Block-3
Uploaded by
Jorge Vega RodríguezBlock-3
Uploaded by
Jorge Vega RodríguezBlock 3
Statistical Analysis
BLOCK 3 INTRODUCTION
Block 3 “Statistical Analysis” deals with the measures of central tendency and dispersion,
sampling distribution and statistical analysis.
Unit 7 “Descriptive Statistics-I” deals with different types of measures of central tendency
and measures of dispersion.
Unit 8 “Descriptive Statistics-II” deals with different types of correlations, regression,
regression analysis and regression coefficients.
Unit 9 “Sampling Distributions” deals with sampling distribution, the concept of standard
error and the Central Limit Theorem.
Unit 10 “Statistical Analysis-I” deals with the concept of hypothesis, the procedure of
testing a hypothesis, and also the test of hypothesis for large samples for the population
mean and the difference between two population means.
Unit 11 “Statistical Analysis-II” deals with the small sample test procedures for a single
population mean and the difference between two population means based on t-
distribution. Also, the unit delves into the chi-square test for the goodness of fit and
independence of attributes. The test for the equality of two variances using F- distribution
is also discussed in the unit.
Unit 12 “Analysis of Variance” delves into the one-way as well as two-way Analysis of
Variance.
Objectives
After studying this block, you should be able to:
explain the different types of measures of central tendency;
explain the measures of dispersion;
explain the concept of correlation between two variables and identify different
types of correlations;
explain regression, regression analysis and linear regression;
explain the salient properties of regression coefficients;
explain the concept of sampling and sampling distribution;
explain the concept of standard error and the Central Limit Theorem;
describe the sampling distribution of the sample mean and the difference
between the two sample means;
describe the procedure of testing a hypothesis;
describe the small sample test procedures for a single population mean and
the difference between two population means based on t-distribution;
explain the chi-square test for the goodness of fit and independence of
attributes;
describe the test for the equality of two variances using F-distribution; and
explain the Analysis of Variance technique;
We hope that after studying this block you will acquire an understanding of statistical
analysis.
Wishing you success in this endeavour!
Descriptive Statistics-I
UNIT 7 DESCRIPTIVE STATISTICS-I
Structure
7.1 Introduction
7.2 Objectives
7.3 Measures of Central Tendency
7.3.1 Significance of the Measure of Central Tendency
7.3.2 Properties of a Good Average
7.3.3 Different Measures of Central Tendency
7.4 Arithmetic Mean
7.4.1 Calculation of Simple Arithmetic Mean
7.4.2 Combined Arithmetic Mean
7.5 Median
7.5.1 Finding the Median for a Set of Data
7.6 Mode
7.6.1 Calculation of Mode
7.7 Measures of Dispersion
7.8 Range
7.9 Mean Deviation
7.10 Standard Deviation and Variance
7.10.1 Root Mean Square Deviation (RMSD)
7.10.2 Standard Deviation (S.D.)
7.10.3 Variance
7.11 Coefficient of Variation
7.12 Let Us Sum Up
7.13 Key Words
7.14 Suggested Further Reading/References
7.15 Answers to Check Your Progress
7.1 INTRODUCTION
A frequency distribution for a given set of numerical data and different
graphs using the raw data in hand; which, in fact, is a method of
summarization of data for having a rough or approximate idea about the
nature, properties and pattern of variation among the values of the variable.
But what, if we need to observe whether there is some tendency in the data
set to concentrate around a single numerical figure or, in other words, what
would be the appropriate single representative value of the variable,
representing most of the part of the data set? Obviously, then we have to find
such a numerical figure based upon all the values of the characteristic and not
only with only one or few of the values, in order to make it the best 151
Statistical Analysis representative figure. The choice of such a single number or summary
statistics that we choose to summarize a given data depends upon the
particular characteristics we want to describe. The statistical measures which
describe such characteristics are called measures of central tendency or
measures of location.
In this unit, we are going to discuss about different measures of central
tendency. Measures of central tendency in Statistics play a very important
and powerful role; since, such a measure reflects most of the features of the
entire mass of data as regards to its tendency to concentrate around a single
value. Generally, any measure of central tendency is called an “average”.
Average of a set of data is actually found to locate the value (or point of
location of it) to which the average value coincides in the range of values of
the characteristic.
Different measures of central tendency provide us an idea about the central
value around which most of the values of the data set are concentrated.
However, being values of a variable and not of a constant, all of them are not
of the same magnitudes; rather they are scattered or dispersed over a given
range of values. This scatteredness of variate-values is a peculiar and natural
characteristic of all the variables; whether they may be less scattered or may
be highly scattered. Thus, even if central values of two different sets of data
are same, the scatteredness will be different of all the data sets. In this
context, we should have different measures of the second property, namely,
property of scatteredness or spreadness of data; or what is popularly called,
“property of dispersion”.
Section 7.3 highlights the meaning of measure of central tendency along with
its significance for the preliminary analysis of the data, important properties
which a good measure (average) is supposed to be satisfied. Section 7.4
describes different types of arithmetic mean which can be computed over
data sets which are tabulated in different manners along with the merits and
demerits of it. Section 7.5 describes the methods of computing median from
data sets which are available in different forms of tabular representation. The
merits and demerits of it have also been mentioned in the section. Section 7.6
describes the methods of computing it from data sets which are available in
different forms of tabular representation. The merits and demerits of it have
also been mentioned in the section. Section 7.7 will present a discussion on
different measures of dispersion, like, range, mean deviation, standard
deviation, etc. Also, we shall highlight the essential properties of such
measures which are desirable for a good measure of dispersion. Section 7.8
deals with the simplest kind of measure of dispersion: Range, its calculation
and merits and demerits. Section 7.9 is devoted to the description of Mean
Deviation as a measure of dispersion along with its computation and its
merits and demerits. Section 7.10 describes Standard Deviation and Variance
as very important and most popular measures. The methods of computing of
both are explained using different kinds of data.
152
Descriptive Statistics-I
7.2 OBJECTIVES
After studying this unit, you should be able to
explain the concept of central tendency;
define an average;
explain the significance of a measures of central tendency;
explain the properties of a good average;
know the different types of measures of central tendency;
calculate arithmetic mean, median and mode for different types of
classification of data; and
calculate range, mean deviation and standard deviation for different
types of data.
7.3 MEASURES OF CENTRAL TENDENCY
It is a bare fact that, whatever be the characteristic under study and the
number of values of the characteristic measured; there is always a tendency
of the data set, so obtained, to concentrate or cluster around a particular
value, which is termed as “central value”. This property of the data tending
towards the central value is known as “property of central tendency”. For
example, if the data on age of primary school children is considered, most of
the ages would be around 5 to 6 years, whereas, most of the ages of
undergraduate students would be around 17-18 years. Thus, central value in
the first data set is 5 years and for second data set is about 18 years.
Central values of data sets are generally those figures which might be
considered to be representative figures to describe the ‘property of central
tendency’ of the data sets. In this sense, these are also known as “measures of
Central tendency” or more popularly the “average”.
According to Professor Bowley, averages are “statistical constants which
enable us to comprehend in a single effort the significance of the whole”.
They throw light as to how the values are concentrated in the central part of
the distribution. In other words, an average is a single value which is
considered as the most representative for a given set of data. Measures of
central tendency show the tendency of some central value around which data
tend to cluster.
7.3.1 Significance of the Measure of Central Tendency
There are two main reasons for studying an average.
1. To get a single representative
A measure of central tendency provides us the idea of that value in the
entire range of the data, to which most of the values of the data set
153
Statistical Analysis tends to cluster and, thus, provides a representative figure for the mass
of data. It is easy to remember this representative figure of any data set
whereas all the values of the set are impossible to remember. For
example, in a class of 50 students, it is not possible to remember all the
heights, but the average height of these students, being a single number,
is easy to remember.
2. To facilitate comparison
Measures of central tendency enable us to compare two or more
populations by deducing the mass of data in one single figure. The
comparison can be made either at the same time or over a period of
time. For example, if a subject has been taught in more than two
classes, obtaining the average marks of students of those classes,
comparison of average performance of students over classes can be
made.
7.3.2 Properties of a Good Average
Since, an average or a measure of central tendency is obtained for any data
set and is considered to be a true representative of all the values of a data set;
it is expected to possess some properties so that one can choose the best
measure depending upon the nature of the data in hand. The following are the
properties of a good measure of an average:
1. It should be simple to understand
Since we use the measures of central tendency to simplify the
complexity of a data so an average should be understandable easily
otherwise its use is bound to be very limited.
2. It should be easy to calculate
An average not only should be easy to understand but also should be
simple to compute so that it can be used as widely as possible.
3. It should be rigidly defined
A measure of central tendency should be defined properly without any
ambiguity. It should have a rigid definition as well as an appropriate
interpretation. It should also have an algebraic formula so that if
different people compute the average from same figures, they get the
same answer.
4. It should be liable for algebraic manipulations
A measure of central tendency should be liable for the algebraic
manipulations, that is, it must have some good properties which widen
its application in different situations. For example, if there are two or
more sets of data and the individual information are is available for
these sets, the formula of the concerned measure can be manipulated in
such a way that the formula of same measure can be obtained for the
154 combined set.
5. It should be least affected by sampling fluctuations Descriptive Statistics-I
A measure of central tendency (an average) must be robust (almost
stable) in the sense that it should not be affected too much if there is a
large variation in samples selected from a particular population. This
means that if we select 10 different groups of observations from same
population and compute the average of each group, then we should
expect to get approximately the same values. However, whatever little
difference between the averages values appear must be because of the
sampling fluctuation only.
6. It should be based on all the observations
If any measure of central tendency is used to analyze the data, it is
desirable that the measure is based upon each and every observation for
its calculation.
7. It should be possible to calculate even for open-end class intervals
If a measure of central tendency is computed for a grouped data, that is,
for a data set arranged in the form of a frequency distribution, then, if
possible, it should be in such a form that it can be calculated for open
end classes also.
8. It should not be affected by extremely small or extremely large
observations
As described in the property 6 above, an average must be dependent on
all the observations of the data and, hence, each and every observation
influences the value of the average. As a representative figure,
therefore, an average should not be highly affected by some of the very
large and very small values which generally occur in any set of data.
7.3.3 Different Measures of Central Tendency
As we described earlier, generally every set of data has a central value around
which the values of the variable tend to cluster. If we wish to know the
central value in a set of data, some measures of central tendency (also known
as measures of location) are computed. These measures are popularly known
as “average”.
In Statistics, some of the measures of central tendency which are very much
popular are as follows:
1. Arithmetic Mean or simply Mean;
2. Median;
3. Mode;
4. Geometric Mean;
5. Harmonic Mean.
We shall discuss Arithmetic Mean, Median and Mode only out of these
measures one at a time in next sections.
155
Statistical Analysis
7.4 ARITHMETIC MEAN
“Arithmetic mean” (also called “mean”) of any set of data is defined as the
sum of all the observations of the data set, divided by the total number of
observations in the set. It is the most popular and widely used measure of
central tendency.
As per above definition of mean, mathematically it can be written as
Sum of all the observations of the data set
Mean = … (7.1)
Number of observations in the set
This is the only basic formula for computation of mean. However, you will
come across with some other formulae for computing mean in further text;
but these formulae in no way provide any new method of computation of
mean; rather all these formulae are, in fact, reducible to this basic formula on
simplification.
We have seen that in Statistics, generally the raw data are summarized in a
number of forms, which are
1. Individual series or “ungrouped data”;
2. Discrete frequency distribution and
3. Continuous frequency distribution.
You might be knowing that the data summarized in the forms 2 and 3 are
popularly termed as “grouped data”. Accordingly, you will see below that the
basic formula for the computation of mean, which is used for ungrouped data,
is presented in a different form for the grouped data.
7.4.1 Calculation of Simple Arithmetic Mean
(I) For Ungrouped Data
(a) Direct Method
If individual values of a variable are given in the form of individual series,
that is, the data are not grouped into classes; the basic formula is directly
applicable for the computation of mean.
Thus, for a given set of data with N values x1, x 2 , , x N ; the mean would be
N
x x 2 x 3 xN
X 1
x
i1 i
. … (7.2)
N N
where X stands for the arithmetic mean or mean of the set; X being the
characteristic (variable) under study.
(b) Short-Cut Method
Under this method, we arbitrarily choose two constants, say A and h; A is
used to change the origin of the values and h for changing the scale of values.
156
The constant A is frequently called “assumed mean”. This technique is very Descriptive Statistics-I
much used in daily life.
If x1, x 2 , x3 ,..., xN are the N observations for the variable X and let A and h be
the chosen constants, where A be the assumed mean, then let deviations of
x -values from A respectively be d1,d2 ,d3 ,...,dN where
di xi A , i 1, 2,, N.
di
Further, let ui = , where u1, u2, …, uN are N values of the transformed
h
variable U. Then, we have
N N N N
X =
x
i1 i
A d
i 1 i i1
A h u
i1 i
N N N
N
NA
h
u
i1 i
A hU;
N N
N
where, U =
u
i1 i
being the mean of the transformed variable U.
N
Thus, the relation between the mean of the original variable X and the mean
of the transformed variable U is obtained as
X A hU. … (7.3)
We see that the mean is affected by change of origin and scale both, since the
relation obtained above, involves both the constants, A, used to change the
origin and h, the constant used for changing the scale.
Example 1: Calculate the mean of runs scored, as given below, by ten
players in 20-20 cricket match:
Player 1 2 3 4 5 6 7 8 9 10
Score 15 20 25 19 12 11 13 17 18 20
Solution: We have been given the scores of 10 players in a 20-20 cricket
match. Therefore, we have N = 10.
Now for calculating the arithmetic mean or average score of 10 players, we
need to follow the following steps:
1. Calculate the sum, x , by adding the scores of 10 players
2. Divide the total by N, the number of observations,
In symbols:
15 20 25 19 12 11 13 17 18 20 170
X 17 .
10 10
Thus, the mean score of the 10 players is 17 runs.
Example 2: Use short-cut method to calculate arithmetic mean for the data
given in Example 1.
Solution: We follow the following steps for calculating the arithmetic mean
or average score of the 10 players using short-cut method as: 157
Statistical Analysis 1. Decide the value of the assumed mean A and the constant h.
xi A
2. Calculate the values ui (i=1, 2, … , 10) of the transformed
h
variable U.
10
3. Obtain the sum, i1ui and hence, the value of U. Use the formula (1.3) to
get the mean X .
In this series of observations, since the values range from 11 to 25; therefore,
18, a neat round value in the middle of 11 and 25, may be taken as assumed
mean, that is we take A =18. Further, let h = 1.
With these values of A and h, we have the following table:
Players Scores (X) U = (X – 18)/1
1 15 –3
2 20 +2
3 25 +7
4 19 +1
5 12 –6
6 11 –7
7 13 –5
8 17 –1
9 18 0
10 20 +2
10
u
i 1 i
= - 10
Using formula (7.3), we get
X A hU A h
u i i
18 1
10
17 runs.
N 10
It can be observed in the example that we have taken h = 1, since, the
transformed values are one-digit numbers, which are easy to tackle in the
computation process. This is the wise choice of the constant h; otherwise, we
take h = 10 or 100, if the transformed values would be two-digit or three-digit
figures, which are not so easy to tackle.
(II) For Discrete Frequency Distribution
By a discrete frequency distribution, we mean a grouped data where all the
values of the concerned variable are given separately associated with the
number of times they occur in the data (you know that the number of times a
value occur in the data is called the “frequency” of the value). Thus, in fact, it
is also a form of the grouped data. For example, suppose that the k distinct
values in the data are x1, x 2 , x3 ,..., xk with associated frequencies f1,f2 ,f3 ,...,fk
respectively.
Since the value x i appears fi times (i = 1, 2, .., k) in the data, the sum of all
observations belonging to the set is equal to
158
k Descriptive Statistics-I
x x1 x1 x1 + x 2 x 2 x 2 + … + xk xk x k
i1 i
f1 times f2 times fk times
k k
x f1x1 f2 x 2 fk xk
i 1 i fx.
i1 i i
The total number of observations in the set is obviously given by
k
f1 f2 fk f.
i 1 i
Therefore, using the basic formula (5.1), we have the formula for computing
mean of a discrete frequency distribution as
k
Mean =
fx .i1 i i
… (7.4)
k
f i 1 i
Both the methods, namely, direct method and short-cut method for
computing the mean are also applicable in this case without any new concept.
Therefore, without adding any new thing here, we present Examples 3 and 4
below to illustrate how these methods can be applied for computing mean of
a discrete frequency distribution.
(a) Direct Method
Let us consider the following discrete frequency distribution for illustration:
Example 3: The following table gives the data relating to the marks (out of
10) of 100 students in a statistics test:
Score(X) 0 1 2 3 4 5 6 7 8 9 Total
Frequency (f) 1 2 4 9 15 21 20 15 9 4 100
Find the average marks of the students:
Solutions: We can use directly the formula (1.4) for the purpose. For this, we
prepare the following table for finding the numerator and denominator of the
formula:
Frequency
Score (X) fX
(f)
0 1 0
1 2 2
2 4 8
3 9 27
4 15 60
5 21 105
6 20 120
7 15 105
8 9 72
9 4 36
Total 100 535
159
Statistical Analysis Thus, by using the formula we get
X
fxi i
535
5.35 ;
f i 100
which provides the average marks of the students in statistics test.
(b) Short-cut Method
The short-cut method of calculating the mean, as described in sub-section
7.4.1 (Ib) can also be applied in this case with the same aim to reduce the
complexity of the computation process. We may similarly choose appropriate
constants A and h, where A is the assumed mean. Defining the transformed
variable U through the transformation U =
X A , we can find the u values
i
h
for each xi values and hence, we can find the mean U of the transformed
variable in order to apply the formula (1.3).
Let us describe the steps which we follow in short-cut method while applying
it to a discrete frequency distribution.
Step I: Determine the appropriate values of assumed mean A and the
constant h.
Step II: Find the values ui =
xi A corresponding to each x .
i
h
Step III: Find the values fiui for each i and hence, find their sum, f u .
i
i i
Step IV: Find the mean of the transformed variable U, using formula
U
fu .
ii i
f ii
Step V: Use the formula X A hU for computing the mean of the variable X.
Now let us solve the following exercises for illustrating the calculation
process:
Example 4: Following table gives the wages paid to 125 workers in a
factory. Calculate the arithmetic mean of ways for the data by using short-cut
method.
Wage (in Rs) 200 210 220 230 240 250 260
No. of Workers 5 15 32 42 15 12 4
Solution: In this case, assume mean A is taken as 230 because it is neatly at
the centre of the range of the variable X. Let us take h = 1.
The other needed calculations are shown in the following table:
160
Wages No. of Deviations fU Descriptive Statistics-I
(X) Workers U = (X – A)/h
(f)
200 5 –30 –150
210 15 –20 –300
220 32 –10 –320
230 42 0 0
240 15 10 150
250 12 20 240
260 4 30 120
125 f u 260
i i
i
Using the formula (1.3), the arithmetic mean for the variable X will be
X A hU A h
fu
ii i
230
260
1 230 2.08 227.92.
f ii
125
(III) For Continuous Frequency Distribution
We can compute the mean of a continuous frequency distribution either using
the direct method or by using the short-cut method. The only thing in both the
methods is that, given the class intervals and corresponding class frequencies;
we obtain first the mid values of each class intervals and consider these mid
values as the k values x1, x2, …, xk of the variable X and the given class
frequencies f1, f2, …, fk as frequencies associated with the respective values
x1, x2, …, xk.
In order to illustrate the procedure of calculation, we consider the following
two examples:
(a) Direct Method
Let us consider the following example:
Example 5: Using direct method, calculate arithmetic mean for the following
data:
Class 20-25 25-30 30-35 35-40 40-45 45-50 50-55
Frequency 8 10 12 20 11 4 5
Solution: For computing arithmetic mean for the given continuous frequency
distribution, we first obtain mid values (X) for all the classes. After
determining these mid-values, formula (1.4) is used for calculating X. The
calculations needed are shown in the following table:
Classes Mid-values Frequency(f) fX
(X)
20-25 22.5 8 180.0
25-30 27.5 10 275.0
30-35 32.5 12 390.0
161
Statistical Analysis 35-40 37.5 20 750.0
40-45 42.5 11 467.5
45-50 47.5 4 190.0
50-55 52.5 5 262.5
f = 70
i
i fxi i 2515
Using formula (1.4), the Arithmetic Mean will be
X
fx
i i
2515
35.93 .
f i 70
(b) Short-Cut Method
Considering the mid values of the given class intervals as the values Xi,
follow the process as shown in the Example 4 with suitable choices of A
and h.
Consider the following example:
Example 6: Calculate arithmetic mean for the data given in Example 5, using
short-cut method.
Solution: In this case, let the assumed mean A = 37.5 and h = 1. Rests of the
calculations are presented in the following table:
Classes Mid Values Frequencies (f) Deviations fU
(X) U= (Xi –
A)/h
20-25 22.5 8 –15 –120
25-30 27.5 10 –10 –100
30-35 32.5 12 –5 –60
35-40 37.5 (A) 20 0 0
40-45 42.5 11 5 55
45-50 47.5 4 10 40
50-55 52.5 5 15 75
f 70
i f u 110
i
i i
Using the formula (7.3), the arithmetic mean can be obtained as
X A hU A h
fuii i
37.5
110
37.5 1.57 35.93.
f ii
70
7.4.2 Combined Arithmetic Mean
If the arithmetic means and the number of observations of two groups are
known, we can calculate the combined arithmetic mean of these groups. The
combined mean formula for two groups is as under
N1X1 N2 X2
X12
N1 N2
162
Here, X12 combined mean of two groups; Descriptive Statistics-I
(N1, X1 ) : Number of observations and mean of the first group;
(N2, X2 ) : Number of observations and mean of the second group;
Similarly for the three groups, the formula becomes
N1X1 N2 X2 N3 X3
X123 .
