale.

Also, we can be 95 percent certain that the true population mean lies + two
standard errors from the sample mean, i.e. 136.81 + 11.96, or between 124.85
and 148.77 on the survey instrument scale.

Source: Jackson (1989)

5
For a normal, or approximately asymmetrical distribution, the 95 percent confidence limits of the
0.5
population true mean can be computed: sample mean + 1.96s/(n ).

68
 Research Methods and Statistics

In a study to estimate a characteristic in a population of 1000 subjects,

a sample of 50 is chosen at random. In responding to a question there
is only two dichotomous values: 0 (No) and 1 (Yes). In the sample 15
respondents say ‘yes’ and 35 say ‘no’. Estimate the standard error.

The mean of the sample response is: x/n = (15/50) = 0.3 (in other words, a
proportion of thirty percent of the sample claim the characteristic).

The standard error of a proportion = [p(1-p)/n]0.5 = [0.3(1-0.3)/50]0.5 = 0.0648

Thus, we can be 67 percent certain (i.e. + one standard deviation from the
mean) that the value of the true population characteristic lies between one
standard error of the sample mean: 0.3 + 0.0648, or between 23.52 and 36.48
percent of the study population. Also, we can be 95 percent confident that the
true population mean lies + two standard errors of the sample mean: 0.03 +
0.1296, or between 17.04 and 42.96 percent of the study population.

Summary of Sampling Distribution Formulae

1. Calculation of Confidence Limits for an Unknown Percentage Mean

0.5
 lies between the sample mean percentage + z[p(100-p)/n]

where, n = sample size

p = sample estimate of the mean
z1 = 1.96 for 95 percent confidence limits
z2 = 2.58 for 99 percent confidence limits

Out of a random sample of 1000 subjects, 500 say they have a positive attitude
towards the object or subject of the study. What conclusions can we draw about
the percentage of the total population who may hold this view?

Our sample estimate (p) is, in effect, a random selection from a normal distribution (n>30) with
mean =  and a standard error (= standard deviation) of [(100 -)/n] . Accordingly we can
0.5

state with 95 percent confidence that the true population mean lies between:

0.5
 + [1.96[(100-)/n]] = 50 percent + [1.96[50(100-50)/1000] = 3.1 percent

Conclusion We can state with 95 percent confidence that the true proportion of people
who hold this view is within the range of 46.9 to 53.1 percent

69
 Research Methods and Statistics

2. Calculating the Sample Size for Established Accuracy from a Large Pilot-
Study

ns = 4(100-)/L2

where ns = required sample size

 = sample proportion estimate of the mean
L = specified percentage of accuracy

In order to evaluate the success of an attitudinal change programme, a company

interviewed 400 workers exposed to the change programme. Some 120 workers
displayed evidence of the required positive attitudinal change. How accurately does
this survey reflect the percentage of workers who have undergone a positive
attitudinal change? How many more workers must be surveyed in order to establish
this percentage with + 2 percent accuracy with 95 percent confidence?

In the survey, ns = 400 and  = (120/400) = 30 percent and provides the computed accuracy at
the 95 percent confidence limits:
 lies between  + 1.96 [(100-)/n]
0.5

= 30 percent + 1.96[30(100-30)/400]
0.5

= 30 percent + 4.6 percent

To compute the sample size necessary to reduce the standard error of + 4.6 percent to one of +
2 percent, with L = 2 and  = 30:
2
ns = 4(100-)/L
2
= 4(30)(100-30)/2
= 2100

Conclusion In order to be establish that 30 percent of workers have undergone a positive

attitudinal change, within + 2 percent accuracy, we need a sample of 2100 workers to be 95
percent confident.

70
 Research Methods and Statistics

V
The Normal Distribution

The most important continuous probability distribution in the field of statistics is

the normal distribution:

 empirical data is often normally, or approximately normally,

distributed

 the assumption of normality allows for the application of powerful

statistical analyses

 the distribution of many sample statistics tends to normality as the

sample size increases (>30)

 many population distributions can be readily transformed to

normality

The properties of the normal distribution are:

 observations tend to cluster at the mean: mean = mode = median

 the distribution of observations is symmetrical about the vertical

axis through the mean

 the total area under the curve is equal to unity, i.e. 1.0000

 the normal curve continues to decrease in height as one proceeds

in either direction away from the mean, but never reaches the
horizontal axis, i.e. there is a presumption of negative and positive
infinity

 the area under the curve between two ordinates, X = a and X = b

where a < b, represents the probability that X lies between a and b
and can be expressed by the probability of a < X < b

 when the variable X is expressed in standard units, z = (X - )/

71
 Research Methods and Statistics

-3 -2 -1  +1 +2 +3

68.27 percent

95.45 percent

99.73 percent

Some Properties of the Normal Distribution

How to Find the Area of the Standardized Normal Distribution

Statistical tables by White, Yeats and Skipworth (1991) and Murdoch and Barnes
(1985) of the areas of the standardized normal distribution give the probability
that a random variable will be greater than the mean (U), i.e. the area in the tail.
Hence, for the z = 1.2 we are given a value of 0.11507, but this represents the
probability of a random sample being greater than the mean, i.e. the probability
of being in the tail of the distribution. To find the probability of a random sample
falling in the acceptance region (z = 0 to z = 1.2), we must subtract this value
from 0.5:

Probability (z = 0 to z = 1.2) = 0.50000 – 0.11507

= 0.38493 (say, 38.5 percent)

72
 Research Methods and Statistics

The intelligence quotient of a sample of 5000 British adults provided a

mean of 100 with a standard deviation of 15. Assuming that
intelligence quotient is normally distributed, how many adults would
we expect to have an IQ between 120 and 130?

Computation

The intelligence quotient recorded as 120 and 130 can actually have a value
from 119.5 and 130.5 assuming that the IQ is recorded to the nearest unit of
measurement. Hence:

119.5 in standard units = (x-)  (119.5-100)/15

= +1.3000
150.5 in standard units = (x-) = (130.5-100)/15
= +2.0333

< Pr >

Intelligence Quotient: 100 120 130

Standard Score (z): 0 +1.30 +2.03

The required proportion of standard units = (area between z = 1.3000) + (z =

2.0333). Using White, Yeats and Skipworth (1991), Tables for Statisticians,
pp.14:
Area between z = 0 to z = 1.3000 . . . ... = 0.09680
Area between z = 0 to z = 2.0333 . . . ... = 0.02118
Subtract: 0.07562

Conclusion

The number of British adults in the sample having an IQ of between 120 and
130 = 5000 (0.07562) = 378.1 (say, 378)

73
 Research Methods and Statistics

VI
Statistical Decision Theory:
Tests of Hypothesis and
Significance
Very often in practice we are called upon to make decisions about populations on
the basis of sample information. Such decisions are called statistical decisions.
For example, we may wish to decide on the basis of sample data whether one
psychological procedure is better than another, whether the findings from survey
data are representative of the population, whether the conclusions reached as to
an experiment are valid, etc.

What is a Hypothesis?

A hypothesis refers to conjectures that can be used to explain observations; a

hunch, or educated guess, which is advanced for the purpose of being tested. A
hypothesis provides us with a guiding idea to determine what was relevant and
what was irrelevant. This mode of accounting for problems possesses three
steps:

1. the proposal of a hypothesis to account for a phenomenon

2. the deduction from the hypothesis that certain phenomena should

be observed in given circumstances

3. the checking of this deduction by observation.

Example

We may reason that a deprived family background causes low reading

attainment in children. We may then produce empirical evidence that low
family income and overcrowding are associated with poor reading
attainment. If no such evidence was forthcoming, then the hypothesis
must be decisively rejected. But if the predicted relationship was found,
could we conclude that the hypothesis was correct, i.e. poor family
background does cause low reading attainment? The answer must be

74
 Research Methods and Statistics

‗no‘. It might equally be the case that a low level of school resources is
also to blame. The main point to note is that the scientific process never
leads to certainty in explanation, only to the rejection of existing
hypotheses and the construction of new ones.

How do we Formulate Hypotheses?

Generally, a hypothesis is derived from theory or the literature on the problem.

But regardless of source, a hypothesis must meet one criterion: it must be stated
in such a way that it can be confirmed or refuted. In other words, it must be
amenable to test.

Example ―Is a student’s authoritarian behaviour directly related to his or her

attitudes concerning punishment received from his or her parents?‖

Although the statement of purpose is quite detailed, it conveys little information

unless the conceptual propositions are detailed: what do we mean by
authoritarian behaviour, attitudes concerning punishment and received?
Although most individuals may know the meanings, they lack scientific precision.

Once the conceptual propositions have been established (i.e., the meanings in
scientific terms) we then need an operational proposition that defines the
concepts in such a way that they can be observed and measured. This may be
derived from a score achieved on a particular scale of authoritarianism; indirectly,
we may study the relationship between childhood aggression and exposure to
violent television programmes, but we still need to define both the variables
under study – aggression and television violence – in operational terms. The
former might be simply a tally of aggressive acts such as hitting, fighting,
damaging property, etc. Or it might be based on the analysis of projective test
material (Thematic Apperception Test). A panel of judges may be used to
develop an operational definition of aggression by watching a child in a free-play
situation and then rate the child‘s aggressiveness on a five-point scale.
Alternatively, we could observe children as they play with a selection of toys we
had previously classified as aggressive (guns, tanks, knives, etc,) and toys
classified as non-aggressive (cars, dolls, puzzles, etc.).

Defining violence may be a little more difficult to agree on. What constitutes
television violence? The problem here is both cultural and the difference in
precision between what the general public will accept in defining a term and what
researchers will accept. To operationalize the concept of television violence we
could use a checklist of items, such as ―Was there physical contact of an
aggressive nature?‖ ―Has an illegal act taken place?‖ etc. Perhaps you can
establish a criterion that a violent television programme must have five or more
items checked ‗yes‘ for it to be considered violent.

75
 Research Methods and Statistics

From the General to the Operational

Problem or General Hypothesis: You expect some children to read better than
others because they come from homes, in which there are positive values and
attitudes to education.

Research Hypothesis: Reading ability in nine-year-old children is related to

parental attitudes towards education.

Operational Hypothesis: There is a significant relationship between reading

ability for nine-year-old children living in Carlisle as measured by standardized
reading test (state test) and parental attitudes to education as measured by the
attitudinal scale derived from test (state test).

Criteria for Judging Hypotheses

1. Hypotheses should be clearly stated. General terms such as personality,

self-esteem, moral fibre, etc. should be avoided: the statement demands
concise, scientific definition of the concepts and terms:
―Personality as measured by the Eysenck‘s Personality Inventory . . .‖

2. Hypotheses predict an outcome, an obvious necessity is that instrument

exists to provide valid and reliable measures of the variables involved.
3. Hypotheses should state differences or relationships between variables.
A satisfactory hypothesis is one in which the expected relationship is
made explicit.

4. Hypotheses should be limited in scope. Hypotheses of global significance

are not required. Those that are specific and relatively simple to test are
preferable.

5. Hypotheses should not be inconsistent with known facts. All hypotheses

should be grounded in past knowledge. The hypothesis should not lead
the cynical reader to say: ―Whatever led you to expect that?‖ or ―You
made this one up after you collected the data.‖

Unconfirmed Hypotheses

Does an unconfirmed hypothesis invalidate prior knowledge or the literature?

Well, either the hypothesis is false, or some of the previous information is
erroneous, or other information has been overlooked, or some information may
have been misinterpreted b the researcher, or the experimental design might
have been incorrect. A new hypothesis may need to be formulated and tested
using a different study – scientific progress in developing alternative paradigms!
Even if the hypothesis is refuted, knowledge is advanced

76
 Research Methods and Statistics

Statistical Hypotheses

In attempting to reach decisions, it is useful to make assumptions about the

population involved in the study. Such assumptions, which may or may not be
true, are called statistical hypotheses and in general are statement about the
probability distributions of the population. In many instances we formulate a
statistical hypothesis for the sole purpose of rejecting or nullifying it. For
example, if we want to decide whether one psychological procedure is better
than another, we formulate the hypothesis that there is no significant difference
between the procedures (i.e. any observed differences are merely due to
sampling error from the same population). Such hypotheses are called null
hypotheses and are denoted by the symbol H0. Any hypothesis that differs from
a given hypothesis is called an alternative hypothesis. A hypothesis alternative
to the null hypothesis is denoted H1.

Tests of Hypotheses and Significance

If on the supposition that a particular hypothesis is true we find that results

observed in a random sample differ markedly from those expected under the
hypothesis. On the basis of chance using sampling theory, we would say that
the observed differences are significant and we would be inclined to reject the
hypothesis (or at least not accept it on the basis of the evidence obtained).
Procedures which enable us to decide whether to accept or reject hypotheses or
to determine whether observed samples differ significantly from the expected
results are called tests of hypotheses, tests of significance or rules of decision.