N1 N2 N3
The formula can also be extended for k – groups
N1X1 N2 X 2 ... Nk Xk
X12...k .
N1 N2 ... Nk
Let us consider the following examples:
Example 7: The mean marks of the 120 students in section A is 80 and mean
marks of 80 students in section B is 90. Find the combined mean of both the
sections.
Solutions: Here N1 120, N2 80, X1 80 and X2 90.
Using formula, the combined mean of all the 200 students will be
N1X1 N2 X2 120 80 80 90
X12
N1 N2 120 80
9600 7200 16800
X12 84marks
200 200
CHECK YOUR PROGRESS 1
Note: i) Check your answers with those given at the end of the unit.
1) Find the arithmetic mean of the observations 5, 8, 12, 15, 20, 30.
2) For the following discrete frequency distribution, find arithmetic
mean.
Wages (in Rs) 20 25 30 35 40
No. of workers 4 8 20 12 6
3) Find arithmetic mean of the distribution of marks given below.
Marks 0-10 10-20 20-30 30-40 40-50
No. of
6 9 17 10 8
students
163
Statistical Analysis
7.5 MEDIAN
Generally, in a raw data set, values of the study variable are given in a
haphazard way, that is, in the order in which values are collected in the
experiment or survey and, therefore, it does not follow any rule of
arrangement. However, for any statistical analysis, it is sometimes advisable
to arrange the data in the form of an ‘array’, that is, arranging the values
either in an ascending or descending order of magnitude. In an array, every
observation holds a certain rank. For example, when arranged in ascending
order of magnitude, the very first value is the smallest one, the second value
is the second smallest value, and so on; but when arranged in the descending
order, the very first value is the largest value, the second value is the second
largest value and so on. Thus, it is the way of arrangement which changes the
rank of values in the series.
“Median” of a set of data is a positional value. The Median is defined as the
middle-most value of the variable, when values are arranged either in
ascending or descending order of magnitude. According to Connor, “The
median is the value of the variable which divides the group or series into two
equal parts, one part comprising all values greater and the other part, value
lesser than the Median”. Since the median denotes the central position in a
series of values, it is called a position average.
7.5.1 Finding the Median for a Set of Data
(I) For Ungrouped Data
Let N values of a variable X be given in the form of raw data, that is, each
individual value is given separately in the order they are recorded. In order to
find the median of the series of values, we shall consider the following two
cases:
Case I: When there are odd number of values
Let N be an odd number, that is, we can represent N as N = 2k + 1, where k =
0, 1, 2, … . Then, let all the N values be arranged in the ascending order of
magnitude and, thus, we have the values as
x1 x2 x3 … xk xk+1 xk+2 xk+3 … x2k+1.
Clearly, in this series the middle-most value is xk+1, and there are equal
number of values on both the sides of it. According to the definition of
median, therefore, xk+1 is the median of the series. Thus, if there are 2k+1
values (an odd number of values) in the data, the median is the value which
occupies the (k +1)th place or, equivalently, [(N + 1)/2]th position, when
values are arranged either in ascending or descending order of magnitude.
Example 8: Let us consider the following set of raw data:
5, 12, 45, 8, 21, 32, 5, 36, 16, 8, 20, 24, 54, 18, 44, 56, 20, 30, 44
164
Solution: Here, N = 19 = 2x9 +1, means k = 9. Descriptive Statistics-I
Arranging these values in ascending order of magnitude, we get the series as:
5, 5, 8, 8, 12, 16, 18, 20, 20, 21, 24, 30, 32, 36, 44, 44, 45, 54, 56
Then, since the value 21 occupies the middle-most position (that is, 10th
position), it is the median of the set.
Case II: When there are even number of values
Let there be an even number of observations. Then, N can be written as N =
2k for k = 1, 2, 3, … . If these values are arranged in ascending order of
magnitude as
x1 x2 x3 … xk xk+1 xk+2 xk+3 … x2k.
In such situations, we observe that there is no single middle-most value in the
series, rather, we will get two middle-most values which would be xk and xk+1
occupying the (N/2)th and [(N/2)+1]th positions. Following the definition of
the median, therefore, we consider neither of the values as the median, but we
take the arithmetic mean of both the middle-most values as the median of the
series. In other words, in such situations the median of the series, denoted by
Md, will be given by
th th
N N
2 value 2 1 value
Md
2
Example 9: Calculate the median for the following data:
7, 8, 9, 3, 4, 10
Solution: First we arrange given data in ascending order as 3, 4, 7, 8, 9, 10
Here, (N) = 6 = 2x3 (even). So, here k = 3. Therefore, we get the median as
3rd observation 4th observation 7 8 15
Md 7.5 .
2 2 2
You can see that in such cases, median does not correspond to any of the
values of the series, rather, it is the mean of two middle-most consecutive
values of the series.
(II) For Discrete Frequency Distribution
In case of discrete frequency distribution, the procedure of determining
median is fundamentally the same as that for an individual series. The
procedure consists of the following steps:
(i) Let the number of distinct values of the variable be k. Arrange these
values in ascending or descending order of magnitude, each associated
with their corresponding frequency. Let fi (i = 1, 2, …, k) denotes the
frequency associated with the ith value. Further, we have f0 = 0 and
fk+1=0;
165
Statistical Analysis (ii) Obtain the “less than type” cumulative frequencies F1 = f1, F2 = f1+f2,
… , Fk = f1+f2+…+fk corresponding to each value, and list them in a
separate column. We have, F0 = 0 and Fk+1 = Fk. The last cumulative
frequency will coincide with the total number of values, that is, N;
Now, as before, we shall consider two cases: N is odd and N is even.
(iii) If N is odd, then, find the value of [(N +1)/2]. If the inequality Fi-1 (N
+1)/2 Fi is satisfied for the ith value (i = 1, 2, …, k), then the ith value
will be the median of the data.
(iv) If N is even, then we know that there will be two middle-most values,
the mean of which would be the median. Find the values of (N/2) and
[(N/2)+1]. The first middle-most value (ith value) will be obtained using
the inequality
Fi-1 Fi. Now, the second middle-most value (jth value) will be
obtained by using the inequality Fj-1 1 Fj, where j = 1, 2, …
, k.
Note: Since, for the case of N even, there would be two middle most values,
therefore, care must be taken in finding the second middle-most value (i+1)th
value after finding the first middle-most value (the ith value). It may happen
that both the jth and ith values are the same value, that is, j = i or j i 1.
Example 10: Find the median size of the following series:
Size (X) 4 5 6 7 8 9 10
Frequency (f) 06 12 15 28 20 14 05
Solution: Values are already arranged in ascending order of magnitude,
therefore, we compute less than cumulative frequencies as shown in the table:
Cumulative Frequency
Size X f
(Fi)
4 6 6
5 12 18
6 15 33
7 28 61
8 20 81
9 14 95
10 5 100
100
Here, N is even, so we compute (N/2) and [(N/2) +1], which are 50 and 51.
The first value (N/2) satisfies the condition Fi-1 Fi for F3 F4;
F3 and F4 are respectively, 33 and 61. Also, we see that the value [(N/2) + 1]
satisfies the condition for F3 F4. This implies that both (N/2) and
[(N/2) +1], correspond to the value 7. There mean will also be 7. Thus,
median is 7.
166
(III) For Continuous Frequency Distribution Descriptive Statistics-I
In the case of continuous frequency distribution, we first locate the median
class, that is, the particular class which contains the exact median which
would have been obtained if the data was given in the form of ungrouped
series of values. Obviously, after locating the median class, finally the
approximate value of median within the class is determined by using an
interpolation formula. The procedure of computation involves the following
steps:
(i) Compute less than type cumulative frequency for each of classes given.
Let there be k classes in all.
(ii) Find the value (N/2).
(iii)Now, find the class for which the condition Fi-1 N/2 F ; where Fi is
th
the cumulative frequency corresponding to i class (i =1, 2, …, k), is
satisfied. Then, the ith class is said to be “Median Class”.
(iv) Obtain the median value by applying the formula;
N
C
Median L 2 h
f
where, L = Lower class limit of the Median class
C = cumulative frequency of preceding class of the median
class
f = frequency of the median class
h = class interval (class width) of the median class.
Example 11: 100 salesmen employed by a company have booked the
following number of orders for a newly introduced FAX machine during the
last six months:
No. of Orders Booked 10-20 20-30 30-40 40-50 50-60 60-70 70-80
No. of Salesmen: 4 12 25 30 15 8 6
Calculate median of the above data and account for the difference, if any.
Solution: We compute the cumulative frequencies, which are:
No. of Orders No. of Cumulative
Booked Salesmen (f) Frequency (F)
10-20 4 4
20-30 12 16
30-40 25 41
40-50 30 71
50-60 15 86
60-70 8 94
70-80 6 100
f 100 167
Statistical Analysis Here, (N/2) = 50; so we see that the condition Fi-1 N/2 F is satisfied for
the values Fi-1=41 and Fi=71. Therefore, median class is 40 – 50.
Now, we use the formulae to calculate the median as:
N
C
50 41 9
Median L 2 h 40 10 40 10 43 .
f 30 30
CHECK YOUR PROGRESS 2
Note: i) Check your answers with those given at the end of the unit.
4) Find the median of the following series of values:
(i) 10, 6, 15, 2, 3, 12, 8
(ii) 10, 6, 2, 3, 8, 15, 12, 5
5) Find the median of the following frequency distribution
Marks 5 10 15 20 25 30 35 40 45 50
No. of 2 5 10 14 16 20 13 9 7 4
Students
6) Find Median for the following frequency distribution
Marks 0-10 10-20 20-30 30-40 40-50 50-60 60-70
No. of students 5 10 15 20 12 10 8
7.6 MODE
The Word ‘Mode’ comes from the French word ‘La Mode’ which means the
fashion. In statistical language, the Mode is that observation in a distribution
which occurs most often, i.e., the value which is most typical. According to
Croxton and Cowden, the mode of a distribution is the value around which
the items tend to be most heavily concentrated. Thus, mode or the modal
value is the value in a series of observations which occurs with highest
frequency. For example, when we say that the average size of shoes which
are sold in maximum number in a shop is 7, we talk about the modal size of
shoes.
In real life situations, generally we observe data with three types of mode. A
data set may be either (i) unimodal that is, having only one mode; (ii)
bimodal, that is, having two modes and (iii) multi-modal, that is, more than
two modes. To illustrate this, let the scores of 10 batsmen be 40, 45, 58, 46,
48, 48, 58, 58, 70 and 81. Then the only mode for the data set is 58 runs,
because it appears three times. So, it is a unimodal data set or unimodal
distribution. Again, let the runs scored by 10 batsmen of other team be 42,
47, 58, 79, 72, 81, 95, 50, 56 and 69. Then, there is no mode in the above
series of scores as every score has equal frequency.
168
Frequency 5 Descriptive Statistics-I
4
3
2
1
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13
X
Consider then a third series of scores, say: 86, 62, 58, 58, 48, 47, 48, 81, 59
and 50. Here, we observe that there are two modes 48 and 58, as they both
occur twice while other values occur only once. In such a case the
distribution is called bimodal. But still, in case of a bimodal distribution, the
two modes need not always have the same frequency as shown in the figure
given below:
7
6
Frequency
5
4
3
2
1
0
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
X
However, this is clear from the above discussion that in some cases mode
may be absent while in others there may be one or more modes. If the
frequency curve of the frequency distribution is drawn, the mode is the value
of the variable at which curve reaches its maximum, i.e., it is the value where
the concentration of observations is maximum. In case of a bimodal
distribution, the concentration of observation occurs at two points and thus,
we have two modes.
Here, it is important that the occurrence of more than one mode in a
distribution may be useful for further statistical analysis, but the mode as a
measure of central tendency has little significance in the case of bimodal or a
trimodal data.
Mode is an important ‘average’ in many situations specially in data related to
marketing studies. For example, when we talk about, a brand of shirt which is
most popular amongst younger generation, a song which is listened by the
largest section of societies, etc.
7.6.1 Calculation of Mode
(I) For Individual Series
For an individual series the mode can be located simply by inspection or
observing the values or observations. Here the values of variable under 169
Statistical Analysis consideration may first be arranged in the form of a discrete frequency
distribution and then, by looking the frequencies, we can find the value
possessing highest frequency. The value with highest frequency would be the
mode of the data.
Mathematically, if x1, x2,…, xn are the n observations and if some of the
observation are repeated in the data, say x i is repeated highest times then we
can say that x i would be the mode value.
Thus, in fact, it would be necessary to list the individual values of the
variable with their corresponding frequencies. Only then, it would be possible
to locate the modal value of the data. We present here an example to illustrate
how to find the mode of an individual series of values:
Example 12: Find the mode of the following observations given in an
individual series of scores:
9, 9, 10, 11, 11, 11, 12, 14, 14, 17
Solution: Here there are only ten observations in the series we need not put
the data in the form of a frequency distribution and mode can be determined
by just inspection only. In this case, the observation 11 has the maximum
frequency 3. Therefore, 11 is the mode.
(II) For Discrete Frequency distribution
In case of a discrete frequency distribution; mode can be located simply by
inspection. Here the variable having the maximum frequency will be taken as
mode.
Example 13: Determine mode for the following distribution:
X 10 12 14 16 18 20 22
f 4 6 10 11 21 10 5
Solution: In the above discrete frequency distribution the value 18 has the
maximum frequency 21. Therefore, 18 is the mode of the given discrete
distribution.
CHECK YOUR PROGRESS 3
Note: i) Check your answers with those given at the end of the unit.
7) Find the mode for the following items
4, 7, 6, 5, 4, 7, 8, 3, 7, 9, 2, 7, 6, 1, 2, 5, 10, 8, 4, 7, 1, 4, 2, 1, 3, 7, 9, 3,
7, 6.
8) Find mode value for the given data
2, 2, 3, 4, 9, 7, 9, 10, 12, 12, 10, 9, 12, 9, 7, 10, 9, 7, 9, 12, 7, 3, 2, 7, 9,
9, 7, 4, 12, 10, 9, 7, 12.
170
(III) For Continuous Frequency Distribution Descriptive Statistics-I
We know that in continuous frequency distribution, instead of listing
individual values of the variable, values are grouped into various class
intervals along with the number of values lying within the class. Therefore,
frequency of each and every value in such distributions are not known, hence,
the previously stated methods of finding mode is not applicable here. Due to
this difficulty, the exact modal value cannot be obtained in such distribution.
What then we can do is simply to find an approximate modal value for the
distribution using some technique or using some formula for finding
approximate mode. There are two approaches for this, (i) using the graphical
method, or (ii) using some approximate formula developed from an
appropriate diagram under certain assumptions. For this purpose, it is seen
that histogram is the appropriate figure which can be utilized. However, if
class intervals are made narrower and narrower, the histogram is can be used
to locate the modal value by joining mid values of all such classes. Since, the
graphical method is practically not a simple method to adopt, mode of
continuous frequency distributions is computed using a mathematical
formula, which is derived from a histogram under certain assumptions. We
shall not provide the proof of the formula here, rather, we shall illustrate its
application through examples.
The formula is as follows:
f1 f0
Mode L h
2f1 f0 f2
where L = lower limit of the modal class,
f1 = frequency of the modal class,
f0 = frequency of the pre-modal class,
f2 = frequency of the post-modal class, and
h = width of the modal class.
Modal class is that class which has the maximum frequency.
Remark 1: For determining mode, the distribution must have continuous
classes.
Example 14: Obtain mode of the following distribution:
Classes: 10-20 20-30 30-40 40-50 50-60 60-70
Frequency: 08 12 25 45 11 09
Solution: In this example, the highest frequency 45 lies in the class (40-50)
hence, this is the modal class. For determining the mode value in the modal
class, we use the formula as:
f1 f0
Mode L h
2f1 f0 f2
171
Statistical Analysis Here, f0 25, f1 45, f2 11, L 40 and h 10.
45 25
Thus Mode 40 10 40 3.70 43.70
2 45 25 11
CHECK YOUR PROGRESS 4
Note: i) Check your answers with those given at the end of the unit.
9) For the following data, calculate mode:
Class Interval Frequency
0-10 3
10-20 5
20-30 7
30-40 9
40-50 4
7.7 MEASURES OF DISPERSION
According to Spiegel, “the degree to which numerical data tend to spread
about an average value is called the variation or dispersion of data”. Since,
Statistics mostly deals with those characteristics which are variable in nature,
that is, varies over unit to unit; dispersion is always present in any set of data
and, therefore, in order to explore the nature of the data, the degree of
dispersion should be measured by some means. The degree of spreadness or
scatteredness of data about an average value is generally explored through
some measures of dispersion. Obviously, intuitively it can be thought that
such measures must assume value zero if there is no variation in the values of
the variable or in other words the characteristic is a constant instead of a
variable. The measures of dispersion may be of two types which are
mentioned below:
(i) Absolute Measures of Dispersion
Absolute measures of dispersion are those measures which depend upon the
unit(s) in which the values of the variable are expressed. This means that if
two variables are recorded in different units, then corresponding absolute
measures of dispersion are not helpful in comparing their degree of
scatteredness. However, if units of these two or more variables are same, they
can be compared with each other in respect of their dispersion. Thus, absolute
measures of dispersion are useful for comparing variation in two or more
distribution where units of measurement are the same. Such measures are not
suitable for comparing the variability of the distribution expressed in
different units of measurement.
172
(ii) Relative Measures of Dispersion Descriptive Statistics-I
Owing to the problem of comparison of two or more variables measured in
different units, as described above, we may think of developing some relative
measures of dispersion which are free from units of the variables. Such
measures are popularly known as “relative measures of dispersion”. Relative
measures of dispersion are expressed as ratio or percentage or the coefficient
of the absolute measures of dispersion. In this way, relative measures are
pure unit free less numbers. Relative measures are useful for comparing
variability between two or more distributions where units of measurement
may be different.
Following are some important measures of dispersion:
1. Range
2. Mean Deviation (M. D.)
3. Standard Deviation (S. D.) and Variance
4. Coefficient of Variation (C. V.)
Different measures of dispersion are discussed in following sections:
7.8 RANGE
Range is the simplest measure of dispersion. It is defined as the difference
between the largest value (L) and the smallest value (S) of the variable in the
distribution. Its merit lies in its simplicity. It can be defined as
Range(R) L S
where, L: Largest value of the variable, and
S: Smallest value of the variable, as give in the data set.
It should be mentioned here that whether the data are given in the form of a
series of the values or in the form of grouped data, in both the cases range is
well defined and, therefore, can be computed easily.
Example 15: Find the range of the distribution 6, 8, 2, 10, 15, 5, 1, 13.
Solution: For the given distribution, the largest value of variable is 15 and
the smallest value is 1.
Hence, Range (R) = Largest value (L) - Smallest value (S) =15-1= 14.
Example 16: Marks of 10 students in Mathematics and Statistics are given
below:
Marks in Mathematics 25 40 30 35 21 45 23 33 10 29
Marks in Statistics 30 39 23 42 20 40 25 30 18 19
Compare the range of marks in the two subjects.
Solution: The lowest and the highest marks in mathematics are 10 and 45
respectively. Also, the lowest and highest marks in statistics are 18 and 42
respectively.
173
Statistical Analysis Thus, Range (R) of marks in mathematics = L – S = 45 – 10 = 35
Range (R) of marks in Statistics = L – S = 42 – 18 = 24
CHECK YOUR PROGRESS 5
Note: i) Check your answers with those given at the end of the unit.
10) Calculate the range for the following data:
60, 65, 70, 12, 52, 40, 48
11) Calculate range for the following frequency distribution:
Class 0-10 10-20 20-30 30-40 40-50
Frequency 2 6 12 7 3
7.9 MEAN DEVIATION
As observed, the range is positional measure of dispersion as its computation
is not based on all the observations. More so, it is not measure of dispersion
in the strict sense of the term as they do not measure scatteredness in
observations around an average. As a first measure, mean deviation (M. D.)
is defined for meeting the objective. M. D. of a distribution is the arithmetic
mean of the absolute deviations of various values from some central value,
such as mean, median or mode. Since, the sum of deviations of values from
their mean is always zero (an algebraic property of mean); the mean
deviation would be always zero even if there is a greater variability in the
data. That is why, in defining M. D. the average of the absolute deviations
[ignoring plus (+) and (–) minus signs of actual deviations] from some central
value are taken. In general, the central value can be any measure of central
tendency or even some arbitrarily chosen constant, but usually; being the
mostly used measure of central tendency, mean is taken as the central value
for taking deviations.
The deviation of an observation xi from the constant A is defined as (xi – A),
which might be either positive, negative or zero. In order to define M. D.,
therefore, we ignore the sign of the deviations and consider the absolute
values of the deviations which are denoted by |xi – A|.
Thus, the mean deviation (M. D.) defined as
N
1
M. D.
N
x
i1
i A
N
where, x
i 1
i A Sum of absolute deviations of values taken from the
constant A.
As mentioned above, it can be shown that if we take A = median of the data,
the quantity xi – Ais least as compared to any other choice of A. But in
174 practice, the mean deviation is computed by taking the constant A as the
mean of the data, since, mean is the most frequently used measure of central Descriptive Statistics-I
tendency. If A is the mean of the data, the M. D. would be given by
N
x X
i 1
i
M. D.= ,
N
N
where x
i 1
i X Sum of absolute deviations taken from mean.
However, if A is the median, Md; M. D. would be given by
N
x M
i 1
i d
M. D. =
N
N
Where x M
i1
i d = Sum of absolute deviation taken from the median.
In case the data are given in the form of frequency distribution, the formula
for M. D. can be put as:
k k
f
i 1
i xi X f
i 1
i xi Md
M. D. = k
and M. D. = k
f
i 1
i f
i1
i
Without any confusion, in such situations, Xi represents the mid-value of the
ith class interval and fi, the frequency of this class, where there are k such
class intervals.
Example 17: Find mean deviation for the given data
1, 2, 3, 4, 5, 6, 7
Solution: First of all we find Mean
1 2 3 4 5 6 7 28
X 4
7 7
So, we have |xi - X| as 3, 2, 1, 0, 1, 2, 3 and ∑ |x X|= 12.
12
Therefore, MD 1.71
7
Example 18: Find mean deviation for the following data:
X 1 2 3 4 5 6 7
F 3 5 8 12 10 7 5
Solution: We do the necessary calculations in the following table:
X f fX x X f x X
1 3 3 3.24 9.72
2 5 10 2.24 11.20
175
Statistical Analysis 3 8 24 1.24 9.92
4 12 48 0.24 2.88
5 10 50 0.76 7.60
6 7 42 1.76 12.32
7 5 35 2.76 13.80
Total 50 212 12.24 67.44
From the values obtained in the columns of the table, we have
k
f x
i 1
i i
212
X k
4.24
50
f
i 1
i
f
i1
i xi X
67.44
and M.D k
1.348
50
f
i1
i
Example 19: Compute the mean deviation (M.D.) from the following data.
Classes: 0–20 20–40 40–60 60–80 80–100 100–120
Frequency: 5 50 84 32 10 6
Solution: We use the technique of change of origin and scale for calculating
the mean of the data. We choose the assumed mean as 50 and the constant H
for changing the scale, as 20. The transformed values, denoted by d’ are then
computed and presented in the fourth column of the table. We do other
necessary calculations in the following table:
x 50
Classes X f d fd xX f xX
20
0–20
10 5 –2 –10 41.07 205.35
20–40
30 50 –1 –50 21.07 1053.50
40–60
50 84 0 0 1.07 89.88
60–80
70 32 1 32 18.93 605.76
80–100
90 10 2 20 38.93 389.30
100–
110 6 3 18 58.93 353.58
120
N
= fd 10 f x X 2697.37
187
176
X A h
f d' 50 20 10
50 1.07 51.07
Descriptive Statistics-I
N 187
1 2697.37
Mean Deviation
N
f x X 187
14.42
CHECK YOUR PROGRESS 6
Note: i) Check your answers with those given at the end of the unit.
12) Following are the marks of 7 students in statistics. Find the mean
deviation.
16, 24, 13, 18, 15, 10, 23
13) Find the mean deviation for the following distribution:
Class 0-10 10-20 20-30 30-40 40-50
Frequency 5 8 15 16 6
7.10 STANDARD DEVIATION AND VARIANCE
In the previous section, we have seen that while calculating the mean
deviation, negative deviations are straightaway made positive by finding the
absolute value of all the negative deviations. But, since the absolute values
are generally difficult to operate and are not amenable to algebraic treatments
easily, another way may be thought of converting negative values to positive
values. One method may be thought to convert all the deviations into positive
quantities by considering the squared values of deviations, instead of taking
their absolute values. In this section, we shall use this method for defining
some other measures of dispersion.
7.10.1 Root Mean Square Deviation (RMSD)
Before describing other measures of dispersion, such as, Standard Deviation
(S.D.), Variance and Coefficient of Variation (C.V.), we shall describe first a
general form of measures of dispersion, which is called “Root Mean Square
Deviation (RMSD)”. The RMSD is defined as the “square root of the mean
of squared deviations of all the variate-values taken from an arbitrarily
chosen constant, say A”.
Thus, if xi is the value of the variable as recorded on the ith unit (i = 1, 2, …,
N), with corresponding mean X, then RMSD is defined as
N
1 2
RMSD =
N
(x A)
i 1
i
Where, A may be any number. Since, deviations are taken from the constant
A, we call this formula as RMSD with respect to the value A.
177
Statistical Analysis 7.10.2 Standard Deviation (S.D.)
Standard deviation (S. D.), popularly denoted by the letter σ, is a special case
of RMSD. It is defined as
N
1 2
S.D.( )
N
(x X)
i 1
i
Where, X is the mean of the data set. It is defined as the positive square root
of the mean of the squared deviations taken from the respective mean of the
data. You can observe that S.D. is same as RMSD when A = mean of the data
(a constant).
The question is that why the constant A has been taken as mean of the
distribution and not any other measure of central tendency. The reason is that
RMSD is the least when A = X . Let us show this fact mathematically as
follows:
We have
N N
1 1 2
RMSD =
N
i 1
(xi A)2
N
(x X ) X A
i 1
i
N N N
1 1 1
N
(xi X)2
N
(X A)2 2
N
(x i
X) X A
i1 i1 i 1
N
1 2
Now, if A = X, it reduces to
N
(x X)
i 1
i , which is defined as S.D. So, we
see that RMSD attains its least value when constant A is taken as the mean of
the data, and then RMSD is called Standard Deviation (σ).
The concept of standard deviation was first introduced by Karl Pearson. It
satisfies most of the properties of a good measure of dispersion. However, it
can be seen that S.D. is not unit free. Its unit is same as that of x-values.
Alternative formula for computation of Standard Deviation:
In order to simplify the calculation process and also to reduce the error of
rounding to the minimum extent, we can use an alternative formula for the
calculation purpose of Standard Deviation, which is obtained below:
We have
N N N N
1 1 2 2
Standard Deviation (σ) (xi X)2 x X i x .X
2 i
N i1 N i 1 i 1 i 1
Thus, we have an alternative form of the formula of S.D. as
N N
1 1
Standard Deviation (σ) xi2 NX2 x 2
i X2
N i1 N i1
If the data are given in the grouped data form, either in the discrete or
continuous frequency distribution, the formula for S. D. becomes:
178
1 k Descriptive Statistics-I
S tan dard Deviation( ) fi (x i X)2 i 1,2,...,k
N i 1
k
where, there are k class intervals; N fi and fi is the frequency of the ith
i1
class interval.
Proceeding in the same manner as we did above to find the alternative
formula of standard deviation; we can derive alternative formula for standard
deviation when data are available in the grouped form, which is given as
k k
1 2
S tan dard Deviation ( ) fx i i X2 where, N fi
N i 1 i1
7.10.3 Variance
Variance is another frequently used measure of dispersion. Actually, variance
is nothing but the square of the standard deviation and hence, is denoted by
σ2 .
Variance is the average of the square of deviations of the values taken from
the respective mean. Thus, variance is defined as
N
1
Variance (σ2 ) (x X) i
2
N i1
and for a frequency distribution, the formula is
k
1
Variance (σ2 ) = f (x X)
i i
2
N i 1
where, all symbols have their usual meanings as given for the formula of
standard deviation above. The unit of the variance will be the square of the
variable X.
Alternative formula for computation of Variance:
Since, variance is the square of standard deviation; the same problem arises
for variance also. Therefore, we have to use an alternative formula for
variance also which must be used for computation purpose.
Squaring the alternative formula for standard deviation, we have alternative
formula for calculation of the variance, which is given by
N
1
2 x 2
i X2
N i 1
Similarly, the alternative formula for computing σ when the data are given
in the grouped form, will be
k
2 1 k fx i i k
fi x i2 X 2
where, X i 1
k
and N fi
N i1 i1
f
i 1
i
k being the number of class intervals. 179
Statistical Analysis Example 20: Calculate S.D. for the following set of observations:
X: 10 11 17 25 7 13 21 10 12 14
Solution: There are 10 values of the variable, therefore, n = 10.
Here, we have ∑ x = 102+112+ ….+142 = 2234. Also, ∑ x = 140.
N
x
i 1
i
140
Therefore, X 14 ;
N 10
2234 2
and hence, S.D.( ) 14 223.4 196 27.4 5.2345
10
Example 21: Calculate the S.D. of the following discrete frequency
distribution:
Size (X) 4 5 6 7 8 9 10
Frequency 6 12 15 28 20 14 5
Solution: For rest of the computation, we prepare the following table:
Size (X) 4 5 6 7 8 9 10 Total
Frequency 6 12 15 28 20 14 5 100
(f)
fX 24 60 90 196 160 126 50 706
fX2 96 300 540 1372 1280 1134 500 5222
Therefore, standard deviation is
5222 2
S.D.( ) 7.06 52.22 49.8436 2.3764 1.542
100
Example 22: Calculate standard deviation from the following data:
10– 12– 14– 16–
Age (in years) 4–6 6–8 8–10
12 14 16 18
No. of Students 30 90 120 150 80 60 20
Solution: Let us use the short-cut method for the calculation. We choose A =
11 and H = 2. Then, the calculations are shown in the table given below:
Frequency Mid – X 11
Age d fd fd2
f Value 2
4–6 30 5 –3 –90 270
6–8 90 7 –2 –180 360
8–10 120 9 –1 –120 120
10–12 150 11 0 0 0
12–14 80 13 1 80 80
180
14–16 60 15 2 120 240 Descriptive Statistics-I
16–18 20 17 3 60 180
2
N = 550 fd 130 fd 1250
We know that (σ hσ . So, the S. D. of X (the original variable) will be
2
2
X h
f d
f d
N N
2
1250 130
2 2 2.272 0.055 2.977
550 550
CHECK YOUR PROGRESS 7
Note: i) Check your answers with those given at the end of the unit.
14) Find the standard deviation for the following data.
10 27 40 60 33 30 10
15) Calculate standard deviation for the following data:
Class 0-10 10-20 20-30 30-40 40-50
Frequency 5 8 15 16 6
7.11 COEFFICIENT OF VARIATION
It is now clear that the standard deviation or the variance as measures of
dispersion, give us an idea about the extent to which observations are
scattered around their mean. Therefore, two or more distributions can be
compared directly for their variability of such characteristics of two or more
distributions, which are measured in same unit with the help of corresponding
standard deviations and a distribution having the largest (smallest) value of S.
D. or variance as compared to the same of the other distributions, may be said
to be most dispersed (least dispersed) distribution than other distributions.
However, mostly, when we wish to compare the variability of such
characteristics of two or more distributions, which are measured in different
units or even to compare two or more characteristics in the same distribution
such that the characteristics are measured in different units; a problem arise
due to the non-conversion of units. We know that any unit of height
measurement is not convertible to any unit of weight measurement. So,
variability of heights is not directly comparable on the basis of S. D. or
variance. Similar may be the case with other characteristics also. Then, the
question is how to compare variability’s of such characteristics which are
measured in different units.
As before, we define a relative measure of variability with the help of S. D.
for this purpose. For making comparisons in the above two situations, we use
181
Statistical Analysis a relative measure of dispersion, called coefficient of variation (C.V.) which
is defined as
CV = 100
x
It is easy to see that the unit of the numerator is same as that of the
denominator and, hence, it is a unit free number and, therefore, it is called to
be a coefficient. In this sense it is a relative measure of variability. If we are
comparing the two data series, the data series having smaller CV will be
more consistent. The distribution having greater C.V. is considered to be
more dispersed than the other distribution with lesser C.V.
Example 23: Suppose batsman A has mean 50 with SD 10 of runs he scored
and Batsman B has mean 30 with SD 3 for the runs he scored. What do you
infer about their performance?
Solution: Obviously, in cricket, a batsman is said to be more consistent or
reliable in his performance if the variability of runs he scored from match to
match is less.
A batsman is not said to be a good performer if the variability of scores is
larger. However, since A has higher mean of runs than B, this means A is a
better run maker.
However, B has lower CV (3/30 = 0.1) than A (10/50 = 0.2) it means that B
is more consistent.
Example 24: Two workers on the same job show the following results over a
long period of time:
Worker A Worker B
Mean time of completing the job
25 20
(minutes)
36 16
Variance
i) Which worker appears to be more consistent in the time he requires to
complete the job?
ii) Which worker appears to be faster in completing the job?
Solution: (i) Given that X1 25, X2 20, 12 36, 22 16.
1 6
For worker A, Coefficient of Variation 100 100 24%
X1 25
2 4
For worker B, Coefficient of Variation 100 100 20%
X2 20
i) The comparison of C.V.s led to the conclusion that worker B appears
more consistent in the time needed for completing the job.
ii) Since the mean time taken to complete the job is less for worker B, so
182 worker B appears to be faster in completing the job.
CHECK YOUR PROGRESS 8 Descriptive Statistics-I
Note: i) Check your answers with those given at the end of the unit.
16) The following table shows the results of an analysis relating to the
prices of two shares X and Y, that were recorded during the year
2000:
Share Arithmetic mean (Rs.) Standard Deviation (Rs.)
X 45.00 9.00
Y 62.50 18.75
Which of the two shares, X or Y, have more variability in prices?
2
17) If n=10, x 24, x 200 , find the coefficient of variation.
7.12 LET US SUM UP
In this unit, we have discussed:
1. Some basic measures of central tendency and methods of calculation of
the Arithmetic mean, Median and Mode for different kinds of data.
2. properties of a good measure of measures of central tendency and
dispersion which are supposed to be satisfied.
3. different types of measures of dispersion; like, Range, Root Mean
Square Deviation, Standard deviation, Variance and Coefficient of
Variation.
4. methods of calculation of Range, Standard deviation, Variance and
Coefficient of Variation for different kinds of data.
7.13 KEY WORDS
Measures of : A general term for the various averages that attempt to
Central Tendency describe the middle or typical value in a distribution.
Measures of : A general term for various measures of the amount by
Variability which scores are dispersed or scattered.
7.14 SUGGESTED FURTHER READING/
REFERENCES
Witte, R., & Witte, J. (2017). Statistics. Hoboken, NJ: John Wiley & Sons.
183
Statistical Analysis
7.15 ANSWERS TO CHECK YOUR PROGRESS
1) For calculating the arithmetic mean, we add all the observations and
divide by 6 as follows:
X
x i
5 8 12 15 20 30
15
n 6
2) For calculation of mean, the following table is made:
Wages No. of Workers (f) fX
20 4 80
25 8 200
30 20 600
35 125 420
40 6
i 1
fi xi
1540 240
X 5
30.8
50
f f 50
i 1
i fx i i 1540
3) Using short cut method
X 25
Marks f X d fd
10
0-10 6 5 2 12
10-20 9 15 1 9
25=
20-30 17 0 0
A
30-40 10 35 +1 +10
40-50 8 45 +2 +16
f 50 i
fd 5
i i
Now X A
fidi 5
h X 25 10 = 26
N 50
4) (1) Here, N = 7, an odd number. After arranging the values in
ascending order, we get 2, 3, 6, 8, 10, 12, 15 and, therefore, the
Median value will be
Median = value of [(N+1)/2]th item= value of 4th item = 8.
(2) Here N= 8, an even number, so arranging values in ascending
order we get the values as 2, 3, 5, 6, 8, 10, 12, 15 and, therefore,
Median = Mean of [N/2]th and [(N/2) +1]th values = (6+8)/2 = 7.
184
5) For the given data, we compute less than cumulative frequencies as Descriptive Statistics-I
follows:
Marks 5 10 15 20 25 30 35 40 45 50
No. of Students 2 5 10 14 16 20 13 9 7 4
Cumulative 2 7 17 31 47 67 80 89 96
Frequency 100
Since, here N is even, so there will be two middle-most values, the
first is (N/2)th value and the second is [(N/2) + 1]th value.
Here, N/2 = 50 and [(N/2) + 1] = 51. Thus, median is equal to the
mean of the 50th and 51st values. We see that (N/2)th value satisfies the
condition F5< (N/2)< F6.
So, (N/2)th value will be equal to 30. Also, [(N/2)+1] value is 30.
Therefore, their mean is also 30. So, median is 30.
6) First we shall calculate the cumulative frequency distribution
Cumulative
Marks f
Frequency
0-10 5 5
10-20 10 15
20-30 15 30=C
30-40 20=f 50
40-50 12 62
50-60 10 72
60-70 8 80
N= 80
Here N/2 = 40. Since, 40 is not in the cumulative frequency so, the
class corresponding to the next cumulative frequency 50 is median
class. Thus 30-40 is median class.
N
2 C
Median L h = 30 40 30 10 = 35
f 20
7) For calculation of mode, the following table is made:
Size 1 2 3 4 5 6 7 8 9 10
Frequency 3 3 3 4 2 3 7 2 2 1
From the frequency distribution we see that size 7 occurs with
maximum frequency of 7 hence mode is 7.
185
Statistical Analysis 8) First we prepare frequency table as
X 2 3 4 7 9 10 12
f 3 2 2 7 9 4 6
This table shows that 9 have the maximum frequency. Thus, mode is 9.
9) From the given frequency distribution corresponding to highest
frequency 9 modal class is 30-40 and, therefore, we have L=30, f1 =9,
fo 7 , f2 4 , h 10 Applying the formula,
97
Mode 30 10 30 2.86 32.86
29 7 4
10) Range R XMax XMin = 70-12 = 58
11) First convert the given inclusive class limits continuous classes
(Exclusive classes) as shown in the table:
Inclusive Continuous
Frequency
Classes Classes
6–10 5.5–10.5 7
11–15 10.5–15.5 8
16–20 15.5–20.5 15
21–25 20.5–25.5 35
26–30 25.5–30.5 18
31–35 30.5–35.5 7
36–40 35.5–40.5 5
Here, L = 40.5 and S = 5.5
Then we have R = 40.5 – 5.5 = 35
12) For calculation of mean deviation, the following table is made:
x (x-17) xX
16 -1 1
24 +7 7
13 -4 4
18 +1 1
15 -2 2
10 -7 7
23 +6 6
x 119 x i X 28
186
N N Descriptive Statistics-I
x
i1
i
119
x X
i1
i
28
X 17 And M.D 4
N 7 N 7
13) For calculation of mean deviation, the following table is made:
Class x f xf x X xX f xX
0-10 5 5 25 -22 22 110
10-20 15 8 120 -12 12 96
20-30 25 15 375 -2 2 30f
30-40 35 16 560 8 8 128
40-50 45 6 270 18 18 108
50 1350 f x X 472
k k
i1
f i xi
1350
f
i1
i xi X
472
X k
27 and M.D k
9.44
50 50
f
i 1
i f
i1
i
14) For calculation of standard deviation, the following table is made:
X X2
10 100
27 729
40 1600
60 3600
33 1089
30 900
10 100
2
x 210 x 8118
X
x 210 30
N 7
N
1 2 8118
N
x X
i 1
2
7
900 1159.7 34.61
15) Let us take A=25
Class x f d = x-A f d f d2
0-10 5 5 -20 -100 2000
10- 15 8 -10 -80 800
187
Statistical Analysis 20 25 15 0 0 0
20- 35 16 10 160 1600
30 45 6 20 120 2400
30-
40
40-
50
2
N=50 fd 100 fd 6800
2
k
1 k fi di
1 100
2
fidi
2
i 1 = 6800
N i1 N 50 50
136 4 132 11.49
16) We are given that X 45, Y 62.50, X 9, Y 18.75
X 9
C.V. for share X 100 100 20%
X 45
Y 18.75
C.V. for share Y 100 100 30%
Y 62.50
On comparing the coefficients of variation for two shares, we find
that share Y shows greater variation in prices.
17) We have
N
1 1 2
2 x 2
i X2 200 4 20 16 4
N i 1 10
2 4 2
Therefore, coefficient of variation
2
C.V. X 100 100 50%
X 4
188
Descriptive Statistics-II
UNIT 8 DESCRIPTIVE STATISTICS-II
Structure
8.1 Introduction
8.2 Objectives
8.3 Correlation Analysis
8.3.1 Types of Correlation
8.3.2 Measures of Correlation
8.4 Scatter Diagram
8.5 Karl Pearson’s Correlation Coefficient
8.5.1 Properties of Correlation Coefficient
8.5.2 Calculation of Correlation Coefficient
8.6 Spearman’s Rank Correlation Coefficient
8.6.1 Method of Calculation of Rank Correlation
8.7 Concept of Regression
8.8 Lines of Regression
8.8.1 Single Variable linear Regression Lines
8.8.2 Calculation of Regression Lines
8.9 Regression Coefficients
8.9.1 Properties of Regression Coefficients
8.10 Let Us Sum Up
8.11 Key Words
8.12 Suggested Further Reading/References
8.13 Answers to Check Your Progress
8.1 INTRODUCTION
In the previous unit, you have studied various measures, such as, measures of
central tendency and measures of dispersion which are virtually used for
analyzing the nature of a single variable, say X, and its distribution in many
aspects. These measures, in fact, disclosemany such aspects of the variable,
which are considered to be basic properties of the variable for its further
study. However, when we come across with simultaneous study of two or
more variables, besides the separate study of these variables, one thing is
more important and necessary to note. This is, whether the variables exhibit
some kind of relationship between them or whether they do not exert any
kind of impact on others, that is, variables are supposed to be independent to
each other. For example, take the case of price, demand and supply of 50
commodities in a market. We know that, according to laws of economics,
variables ‘price’ and ‘demand’ exhibit one kind of direct relationship, since, 189
Statistical Analysis increasing demand results to increase in price also. On the other hand,
‘supply’ and ‘price’ have an indirect relationship, since, more the supply, less
would the price of commodities. In statisticalterminology, if there exist any
type of relationship between two variables, we say that they are mutually
‘correlated’ to each other. Thus, the word ‘correlation’, in fact, describes a
specific property of two variables.