Type I and Type II Errors

If we reject a hypothesis when it should be accepted, we say a Type I error has

been made. On the other hand, if we accept a hypothesis when it should be
rejected, we say that a Type II error has been made. In either case a wrong
decision has been made or an error of judgement has occurred. In order for any
tests of hypotheses to be sound, they must be designed so as to minimize these
errors of judgement. The probability of making a Type I error is determined
before the experiment. A Type I error can be limited by properly choosing a
level of significance (): a level of significance of 0.05 implies a 5 percent chance
of making the wrong decision. The probability of making a Type II error cannot
be determined before the experiment. It is possible to avoid risking a Type II
error altogether by simply not making them, which amounts to never accepting
hypotheses. However, for any given sample size, an attempt to reduce one type
of error is often accompanied by an increase in the other type of error. The only
way to reduce both types of error is to increase the sample size, which may or
may not be possible.

77
 Research Methods and Statistics

One-Tailed and Two-Tailed Tests

Often we display interest in extreme values of the statistic X or its corresponding

standard z score on both sides of the mean, i.e. in both ―tails‖ of the distribution.
For this reason such tests are called two-tailed tests. In such tests the area
covered by critical regions in both tails is equal to the level of significance, i.e.
/2. However, we may be interested only in extreme values to one side of the
mean, for example when we are testing the hypothesis that one procedure is
better than another. Such tests are called one-tailed tests and the critical region
to one side of the distribution has an area equal to the level of significance. The
critical values of z for levels of significance are found by use of tables of the
normal distribution:

Level of significance  0.10 0.05 0.01 0.005 0.002



Critical values of z for -1.28 -1.64 -2.33 -2.58 -2.88

one-tailed test or or or or or
+1.28 +1.64 +2.33 +2.58 +2.88

Critical values of z for -1.645 -1.96 -2.58 -2.81 -3.08

two-tailed test and and and and and
+1.645 +1.96 +2.58 +2.81 +3.08

78
 Research Methods and Statistics

Tests Involving the Normal Distribution

Suppose that under a given hypothesis the sampling distribution of a statistic X is

normally distributed with mean  and a standard deviation , then the
distribution of the standardized score (z) is given by (X - )/

In an experiment on extra-sensory perception a subject in one room is asked

to state the colour (red or blue) of a card chosen from a deck of 50 well-
shuffled cards by an individual in another room. It is unknown to the
subject how many red or blue cards are in the deck. If the subject identified
35 cards correctly, determine whether the results are significant at the 0.05
level of significance.

Step 1
A statistical or Null Hypothesis is set up

This is the initial assumption is almost invariably an assumption about the value
of a population parameter: the probability (p) of the subject choosing the colour of
the card correctly is 0.5, hence the subject is merely guessing and the
experimental results are due to chance:

H0 : p = 0.5

Step 2
An Alternative Hypothesis is defined that is accepted if the test
permits us to reject the null hypothesis

The subject is not guessing and the experimental results are indicative of the
subject having powers of extra-sensory perception:

H1 : p > 0.5

Step 3
An appropriate level of significance () is established

By convention, a level of significance of 0.05 (or 5 percent) is sufficient in most

situations. However, should the consequences of wrongly rejecting the null
hypothesis (H0) be serious enough to warrant it, a 0.01 level of significance can
be applied. We choose a one-tailed test, since we are not interested in the
ability to obtain extremely low scores but rather the ability to obtain high scores:

 = 0.05

79
 Research Methods and Statistics

Step 4
The appropriate sampling distribution is defined and the
critical values established to identify the accept/reject regions

If the null hypothesis (H0) is true, the mean and standard deviation of the number
of cards identified correctly is given by:

 = (number of cards in the pack * probability) = Np = 50(0.5) = 25

 = [Np(1-p)]0.5 = [50(0.5)(1-0.5)]0.5 = (12.5)0.5 = 3.54

Z1

Critical Region
 = 0.05

Accept H0 Reject H0


Step 5
The decision rule is established

For a one-tailed test at a level of significance of 0.05 we must choose z1 so that

the area of the critical region of high scores is 0.05. Then the area between 0
and z1 = 0.4500, and z1 = 1.645 (read from statistical tables). Thus our decision
rule or test of significance is:

(i) If the z score observed is greater than 1.645, the results are significant at
the 0.05 level and the subject has powers of extra-sensory perception:
accept H0

(ii) If the z score is less than 1.645 the results are due to chance, i.e. not
significant at the 0.05 level, and reject H0.

80
 Research Methods and Statistics

Step 6
The position of the Sample result is computed

If the null hypothesis is valid, the sampling distribution will have a mean of  = Np
= 50(0.5) = 25, and a standard deviation of  = [(Np(1-p)]0.5 = 50(0.5)(0.5) = 3.54

However, we need to apply a continuity correction, since 32 on a continuous

scale is between 31.5 and 32.5, hence

z = (X-)/ = (31.5-25)/3.54 = 1.84

Step 7
Conclusion

Since 32 in standard score = 1.84, which is greater than 1.645, decision (i) holds,
i.e. we accept H0 : we conclude that the subject has powers of extra-sensory
perception. It does not follow from this that the null hypothesis is true, it merely
means that there is insufficient evidence to reject it (this is equivalent to a ―not
proven‖ verdict).

Tests Involving Differences in Means and Proportions

Let m1 and m2 be the sample means obtained from a large sample of sizes n1
and n2 drawn from respective populations having means and  and standard
deviations andConsider the null hypothesis that there is no difference
between the population means, i.e.  =or the two samples are drawn from
populations having the same mean.

An aptitude test was given to two groups of people consisting of 40 and

50 subjects respectively. In the first group the mean score was 74 with a
standard deviation of 8, while in the second group the mean score was 78
with a standard deviation of 7. Is there a significant difference between
the performance of the two groups at the 0.05 level of significance?

Suppose the two groups come from two populations having the respective
means of  and Then we have to decide between the hypotheses:

H0 :  =  and the difference is merely due to chance

H1 : < , or > and there is a significant difference

between the two groups

81
 Research Methods and Statistics

Under the null hypothesis it is assumed that both groups come from the same
population. Since the population standard deviation is not known, an estimate of
the population standard deviation is computed by pooling the two sample
standard deviations:

est = [(/n1) + (/n2)]0.5 = [(82/40) + (72/50)]0.5 = 1.606

Then z = (m1 – m2)/est = (74 – 78)/1.606 = -2.49

For a two-tailed test the results are significant at the 0.05 level if z lies outside the
range –1.96 to +1.96. Hence we conclude that at a 0.05 level there is a
significant difference in performance of the two groups and that the second group
is probably better.

Two groups, A and B, consist of 100 subjects each who have a phobia
towards spiders. A psychoanalytical technique is administered to group
A but not group B (which is called the control group); otherwise the two
groups are treated identically. It is found that in groups A and B, 75 and
65 subjects, respectively, manage the phobia effectively. Test the
hypothesis that the psychoanalytical technique is effective using a 0.05
level of significance.