If we state that the relationship between the two variables under consideration
is linear, it means that the this relation may be represented in the
mathematical form Y = b0 + b1X; which is the equation of a straight line,
where Y and X are respectively the dependent and independent variables and
b0 and b1 are some suitable constants (we shall use the word ‘parameter’ for
such constants in further text) so chosen that the straight line passes through
maximum number of points (xi, yi) for i = 1, 2, …, N; on the scatter diagram.
This is same as to say that if our purpose is to estimate or to find the value of
the dependent variable Y for a given value of the independent variable X
through the straight line obtained, then the actual value Y and the
corresponding estimated value of it must be as close to each other as possible,
that is, the difference of actual and estimated values must be minimum for
each Y-values. Such a line used for this purpose is generally called a
“Regression Line” and the theory behind regression lines and their
applications to practical problems is termed “Linear Regression Analysis”.
Section 8.3 of this unit discusses the conceptsand different type
ofcorrelationssupported with a number of practical examples that describe the
situationsunder which they might be observed.Section 8.4 describes the
‘scatter diagram’ which helps us diagrammatically to explore the nature of
the relationship between the variables. Section 8.5 describes a measure of
correlation, popularly known as “correlation Coefficient”, which provides the
type of correlation and the magnitude of the correlation between two
variables. Some of the salient properties of it are mentioned and theoretically
proved. Section 8.6 discusses rank correlation coefficient, which provides the
magnitude of the correlation between two attributes. Section 8.7 here
explains the basic meaning of the word ‘regression’. Section 8.8 describes the
mathematical procedure of obtaining all the linear regression lines in case of
two variables. Some of the properties of linear regression lines with one
independent variable are also discussed. Section 8.9 provides the meaning of
regression coefficients in regression analysis and their significance. Some
important properties of regression coefficients are also discussed.
8.2 OBJECTIVES
After studying this unit, you should be able to:
explain the concept of correlation between two variables and identify
different types of correlations;
describe the scatter diagram and its use in correlation analysis;
190
explain the computation method and describe the important properties Descriptive Statistics-II
of correlation coefficient;
explain regression, regression analysis and of linear regression;
derive expressions of lines of regression of Y on X and X on Y
mathematically on the basis of least squares principle;
explain how to apply both the regression lines in practical problems
for predicting the value of one variable given the value of the other;
and
explain the salient properties of regression coefficients.
8.3 CORRELATION ANALYSIS
In previous units we were confined to the study of nature and properties of
only one variable. What kind of data in such case are given to us were values
of the variable as measured over a number of objects, that is, data were
univariate data. However, in many practical situations, we may have data in
which more than one variable can be associated with each unit of
observation. For example, for providing the information about the marks
obtained by the students of a class in two subjects, say,Statistics and
Mathematics; the values associated with each student are the marks obtained
by them in Mathematicsas well as in Statistics. A practical example of such a
data is depicted in the following table:
Number of Student 1 2 3 4 5 6 7 8
Marks in Statistics (X) 90 88 52 48 67 85 56 65
Marks in Mathematics
95 70 64 75 38 69 86 54
(Y)
Here, each student assumes values on two variables X (marks in Statistics)
and variable Y (marks in Mathematics) simultaneously and, as such, we get a
‘bivariate distribution’. You may also recall here that in a bivariate
distribution, we have two variables of observation on which values are
recorded for each unit of observation. Similarly, the distributions involving
more than two variables are termed as “multivariate distributions”. In this
unit we confine ourselves to the study of bivariate distributions only in which
two variables may be inter-dependent.
Let us consider the example of sales revenue and expenditure on advertising
in business. A natural question arises in mind that is there any connection
between sales revenue and expenditure on advertising? Does sales revenue
increase or decrease as expenditure on advertising increases or
decreases?Similarly, if we see the example of time spent on study and marks
obtained by students, a natural question appears whether marks increase or
decrease as time spent on study increases or decreases.
191
Statistical Analysis In all these situations and some other situations, we try to find out a type of
relation, if any, between two variables and its magnitude. Then, correlation
analysis comes up in the way for answering the question, if there is any
relationship between one variable and the other.
For instance, generally, the weight of a person is seen to increase with height
up to a certain age; prices of commodities seem to vary with supply; pressure
exerted on a flexible body generally reduces its volume and, thus, are seen to
have been inversely related; cost of industrial production varies with the cost
of raw material and so on. In all these cases, the change in the value of one
variable appears to be accompanied by a change in the values of other
variable. If it happens, variables are said to be correlated, and then this
relationship is called correlation or covariation. The study and measurement
of the extent or degree of relationship between the two or more than two
variables along with the nature of relationshipbetween them is called
correlation analysis.
8.3.1 Types of Correlation
Depending upon the nature of different types of characteristics (variables)
seen in practice, correlation can be categorized into following three types of
cases:
i) Positive, Zero and Negative Correlation
ii) Simple, Multiple and Partial Correlation
iii) Linear and Non-Linear Correlation.
Let us discuss these types of correlation one at a time as follows:
i) Positive, Zero and Negative Correlation
When we compute the magnitude of relationship between two variables using
some measure, we come across with two types of values; one, positive value
of correlation and, two, negative value of the correlation. We should to
understand why this happens.
Since, as mentioned above, the coefficient of correlation measures the
magnitude of the linear relationship, the magnitude may be either a positive
or a negative quantity. Accordingly, the correlation may be positive or
negative according to the direction of change in the two variables. Positive
correlation refers to the change in the values of both the variables in the same
direction, that is, the values of both the variables increase or decrease in the
same direction. One such relationship is seen between variables ‘income’ and
‘expenditure’. Generally, with increasing income, expenditure of families
also seen to increase but with decreasing income, expenditure is also seen to
decrease.
Similarly, Negative correlation refers to the change of the values of the
variables in opposite directions, that is, if the values of one variable increase
the values of the other variable decrease (on an average) and if the values of
192
one variable decrease, then the values of other variable increase. For Descriptive Statistics-II
example, the price of a product and demand in the market, volume and
pressure of perfect gas, etc.
Sometimes, when we find correlation between two variables, we get it very
close to zero or exactly zero. It is an indication that the two variables are not
affecting each other at all or it is because of the reason that they are not
linearly related. This situation is termed as the case of zero correlation.
ii) Simple, Multiple and Partial Correlation
Based on the number of variables, correlation may be classified in the
following three types of classes:
a) Simple Correlation:Simple correlation is the study of correlation
between only two variables, that is, if we measure the degree or intensity
of the relationship between two variables only, it is called simple
correlation. For example, when we study the correlation between the
price and demand of a product, it is a problem of simple correlation.
b) Multiple Correlation:Multiple correlation is defined and computed
when we consider more than two variables at a time. Let there be (n + 1)
variables Y and X1, X2, X3, …, Xn and we use a relationship between Y
and all other n variables as
Y = a 0 a 1 X1 a 2 X 2 a 3 X 3 a n X n ;
where ai for i = 0, 1, 2,…, n are (n + 1) unknown constants.Since, in this
relationship Y seems to be dependent on n independent variables Xi for i =
1, 2, …, n;obviously, Y- variable can be estimated on the basis of these n
variables, which may be denoted by Ŷ .Then, if we find the simple
correlation between the two variables; namely, variable Y and the
estimated variable Ŷ ; it is termed as “Multiple correlation”. In this sense,
multiple correlation is also a simple correlation. For example, if we study
the relationship between the observed yield of rice(Y),observed on n fields
and the estimated yield ( Ŷ), when the amount of rainfall (X1), fertility of
soil of fields (X2), amount of fertilizers used(X3), etc., were also recorded
and taken into consideration in the relationship, for estimating the yield,
then it becomes the case of multiple correlation. We shall discuss it
separately afterwards.
c) Partial Correlation:In b) above, we observed that sometimes we may
consider more than two variables for finding a particular type of
correlation, which is multiple correlation. Now, think the situation in
another way for the relationship given between Y and n X-variables there.
Since, yield of a crop is always thought to be affected by the variables-
amount of rainfall (X1), soil fertility(X2), amount of fertilizer used (X3)
and other variables; we can say that the observed value of the yield (Y) is
the effect of a number of multiplicity of causes. Therefore, if anyhow we
can find that value of the yield (say, Y) , which could be obtained by
eliminating the effect of these variables on the yield and find the 193
Statistical Analysis correlation between Y and Y, it would have produced another kind of
correlation. This type of correlation is termed as Partial correlation. The
study of it is again a separate subject matter, which we shall consider
afterwards.
iii) Linear and Non-Linear Correlation
a) Linear correlation: In the section 10.2 above, we have mentioned the
meaning of linear and non-linear relationships between two variables. A
relation between the variables Y and X is said to be ‘linear’ if points of
values (yi, xi); i = 1, 2, 3, …, N, when plotted on a graph paper, exhibit a
straight line. For example, the following set of Y and X values represent a
linear relationship:
X: 10 20 30 40 50
Y: 45 85 125 165 205
We see that the relationship between Y and X may be expressed as Y =
4X +5, which is a straight-line representation of the above values. Now, if
the values of the variables Y and X exhibit an exact linear relationship or
are very near to a linear relationship; the correlation between the variables
is called to be ‘linear correlation’. We have already pointed out that this
unit is devoted to the study of correlation analysis when we have a linear
relationship. In other words, when the amount of change in one variable
tends to bear a constant ratio to the amount of change in the other variable,
the variables are said to have linear correlation.
b) Non-Linear Correlation: When the amount of change in one variable
does not bear a constant ratio to the amount of change in the other
variable, it is called a non-linear relationship, since the graph of such a
relation is not a straight line. Some examples of such relations are, Y = a +
bX + cX2 + dX3; Y = abX and Y = aebX; etc.
Whenever, the values of the variables follow a non-linear relationship, the
correlation between the two variables is a non-linear correlation. For
example, if the data are given as
X: 0 2 3 6 5
Y: 5 45 135 3645 1215
the relationship cannot be expressed by a straight line; rather it is
represented by the relation Y = 5(3X), which is non-linear. The curve
drawn along with these values it will show a curvilinear graph.
There are available some methods in the literature which measure the
correlation for curvilinear relationship. However, these will not be discussed
in this unit.
8.3.2 Measures of Correlation
In order to know have an idea of the type of relationship between two
variables and to find the extent of correlation between them, there are some
194 methods, which are listed below:
i) Scatter Diagram Method Descriptive Statistics-II
ii) Karl Pearson’s Coefficient of Correlation
iii) Spearman’s Coefficient of Rank Correlation
We shall be discussing these methods one by one in the following sections.
CHECK YOUR PROGRESS 1
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
1. What do you mean by Correlation?
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
2. Mention few examples of different types of correlations.
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
8.4 SCATTER DIAGRAM
Scatter diagram is a statistical tool for determining the potentiality of
correlation between dependent and independent variables. However, Scatter
diagram does not tellus about the exact relationship between them, neither it
is helpful in providing the degree of relationship of correlation. It is simply
helpful to observe whether they are correlated or not.
Scatter Diagram is graphical device for drawing certain conclusions about the
correlation between two variables. In preparing a scatter diagram, the
observed pairs of observations are plotted on a graph paper in a two-
dimensional space by taking the measurements on variable X along the
horizontal axis and that on the variable Y along the vertical axis. The pairs of
values are represented by dots on the graph. The diagram of dots so obtained
is known as “scatterdiagram”. The pattern and the direction of the
scatteredness of points provides some rough idea about the nature and degree
of the relationship.
Let x i , yi ; i 1, 2,..., N be the bivariate distribution under consideration. By
plotting the N pairs of observations x1 , y1 , x 2 , y 2 ,..., x N , y N on a graph
paper, we have a scatter diagram from which we can draw certain rough
conclusions concerning the extent to which the points cluster about a straight
line as follows: 195
Statistical Analysis 1. If all the plotted points, representing a given data, lie on a straight line,
or seems to be very closer to certain straight line, then it indicates that
the relationship is linear and not curvilinear.
2. If dots are in the shape of a line and line rises from left bottom to the
right top it indicates that the variablesdo have a linear relationship and
both the variables are having a positive correlation.
3. If all the points lie on a straight line or seems to be very close to certain
straight line such that it has a negative slope, then it indicates again a
linear relationship but a negative correlation, that is, an inverse
relationship between the variables.
4. In case of linear relationship, described in 2 and 3, if all the points
exactly lie on a certain straight line, the degree of correlation is said to
be ‘perfect’. If it is point2, the variables are said to have ‘positive perfect
correlation’ and if it is point 3, the variables are said to have ‘negative
perfect correlation’.
5. If plotted points on scatter diagram do not show any trend,it is an
indication thatthere is almost no correlation or a weak correlation
between the variables.
A discussion on the usefulness of scatter diagram in correlation analysis was
made in this section, which makes it clear that such a diagram is useful only
to have a rough idea about the trend of the values and particularly to know
whether the relationship is linear or not, so that the measurement of degree of
correlation can be made with fruitful results. The diagram is not helpful in
any way to state that the relationship is strong, moderately strong, average or
poor.
8.5 KARL PEARSON’S COEFFICIENT OF
CORRELATION
We mentioned in the above sections a number of times that when we study
two variables simultaneously, we are interested to know whether they have
some correlation between them and if yes, what is the nature of relationship
and how much they are related, that is, what is the degree of correlation.
Scatter diagram, though, tells us whether variables are correlated or not;but it
does not indicate the extent to which they are correlated.
No doubt, in order to know the degree of correlation (relationship), we need
some measure which can reflect the degree of relationship in terms of some
numerical value. Then, the value of the measure may numerically provide us
the idea of closest relationship, high relationship, moderate relationship,
weak relationship or no relationship between the variables. Fortunately, in
literature of Statistics, such a measure exists. It is known as “Coefficient of
Correlation”.
Coefficient of correlation given by British biometrician Karl Pearson (1867-
196 1936) gives the exact idea of the extent to which two variables are
correlated.The Karl Pearson’s coefficient of correlation is also known as Descriptive Statistics-II
“product moment correlation coefficient”. It is the most widely used method
of determining theextent and nature of linear relationship between two
variables.Coefficient of correlation measures, in numerical terms, the
intensity or degree of linear relationship between two variables.
Notationally, if X and Y are two random variables,then “Karl Pearson’s
correlation coefficient”or“Product Moment Correlation Coefficient” between
X and Y is generally denoted byletter r orby Corr(X, Y) and is defined as
Cov X, Y
r Corr X, Y ; …(1)
V X V Y
where, Cov X,Y stands for the covariance between X and Y which is
definedas:
1 N
Cov X, Y x i X yi Y
N i 1
and V X , V(Y) stand for the variance of X, Y respectively. Here, V(X) is
defined as:
1 N
V X i 1
(x i X)2 .
N
Similarly, V Y is defined as
1 N
V Y i 1
(yi Y)2
N
where, N is number of paired observations.
Using the definitions of covariance, and variance as given above,the
correlation coefficient “r” may be defined as:
1 N
x i X yi Y
N i 1
r Corr X,Y . … (2)
1 N 2 1 N 2
i 1( x i X) i 1(yi Y)
N N
If V(X) and V(Y) are denoted by the notations σ2X and σ2Y respectively, then
the formula of coefficient of correlation is reduces to
1 N
i1 xi X yi Y
r Corr X,Y N . … (3)
σx σy
It can be seen that given the N values of the variable Y, that is, y1, y2, …, yN
and corresponding values of the variable X; x1, x2, …, xN; it is easy to
N N 2 N 2
compute the quantities x
i 1 i X yi Y , y
i 1 i Y and x
i 1 i X
and hence to compute the value of r.
The Karl Pearson’s correlation coefficient r is seen to be a unit free quantity,
the reason being that the numerator and denominator have same units of
measurement. If X is measures in kilograms and Y in meters, then the unit of 197
Statistical Analysis numerator will be “kilogram meter” whereas in the denominator, the unit of
standard deviation, σ x , would be kilogram and that of σ y would be meter.
Hence, unit of numerator is same as that of denominator and hence cancels
each other. Because r does not depend on any unit, it is called a coefficient.
CHECK YOUR PROGRESS 2
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
3. Define product moment correlation coefficient.
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
8.5.1 Properties of Correlation Coefficient
Before going into the problem of computing r using different methods,we
shall discuss here some of the salient properties of it.
Property 1 (Limits of r): The value of product moment correlation
coefficient(Karl Pearson’s coefficient of correlation) always lies in the range
[–1 to 1], where–1 and 1 are inclusive in the range.
Proof: Consider the inequality
2
1 N x i X yi Y
N i 1 σ x σ y
0 ,
which is always true for all values of xi and yi, since, the left-hand side
(L.H.S.) expression is a squared quantity. Now, considering first the plus sign
between the two terms and expanding the L.H.S., we get
2 2
1 N xi X 1 N yi Y 2 N x i X yi Y
N i 1 σ x N i 1 σ y N i 1 σ x σ y
0
N 2 N 2
1 i 1 x i X 1 y
i 1 i Y 2 N x i X yi Y
or,
N σx2
N σ 2
N
i 1
0. … (4)
y σx σy
But, we know that
1 N 2 1 N 2
N
i 1
x i X V X σ 2x ; and
N
i 1
yi Y V Y σ 2y
1 N
i1 x i X yi Y
andalso N r, from (3).
σxσy
Therefore, (4) becomes
1 + 1 + 2r 0 2 + 2r 0 r 1. … (5)
198
Now, considering the negative sign in the expression Descriptive Statistics-II
2
1 N x i X yi Y
N i 1 σ x σ y
0 ,
we have
2 2
1 N xi X 1 N yi Y 2 N x X yi Y
i 1 i 1 i 1 i
0
N σx N σy N σx σy
or, 1 + 1 – 2r 0 that is 2 - 2r 0 that is, r 1. … (6)
Therefore, combining (5) and (6), we get the result
1 r 1. … (7)
which gives the range of the coefficient of correlation.This completes the
proof of property 1.
Some specific values of r:
We observe thatr 1, 1 , this means that depending upon the values of the
two variables, it may assume any value in between –1 to +1, both values
inclusive. Three important values of r are important to consider in this range.
These are –1, 0 and +1.
(i) It assumes value –1, when there is a perfect negative linear relationship
between the variables, that is, asone variable increases in its values, the
other variable decreases in its values through an exact linear
relationship rule.
(ii) Similarly, it assumes value +1 when there is a perfect positive linear
relationship between the variables, that is, as one variable increases in
its values, the other variable also increases in its values through an
exact linear rule.
(iii) The r assumes value zero, when there is no linear relationship between
the variables. However, if there exists a linear relationship, then a zero
value of r indicates that the variables are not related at all.
(iv) However, other values of r also indicate the extent of the linear
relationship.
Being maximum values of r; –1 and +1 on extreme left side and extreme right
side of the range respectively and zero at the middle of the range; it seems
that r becomes weaker and weaker as it approaches to middle most point of
the range from both the sides. Depending upon the values of r in its range,
we, therefore, define weak, moderate and strong relationships as follows:
(a) The linear relationship is said to be a “weak positive (negative)
relationship” if the value of r lies between 0.0 and 0.3 (–0.0 and - 0.3).
This happens when the linear rule, showing the relationship is
represented by a shaky linear rule.
(b) Values between 0.3 and 0.7 (–0.3 and –0.7) indicate a “moderate positive
(negative) linear relationship” through a fuzzy-firm linear rule. 199
Statistical Analysis (c) Lastly, values between 0.7 and 1.0 (–0.7 and –1.0) is an indication of a
“strong positive (negative) linear relationship” through a perfect linear
rule.
Property 2: The coefficient of correlation, r, is independent of change of
origin and scale of the variables. This indicates that if rXYbe the coefficient of
correlation between two variables X and Y and rUV be the same between the
variables U and V which are respectively obtained by changing the origin and
scale of the variables X and Y, then we haverUV= rXY.
1 N
Cov. X, Y i1 x i X yi Y
Proof: We know that rXY= N .
σXσY σx σy
Xa
Now, let us define the variable U , which is obtained by changing the
b
origin and scale of the variable X; constant a changes the origin of X and
constant b is used to change the scale of X. Similarly, we define the variable
Yc
V ,which is obtained by changing the origin and scale of the variable
d
Y. From these relations, we get X= a + bU and Y = c + dV.
Hence, we see that X a bU and Y c dV . Therefore, substituting these
values of X, Y, X and Y in the formula of rXY,,we have
1 N
i 1
b u i U d vi V
rXY N … (8)
bσ U .dσ V
The denominator is obtained due to the results that σ X bσ U and σ Y dσ V ;
which are properties of standard deviation(already proved in the unit
discussed the measures of dispersion).
Now, (8) becomes
1 N 1 N
bd i 1 u i U v V
i i 1
u i U vi V
rXY N N ;
bdσ U σ V σ UσV
which, by definition, is the coefficient of correlation between variables U and
V, denoted by rUV. So, we have the relation rXY = rUV. Hence the proof is
complete.
This property is very much useful in computingthe coefficient of correlation
when X and Y values are large enough in magnitudes. Using suitably chosen
constants a, b, c and d; the magnitudes of X and Y values can be reduced in
magnitude so that U and V values are comparatively smaller in magnitudes.
Thus, the calculations steps may be made simpler.
8.5.2 Calculation of Correlation Coefficient
In this sub-section, we shall describe how to calculate the value of r using
different methods of calculations:
200
(a) Using the direct formula of r: Descriptive Statistics-II
We shall show here how to calculate the Karl Pearson’s Coefficient of
Correlation for ungrouped data using the actual formula of r.
The direct (actual) formulaof r given by (2) is
1 N
x i X yi Y
N i 1
r Corr X,Y ;
1 N 2 1 N 2
i 1( x i X) i 1(yi Y)
N N
which needs the computation of
1 N 1 N
Cov X, Y
N
i 1
x i X yi Y ; V X i 1( x i X)2 and
N
1 N
V Y (yi Y)2 .For a given set of pairs of values (xi, yi), these
N i 1
expressions can be obtained in the form of a table. By taking the following
example, we will show how these quantities can be calculated:
Example 1: Find the correlation coefficient between advertisement
expenditure and profit for the following data:
Advertisement expenditure 30 44 45 43 34 44
Profit 56 55 60 64 62 63
Solution: In order to find the value of r from this formula, it is clear that we
N N 2
need the values of the quantities xi 1 i X yi Y , y
i 1 i Y and
N 2
xi 1 i X . For this, we first need the values of X and Y. The following
table provides rest of the necessary calculations:
Taking Advertisement expenditure as variable X and Profit as variable Y, we
1 6 240 1 6 360
see that X
6
x
i 1 i
6
40 and Y
6
y
i 1 i
6
60.
X Y X X (X X) 2 Y Y (Y Y) 2 X X Y Y
30 56 −10 100 −4 16 40
44 55 4 16 −5 25 −20
45 60 5 25 0 0 0
43 64 3 9 4 16 12
34 62 −6 36 2 4 −12
44 63 4 16 3 9 12
6 6 6 6 7 6
xi (x i x) 2 (y i y) 2
y
i 1 i x
i 1 i x i 1 y i 1 i y i 1 x i 1 i x yi y
240 360 0 202 0 70 32
Substituting values from the table in the formula of r, we get
32 32 32
r 0.27 .
202 70 14140 118.91
201
Statistical Analysis Hence, the correlation coefficient between expenditure on advertisement and
profit is 0.27. Since the value of r is 0.27, the relationship between the two
variables is positive but weak.
(b) Using the extended formula of r:
In the above examples, we used the formula of r as it was originally defined
in the formula (2). However, we mentioned a number of times in earlier units
that in the process of calculations, we need to avoid terms involving
deviations of values of the variables from its respective means, since,
calculation of these terms istroublesome on the one hand and generate
rounding errors which affect the actual value of the quantity. We, therefore,
used formulae for variance, standard deviation and many such terms by
expanding the deviation-type terms in order to simplify the calculations as
well as minimizing the rounding errors as far as possible. We should develop
a formula for it by expanding these terms which would be preferable over the
direct formula for calculation purpose. Let us find it.
We have already expanded formulae of standard deviations σ X and σ Y
under the previous unit discussing measures of dispersion above. These are
1 N 1 N
σ X i 1x i2 X 2 and σ Y i 1yi2 Y 2
N N
which are the terms in the denominator of (2).
Now, the numerator of (2) is
1 N
Cov (X, Y) = xi X yi Y
N i 1
1 N N N N
x y Y i 1x i X i 1yi i 1XY
i 1 i i
N
1 N
x i yi NXY NXY NXY ;
N i 1
N N N 1 N
since, x NX,
i 1 i y NY and
i 1 i i 1
XY NXY xi yi XY.
N i 1
Thus, the extended formula for r is given by
1 N
x i yi XY
N i 1
r . … (9)
1 N 2 2 1 N 2 2
i 1x i X i 1yi Y
N N
It can further be reduced to the formula
N N N
N i 1x i y i i 1x i i 1y i
r . … (10)
2 2
N 2
N i 1x i x N
i 1 i
N 2
N i 1yi
y
N
i 1 i
It can be seen that in this formula of r, all the deviation-type terms have been
202 removed and all the terms can be calculated using the exact values of xi and
yi. Thus, this form of the formula is suitable for calculation purpose, since, it Descriptive Statistics-II
minimizes the rounding error in each the terms.
Let us now illustrate the steps which should be followed for computing the
value of r using the extended formula as obtained in (10). We use the
following example for this purpose:
Example3: Calculate Karl Pearson’s coefficient of correlation between price
and demand for the following data:
Price 17 18 19 20 22 24 26 28 30
Demand 40 38 35 30 28 25 22 21 20
Solution:Since, formula for r is
N N N
N i 1x i y i i 1x i i 1y i
r ;
2 2
N 2
N i 1x i x
N
i 1 i
N 2
N i 1yi
y
N
i 1 i
N N N N 2
therefore, we will compute the values of x ,
i 1 i y ,
i 1 i x yi ,
i 1 i x
i 1 i
N 2
and y from the given values of xi and yi. We present the calculations for
i 1 i
these terms in the following table:
X Y X2 Y2 XY
17 40 289 1600 680
18 38 324 1444 684
19 35 361 1225 665
20 30 400 900 600
22 28 484 784 616
24 25 576 625 600
26 22 676 484 572
28 21 784 441 588
30 20 900 400 600
x i 204 yi 259 xi2 4794 yi2 7903 x i yi 5605
Therefore,
r
9 5605 204 259
9 4794 204 204 9 7903 259 259
50445 52836
r
43146 41616 71127 67081
2391 2391
r 0.96 .
1530 4046 2488.0474
203
Statistical Analysis
CHECK YOUR PROGRESS 3
Note: i) Check your answers with those given at the end of the unit.
4) Calculate coefficient of correlation between X and Y for the following
data:
X 1 2 3 4 5
Y 2 4 6 8 10
5) Find the coefficient of correlation for the following ages of husband
and wife.
Husband’s 23 27 28 29 30 31
age
Wife’s age 18 22 23 24 25 26
8.6 SPEARMAN’S RANK CORRELATION
COEFFICIENT
For the calculation of product moment correlation coefficient characters must
be measurable. In many practical situations, characters are not measurable.
Sometimes, we are given a distribution of items where no numerical measure
can be made but where best and worst or most favoured and least favoured
can be identified. They are quality characteristics and individuals or items
can be ranked in order of their merits. This type of situation occurs when we
deal with the qualitative study such as honesty, beauty, voice, etc. Rankings
are often applied in these situations to put the series into an order. For
example, contestants of a singing competition may be ranked by judge
according to their performance. In another example, students may be ranked
in different subjects according to their performance in tests.
Arrangement of individuals or items in order of merit or proficiency in the
possession of a certain characteristic is called ranking and the number
indicating the position of individuals or items is known as rank. If ranks of
individuals or items are available for two characteristics then correlation
between ranks of these two characteristics is known as rank correlation. If we
have a group of individuals ranked according to two different characteristics,
it is natural to ask that is there any association between the ranks? To answer
this question, we need to use the spearman’s rank correlation coefficient.
The formula for the determining the degree of relationship between two
ranked characteristics is given as follows:
6 di2
r 1
N(N 2 1)
204
where di is the difference between the two ranks given to each individual and Descriptive Statistics-II
N is the number of observations.
With the help of rank correlation, we find the association between two quality
characteristics. As we know that the Karl Pearson’s correlation coefficient
gives the intensity of linear relationship between two variables, Spearman’s
rank correlation coefficient gives the concentration of association between
two quality characteristics. In fact, Spearman’s rank correlation coefficient
measures the strength of association between two ranked variables.
8.6.1 Method of Calculation of Rank Correlation
Let us discuss some problems on rank correlation coefficient. We shall
consider three cases to compute the rank correlation coefficient as follows:
Case I: When Actual Ranks are given
Case II: When Ranks are not given
Case III: When Ranks are repeated.
Case I: When Actual Ranks are given
In this case the following steps are involved:
i) Compute d i R X R Y , i.e., the difference between the two ranks given
to each individual
ii) Compute d i2 , i.e., the squares of differences for each individual.
iii) Apply the formula of the rank correlation coefficient by putting the value
of sum of the squares of the difference of the each individuals and N:
Example 4: Suppose we have ranks of 8 students of B.Sc. in Statistics and
Mathematics. On the basis of rank we would like to know that to what extent
the knowledge of the student in statistics and mathematics is related.
Rank in Statistics 1 2 3 4 5 6 7 8
Rank in 2 4 1 5 3 8 7 6
Mathematics
Solution: Let us denote the rank of students in statistics by R x and rank in
mathematics by R y . For the calculation of rank correlation coefficient we
N
2
have to find d
i 1
i which is obtained through the following table:
Rank Rank
Difference of ranks
inStatistics inMathematics d i2
di R x R y
Rx Ry
1 2 −1 1
2 4 −2 4
3 1 2 4
205
Statistical Analysis 4 5 −1 1
5 3 2 4
6 8 −2 4
7 7 0 0
8 6 2 4
2
d i 22
Here, N = number of paired observations =8
6 d i2 6 X 22 132 372
rs 1 1 1 0.74
N (N 2 1) 8 X 63 504 504
Thus, there is a positive association between ranks of Statistics and
Mathematics.
Case-II: When Ranks are not given
Sometimes not have rank not given but actual values of both variables are
available. If we are interested in rank correlation coefficient it is necessary to
have ranks so, we a sign ranks to the given values. Ranks can be assigned by
taking either the highest value as 1 or the lowest value as 1. The next highest
or the next lowest value is given the rank 2 and so on. But whether we start
with the highest value or lowest value we must follow the same method in
case of both the variables. Considering this case, we are taking a problem and
try to solve it.
Example 5: Calculate rank correlation coefficient from the following data:
X 78 89 97 69 59 79 68
Y 125 137 156 112 107 136 124
Solution:
Rank of X Rank of Y
X Y d = Rx-Ry d2
(Rx) (Ry)
78 125 4 4 0 0
89 137 2 2 0 0
97 156 1 1 0 0
69 112 5 6 -1 1
59 107 7 7 0 0
79 136 3 3 0 0
68 124 6 5 1 1
N
d =2
i 1
2
i
206
Spearman’s Rank correlation formula is Descriptive Statistics-II
N
6 d i2
i 1 62 12
rs 1 2
1 1 0.96
N(N 1) 7(49 1) 7 48
Case-III: When Ranks are repeated
In previous section it was assumed that two or more individuals or units do
not have same rank. But there might be a situation when two or more
individuals have same rank in one or both characteristics, then this situation
is said to be tied.
If two or more individuals have same value, in this case common ranks are
assigned to the repeated items. This common rank is the average of ranks
they would have received if there were no repetition. For example, we have a
series 50, 70, 80, 80, 85, 90 then 1st rank is assigned to 90 because it is the
biggest value then 2nd to 85, now there is a repetition of 80 twice. Since both
values are same so the same rank will be assigned which would be average of
the ranks that we would have assigned if there were no repetition. Thus, both
80 will receive the average of 3 and 4 i.e. (Average of 3& 4 i.e. (3+4)/2= 3.5)
3.5 then 5th rank is given to 70 and 6th rank to 50. Thus, the series and ranks
of items are
Series: 50 70 80 80 85 90
Ranks: 6 5 3.5 3.5 2 1
In the above example 80 was repeated twice. It may also happen that two or
more values are repeated twice or more than that.
For example, in the following series there is a repetition of 80 and 110. You
observe the values, assign ranks and check with following:
50 70 80 90 80 120 110 110 110 100
10 9 7.5 6 7.5 1 3 3 3 5
m (m 2 1)
When there is a repetition of ranks, a correction factor is added to
12
d 2 in the Spearman’s rank correlation coefficient formula, where m is the
number of times a rank is repeated. It is very important to know that this
correction factor is added for every repetition of rank in both characters.
In the first example correction factor is added once which is 2(4-1)/12=0.5,
while in the second example correction factors are 2(4-1)/12=0.5 and 3(9-
2
1)/12=2 which are aided to d .
Thus, in case of tied or repeated rank Spearman’s rank correlation coefficient
formula is
207
Statistical Analysis m(m 2 1)
6 d 2 ...
12
rs 1 2
N(N 1)
Example 6: Calculate rank correlation coefficient from the following data:
Expenditure
on 10 15 14 25 14 14 20 22
advertisement
Profit 6 25 12 18 25 40 10 7
Solution: Let us denote the expenditure on advertisement by X and profit by
Y
X Rank of X Y Rank of Y d = Rx-Ry d2
(RX) (RY)
10 8 6 8 0 0
15 4 25 2.5 1.5 2.25
14 6 12 5 1 1
25 1 18 4 −3 9
14 6 25 2.5 3.5 12.25
14 6 40 1 5 25
20 3 10 6 −3 9
22 2 7 7 −5 25
d 83.50
2
m(m 2 1)
6 d 2 ...
12
rs 1 2
N(N 1)
Here rank 6 is repeated three times in rank of X and rank 2.5 is repeated
twice in rank of Y so the correction factor is
3(32 1) 2(2 2 1)
12 12
Hence, rank correlation coefficient is
3(32 1) 2(2 2 1) 3 8 2 3
683.50 6 83.50
12 12 12 12
rs 1 1
8(64 1) 8X 63
6(83.50 2.50)
rs 1 1 1.024 0.024
504
There is a negative association between expenditure on advertisement and
profit.
208
CHECK YOUR PROGRESS 4 Descriptive Statistics-II
Note: i) Check your answers with those given at the end of the unit.
6) Calculate Spearman’s rank correlation coefficient from the following
data:
x 20 38 30 40 50 55
y 17 45 30 35 40 25
7) Calculate rank correlation coefficient from the following data:
x 70 70 80 80 80 90 100
y 90 90 90 80 70 60 50
8.7 CONCEPT OF REGRESSION
Before we proceed to discuss the concept and application of regression
analysis mainly used for prediction purpose, let us describe first the literal
meaning of this word and how it was used for the first time in the subject.
Literally, regression means to return back to a previous and less advanced or
less developed or worse state/condition. It is an act of going or coming back
to or moving backward to a previous worse or more primitive state or
condition.
‘Regress’ is the opposite of ‘progress’. Therefore, to regress is not a good
thing in many situations whereas it may be good in other situations. For
example, if the economy of a country regresses, it is not a good sign for the
economic condition of the country; whereas on the other hand, if the mental
performance of mentally retarded patients starts improving, we can hope that
they are not regressing, which is a good thing for such patients.
The term “regression” was first coined in the nineteenth century by Sir
Francis Galton (1822-1911), a British mathematician, statistician and
biometrician. He used this word while he was working on a biological
phenomenon regarding the height of parents and their offsprings. In fact,
during experimentation he observed the particular biological phenomenon
that if parents in families were very tall, their children tended to be tall but
shorter than their parents. On the other hand, if parents were very short, their
children tended to be short but taller than their parents were. This
phenomenon he named as "regression to the mean," with the word
"regression" meaning to come back to. Although, Galton used the word
regression only for this biological phenomenon, later on, his work was
extended by some other statisticians, like, Udny Yule, Karl Pearson and R. A.
Fisher to a more general statistical context.
209
Statistical Analysis CHECK YOUR PROGRESS 5
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
8. What do you understand the word ‘regression’?
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
8.8 LINES OF REGRESSION
Let us now discuss the Linear Regression Models. As described above, in
linear regression model, we will have some straight-line equations as
regression models.
Considering first the simplest case of linear regression, that is, single
variable linear regression model, our aim here would be to explain the
method of finding the best-fitted regression lines which would be the most
appropriate straight lines expressing the average relationship between
variables.
8.8.1 Single Variable linear Regression Lines
As you know, by single variable regression model we mean in the linear
regression, only one independent variable is included other that one
dependent variable. Such a regression model may be expressed as Y = a + bX
or Y = b0 + b1X1, where Y represents the dependent variable or ‘output’
variable; X or X1 represent the only independent variable or ‘predictor’ and a
and b or alternatively b0 and b1 are the constants (parameters) of the
corresponding regression line. For generalization purpose later on, we shall
write our model as
Y = b0 + b1X1. … (11)
i) Regression Line of Y on X1:
Let us first consider the single variable linear regression model given by
equation (11). This model, as explained earlier, is used to estimate (predict)
the dependent variable Y on the basis of independent variable X1, where we
assume that variables are linearly related to each other with some degree of
correlation. Such a straight-line equation which is used to predict the output
variable Y for given predictor values X1is termed as Regression Line of Y
on X1. We shall show how we can get the straight line (11) which fits best to
a set of pairs of values (yi, x1i); i = 1, 2, …, N.
As far as the problem of finding the most appropriate line which fits best to a
set of data (yi, x1i); i = 1, 2, …, N is concerned; it seems to be equivalent to
obtaining best-fitted straight-line equation using least square principle. We
210
know that the least square principle provides us a curve/line which when Descriptive Statistics-II
fitted to any set of bivariate data; (yi, x1i); it passes through maximum
number of points and rest of the points which do not lie exactly on the
curve/line, the difference between actual and estimated values of the
independent variable are minimum. No doubt, thus, least square principle
provides us a method of finding most appropriate curve/line which is
intuitively a scientifically-justified method and suffices our purpose.
We shall, therefore, use the same principle here for finding the Regression
Line of Y on X1. If we follow the method of least squares, we can write the
sum of squares of residuals as
N 2 N 2 N 2
U y i 1 i b 0 b1x1i i 1 yi b 0 b1 x1i i 1 y i ŷi … (12)
where yi and ŷi respectively denote the actual and predicted values of the
variable Y.
The two normal equations, obtained by minimizing U simultaneously with
respect to parameters b0 and b1 are
N N
y Nb0 b1 i 1x1i ;
i 1 i
… (13)
N N N
i 1 i
y x1i b 0 i 1x1i b1 i 1x1i2 … (14)
N
Multiplying the equation (13) by x and the equation (14) by n and then
i 1 1i
subtracting the resultant second equation from the resultant first equation, we
have
2
y x N y x b x
N
i 1 i
N
i 1 1i
N
i 1 i 1i 1
N
i 1 1i
N
N i 1x1i2
N y x y x
N N N
i 1 i 1i i 1 i i 1 1i
or, b 1 2
… (15)
N x x
N 2 N
i 1 1i i 1 1i
This is the estimated value of the parameter b1, denoted by b̂1 , obtained on the
basis of given (yi, x1i ) values.
N N
Now, since we know that y NY; i 1x1i NX1 , we have
i 1 i
1 N
y x X1Y Cov X , Y
i 1 i 1i 1 rσ Y σ X1
bˆ 1 N
1 N 2 2 V X σ X2 1
N i 1
x 1i X1
1
σY
or, b̂1 r ; r being the product moment correlation between X1 and Y and
σ X1
σ Y and σ X1 ,respectively, the standard deviation of Y and X1 variables.
Substituting the value of b̂1 in equation (13) and solving for b0, we get
σY
b̂0 = Y r X1.
σ X1 211
Statistical Analysis Finally, substituting these estimated values of constants b0 and b1 in the
regression model (11), we get
σY
Y Y r σ X 1 X1 . … (16)
X1
This is the single variable linear regression model, including the single
independent variable X1 and the dependent variable Y. It represents the linear
relationship between the variables Y and X1 where the degree of correlation
is given by r. The equation can be converted into a linear equation as
σ σ
Y Y r Y X1 r Y X1 bˆ 0 bˆ 1X1 ; … (17)
σ X1 σX1
where b̂0 and b̂1 are the estimated values of the parameters b0 and b1
respectively, obtained through the given data set (xi, yi), such that the
regression line (17) provides the best fit to the data, in the sense that the
predicted values of Y would be closest to its corresponding actual values.
Since, through this regression, we predict the values of the dependent
variable Y for given values of the variable X1; it is popularly termed as the
“Regression Line of Y on X1”.
ii) Regression Line of X1 on Y:
As described above, regression line of Y on X1 is used to predict the values
of the dependent variable Y when values of the independent variable X1 are
given. You may argue that for predicting the values of the X1 variable, given
Y values, the same regression line, (12) can be used if we convert the line in
the form
1 bˆ
X1 Y 0 or X1 b*0 b1*Y ;
bˆ
1 bˆ 1
bˆ 1
where b*0 0 and b1* .
ˆ b̂1
b1
However, mathematically it is not incorrect, but statistically this method is
not justified from the point of view of least square principle. You know that
in order to obtain the best-fitted line using this principle, we use the concept
of minimizing the sum of squares of residuals given in (12) as
N 2 N 2
U i 1 y i b 0 b1 x1i i 1 yi yˆ i .
If you observe, in fact, yi ŷi s are the vertical distances of points for
given X-values. Thus, U is the sum of squares of all such vertical distances of
points between the actual and predicted Y-values, which has been minimized
with respect to all the parameters.
Now, if instead of Y values, X1-values are to be predicted for known Y-
values, then, according to least square principle, we should minimize the sum
212
of squares of residuals given by x1i x̂1i s which would be horizontal Descriptive Statistics-II
distances between the actual and predicted X1-values for all the points.
Certainly then, we will get a new regression line which will be different than
X1 b*0 b1*Y, because the method of estimation of parameters b0 and b1 in the
two methods are quite different.