Let p1 and p2 denote respectively the population proportions effectively managing

their phobia (a) using psychoanalysis, (b) not using psychoanalysis. We must
decide between two hypotheses:

H0 : p1 = p2 and observed differences are due to chance

i.e. psychoanalysis is ineffective

H1 : p1 > p2 and psychoanalysis is effective

Under the hypothesis H0 the average proportion of success is given by:

 = (p1 + p2/n1 + n2) = (75+65)/(100+100) = 70

and an estimate of the population standard deviation is given by:

est = [p(1-p)(1/n1 + 1/n2)]0.5 = [(0.70)(0.30)(1/100+1/100)]0.5

= 0.0648

Then z = (p1 – p2)/est = (0.75 – 0.65)/0.0648

= +1.54

82
 Research Methods and Statistics

On the basis of a one-tailed test at a 0.05 level of significance, we would reject

the null hypothesis only if the z score were greater than 1.645. Since the z
score is 1.54, we must conclude that the results are due to chance at this level of
significance, and that the psychoanalytical technique is no more effective in
managing arachnophobia than a placebo.

It should be noted that our conclusions depend on how much we are willing to
risk being wrong. If the results are actually due to chance and we conclude that
the psychoanalytical treatment is effective (Type I error), we might proceed to
treat large numbers of people only to find then that it is ineffective. However, in
concluding that the psychoanalytical technique does not help when it actually
does (Type II error), may be a dangerous conclusion especially if people‘s well
being is at stake.

Associated Normal Approximations

The Binomial Distribution

The binomial distribution is a discrete probability of an event occurring in any

single trial (called the probability of success). Probabilities associated with the
binomial distribution are readily available from the formula:

px(1-p)n-x
p(X) = [n!/x!(n-x)!]

When n is large the computation becomes extremely tedious to operate.

However, if n > 50 and p > 0.1 and np > 5 the normal distribution can be used to
approximate the binomial:

It has been decided to administer a multiple-choice questionnaire to the

first year candidates studying for a degree in psychology. The
examination will comprise of 100 questions, and the candidates are asked
to choose the correct answer from four possible answers for each
question. However, some academic colleagues, who claim that a pass
score of 40 could be easily achieved by guessing, greet this departure
from the traditional examination with skepticism. Test this claim at the
0.05 level of significance.

In this case n = 100 questions and the probability of successfully answering each
question is 1 in 4 = 0.25. By substitution:

 = np = 100(0.25) = 25

 = [np(1-p)]0.5 = [(100)(0.25)(0.75)]0.5 = 4.33

83
 Research Methods and Statistics

Then z = (x - )/ = (40 – 25)/4.33 = 3.364

From tables of the normal distribution we can determine the probability of a

candidate deriving a standard score:

z = +3.364 = 0.00039 (or, less than 0.04 percent).

We can conclude that the multiple-choice questionnaire is reliable if we are

willing to accept one candidate in 2500 to obtain 40 or more correct answers
(and obtain a pass grade) simply by guessing.

84
 Research Methods and Statistics

VII
Small Sampling Theory

For samples of size n>30, called large samples, the sampling distributions of
many statistics are approximately normal. In such cases we can use the statistic
z to formulate decision rules or tests of hypotheses and significance. In each
case the results hold good for infinite populations. In a test of means in
sampling without replacement from finite populations, the z score is given by:

z = m - 
(n-1)0.5

where z the statistic for the normal distribution

m sample mean
 population mean
 population standard deviation
n sample size
and, /(n-1)0.5 is the standard error

However, for sample of size n<30, called small samples, this particular
approximation is not good and becomes worse with decreasing sample size.
Hence, exact sampling with appropriate modifications must be made. If we
consider sample of size n drawn from a normal, or assumed normal, distribution
with a population mean of , using the sample mean m and sample standard
deviation s, the sampling distribution for the t statistic can be expressed:

t = m - 
s/(n-1)0.5

= [(m -s] n-10.5

where t the statistic for Student‘s t-distribution

m sample mean
 population mean
s sample standard deviation
n sample size

85
 Research Methods and Statistics

Tests of hypotheses and significance are easily extended to problems involving

small samples, the only difference is that the z statistic is replaced by the more
suitable t statistic.

Degrees of Freedom

In order to compute the statistic or population parameter such as an estimate of

the population standard deviation, it is necessary to use observations obtained
from a sample. The number of degrees of freedom of a statistic generally
denoted by  is defined as the number of independent observations in the
sample (i.e. the sample size = n) minus the number of population parameters (k)
which must be estimated from sample observations. In other words  = n – k.

Example From a sample of ten observations we have derived the sample

mean and sample standard deviation, m and s respectively, in
order to compute an estimate of the population mean However,
since we must estimate the population parameter , k = 1 and so 
= (n – 1) = (10 – 1) = 9. The degrees of freedom ( is used to
derive the value of t from the tables of the t distribution: the
confidence limits for  = 0.4 to 0.0005 is read horizontally at the top
of the table, the degrees of freedom = 1 to >120 is read vertically.
Hence, for a one-tailed test at the 0.05 level of significance for ten
observations we find t 0.05 for (10-1) = 9 degrees of freedom to be
1.833; and, for a two-tailed test, we find t0.025 for (10-1) = 9 degrees
of freedom to be 2.262

Confidence Limits

In a similar manner to the normal distribution, we can define 95 percent or other

confidence limits by using the table of the t distribution. Likewise, we can
estimate within specified limits of confidence the population mean . For
example, for a 95 percent confidence interval the values are +t0.975 (i.e. we have
0.025 of the area in each tail of the t distribution):

est = sample mean + tcalc s/(n-1)0.5

86
 Research Methods and Statistics

A survey sample of 10 subjects gives a mean of 30 with a standard

deviation of 4. Estimate the population mean with 95 percent
confidence.

1. We are dealing with a small sample of n<30 and the population standard
deviation  is unknown. Hence, the sample mean of 30 is a random
selection from a t distribution with a population standard deviation =
standard error = s/n0.5. Thus, we can state that the sample mean is no
more than t0.025 standard errors from the population mean:

 lies between 30 + t0.025 s/n0.5

2. From tables of the t distribution the value for t0.025 based on (n-1) = 8
degrees of freedom = + 2.306. Hence, substituting in the formula:

 lies between 30 + 2.306(4/90.5) = 30 + 3.07

3. We can conclude that we are 95 percent confident that the population

mean score lies between 26.93 and 33.07

One- or Two-Tailed Tests

To test the hypothesis that the observed sample mean (m) is equal to the
population mean (), or that the observed mean (m1) from one sample is equal to
the observed mean of another sample (m2):

H0 : m = 
or, H0 : m1 = m2 ... ... two-tailed test

To test the hypothesis that the observed sample mean (m) is greater than, or
less than, the population mean ():

H0 : m > 
or, H0 : m < 
or, H0 : m1 > m2
or, H0 : m1 < m2 ... ... one-tailed test

Similarly, to test the hypothesis that an observed sample mean is not equal to the
population mean, or that an observed sample mean is not equal to the observed
mean of another sample, we use a two-tailed test. In a two-tailed test, the area
representing the level of significance must be distributed between the two tails.

87
 Research Methods and Statistics

Thus, for a level of significance of  = 0.05 the area in each tail is equal to /2 =
0.025:

< 0.95 >

/2=0.025   /2=0.025
 < 0.95 >  = 0.05
CV1 CV2 0 CV
t0.025, v 0 +t0.025, v +t0.05, v

Two-Tailed Test (x = ) One-Tailed Test (x < )

Tests of Hypotheses and Significance

Comparative Test of Means

To test the hypothesis H0 that m =  or m1 = m2 from a population with unknown

standard deviation:

Step 1
Establish the Null Hypothesis

The mean scores achieved in the post-test by subjects following the traditional
learning programme are equal to the mean scores achieved in the post-test by
subjects exposed to the computer-aided learning programme,

i.e. H0 : m1 = m2

Alternative Hypotheses

The mean scores achieved in the post-test by subjects following the traditional
learning programme are not equal to the mean marks achieved in the post-test
by subjects exposed to the computer-aided leaning programme.

88
 Research Methods and Statistics

H1 : m1 > m2
or, H1 : m1 < m2
i.e. H1 : m1 is not equal to m2

Step 2
Decision Rule

For a level of significance  = 0.05 with v = (n1 + n2 – 2) = (40+20-2) = 58

degrees of freedom:

If the t calculation is less than –2.000, or greater than +2.000, then reject the null
hypothesis and accept the alternatives hypotheses, otherwise accept H 0.

Step 3

Graphically describe the distribution showing the critical values and the region of
acceptance/rejection.

CV1 CV2
0.95
/2=0.025 /2=0.025

-2.000 0 +2.000

Reject H0 Reject H0
Accept H0

89
 Research Methods and Statistics

Step 4
Collect Sample Data and Compute the Result of the Test

Post-Test Score Post-Test Score

Subject Pre-Test Traditional Computer-Aided
Score Programme Programme
111011 23 64
111219 40 95
111232 27 93
111234 41 90
111235 15 41
111241 37 61
111243 37 58
111247 47 84
111248 31 68
111250 36 44
111251 34 82
111252 49 71
111254 10 37
111255 35 73
111257 47 93
111258 29 64
111260 49 50
111261 29 68
111266 19 77
111269 43 84
111270 28 63
111271 21 84
111272 25 64
111275 34 71
111277 51 87
111278 44 78
111279 21 79
111280 31 42
111281 17 58
111282 14 52
111283 19 35
111284 18 60
111285 26 49
111286 31 82
111287 34 85
111290 20 63
111291 14 73
111292 27 63
111293 32 53

90
 Research Methods and Statistics

111294 16 76
111295 15 54
111296 31 83
111297 17 70
111298 48 74
111299 18 54
111300 32 67
111301 11 57
111302 43 94
111303 30 91
111310 42 60
111311 49 84
111312 33 48
111315 17 56
111318 29 65
111321 24 63
111322 24 66
111323 14 45
111324 16 70
111325 23 69
111326 31 75

From which we derive:

Traditional Learning Programme: n1 = 40, m1 = 63.50, s1 = 15.41228

Computer-Aided Learning Programme: n2 = 20, m2 = 75.95, s2 = 11.40395

Since the population standard deviation is unknown, a calculated estimate can

be obtained by pooling the two sample standard deviations whilst
accommodating for two unequal sample sizes:

 = [(n1s12 + n2s22 )/(n1 + n2 – 2)]0.5

= [(40)(15.412282)+ (20)(11.403952)]/(40+20-2)

= 17.8462
Hence:

tcalc = (m1 – m2)/[(1/n1 + 1/n2)]0.5

= (63.50 – 75.95)/[17.8462(0.025+0.05)]0.5

= -2.5474

91
 Research Methods and Statistics

Step 5
Conclusion

There is a significant difference between the two sample means.

Since tcalc = -2.5474 which is less than the critical value of –2.000 we must reject
the null hypothesis at the 0.05 level of significance, and accept the alternative
hypothesis. The mean score obtained by the sample of subjects exposed to the
computer-aided learning package is not equal to the mean score by the sample
of subjects following the traditional learning programme.

Indeed, one can claim that the mean score of subjects exposed to the traditional
learning programme is less than the mean score of subjects following the
computer-aided learning programme.

Testing a Specified Population Mean from a Small Sample

Using the standard procedure for significance testing, for example can derive the
appropriate test of a specified population mean:

The population mean score of a proprietary test of occupational aptitude

is 100 with an unknown standard deviation. A sample of eight
experienced and competent subjects are given the test and provide a
mean score of 120 with a standard deviation of 5. How likely is it that
the population mean score is an adequate statistic for the test of
aptitude?

1. Null Hypothesis
H0 :  = 100

Alternative Hypothesis
H1 :  > 100

2. Level of Significance
= 0.05 is considered to be adequate

3. Critical Value

Since the sample size n<30 and the population standard deviation is unknown,
then t is the appropriate sampling distribution, with a mean of 100 and a standard
error of s/(n0.5) = 5/(80.5) = 1.7678.

92
 Research Methods and Statistics

The reject region will be t0.05 standard errors above the population mean; from
tables for t0.05 with (n-1) = (8-1) = 7 degrees of freedom the critical value is
established as +1.895

4. Decision Rule

If the tcalc for the sample mean is less than +1.895 we must accept the null
hypothesis. However, if the tcalc for the sample mean is greater than +1.895 then
the null hypothesis must be rejected and the alternative hypothesis accepted.

Displayed graphically:


CV=1.895



     

    Accept H0 Reject H0


5. Computation of t
Accept Reject
To find the position of the observed sample mean in the t distribution:

t = (m - )/standard error
= (110 – 100)/1.7678
= +5.657

6. Conclusion

Since the t score of +5.657 is greater than the critical value of +1.895 we must
reject the null hypothesis at the 0.05 level of significance, and accept the
alternative hypothesis. The population mean for the proprietary test of aptitude
is not 100; or, a very rare event has occurred. Further sampling would be
advised to confirm this conclusion.