Now, since we wish to treat X1-variable as dependent and Y-variable as
independent in order to obtain regression line of X1 on Y using least square
principle, we write the single variable linear regression line of X1 on Y as
X1 c 0 c1Y … (18)
Using the principle of least squares, the sum of squares of residuals
N 2 N 2
V i 1 x1i c 0 c1 y i i 1 x1i x̂1i
is to be minimized with respect to parameters c 0 and c1 . The corresponding
two normal equations will be
N N
x Nc0 c1 i 1yi ;
i 1 1i
… (19)
N N N
y x1i c0 i 1yi c1 i 1yi2
i 1 i
… (20)
Finally, solving the above two equations, we get
c1
N
N i 1yi x1i y x
N
i 1 i
N
i 1 1i
… (21)
2
y y
N 2 N
N i 1 i i 1 i
This is the estimate of the parameter c1 , say ĉ1 , as obtained on the basis of
the given set of data.
The expression (16) can further be written as
Cov X1 , Y rσ X1 σ Y σ X1
cˆ 1 2
2
r .
σ Y σ Y σY
Further, substituting the value of ĉ1 in the equation (19) and solving for c0 ,
we get
σ X1
cˆ 0 X1 r Y
σY
Hence, the regression line of X1 on Y is given by
σ X1
X 1 X1 r
σY
Y Y . …(22)
It can alternatively be written as
σX σX
X1 X1 r 1 Y r 1 Y cˆ 0 cˆ1Y … (23)
σY σY
which is the regression line of X1 on Y.
213
Statistical Analysis Thus, there are always two lines of regression, one, regression line of Y on
X1, for predicting the values of dependent variable Y given the values of
independent variable X1 and the other, regression line of X1 on Y, for
predicting the values of dependent variable X1 given the values of
independent variable Y.
8.8.2 Calculation of Regression Lines
After obtaining the two regression lines (16) and (22), the next step would be
to fit these lines to a given set of data (yi, x1i) for i = 1, 2,.., N; and then
consequently to use them for prediction purposes. We observe that for this
purpose, we have to find the values of Y, X1 , σ Y , σ X and r. Given yi and x1i
1
values, it is easy to compute the values of means, Y, X1 . However, although
the values of σ Y , σ X and r can be computed using their respective direct
1
formulae, but we avoid to use them for computation purpose since they are
not free from error of approximations. We have mentioned there that
alternative formulae for them should be used which are derived from their
direct formulae. For readiness of the text in this unit, we present below these
formulae which should be used for calculation
1 N
1 N 1 N x1i yi X1 Y
N i 1
Y i 1yi ; X1 i 1x1i ; r ;
N N 1 N 2 2 1 N 2 2
i 1x1i X1 i 1yi Y
N N
1 N 2 1 N 2
σY y Y 2 ; σ X1
i 1 i x1i X12 .
N N i 1
CHECK YOUR PROGRESS 6
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
9. What are lines of regression? What are their purposes?
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
10. Explain why there are two lines of regression. Obtain the expression of
regression line of Y on X1 where Y and X1 are respectively the dependent
and independent variables.
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
214
Descriptive Statistics-II
8.9 REGRESSION COEFFICIENTS
We have seen that the regression line of Y on X1 is given by
σ σ
Y Y r Y X1 r Y X1
σ X1 σ X1
Where as that of X1 on Y is
σX σX
X1 X1 r 1 Y r 1 Y .
σY σY
σY σX
The multipliers r of X1 in the line of Y on X1 and r 1 of Y in the
σX σY
1
line of X1 on Y are respectively called the “Regression Coefficient of Y on
X1” and “Regression Coefficient of X1 on Y. Symbolically, these are
σY
respectively denoted by letters b YX and b X Y. ; that is, b YX r ,the
1 1 1
σ X1
σ X1
regression coefficient of Y on X1 and b X Y. r , the regression coefficient
1
σY
of X1 on Y.
It can be seen that, in fact, regression coefficients give us the slopes of the
σY
respective two regression lines. Therefore, regression coefficient b YX r 1
σ X1
gives the change in the dependent variable Y as a unit change in the
independent variable X1. As for example, if we are considering the
production of a crop by Y and amount of rainfall by X1, then regression
coefficient of Y on X1 represents the change in production of crop which
occurs due to a unit change in rainfall.
σ X1
Similar interpretation can be given for the regression coefficient b X Y r 1
σY
of X1 on Y.
8.9.1 Properties of Regression Coefficients
Property 1: The geometric mean of the two regression coefficients is the
coefficient of correlation r.
σY σ
Proof: We know that b YX r and b X1Y r X1 . Therefore, we have
1
σ X1 σY
b YX1 . b X1Y r 2 .
This indicates that r b YX . bX Y ;1 1
that is, the coefficient of correlation, r is the geometric mean of the two
regression coefficients. Thus, given the magnitude of regression coefficients,
the value of the correlation coefficient can be obtained. However, as far as
215
Statistical Analysis the sign of the correlation coefficient is concerned, we can observe that
whenever r is negative, both the coefficients b YX and b X Y would be 1 1
negative, whereas if r is positive, both the coefficients would be positive.
Therefore, we take r positive if both the regression coefficients are positive,
otherwise r would be negative.
Property 2: If one of the regression coefficients is greater than one, then
other must be less than one.
Proof: Let b YX the regression coefficient of Y on X1 is greater than one, that
1
is,
1
b YX1 1 or 1.
b YX1
We know that
b YX1 . b X1Y r 2 1 (From Property 1)
1
b X1Y 1.
b YX1
Thus, if b YX is greater than one then b X Y is less than one.
1 1
Property 3: Arithmetic mean of the regression coefficients is greater than the
b YX1 b X1Y
correlation coefficient, that is, r subject to the condition r > 0.
2
Proof: Suppose that arithmetic mean of regression coefficients is greater
than the correlation coefficient. Then, we have
b YX1 b X1Y
2
r b YX1 b X1Y 2r
b YX1 b X1Y 2 b YX1 . b X1Y since, r b YX . bX Y
1 1
Therefore, we have
b YX1
b X1Y 2
b YX1 . b X1Y 0
2
b YX1 b X1Y 0,
which,being a squared quantity, is always true.Therefore, assuming that
b YX1 b X1Y
r;
2
that is, arithmetic mean of regression coefficients is greater than the
correlation coefficient, is always true.
Let us now solve some numerical problems on what we have discussed in this
unit:
216
Example 7: Height of fathers and sons in inches are given below: Descriptive Statistics-II
Height of Father 65 66 67 67 68 69 70 71
Height of Son 66 68 65 69 74 73 72 70
Find two lines of regression and calculate the estimated average height of son
when the height of father is 68.5 inches.
Solution: Let us denote the father’s height by X1 and son’s height by Y, then
for finding the two lines of regression, we do necessary calculations in the
following table:
X1 Y X12 Y2 X1Y
65 66 4225 4356 4290
66 68 4356 4624 4488
67 65 4489 4225 4355
67 69 4489 4761 4623
68 74 4624 5476 5032
69 73 4761 5329 5037
70 72 4900 5184 5040
71 70 5041 4900 4970
2 2
x 1i 543 y 557 x 1i 36885 y i 38855 x 1i yi 37835
From the values obtained in thetable, we have
543 557
X1 67.88; Y 69.62.
8 8
1 N 1 N 2
Standard Deviation of X X (x1i X1 )2 x1i X12
1
N i 1 N i 1
1
36885 (67.88) 2
8
4610.62 4607.69 2.93 1.71
1 N 1 N 2
Similarly, Standard deviation of Y Y (yi Y) 2 yi Y 2
N i 1 N i 1
1
38855 (69.62)2
8
4856.88 4846.94 9.94 3.15
217
Statistical Analysis Now correlation coefficient
N N N
N x1i yi ( x1i )( y i )
i 1 i 1 i 1
r Corr(X1 , Y)
N N N N
2 2 2
1i 1i i
N x ( x ) N y ( yi )2
i 1 i 1 i 1 i 1
8 37835 (543) (557)
(8 36885 (543) (8 38855) (557)
2 2
302680 302451 229
(295080 294849310840 310249 231 591
229 229
0.62
136521 369.49
Substituting the value of X1 , Y, X , Y and r in regression equations, we
1
get regression line of Y on X1 as
3.15
(Y 69.62) 0.62 (X1 67.88)
1.71
Y 1.14X1 77.38 69.62
Y 1.14X1 7.76
and regression line of X1 on Y
1.71
(X1 67.88) 0.62 (Y 69.62)
3.15
X1 0.34 Y 23.67 67.88
X1 0.34 Y 44.21
Obviously, estimate of height of son for the height of father = 68.5 inch is
obtained by the regression line of Y on X1
Y 1.14X1 7.76 .
Putting X1 = 68.5 in the above regression line, we have
Y 1.14 68.50 7.76 78.09 7.76 70.33
Thus, the estimate of son’s height for the father’s height 68.5 inch is 70.33
inch.
CHECK YOUR PROGRESS 7
Note: i) Check your answers with those given at the end of the unit.
11) We have data on variables X1 and Y as
X1 5 4 3 2 1
Y 9 8 10 11 12
218
Calculate Descriptive Statistics-II
(i) both the regression coefficients,
(ii) correlation coefficient,
(iii) regression lines of Y on X1 and X1 on Y, and
(iv) the estimated value of Y for X1 =4.5.
8.10 LET US SUM UP
In this unit, we discussed
1. the concept of correlationbetween two variables and its importance in the
simultaneous analysis of them;
2. the meanings of different types of correlationsuch as, linear correlation,
multiple correlation, partial correlation, non-linear correlation;
3. scatter diagramand how and why it is helpful to have an approximate
idea about the trend of the relationship between the variables;
4. how the formula of Karl Pearson’s coefficient of correlationis defined
and how the formula for computing the coefficient is obtained from it;
5. the different properties of correlation coefficient, like, the range of it,
some significant values of it, effect of change of origin and scale on it;
6. What the word “regression” means and in what sense it was introduced
inthe subject of Statistics;
7. What is the importance and meaning of regression analysis in Statistics;
8. How the two single variable linear regression models are theoretically
obtained using least squares principle. What are regression lines of Y on
X1 and X1 on Y. How the two regression lines can be applied to any set
of data for predicting the dependent variable;
9. What are the important properties of regression lines;
10. What do we mean by regression coefficients in the regression analysis
and what is the significance of these coefficients; and
11. What are some of the important properties of regression coefficients and
how these are useful in order to calculate some parameters of the
regression models.
8.11 KEY WORDS
Scatterplot : A special graph containing a cluster of dots that
represents all pairs of observations.
Pearson Correlation : A number between –1.00 and 1.00 that describes
Coefficient (r) the linear relationship between pairs of quantitative
variables.
219
Statistical Analysis
8.12 SUGGESTED FURTHER READING/
REFERENCES
Witte, R., & Witte, J. (2017). Statistics. Hoboken, NJ: John Wiley & Sons.
8.13 ANSWERS TO CHECK YOUR PROGRESS
1) Please refer to section 8.3.
2) Please refer to section 8.3.1.
3) Please refer to section 8.5.
4) First, we form the table showing needed calculations.
X Y X X (X X) 2 Y Y (Y Y) 2 X X Y Y
1 2 –2 4 –4 16 8
2 4 –1 1 –2 4 2
3 6 0 0 0 0 0
4 8 1 1 2 4 2
5 10 2 4 4 16 8
15 30 0 10 0 40 20
Here x i 15 X 3 and yi 30 Y 6
From the calculation table we observe that
(x i X)2 10 , (yi Y)2 40 and x i X yi Y 20
Substituting these values in the formula
N
x
i 1
i X yi Y
20 20
r 1
N
2
N
2 10 40 20
x
i 1
i X y
i 1
i Y
Hence, there is perfect positive correlation between X and Y.
5) Let us denote the husband’s age as X and wife’s age by Y
X Y X Y2 XY
23 18 529 324 414
27 22 729 484 594
28 23 784 529 644
29 24 841 576 696
30 25 900 625 750
31 26 961 676 806
x i 168 yi 138 xi2 4744 yi2 3214 x i yi 3904
220
We use the formula Descriptive Statistics-II
r
6 3904 168 138 23424-23184
=1
6 4744 168 168 6 3214 138 138 240 240
Hence there is perfect positive correlation between X and Y.Perfect
positive correlation between X and Y is because of the perfect linear
relationship, given by Y = X – 5, exists.
6) First, we form the table showing needed calculations.
Rank of x Rank of y d = RX - 2
X Y d
(Rx) (Ry) RY
20 17 6 6 0 0
38 45 4 1 3 9
30 30 5 4 1 1
40 35 3 3 0 0
50 40 2 2 0 0
55 25 1 5 −4 16
n
2
d
i 1
i =26
6 X 26 26 9
rs 1 1 0.26
6(36 1) 35 35
7) First, we form the table showing needed calculations.
Rank of x Rank of y
x y d = Rx-Ry d2
(Rx) (Ry)
70 6.5 90 2 4.5 20.25
70 6.5 90 2 4.5 20.25
80 4 90 2 2 4
80 4 80 4 0 0
80 4 70 5 −1 1
90 2 60 6 −4 16
100 1 50 7 −6 36
2
d
97.5
Here, rank 4 and 6.5 is repeated thrice and twice respectively in rank
of X and rank 2 is repeated thrice in rank of Y, so the correction
factor is
221
Statistical Analysis 3(32 1) 2(2 2 1) 3(32 1)
12 12 12
and therefore, rank correlation coefficient is
3(32 1) 2(2 2 1) 3(32 1)
697.5
12 12 12 6(102)
rs 1 1 0.82
7(49 1) 336
8) Please refer to section 8.6.
9) Please refer to section 8.7.
10) Please refer to section 8.7.1.
11) For getting necessary values for computation purpose, we present the
calculations in the following table:
X1
x1i 15
3 and Y
y i
50
10
n 5 n 5
S. No. X1 Y (X1 X1 ) (X1 X1 )2 (Y Y) (Y Y )2 (X1 X1 )(Y Y)
1 5 9 2 4 –1 1 –2
2 4 8 1 1 –2 4 –2
3 3 10 0 0 0 0 0
4 2 11 –1 1 1 1 –1
5 1 12 –2 4 2 4 –4
Total 15 50 0 10 0 10 –9
(i) Regression coefficient of X1 on Y,
b X1Y
(x X )(y Y)
1i 1 i
9
0.9
2
(y Y)
i 10
Regression coefficient of Y on X1,
b YX1
(x X )(y Y) 9 0.9 .
1i 1 i
2
(x X )
1i 10 1
(ii) r b YX1 b X1Y (0.9) (0.9) 0.90
Note: We are taking “-“sign because correlation coefficient and
regression coefficients have same sign.
222
(i) To find regression lines we need Y , X1 b YX and b X Y . From the
1 1
Descriptive Statistics-II
calculation in table above we have
Regression line of Y on X1 is (Y 10) 0.90(X1 3) and
Regression line of X1 on Y is (X1 3) 0.90 (Y 10)
(i) To estimate Y we use regression line of Y on X1 which is
(Y 10) 0.9 (X1 3)
putting X1 =4.5 we get (Y 10) 0.9 (4.5 3) Y 8.65
223
Statistical Analysis
UNIT 9 SAMPLING DISTRIBUTIONS
Structure
9.1 Introduction
9.2 Objectives
9.3 Basics of Sampling
9.4 Sampling Distribution
9.4.1 Standard Error
9.4.2 Central Limit Theorem
9.5 Sampling Distribution of Statistics
9.5.1 Sampling Distribution of Mean
9.5.2 Sampling Distribution of Diference in Two Means
9.5.3 Sampling Distribution of Proportion
9.5.4 Sampling Distribution of Diference in Two Proportions
9.6 Exact Sampling Distributions
9.6.1 Chi-square Distribution
9.6.2 Student’s t-Distribution
9.6.3 F-Distribution
9.7 Let Us Sum Up
9.8 Key Words
9.9 Suggested Further Reading/References
9.10 Answers to Check Your Progress
9.1 INTRODUCTION
In general, extracting information from all the elements or items of a large
population may be time consuming and expensive if the population size is
infinitely large. But even then sometimes there are so many problems
attached to large size populations where it becomes necessary to draw
inference about population parameters.For example, one may wish to
estimate the average height of all the two thousand students in a college, a
businessman may be interested to estimate the proportion of defective items
in a production line, a manufacturer of car tyres may want to estimate the
variations took place in the diameter of produced tyres, a pharmacist may
want to estimate the difference of effect of two types of drugs, etc.In all such
cases, there is always an unknown population involved whose characteristics
are described through some parameters.
Due to large sizes of populations, in which we may be interested, for
drawing inferences about the population parameters, generally we draw a
224 sample and determine a function of sample values, which is called statistic.
Selection of a sample, that is, only a part of the population saves a lot of time, Sampling Distributions
money and labour and results drawn on sample values are quicklyavailable
forinterpretation and as good sometimes as obtained on the basis of the entire
population. The process of generalising sample results to the population is
called Statistical Inference. Sincethere might be a large number of samples of
same size drawn from the population, the values of statistic generally vary
from sample to sample and is associated with the probability of selection of
the particular sample. Therefore, the sample statistic is a random variable
following any probability distribution. Therefore, it may be a matter of
interest for a statistician to know what distribution the statistic follows if the
samples are assumed to be selected from a theoretical distribution like, a
normal distribution with given mean and variance; a binomial distribution; a
gamma distribution; a Poisson distribution and so on. In contrast to
theoretical distributions, probability distribution of a statistic in popularly
called a sampling distribution. In this unit we shall discuss the sampling
distribution of sample mean; of sample median; of sample proportion; of
difference between two sample means and sample proportions.
Due to this curiosity, Prof. R.A. Fisher, Prof. G. Snedecor and some other
statisticians worked in this area and obtained exact sampling distributions
which are followed by some of the important statistics. In present unit of this
block, we shall also discuss some important sampling distributions such as 2
(read as chi-square), t and F. Generally, these sampling distributions are
named on the name of the originator, for instance, Fisher’s F-distribution is
named like this after the name of its inventor Prof. R.A. Fisher.
9.2 OBJECTIVES
After studying this unit, you should be able to:
explain the concept of sampling and sampling distribution;
explain the concept of standard error and Central Limit Theorem;
define the sampling distribution; and
describe the sampling distribution of sample mean and difference of
two sample means.
9.3 BASICS OF SAMPLING
Before discussing sampling distributions of different kinds of statistic, in this
section we shall be discussing basic concepts and definitions of some of the
important terms which areverymuch helpful in understanding the
fundamentals of statistical inference and are frequently used in this unit.
Population, Sample and Sampling
Literally, “population” means a well-defined group (collection or bunch) of
some objects (units or elements). Examples of populations and its units are : a
city with some clear-cut territory having its dwellers as units (elements) or
225
Statistical Analysis having houses as its units; a hospital with indoor patients as units or its
doctors as units; a library with its employees as units or books as its units; a
river with its fishes as units; a school with enrolled students as units; a
banyan tree with leaves as units; sky with stars as units and so on. Thus,
population is the collection or group of individuals /items /units/ observations
under study. The total number of elements / items / units / observations in a
population is known as its size and generally denoted by N.
A finite subset of units of a population is called a “sample” and the number
of units belonging to the sample is called the sample size. If n denotes the
sample size, then necessarily we have the condition n < N and then the
sample is said to be a proper sample. However, in some situations, we have n
= N, that is, all the units of the population are selected as sample. As
discussed above, in almost all kinds of statistical studies, a sample is selected
for inferential purposes because of the reasons stated above. However, even
if the population is not large enough or it has a well-defined territory or it is
not of destructive type; a sample is useful since besides it is always less
costly, less time consuming, it helps us to get a handy result in very short
time.
A sample of size n is not selected simultaneously, rather n units are selected
one by one in n draws in order to constitute a sample. The process through
which units of the population are selected in different draws to constitute a
sample of specified size, is known as “sampling”. Thus, sampling is a draw-
to-draw mechanism of selecting some units from a population. Sampling is
mainly of two types, namely,
(a) Probability Sampling or Random Sampling, and
(b) Non-Probability Sampling.
The technique of random sampling is of fundamental importance in the
application of statistics. The Estimation theory is based on the assumption of
random sampling. It is a scientific method of selecting samples accordingly
to some laws of chance in which each unit in the population has some
definite pre assigned probability of being selected in the sample.Simple
Random Sampling, Stratified Sampling, Systematic Random sampling, etc.,
are the examples of random sampling.
In non-probability sampling, the sample is selected with definite purpose
and, hence, the choice of sampling units depends entirely on the discretion
and judgment of the investigator. It is, therefore, not a scientific method and
attracts severe criticisms in many circumstances. Purposive sampling,
Judgement Sampling, etc., are the examples of non-probability sampling.
We know that if n < N, it is a sampling procedure. When n = N the process is
known as “complete enumeration or census”. An example of complete
enumeration is the decennial census of India.
226
Need for Sampling in Statistical Inference Sampling Distributions
As we have discussed in Section 9.1, in many practical and real situations the
population under consideration is either infinitely large in size or it is
unbounded in the sense that its boundaries are not well-defined or it is of
destructive nature. As for example, in ordert to determine average life of two
hundred produced electric bulbs, it would be necessary to light these bulbs
unless they all get fused, resulting destruction of the entire lot of production.
Similarly, in order to estimate the proportion of smokers in a large city,
information have to be gathered from each and every person dwelling in the
city which would be a very difficult task in terms of manpower, money and
time to be required. Sometimes, the population size is not known if its
boundaries are not well-defined, like number of fish in a pond or lake. In all
such cases, it is not feasible to gather information on all the units, rather it
becomes almost impossible. Keeping in view the difficulty in contacting each
and every unit of the population due to these reasons and also to save the
time, money and manpower to be required for this, generally a part of the
population, popularly known as a “sample” is selected in some pre-asigned
manner which in turn is used for drawing the inferences on the population
itself or on the parameters of the population. The results obtained from the
sample are projected in such a way that they are valid for concluding about
the entire population. Therefore, the sample works like a “Vehicle” to reach
(drawing) at valid conclusions about the population. In fact, a sample helps
us to reach to the “whole”(population) from a “part” (sample) in all types of
statistical studies. Thus, statistical inference can be defined as:
“It is a process of concluding (projecting or infering) something desired about
a given population on the basis of sample results”.
Parameter and Statistic
A parameter is a function of population values which is used to represent
certain characteristics of the population.For example, population total,
population mean, population variance, population coefficient of variation,
population proportion, population correlation coefficient, etc., are all
parameters since their calculation involve all the population values.
A statistic is a function of sample values only and does not contain any
unknown population parameter in it. For example, if X1 ,X2 , ..., Xn represent
values of the variable X in a random sample of size n taken from a population
1 n
then sample mean X X i is a statistic.
n i 1
Estimator and Estimate
A statistic which is used to guessor to estimate ( infer) an unknown
population parameter then it is known as estimator and the value of the
estimator based on observed value of the sample is known as estimate of
227
Statistical Analysis parameter. For example, suppose the parameter λ of the Poisson
1 n
populationf(x, λ) is unknown. If we use sample mean X Xi ,
n i 1
calculated on the basis of sample values X1 ,X2 , ..., Xn to estimate λ then X
is an estimator and any particular value of x is called estimateof parameter λ.
9.4 SAMPLING DISTRIBUTION
As discussed under Section 9.1, a statistic is calculated on the basis of a
sample of specific size, selected from the given population for drawing the
inference about the population. But due to the fact that sample size is always
smaller than the population size; (n < N), a number of samples of same size
can theoretically be drawn from the population (theoretically nCN possible
samples). Thus, in fact, we have in all nCN values of the statistic. Moreover,
since in probability sampling, each sample is selected with some pre-defined
probability of selection, it means that all these statistic values have some
probability to appear. In this sense, we can say that a statistic ̂ is a random
k k
variable defined as ˆ , p(ˆ ) , where ̂ and p(ˆ ) are respectively the value
k k
of the statistic ̂ k and the probability of its appearance due to the kth
sample; k = 1, 2,…, nCN.
Generally, in practice onlya single random sample is taken from a given
population and its mean X is considered to be representative of the
population mean µ. This sample mean may or may not represent the
population mean. Since we cannot determine the proximity of sample mean
and population mean on the basis of single random sample, we can use the
concept of sampling distribution to bring the value of sample mean close to
that of Population mean. Being a random variable, the statistic must possess a
probability distribution which may be used to answer some questions
regarding the nature of the statistic, for example, what is the probability that
its value as obtained from the given sample is differing from the parameter
value by amargin of 10? The probability distribution of a statistic is called
sampling distribution of the statistic. Let us illustrate how the sampling
distribution of a statistic can be obtained given a population. For instance, let
we wish to estimate the population mean using the sample mean as an
estimator.
Suppose that a baby-sitter has 5 children under her supervision. The ages of
children are 2, 4, 6, 8 and 10 years. Supposing this group of children as a
population of size 5, we get the population mean as
1 N 2 4 6 8 10
X
N i 1
Xi
5
6
228
Therefore, the variance of this population is given by: Sampling Distributions
2 2 2
1 N
X i X
2 2 2 6 4 6 ....... 10 6 8
N i 1 5
Now, let us take all the possible simple random samples of size 2 without
replacement from this population. There are 5C2 = 10 such possible samples
which are listed below along with their respective means:
Table 9.1: Possible samples
Sample
Sample Sample Mean
No.
1 2, 4 3
2 2, 6 4
3 2, 8 5
4 2, 10 6
5 4, 6 5
6 4, 8 6
7 4, 10 7
8 6, 8 7
9 6, 10 8
10 8, 10 9
Now, suppose that our selected sample is either (2, 4) with mean 3 or it is the
sample (8, 10) with mean 9. In both cases, the sample is not a good
representative of the population since sample mean is far from the population
mean 6. But if the selected samples are coincidently either (2, 10) or (4, 8),
then these are good representatives of the population in the sense that both
have sample means exactly equal to population mean. Thus, this example
illustrates that a single random sample may or may not be representative for
the decision maker to reach meaningful conclusion. However, the grand
mean of the distribution of these ten sample means, is calculated and
observed to be equal to the population mean as follows:
1 10 3 45 65 6 7 89
x
10 i 1
xi
10
6.
Hence, the mean of the sample means can be considered to represent the
population mean for analysis and decition making purposes.
Now let us put sample means along with their probabilities of occurrence as
follows:
229
Statistical Analysis Table 9.2: Probability Distribution
Probability
Sample
Frequency of
Mean
Occurrence
3 1 0.1
4 1 0.1
5 2 0.2
6 2 0.2
7 2 0.2
8 1 0.1
9 1 0.1
Total 10 1.0
This distribution which shows the distribution of probabilities over all the
possible values of the sample mean is refered to as sampling distribution of
the sample mean. Symbolically, it can be denoted as {( x , p( x )}. As
explained above, the sampling distribution of a statistic can be defined as:
“The probability distribution of all possible values of a statistic that would be
obtainedby drawing all possible samples of the same size from the population
is called sampling distribution of that statistic.”
The mean, variance and other measures of a sampling distribution can be
obtained in a similar way as computed in a frequency distribution, taking
probabilities as frequencies. Thus, mean of the above sampling distribution
will be
7
x x i p x i 3 x 0 . 1 4 x 0 . 1 5 x 0 . 2 6 x 0 . 2 7 x 0 . 1 8 x 0 . 1 9 x 0 . 1 6
i 1
This value is same as the population meanµ. The variance of the distribution
is obtained as:
2 2 2
1 N
x i x
2 2 3 6 4 6 ....... 9 6 28 4
x
7 i 1 7 7
9.4.1 Standard Error
The expression obtained above, obviously measures the variability of the
sample mean around the actual mean, that is, how much samplestatistic may
vary from sample to sample.
This is equivalent to population variance which measures the deviation of the
population values around the population mean. If we consider the positive
230 square root of this , it would be equivalent to standard deviation. In order to
differentiate with population standard deviation, it is called as “standard Sampling Distributions
error” (SE) of that statistic. Thus, standard error ofa statistic can be defined
as:
“The standard deviation of a sampling distribution is known as standard
error”.
The computation of the standard error is a tedious process. However, the
formula for calculating standard error of sample mean on the basis of a
sample size n is seen always to be
SE X ,
n
where σstands for the standard deviation of the population.
Standard errors of some otherwell known statistics are given below:
1. The SE of sample proportion (p)is given by
PQ
SE(p) =
n
where, P is population proportion, n is sample size and Q = 1-P.
2. The SE of difference of two sample means is given by
σ12 σ 22
SE X Y =
n1 n 2
where, σ12 and σ 22 are the population variances of two different populations
and n1 and n2 are the sample sizes of two independent samples selected from
the two populations respectively.
3. The SE of difference between two sample proportions is given by
P1Q1 PQ
SE( p1 p 2 ) = 2 2
n1 n2
where, P1 and P2 are population proportions in two different populations from
which samples of sizes n1 and n2 are taken respectively; Q1= 1-P1 and Q2 = 1-
P2.
From all the above formulae, we can understand that standard error is
inversely proportional to the sample size. Therefore, as sample size increases
the standard error decreases.
The standard error is used to express the accuracy or precision of the estimate
of population parameter because the reciprocal of the standard error is the
measure of reliability or precision of the statistic.Standard error also
determines the probable limits or confidence limitswithin which the
231
Statistical Analysis population parameter may be expected to lie with certain level of
confidence.Standard error is also applicable in testing of hypothesis.
9.4.2 Central Limit Theorem
The central limit theorem is the most important theorem of Statistics. It was
first introduced by De Movers in the early eighteenth century. The theorem
states that regardless of the nature of the distribution of the population, the
distribution of the sample mean approaches the Normal Probability
distribution as the sample size increases. In general, the larger the sample
size, the closer the proximity of the distribution of sample mean to the
Normal distribution. However, in practice, sample sizes of 30 or larger are
considered adequate for this purpose. It should be noted, however, that the
sampling distribution of sample mean would be always normally distributed
if the original population is normally distributed.
Therefore, according to the central limit theorem, if X1 ,X2 , ..., Xn is a
random sample of size n taken from a population with mean µ and variance
σ2 thenthe sampling distribution of the sample mean tends to normal
distribution with mean µ and variance σ2/n as sample size tends to large
(n 30) whatever be the form of parent population, that is,
2
X ~ N ,
n
and the variate
X
Z ~ N 0,1
/ n
follows normal distribution with mean 0 and variance unity, that is, the
variate Z follows standard normal distribution.
9.5 SAMPLING DISTRIBUTION OF STATISTICS
Sample mean is one of the most commonly used statistic in any of the
statistical studies in order to study the nature and charactristics of a given
population. Examples are average income in a locality in a social survey,
average life of produced items in a manufacturing process, average
temperature in a day in meterological survey, etc. Oweing to these reasons, it
is important to obtain the sampling distribution of sample mean.
9.5.1 Sampling Distribution of Sample Mean
In the previous section we have elaborated the meaning of sampling
distribution of the sample mean and presented an example how it can be
obtained. On the basis of these, the sampling distribution of sample mean can
be defined as follows: “The probability distribution of sample mean that
232 would be obtainedby drawing all possible samples of same size from the
population is called sampling distribution of sample mean or simply of Sampling Distributions
mean.”
If X1 , X 2 , ..., X n is an independent random sample of size n taken from a
normalpopulation with mean µ and variance σ2 then it has been established
that sampling distribution of sample mean X is also normal. The mean and
variance of sampling distribution of X can be obtained as
X X2 ... X n
Mean of X E X E 1 By defination of X
n
1
[E(X1 ) E(X 2 ) ... E(X n )]
n
Since E(Xi) = mean of Xi for all i=1, 2,…,n; therefore, we have E(Xi) = µ.
Similarly, Var(Xi) = σ2 for all i. We, therefore, have
1
E X ...... (n times) n
n n
and variance
1
Var(X) Var (X1 X 2 ... X n )
n
1
Var(X1 ) Var(X 2 ) ... Var(X n )
n2
1 2 n 2 2
n
2 2 2
2 ...... (n times) 2
n n
We, therefore, conclude that If X i ~ N , 2 then
2
X ~ N ,
n
and
SE X SD X Var X
n
Let us illustrate these results using some examples.
Example 1: Diameter of a steel ball bearing produced on a semi-automatic
machine is known to be distributed normally with mean 12 cm and standard
deviation 0.1 cm. If we take a random sample of 10 ball bearings then find
mean and variance of sampling distribution of mean.
Solution: Here, we are given that
= 12, σ = 0.1, n = 10
Since the sample is taken from the population of ball bearings in which
diameter follows normal distribution N(12, 0.01), we have 233
Statistical Analysis E X 12
2 (0.1) 2
Var X 0.001 .
n 10
Thus, X ~N(12, 0.001).
Example 2: If ages of 5 employees of Account Section of a Company are 58,
66, 64, 62 and 50 years then construct the sampling distribution of average
age of employees by taking all samples of size 2 with replacement.
Solution:Here, we are given that N= 5,n = 2
Hence the population mean:
54 56 60 64 66
X 60
5
Therefore, the variance of this population is given by:
2
54 602 56 602 ....... 66 60 2
104
20.08
5 5
Now, all possible samples (without replacement) are 10 which are shown in
the Table given below:
Table 9.3: Possible Samples
Sample Sample Sample
Number Observation Mean
1 54, 56 55
2 54, 60 57
3 54, 64 59
4 54, 66 60
5 56, 60 58
6 56, 64 60
7 56, 66 61
8 60, 64 62
9 60, 66 63
10 64, 66 65
We, therefore, have the grand mean of the distribution as
1 10 55 57 58 ............... 62 63 65
x
10 i 1
xi
10
60
which is same as the population mean. Now, let us construct sampling
distribution of the sample mean as given under:
234
Table 9.4: Probability Distribution Sampling Distributions
Probability
Sample
Frequency of
Mean
Occurrence
55 1 0.1
57 1 0.1
58 1 0.1
59 1 0.1
60 2 0.2
61 1 0.1
62 1 0.1
63 1 0.1
65 1 0.1
Total 10 1.0
Therefore, the mean of sample means can be obtained by the formula:
7
x x i px i 55x 0.1 (57 x 0.1) 58x 0.1 59 x 0.1 60 x 0.2
i 1
61x 0.1 62 x 0.1 63x 0.1 65x 0.1 60
This value is same as the population meanµ. The variance of the distribution
is obtained as:
2 2 2
1 N
x i x
2 2 55 60 57 60 ....... 65 60 78 8.66
x
9 i 1 9 9
CHECK YOUR PROGRESS 1
Note: i) Check your answers with those given at the end of the unit.
1) If lives of 5 televisions of certain company are 4, 6, 8, 10 and 12 years
then construct the sampling distribution of average life of televisions by
taking all samples of size 2.
2) The weight of certain type of a truck tyre is known to be distributed
normally with mean 200 pounds and standard deviation 4 pounds. A
random sample of 10 tyres is selected. What is the sampling distribution
of sample mean? Also obtain the mean and variance of this distribution.
235
Statistical Analysis 3) The mean life of CFL blubs produced by a company is 2550 hours. A
random sample of 100 CFL bulbs is selected and the standard deviation
is found to be 54 hours. Find the mean and variance of the sampling
distribution of mean.
9.5.2 Sampling Distribution of Difference Between Two
Sample Means
Instead of sample mean of a single population, sometimes one may be
interested into two populations and, hence, in finding the sampling
distribution of the difference of sample means which are obtained on the
basis of the samples drawn from two populations. Such cases arise when we
wish to compare average lives of elecrtric bulbs of two kinds or to compare
the efficiency of two different drugs of same disease or to compare the
average marks of two different sections in a school, etc.
Let the same characteristic measured on two populations, say population-I
and population-II be represented by X and Y variables. Suppose population-I
have mean 1 and variance 12 whereas population-II have mean 2 and
variance 22 . Then,let X be the sample mean based on a sample of size n1
selected from population-I and Y be the sample mean based on a sample of
size n2 selected from the population-II. As before, it is obvious that, if
necessary, one may select all possible samples of sizes n1 and n2 respectively
from the two populations. Then considering all possible differences of means
and sampling distribution of difference of population means can be obtained.
The sampling distribution of difference of sample means can be defined as:
“The probability distribution of all values of the difference of two sample
means would be obtained by drawing all possible samples from both the
populations. Such a distribution is called sampling distribution of difference
of two sample means.”
If both the parent populations are normal, that is,
X ~ N 1 , 12 and Y ~ N 2 , 22
then as we discussed in previous section
2 2
X ~ N 1 , 1 and Y ~ N 2 , 2
n1 n2
If two independent random variables X and Y are normally distributed then
the difference (X Y) also normally distributed. Therefore, the sampling
distribution of difference of two sample means (X Y) also follows normal
distribution with mean
E X Y E X E Y =µ1− µ2
236
and variance Sampling Distributions
12 22
Var X Y Var X Var Y
n1 n 2
Therefore, standard error of difference of two sample means is given by
12 22
SE X Y Var X Y .
n1 n 2
Let us see an application of the sampling distribution of difference of two
sample means with the help of an example.
Example 3: LED Bulbs manufactured by company A have mean lifetime of
2400 hours with standard deviation 200 hours, while LED Bulbs
manufactured by company B have mean lifetime of 2200 hours with standard
deviation of 100 hours. If random samples of 125 LED Bulbs of each
company are tested, find. The mean and standard error of the sampling
distribution of the difference of mean lifetime of LED Bulbs.
Solution: Here, we are given that
1 = 2400, σ1 = 200, 2 = 2200, σ2 = 100 and n1 = n2 = 125
Let X and Y denote the mean lifetime of CFLs taken from companies A and
B respectively. Since n1 and n2 are large (n1, n2 > 30) therefore, by central
limit theorem, the sampling distribution of (X Y) follows normal
distribution with mean
E(X Y) 1 2 2400 2200 200
and variance
2 2
12 22 200 100
Var(X Y) 400
n1 n 2 125 125
Therefore, the standard error is given by
SE X Y Var X Y 400 20 .
Now, continuing our discussion, the sampling distribution of difference of
two means, we consider another situation.
If population variances 12 and 22 are unknown then we estimate 12 and 22
by the values of the sample variances of the samples taken from the first and
second population respectively. For large sample sizes n1 and n2 30 , the
sampling distribution of (X Y) is very closely normally distributed with
s2 s2
mean (µ1− µ2) and variance 1 2 .
n1 n 2 237
Statistical Analysis If population variances 12 and 22 are unknown and 12 22 2 then σ2 is
estimated by pooled sample variance s 2p where,
1
s 2p
n1 n 2 2
n 1s12 n 2 s 22
and variate
t
X Y 2
1
~ t ( n1 n 2 2 )
1 1
sp
n1 n 2
follows t-distribution with (n1 + n2 − 2) degrees of freedom. Similar to
sampling distribution of mean, for large sample sizes n1 and n2 30 the
sampling distribution of (X Y) is very closely distributed as normal with
1 1
mean (µ1− µ2) and variance s 2p .
n1 n 2
CHECK YOUR PROGRESS 2
Note: i) Check your answers with those given at the end of the unit.
4) The average height of Male workers in a hospital is found to be 68
inches with a standard deviation of 2.3 inches whereas the average
height of Female workers in a hospital is found to be 65 inches with a
standard deviation of 2.5 inches. If a sample of 35 Male and 50 Female
mean and standard error of the sampling distribution of the difference
between workers are selected at random, find is the the sample means
of height of Male workers and female workers.
9.5.3 Sampling Distribution of Sample Proportion
In Section 9.5.1, we have discussed the sampling distribution of sample mean
if some quantitative characteristic is taken into consideration..But in many
real word situations,when some qualitatrive characteristic is under
consideration, sample proportion, instead of sample mean is computed on the
basis of a sample and hence, sampling distribution of sample proportion is
needed. Examples of cases in which sample proportions are needed for
analysis are proportion of male births; proportion of defective items in
manufacturing process; proportion of cancer cases in population, etc.
For sampling distribution of sample proportion, we need sample proportion p.
Let a sample of size n is taken from the population, then p is given by
X
p 1
n
where X is the number of observations /individuals / items / units in the
sample which have the particular characteristic under study. For better
238
understanding of the process, we consider the following example in which Sampling Distributions
size of the population is very small:
Suppose, there is a lot of 4 cartons A, B C and D of electric Tubes and each
carton contains 20 Electric Tubes. The number of defective bulbs in each
carton is given below:
Table 9.4: Number of Defective Bulbs per Carton
Carton Number of
Defectives Bulbs
A 2
B 4
C 1
D 3
The population proportion of defective tubes can be obtained as
2 4 1 3 10 1
P
20 20 20 80 8
Now, let us assume that we do not know the population proportion of
defective tubes. So we decide to estimate population proportion of defective
tubes on the basis of samples of size n = 2. There are N C n 4 C 2 6 possible
samples of size 2 with replacement. All possible samples and their respective
proportion defectives are given in the following table:
Table 9.6: Calculation of Sample Proportion
Sample Sample Sample Sample
Carton Proportion(p)
Observation
1 (A, B) (2, 4) 6/40
2 (A, C) (2, 1) 3/40
3 (A,D) (2, 3) 5/40
4 (B, C) (4, 1) 5/40
5 (B, D) (4, 3) 7/40
6 (C, D) (1, 3) 4/40
From the above table, we can see that value of the sample proportion is
varying from sample to sample. So we consider all possible sample
proportions and calculate their probability of occurrence. Since there are 6
possible samples therefore the probability of selecting a sample is 1/8. Then
we arrange the possible sample proportions with their respective probability
in Table 2.7:
239
Statistical Analysis Table 9.7: Sampling Distribution of Sample Proportion
S.No. Sample Frequency Probability
Proportion(p)
1 3/40 1 1/6
2 4/40 1 2/6
3 5/40 2 2/6
4 6/40 1 1/6
5 7/40 1 1/6
Total 1.00
This distribution is called the sampling distribution of sample proportion.
The mean of sampling distribution of sample proportion can be obtained as
1 k k
p ii
K i 1
p f where, K
i 1
fi
1 3 4 5 6 7 30 1 1
1 1 2 1 ... 1
6 40 40 40 40 40 40 6 8
Thus, we have seen that mean of sample proportion is equal to the population
proportion.
If a population whose elements are divided into two mutually exclusive
groups− one containing the elements which possess a certain attribute and
other containing elements which do not possess the attribute, then number of
successes (elements possess a certain attribute) follows a binomial
distribution with mean
E(X) nP
and variance
Var(X) nPQ where Q 1 P
where, P is the probability or proportion of success in the population.
Now, we can easily find the mean and variance of the sampling distribution
of sample proportion by using the above expression as
X 1 1
E(p) E E(X) nP P
n n n
and variance
X 1 Var aX a 2Var X
Var(p) Var 2 Var(X)
n n
1 PQ
nPQ . Var X nPQ
n 2
n
240
Also standard error of sample proportion can be obtained as Sampling Distributions
PQ
SEp Var( p)
n
If the sampling is done without replacement from a finite population then the
mean and variance of sample proportion is given by
E p P
and variance
N n PQ
Var p
N 1 n
where, N is the population size and the factor (N-n) / (N-1) is called finite
population correction.