93
 Research Methods and Statistics

Comparing Two Sample Means from Small Samples

A sample of six subjects in an experiment have an average

performance of 96 on the measurement scale with a standard deviation
of 4. Another sample of five subjects in the same experiment have an
average performance of 92 with a standard deviation of 3.5 on the
measurement scale. Do these samples indicate that the overall mean
in the two sample differ?

m1 = 96 s1 = 4.0 n1 = 6
m2 = 92 s2 = 3.5 n2 = 5

1. Null Hypothesis: There is no significant difference between the sample

mean scores achieved by the two experiments:

H0 : m1 = m2

Alternative Hypothesis: The mean scores achieved by the two samples

are not equal:
H1 : m1 > m2; or, m1 < m2

2. Decision Rule: For v = (n1+ n2 –2) = (6+5-2) = 9 degrees of freedom

at the 0.05 level of significance (/2 for a two-tailed test): if the calculation value
for t in the distribution is less than-2.262 or greater than +2.262 then reject the
null hypothesis and accept the alternative hypothesis, otherwise accept.

2. Compute the Result of the Test:

t = (m1 – m2)/SE

where the standard error = (1/n1+1/n2)0.5

Since the population standard deviation () is unknown, this can be estimated
from pooling the two sample standard deviations (s1 and s2):

 = [(n1s12 + n2s22)/(n1+ n2 –2)]0.5 = [(6)(42) + (5)(3.52)] / (6+5-2)]0.5

= 4.18

Hence, SE = 4.18(1/6+1/5)0.5 = 2.5311

and, t = (m1 – m2) / SE = (96-92) / 2.5311

= 1.58

94
 Research Methods and Statistics

4. Conclusion

Since the calculated value of t is greater than the critical value of –2.262, and is
less than the critical value of +2.262, we must accept the null hypothesis: there is
no significant difference between the two sample means at the 0.05 level of
significance.

Computing the Sample Size from a Pilot Study

We have seen that to be 95 percent certain that the sample mean is not more
than a specified limit from the population, or true, mean we can compute:

Population Mean = Sample Mean + t0.025s/(n0.5)

transposing this formula gives:

Sample Size (n0.95) = [t0.025s/L]2

where L is the specified limit of acceptable error, for example:

We have piloted a survey questionnaire using nine respondents and

have a sample mean of 50 with a standard deviation of 8. We wish to
be 95 percent confident of our estimate of the sample size, whilst
accepting an error of 3 units on the measurement sale

For n = 9, s = 8, and m = 50 the 95 percent confident limits of the sample size,

accepting an error of 3 units on the measurement scale, can be computed:

N = [t0.025s/L]2

= [(2.306)(8)/3]2

= 37.814 (say, 38)

Comparing the Variability of Two Samples (Analysis of Variance)

While the z and t tests are useful for examining single differences, such as
sample means, there can be problems when a set of differences is to be
examined. A significance test considers how likely it is that a given result is due
to sampling error rather than representing a real difference. By convention, we
reject the null hypothesis of ‗no difference‘ if the probability of this is as low as

95
 Research Methods and Statistics

0.05 (i.e. this means that if we do twenty experiments or tests we are very likely
to be making at least one Type I error).

The F-distribution is used to compare the variability of two samples rather than
the sample mean values, for example:

We have designed an instrument to measure an attribute and piloted this

on two samples of the population under study. In a series of ten trials we
compare the results obtained with the measurement instrument with
those obtained from a well validated and reliable measurement instrument
derived from the literature. The results indicate that whilst the sample
exposed to our instrument provides a standard deviation of 4.099, the
sample exposed to the existing instrument gives a standard deviation of
2.011. Does the pilot study indicate that our measurement instrument is
less consistent than the existing measurement instrument?

n1 = 10 s12 = 4.099
n2 = 10 s22 = 2.011

1. Null Hypothesis

There is no significant variability between the two measurement instruments

H0 : 12 = 22

Alternative hypothesis

H1 : 12 > 


2. Level of Significance

A level of significance of  = 0.05 (i.e. at the five percent level)

3. Decision Rule

The critical value for F is derived from table of the F-distribution, the degrees of
freedom being:

s12 (the numerator) is read horizontally = v1 = n1 – 1 = 9

s22 (the denominator) is read vertically = v2 = n2 – 1 = 9

Hence, for F0.05, 9, 9 = 3.18

96
 Research Methods and Statistics

CV
0.99

=0.05

3.18
Accept H0 Reject H0

4. Compute the Result of the Test

F = s22/s12

= 16.802/4.044

= 4.153
5. Conclusion

Since the calculated F value of 4.153 is greater than the critical value of 3.18, we
must reject the null hypothesis and accept the alternative hypothesis 6. There is
a significant difference in the variability of the two tests at the 0.05 level of
significance: the scores derived from our measurement instrument are
significantly more variable than those derived from the existing measurement
instrument. However, we would accept the null hypothesis at the 0.01 level of
significance where the F calculated value of 4.135 is less than Fcv = 5.35

6
Occasionally, we may have no prior knowledge of the variability of two observations and a two-
tailed test may be called for. In this instance, the lower critical value for F is the reciprocal of the
upper critical value: in the example above, for v1 = 9, v2 = 9 at 0.05 level of significance, the upper
critical value 4.03 and the reciprocal is 0.2421.

97
 Research Methods and Statistics

The Chi-Square ( Test of Significance

As we have seen, the results obtained in samples do not always agree precisely
with theoretical results expected according to the rules of probability. With
problems of categorical data, the previous methods of applying the z and t
sampling distributions are unsatisfactory.  is a measure of the discrepancy
existing between observed and expected frequencies. Suppose that in a
particular sample a set of events E1, E2, E3 . . En are observed to occur with
frequencies o1 o2 o3 . . . on called observed frequencies, and that according to
probability rules they are expected to occur with frequencies e1 e2 e3 . . . en called
expected frequencies. Often we wish to know whether observed frequencies
differ significantly from expected frequencies.

Event E1 E2 E3 E4 E5 ... En

Observed o1 o2 o3 o4 o5 ... on
Frequency

Expected e1 e2 e3 e4 e5 ... en
Frequency

If 2 = 0, observed and theoretical frequencies agree exactly, while if  > 0 the

do not agree exactly. The larger the value of , the greater is the discrepancy
between observed and expected frequencies.

The  distribution can be defined as:

2 = (observed value – expected value) / expected value

2

= (o1 – e1)2/e1 + (o2 – e2)2/e2 + (o3 – e3)2/e3 + . . . (on – en)2/e

The chi-square distribution is used to test if a series of observed values differs

significantly from that which is expected. The chi-square distribution is typically
skewed:

98
 Research Methods and Statistics

1- CV

0 Chi-Square



Degrees of Freedom

The number of degrees of freedom of the  statistic, denoted v, is defined as the

number of n independent observations in the sample. However, since we must
estimate the population mean () from samples of the population, by convention
v = n – 1. The degrees of freedom (v) for the  distribution is determined:

(Number of Columns (k) in the Table– 1)(Number of Rows (h) in the Table– 1)

Thus, for a table of six columns and four rows, v = (k-1)(h-1) = (6-1)(4-1) = 15

Confidence Intervals for 

As with the normal and t distributions, we can set a level of significance and
define confidence intervals with tables of the  distribution, for example:


 0.05 0.025 0.01 0.005
v
1 3.84146 5.02389 6.63490 7.87944
5 11.0705 12.8325 15.0863 16.7496
10 18.3070 20.4832 23.2093 25.1882
15 24.9958 27.4884 30.5779 32.8013
20 31.4104 34.1696 37.5662 39.9968

99
 Research Methods and Statistics

Contingency Tables

Corresponding to each observed frequency in a (h)(k) contingency table, there is

an expected or theoretical frequency, which is computed according to the rules of
probability. These frequencies which occupy the cells of a contingency table are
called cell frequencies. The total frequency in each row of column is called the
marginal frequency. To investigate the agreement between observed and
expected frequencies, we compute the statistic

 = (o – e ) /e
n n
2
n

The observed frequencies of a phenomenon are displayed:

Column (k)
I II Total
Row (h)

A a1 a2 NA

B b1 b2 NB

Total N1 N2 N

Applying the cell frequency coding above, the expected frequencies are
computed as follows:

Column (k)
I II Total
Row (h)

A N1Na/N N2Na/N NA

B N1Nb/N N2Nb/N NB

Total N1 N2 N

For example:

100
 Research Methods and Statistics

The table below shows the results of an experiment to investigate the

effect of a hypnotherapy technique on a group of subjects who complained
that they did not sleep well. Some subjects were exposed to
hypnotherapy while others were given a placebo treatment (although they
all thought they were getting hypnotherapy). They were later asked
whether the hypnotherapy helped them sleep well or not. The results of
the responses are shown in the table below:

Slept well Did not sleep well

Exposed to the
hypnotherapy 44 10
technique

Given a placebo 81 35
treatment

Assuming that all subjects told the truth, test the hypothesis that there is
no difference between treatment by hypnotherapy and the placebo at a
level of significance of 0.05

1. Null Hypothesis There is no significant difference between the two

treatments, i.e. the application of the hypnotherapy technique to subjects
is independent of the subjects who claimed to have slept well.

Alternative Hypothesis There is a significant difference between the

subjects exposed to the hypnotherapy technique and claimed that they
slept well, and those subjects exposed to the placebo treatment.

2. Level of Significance A level of significance of  = 0.05 is considered

to be acceptable

3. Decision Rule For k columns and h rows, the degrees of freedom =

v = (k-1)(h-1) = (2-1)(2-1) = 1. Hence, for a one-tailed test, for v = 1, and
0.05 the critical value of  = 3.84146

101
 Research Methods and Statistics

CV = 3.84146
1 -  = 0.95

 = 0.05
0
Accept H0 Reject H0

If the calculated value of is less than 3.84146 we must accept the null
hypothesis, i.e. that there is no difference between the hypnotherapy treatment
and the placebo. Otherwise reject.

4. Computation of 

The table of frequencies expected under H0 is calculated from the observations:

Slept well Did not sleep well Total

Exposed to (125)(54)/170 = (45)(54)/170 =
hypnotherapy 39.71 14.29 54
Exposed to (125)(116)/170 = (45)(116)/170 =
placebo 85.29 30.71 116
Total: 125 45 170

Hence, 2 =  (oj – ej)2/ej = (44-39.71)2/39.71 + (81-85.29)2/85.29 +

(10 - 14.29)2/14.29 + (35 - 30.71)2/30.71
= 2.56643

4. Conclusion

Since the calculated  value of 2.56643 is less than the critical value of 3.84146,
the null hypothesis cannot be rejected at the 0.05 level of significance: there is no
significant difference between the two treatments.

102
 Research Methods and Statistics

VIII
Linear Regression and
Correlation

If X and Y denote two variables under consideration, a scatter diagram shows the
location points X,Y on a rectangular co-ordinate system. If all the points in this
scatter diagram seem to lie near a perceived line, the correlation or fit is termed
linear.

*
*
* *
* ** *
* ** *
* *
* * *
*
* *

Positive Linear Negative Linear

* * * *
* * * * * *
* * * * * *
*
* * * *
* * * *
* * * * * * *
* * * * * *
* *
**
* * * *** *
* * * ** * * *
* * * * * ** *

No Correlation or ‗Fit‘
Curvilinear

103
 Research Methods and Statistics

Linear Regression

Very often in practice a linear relationship is found to exist between two variable
say, X and Y. It is frequently desirable to express this relationship in
mathematical form by determining the regression equation that connects these
two variables. The least square line is one with the best goodness of fit in that
deviation or error is minimum. The least square line can be expressed:

Y = a + bX

Where a is the intersect of the line with the vertical axis and can be computed:

YX2 -XXY
a = NX2 – (X)2

and b, the gradient or slope of the line:

NXY - XY
b = NX2 – (X)2

b = d/c

d
a
c

0 X

104
 Research Methods and Statistics

Compute the regression equation for the following data, using X as the
independent variable:

X 1 3 4 6 8 9 11 14
Y 1 2 4 4 5 7 8 9

Step 1 Construct a scatter diagram to identify the model and get an

estimate of the intersect a

Y 15-

10-
*
*
*
5- *
* *
*
*

0 5 10 15 20 X

The scatter diagram indicates that the estimate for the value a is positive, and
lies between 0 and 5

Step 2 Compute the values for a and b

X Y X2 XY Y
1 1 1 1 1
3 2 9 6 4
4 4 16 16 16
6 4 36 24 16
8 5 64 40 25
9 7 81 63 49
11 8 121 88 64
14 9 196 126 81

Total 56 40 524 364 256

105
 Research Methods and Statistics

Subsequently, the standard equations become:

a = [(40)(524) – (56)(364)] / [(8)(524) – (56)(56)] = 0.545

b = [(8)(364) – (56)(40)] / [(8)(524) – (56)(56)] = 0.636

Hence, the regression equation can be expressed:

Y = 0.545 + 0.636X

An alternative method, particularly useful in manually processing complex

numbers, uses the following equations:

Y = aN + bX

XY = aX + bX2

hence,
40 = 8a + 56b . . . ... ... (1)
364 = 56a + 524b . . . ... ... (2)

by multiplying equation (1) by 7, we can solve the equations simultaneously and

by substitution:

280 = 56a + 392b . . . ... ... (1)

364 = 56a + 524b . . . ... ... (2)

subtract equation (1) from equation (2):

84 = 132b . . . ... ... (3)

hence, b = 84/132 = 0.636

substitute the value for b into equation (1):

40 = 8a + 56(0.636)

= 8a + 35.6364

a = (40 – 35.6364) / 8

= 0.545

The regression equation allows us to compute the value of Y for any value of X.

106
 Research Methods and Statistics

Linear Correlation

The statistic r is called the coefficient of correlation and is defined by the equation
for Pearson‘s Product-Moment formula:

r = NXY – (X)(Y)
[[NX – (X)2][NY2 – (Y)2]]0.5
2

The equation automatically gives the sign of r as well as clearly showing the
symmetry between the variable X and Y. It should be noted that in the instance
of linear regression, the quantum r is the same, regardless of whether X or Y is
considered as the independent variable. Thus, r gives a very good measure of
the relationship between two or more variables and lies between –1 and +1.
The close the value of r is to 0, the less there is a linear relationship between the
variables.

In the example provided to compute linear regression, we have:

N=8 X = 56 X2 = 524 XY = 364 Y2 = 256 Y = 40

Substituting in the equation:

r = (8)(364) – (56)(40) / [[(8)(524) – (56)(56)][(8)(256) – (40)(40)]]0.5

= 672 / [(1056)(448)]0.5

= 0.977

Coefficient of Determination

The ratio of the explained variation to the unexplained variation is termed the
coefficient of determination. If there is no explained variation (i.e. r = 0), this
ratio is zero. If there is no unexplained variation (i.e. r = 1), this ratio is one.
Since the ratio is always positive, we denote the coefficient of determination as:

R = r2 = (Explained variation)/(Unexplained variation)

= (Yestimate – Ymean)2
(Y – Ymean)2

... since r is a dimensionless quantity.

107
 Research Methods and Statistics

For practical purposes, the coefficient of determination is computed as the

square of the coefficient of correlation (i.e. r2). Using the example above, the
coefficient of correlation has been computed to be r = 0.9777, hence R =
0.95589. Subsequently, we can claim that say, 95.6 per cent of the variability of
Y is explained by the variability of X; and, 4.4 percent of the variability of Y is
unexplained.

However, it should be noted that both the coefficient of correlation and the
coefficient of determination do not measure a cause-and-effect relationship, but
merely the strength of association between two or more variables. For instance,
research in the 1930‘s showed a strong positive correlation between the
incidence of prostitution and the output of steel in Pennsylvania. It would be
nonsense to claim that the former ‗causes‘ the latter.

Sampling Theory of Correlation

The n pairs of values (X, Y) of two variables can be thought of as a sample from
a population of all possible such pairs. Hence, this bivariate population can be
assumed to have a bivariate normal distribution. The population coefficient of
correlation may be denoted . Tests of significance concerning  require the
sampling distribution of r:

1. Test of Hypothesis: There is no correlation between the two variables in

the population, therefore any explained variability in the sample is chance:
 = 0

t = r(n – 2)0.5 for v = n – 2 degrees of freedom

(1 – r2)

2. Test of Hypothesis: There is a correlation between the two variables in the

population, therefore any unexplained variability in the sample is chance:
 < 0, or  > 0

z = 0.5loge [(1 + r)/(1 – r)] = 1.1513log10[(1 + r)/(1 – r)]

The former equation (1) is more generally applicable. Hence, for r = 0.9777:

t = 0.9777(8 – 2)0.5 = 54.306

(1 – 0.9559)

For a two-sided test, the critical value for t at the 0.05 level of significance and for
v = (8 – 2) = 6 degrees of freedom, = +2.447. Since the calculated t value of
54.306 is greater than the critical value of +2.447, we must reject the null
hypothesis and conclude that the coefficient of correlation is greater than 0.
Hence, our observation is significant at the 5 percent level (see footnote 7: for v =
6, r must be greater than 0.7067 at the 5 percent level of significance)

108
 Research Methods and Statistics

Point Biserial Correlation

Point biserial correlation is a derivative of the Pearson product-moment

correlation r, similarly, the tables used for the product-moment correlation can
determine the significance. Point biserial correlation (rpbis) is used when one of
the variables is in the form of a continuous score and the other is a categorical
dichotomy. For example, the reaction times in an experiment (times is continuous
data) and the sex of the subject (the dichotomy male/female is categorical data);
levels of job satisfaction (assumed continuous data albeit on a Likert scale) and
the categories of manager/worker, etc.

Point biserial correlation is particularly useful to determine whether a particular

survey (or test) item discriminates between high and low scores on the
questionnaire as a whole. The continuous data, in terms of the overall test
score, of each individual is designated as the X variable and the categorical data
(for instance, yes/no, right/wrong, agree/disagree, male/female) the Y variable.
An item that has a positive correlation with the total survey score is discriminating
between candidates in the same way as the overall survey. An item that has a
low positive or negative correlation is not discriminating between the low and
high scoring individuals in the survey (or test) in the same way as the overall
survey; hence, it would be advisable to exclude such an item from the
questionnaire (or test).

We have conducted a survey to assess the levels of felt job satisfaction in

a department of ten employees. Recognizing that some of the employees
are managers and others workers, we wish to correlate the scores with the
category of the respondents: manager or worker. The individual total
scores and the workers’ perceived level in the hierarchy are given below:

Respondent Overall JDI Perceived Level

(random number) Score (manager = 1, worker = 0)
285 60 0
271 56 0
952 51 0
067 58 0
502 49 1
176 48 0
331 55 1
642 45 1
858 47 1
441 55 0

109
 Research Methods and Statistics

The point biserial correlation is defined as:

rpbis = X0 – X1 [(p)(1-p)]0.5
 x

where
X0 = mean score of X for respondents scoring 0 on Y = 54.67
X1 = mean score of X for respondents scoring 1 on Y = 49.00
x = standard deviation of all scores X = 4.820
p = proportion of respondents scoring 0 on Y = 0.6
(1-p) = proportion of respondents scoring 1 on Y = 0.4

Thus,

rpbis = [(54.67 – 49.00)/4.82)] * [(0.6)(0.4)]0.5

= (5.67/4.82) * (0.24)0.5

= 0.5763

Conclusion: There is a moderately high positive correlation between the

perceived occupational status of the respondents and the overall score in the
level of job satisfaction. However, the correlation is significant at the 0.10 level
of significance (r > 0.5494), but not at the 0.05 level (r > 0.6319).7

7
See White, J., Yeats, A. and Skipworth, (1991), Tables for Statisticians, Table 15, p.29, Critical
values of the product-moment correlation.

110