If sample size is sufficiently large, such that np > 5 and nq > 5 then by central
limit theorem, the sampling distribution of sample proportion p is
approximately normally distributed with mean P and variance PQ/n where,
Q= 1 P.
Let us see an application of the sampling distribution proportion with the help
of an example.
Example 4: A machine produces a large number of items of which 15% are
found to be defective. If a random sample of 200 items is taken from the
population and sample proportion is calculated then find mean and standard
error of sampling distribution of proportion.
Solution: Here, we are given that
15
P= = 0.15, n = 200
100
We know that when sample size is sufficiently large, such that np > 5 and nq
> 5 then sample proportion p is approximately normally distributed with
mean P and variance PQ/n where, Q = 1– P. But here the sample proportion
is not given so we assume that the conditions of normality hold, that is, np >
5 and nq > 5. So mean of sampling distribution of sample proportion is given
by
E ( p ) P 0.15
and variance
PQ 0.15 0.85
Var(p) 0.0006
n 200
Therefore, the standard error is given by
SE p Var p 0.0006 0.025
241
Statistical Analysis CHECK YOUR PROGRESS 3
Note: i) Check your answers with those given at the end of the unit.
5) A state introduced a policy to give loan to unemployed doctors to start
own clinic. Out of 10000 unemployed doctors 7000 accept the policy
and got the loan. A sample of 100 unemployed doctors is taken at the
time of allotment of loan. Find the mean and standard error of the
sampling distribution of proportion of doctors who accepted the policy
and got the loan.
9.5.4 Sampling Distribution of Difference of Two Sample
Proportions
Just like the sampling distribution of difference of two population means,
which has been obtained in sub-Section 9.5.2, the sampling distribution of
difference of two population proportions can be obtained which is required
many times in statistical inference. We shall show how it can be obtained.
Suppose, there are two populations, say, population-I and population-II under
study and the proportions of some attribute in populations I and II are
respectively P1 and P2. For finding sampling distribution of difference of two
sample proportions, let samples of sizes n1 and n2 be selected respectively
from populations I and II and the sample proportions of the attribute in these
samples are p1 and p2 respectively. Since sample proportions obviously vary
from sample to sample, their difference will be a random variable having
some probability distribution, which would be termed as the sampling
distribution of difference of two sample proportions. On the basis of
arguments made for one sampling proportion in the previous section, for the
distribution of sampling proportion p due to central limit theorem, here also
we can assume that if n1p1 5, n1q1 5, n2p2 5 and n2q2 5 then
PQ PQ
p1 ~ N P1, 1 1 and p2 ~ N P2 , 2 2
n1 n2
where, Q1 = 1 P1 and Q2 = 1 P2.
Also, by the property of normal distribution, the sampling distribution of the
difference of sample proportions follows normal distribution with mean
E(p1-p2) = E(p1)-E(p2) = P1-P2
and variance
P1Q1 P2Q 2
Var(p1-p2) = Var(p1)+Var(p2)
n1 n2
That is,
PQ P Q
p1 p 2 ~ N P1 P2 , 1 1 2 2
n1 n2
242
Sampling Distributions
Thus, standard error is given by
P1Q1 P2Q2
SE p1 p2 Var p1 p2
n1 n2
Thus, p1-p2 follows a normal distribution with above mentioned mean,
variance and standard error. Let us see an application of the sampling
distribution of difference of two sample proportions with the help of an
example.
Example 5: In one population, 30% persons had hair colour Black and in
second population 20% had the same hair colour. A random sample of 200
persons is taken from each population independently and calculate the
sample proportion for both samples, then find the mean and variance of the
sampling distribution of the difference between two sample proportions.
Solution: Here, we are given that
P1 = 0.30, P2 = 0.20, n1= n2 = 200
Let p1 and p2be the sample proportions of blue-eye persons in the samples
taken from both the populations respectively. We know that when n1 and n2
are sufficiently large, such that n1p1 5, n1q1 5, n2p2 5 and n2q2 5 then
sampling distribution of the difference between two sample proportions is
approximately normally distributed. But here the sample proportions are not
given so we assume that the conditions of normality hold. So mean of
sampling distribution of the difference between two sample proportions is
given by
E p1 p 2 P1 P2 0.30 0.20 0.10
and variance
P1Q1 P2 Q 2 0.30 0.70 0.20 0.80
Varp1 p 2 0.0019
n1 n2 200 200
Thus, standard error
Standard Error = (p1 p 2 ) Var(p1 p 2 ) 0.0019 0.04
CHECK YOUR PROGRESS 4
Note: i) Check your answers with those given at the end of the unit.
6) In city A, 25% persons were found to be smokers and in another city B,
20% persons were found smokers. If 250 persons of city A and 200
persons of city B are selected randomly, then find the mean and
standard deviation of sampling distribution of the difference in sample
proportions.
243
Statistical Analysis
9.6 EXACT SAMPLING DISTRIBUTION
As we have discussed in Section 9.1, some of the well known statisticians,
like Prof. R. A. Fisher, Prof. G. Snedecor, etc. worked on finding some of the
statistic and determined the exact sampling distributions their properties and
applications in differenct areas. These sampling distributions are named on
the nmame of its originator for example, F- distribution is named as Fisher’s
F-distribution and t-distribution as student’s t-distribution on the name of
Prof. W.S. Gosset. Before describing the Exact Sampling distribution first we
will discuss the term “Degree of Freedom” a very useful concept which is
necessarily to be understood before discussing and understanding the
concepts of Exact sampling distributions. The exact sampling distributions
are described with the help of degrees of freedom.
Degrees of Freedom (df)
The term degree of freedom (df) is related to the independency of sample
observations. In general, the number of degree of freedom is the total number
of observations minus the number of independent constraints or restrictions
imposed on the observations. For example, let x1, x2,…, xn be n independent
observations in a sample. Unless some condition is imposed on these x
values, it would have n df. Now, let one condition x1+x2+…+ xn = 100 be
imposed on this set, then it looses 1 df, that is, now df will be n – 1 since the
last value xn or any other value xi will be dependent on all other remaing
values and therefore, the number of independent values will be n-1. Further,
let x12+x22+…+ xn2 = 4000 be the another condition imposed, then now the df
will be n – 2, etc.
For a sample of n observations, if there are k restrictions among observations
(k < n), then the degrees of freedom will be (n–k).
9.6.1 Chi-square Distribution
The chi-square distribution was first discovered by Helmert in 1876 and later
independently explained by Karl- Pearson in 1900. The chi-square
distribution was discovered mainly as a measure of goodness of fit of any
model on the given frequency distribution or probability distribution. .
If a random sample X1, X2,…, Xn of size n is drawn from a normal
population having mean and variance σ2 then the sample variance can be
defined as
1 n n
s2 ( x i x ) 2 or (x i x ) 2 (n 1) s 2 s 2
n 1 i1 i 1
where, ν = n −1; the symbol ν read as ‘nu’.
s 2
Then, the variate 2 which is the ratio of sample variance multiplied
2
by its degrees of freedom and the population variance follows the 2-
244 distribution with ν degrees of freedom.
The probability density functionof 2-distribution with ν df is given by Sampling Distributions
1 2 / 2 1
f 2 e /2
2
; 0 2 … (1)
2 / 2
2
where, ν = n −1.
It can be seen that 2-distribution is a sampling distribution of a statistic
s 2
2 , the shape of probability distribution of which is shown by
2
probability curve shown in Fig. 3.1 for n = 1, 4, 10 and 22.
Fig. 9.1: Chi-square probability curves for n = 1, 4, 10 and 22
After looking the probability curves of 2-distribution for n = 1, 4, 10 and 22,
one can understand that probability curve takes shape of inverse J for n = 1.
The probability curve of chi-square distribution gets skewed more and more
to the right as n becomes smaller and smaller. It becomes more and more
symmetrical as n increases because as n tends to ∞, 2-distribution converges
to a normal distribution. It is apparent from the figure that even when n= 22,
it is tending to a symmetrical probability curve which is an essential property
of a normal curve.
Example 8: What are the mean and variance of chi-square distribution with 5
degrees of freedom?
Solution: The mean and variance of chi-square distribution with n degrees of
freedom are given by
Mean = n and Variance = 2n
In our case, n = 5, therefore,
Mean = 5 and Variance = 10.
Some of the salient features of 2-distribution are mentioned here without
giving proofs of each:
245
Statistical Analysis 1. The probability curve of the chi-square distribution lies in the first
quadrant because the range of 2 -variate is from 0 to ∞.
2. Chi-square distribution has only one parameter n, that is, the degrees of
freedom.
3. Chi-square probability curve is highly positive skewed for smaller values
of n but becomes a symmetrical curve for larger values of n.
4. Chi-square-distribution is a uni-modal distribution, that is, it has single
mode.
5. The mean and variance of chi-square distribution with n df are n and 2n
respectively.
The applications of chi-square distribution are very wide in Statistics. The
chi-square distribution is used (i) to test the hypothesis that whether the
population variance is same as the specified value or not in parametric test
procedures, (ii) to test the goodness of fit, that is, to judge whether there is a
discrepancy between theoretical and experimental observations or not and
(iii) to test the independence of two attributes. The applications listed above
shall be discussed in detail in Unit 4 of this block.
CHECK YOUR PROGRESS 5
Note: i) Check your answers with those given at the end of the unit.
7) What are the mean and variance of chi-square distribution with 10
degrees of freedom?
8) What are the mean and variance of chi-square distribution with pdf
given below
1 2 / 2 2 3
f 2 e ; 0 2
96
9) List the applications of chi-square distribution.
9.6.2 Student’s t-Distribution
The t-distribution was discovered by W.S. Gosset in 1908. He was better
known by the pseudonym ‘Student’ and hence t-distribution is called
‘Student’s t-distribution’.
If a random sample X1, X2,…, Xn of size n is drawn from a normal
population having mean and variance σ2 then we know that the sample
mean X is distributed normally with mean and variance 2 / n , that is, if
Xi ~ N , 2 then X ~ N , 2 / n . Then as mentioned under sub-section
2.3.2, the variate
X
Z
246 / n
is distributed normally with mean 0 and variance 1, that is, Z ~ N 0, 1 . Sampling Distributions
In general, the standard deviation σ is not known and in such a situation the
only alternative left is to estimate it from a sample. The value of sample
variance (S2) is used to estimate it where,
2 1 n1
s (x i x) 2
n 1 i1
X
But then in this case the variate is not normally distributed whereas it
S/ n
follows t-distribution with (n−1) df, that is,
X
t ~ t ( n 1) … (2)
s n
The t-variate is a widely used variable and its distribution is called student’s
t-distribution on the pseudonym name ‘Student’ of W.S. Gosset. The
probability density function of variable t with (n-1) = ν degrees of freedom is
given by
1
f t 1 / 2 ; t … (3)
2
1 t
B , 1
2 2
1
where, B , is known as beta function.
2 2
The probability curve of t-distribution is bell shaped and symmetric about t =
0 line. The probability curves of t-distribution is shown in Fig. 9.2 at two
different values of degrees of freedom n = 4 and 12.
Fig. 9.2: Probability curves for t-distribution at n = 4, 12 along with
standard normal curve
In the figure given above, we have drawn the probability curves of t-
distribution at two different values of degrees of freedom along with
probability curve of standard normal distribution. By looking at the figure,
one can easily understand that the probability curve of t-distribution is similar 247
Statistical Analysis in shape to that of normal distribution and asymptotic to the horizontal-axis
whereas it is flatter than standard normal curve. The probability curve of the
t-distribution is tending to the normal curve as the value of n increases.
Therefore, for sufficiently large value of sample size n, practically for(> 30),
the t-distribution tends to the normal distribution.
The mean and variance of the t-distribution with n df are given by
n
Mean = 0 and Variance = n>2
n2
Now, we shall discuss some of the important properties of the t-
distribution.The t-distribution has the following properties:
1. The t-distribution is a uni-modal distribution, that is, t-distribution has
single mode.
n
2. The mean and variance of the t-distribution with n df are zero and
n2
respectively. The variance of the distribution exists only if n > 2.
3. The probability curve of t-distribution is similar in shape to the standard
normal distribution and is symmetric about t = 0 line but flatter than
normal curve.
4. The probability curve is bell shaped and asymptotic to the horizontal
axis.
Above we have discussed some important properties of t-distribution without
mentioning their proof. You may be interested to know the applications of the
t-distribution. The t-distribution has wide number of applications in Statistics.
The t-distribution is used (i) to test the hypothesis about the significance of a
population mean, (ii) the the hypothesis of equality of two population means
of two normal populations and (iii) to test the hypothesis whether the
population correlation coefficient is zero or not.
Example 9: The life of light bulbs manufactured by the company A is known
to be normally distributed. The CEO of the company claims that an average
life time of the light bulbs is 300 days. A researcher randomly selects 25
bulbs for testing the life time and he observed the average life time of the
sampled bulbs is 290 days with standard deviation of 50 days. Calculate
value of t-variate.
Solution: Here, we are given that
300, N 25, X 290 and s 50
The value of t-variate can be calculated by the formula
X
t
s/ n
Therefore, we have
290 300 10
t 1
248 50/ 25 10
CHECK YOUR PROGRESS 6 Sampling Distributions
Note: i) Check your answers with those given at the end of the unit.
10) The scores on an IQ test of the students of a class are assumed to be
normally distributed with a mean of 60. From the class, 15 students are
randomly selected and an IQ test of the similar level is conducted. The
average test score and standard deviation of test scores in the sample
group are found to be 65 and 12 respectively. Calculate the value of t-
statistic.
11) What are the mean and variance of t-distribution with 8 degrees of
freedom?
12) Write any three applications of t-distribution.
9.6.3 F-Distribution
As we have mentioned in previous unit, F-distribution was introduced by
Prof. R. A. Fisher and defined as the ratio of two independent chi-square
variates when divided by their respective degrees of freedom. If we draw a
random sample X1 ,X 2 ,..., X n1 of size n1 from a normal population with mean
1 and variance σ 12 and another independent random sample Y1 , Y2 ,..., Yn2 of
size n2 from another normal population with mean 2 and variance 22
respectively then 1s12 / 12 is distributed as chi-square variate with ν1 df, that
1s12
is, 12 2
~ (21 ) … (1)
1
1 n1 1 n1
where, 1 n 1 1, X X i and s12 (X i X ) 2
n1 i1 n1 1 i1
Similarly, 2s 22 / 22 is distributed as chi-square variate with ν 2 df, that is,
2s 22
22 2
~ (2 2 ) … (2)
1
n2
1 1 n2
where, 2 n 2 1, Y
n2
X i and s 22
i 1
(Yi Y) 2
n 2 1 i1
Now, if we take the ratio of the above chi-square variates given in equations
(1) and (2), then we get
12 1s12 / 12 s12 / 12 12 / 1
~ F( 1 , 2 ) … (3)
22 2s 22 / 22 s 22 / 22 22 / 2
In the above expression F stands for the F-distribution. In the suffix, υ1 and υ2
are called the degrees of freedom of the F-distribution.
Now, if variances of both the populations are equal, that is, σ12 σ 22 , then F-
variate is written in the form of ratio of two sample variances which is as
follows:
249
Statistical Analysis s12
F ~ F( 1 , 2 ) … (4)
s 22
The probability density function of F-variate is given as
ν1 / 2
F
ν1 / 2 1
f(F)
ν1/ν 2 ; 0F … (5)
ν1 ν2 / 2
ν ν
B 1 , 2 1 ν1 F
2 2 ν 2
As shown in (5), F-variate varies from 0 to , therefore, it is always positive
so probability curve of F-distribution wholly lies in the first quadrant. ν1 and
ν 2 are the degrees of freedom and are called the parameters of F-distribution.
Hence, the shape of probability curve of F-distribution depends on ν1 and ν 2 .
Probability curves of F-distribution taking ( 1 , 2 ) as (5, 5), (5, 20) and (20,
5) are shown in Fig.2.3 below:
1 5 , 2 20
1 5 , 2 5
1 20 , 2 5
Fig. 9.3: Probability curves of F-distribution for (5, 5), (5, 20) and (20, 5)
degrees of freedom.
As it appears from the figure, F-distribution is uni-modal curve. It can be
seen from the figure that by increasing the first degrees of freedom from
1 5 to 1 20 the mean of the distribution (shown by vertical line) does
not change but probability curve shifs from the tail to the centre of the
distribution whereas increasing the second degrees of freedom from
2 5 to 2 20 the mean of the distribution (shown by vertical line)
decrease and the probability curve shifts from the tail to the centre of the
distribution. One can also get an idea about the skewness of the F-
distribution. We observe from the probability curve that it is positively
skewed curve and it becomes very highly positive skewed if ν 2 becomes
small. Now, we shall discuss some of the important properties of F-
distribution.
The F-distribution has the following important properties:
250
1. The probability curve of F-distribution is positively skewed curve. The Sampling Distributions
curve becomes highly positive skewed when ν2 is smaller than ν1.
2. F-distribution curve extends on abscissa from 0 to .
3. F-distribution is a uni-modal distribution, that is, it has single mode.
4. The square of t-variate with ν df follows F-distribution with 1 and ν
degrees of freedom.
2
5. The mean of F-distribution with (ν1,ν2) df is for 2 2.
2 2
6. The variance of F-distribution with (ν1,ν2) df is
222 1 2 2
2
for 2 4.
1 2 2 2 4
7. If we interchange the degrees of freedom ν1 and ν2 then there exists a
very useful relation as
1
F 1 , 2 ,1
F 2 , 1 ,
F-distribution has a lot of applications in Statistics. It is used to test the
hypothesis of equality of variances of two normal populations, for the
significance of multiple correlation coefficients, correlation ratio in the
population. It is also used in one-way and two-way analysis of variance for
testing the equality of several means at a time.
Example 10: A statistician selects 7 women randomly from the population of
women, and 12 men from a population of men. The table given below shows
the standard deviation of each sample and population:
Population Sample
Population Standard Standard
Deviation Deviation
Women 40 45
Men 80 75
Compute the value of F-variate.
Solution: Here, we are given that
n 1 7, n 2 12, 1 40, 2 80, s1 45, s 2 75
The value of F-variate can be calculated by the formula given below
s12 / 12
F
s 22 / 22
where, s12 & s 22 are the sample variances.
251
Statistical Analysis Therefore, we have
2 2
F
45 / 40
1.27
1.93
2 2
75 / 80 0.88
For the above calculation, the degrees of freedom ν1 for women’s data
are7−1= 6 and the degrees of freedom ν2 for men’s data are 12 −1 =11.
CHECK YOUR PROGRESS 7
Note: i) Check your answers with those given at the end of the unit.
13) For the purpose of a survey 15 students are selected randomly from
class A and 10 students are selected randomly from class B. At the
stage of the analysis of the sample data, the following information is
available:
Class Population Sample
Standard Standard
Deviation Deviation
Class A 65 60
Class B 45 50
Calculate the value of F-variate.
14) Write any five properties of F-distribution.
15) What are the mean and variance of F-distribution with 1 5 and
2 12 degrees of freedom?
16) Write four applications of F-distribution.
We now end this unit by giving a summary of what we have covered in it.
9.7 LET US SUM UP
In this unit, we have covered the following points:
1. The statistical procedure which is used for drawing conclusions about
the population parameter on the basis of the sample data is called
“statistical inference”.
2. A group of units or items under study is known as “Population”
whereas a part or a fraction of it is known as “sample”.
3. A “parameter” is a function of population values which is used to
represent certain characteristics of the population and any quantity
calculated from sample values does not contain any unknown population
parameter is known as “statistic”.
4. Any statistic used to estimate an unknown population parameter is
known as “estimator” and the particular value of the estimator is known
as “estimate”.
252
5. The probability distribution of any statistic is called “sampling Sampling Distributions
distribution”of that statistic.
6. The standard deviation of the sampling distribution of a statistic is
known as “standard error”.
7. The most important theorem of Statistics is “central limit theorem”
which states that the sampling distribution of the sample mean tends to
normal distribution as sample size n tends to large (generally when n >
30).
8. The sampling distribution of sample mean and difference between two
sample means.
9. The sampling distribution of sample proportion and difference of two
sample proportions.
10. The properties, probability curve and applications of χ2, t and F
distributions, and
11. Mean and variance of χ2, t and F distributions.
9.8 KEY WORDS
Standard Error of the : A rough measure of the average amount by
Mean which sample means deviate from the
population mean.
Sampling Distribution : The probability distribution of means for all
of the Mean possible random samples of a given size from
some population.
9.9 SUGGESTED FURTHER READING/
REFERENCES
Witte, R., & Witte, J. (2017). Statistics. Hoboken, NJ: John Wiley & Sons.
9.10 ANSWERS TO CHECK YOUR PROGRESS
1) Here, we are given that
N=5, n=2
Since we have to estimate the average life of Televisions on the basis of
samples of size n = 2 therefore, all possible samples (with replacement) are
N
C N 5 C 2 10 and for each sample we calculate the sample meanas shown
in Table 2.8 given below:
253
Statistical Analysis Table 9.8: Calculation of Sample Mean
Sample Sample Sample
Number Observatio Mean ( X )
n
1 4, 6 5
2 4, 8 6
3 4,10 7
4 4, 12 8
5 6, 8 7
6 6, 10 8
7 6, 12 9
8 8, 10 9
9 8, 12 10
10 10, 12 11
Since the arrangement of all possible values of sample mean with their
corresponding probabilities is called the sampling distribution of mean,
thus,we arrange every possible value of sample mean with their respective
probabilities in the following Table 9.9 given below:
Table 9.9: Sampling distribution of sample means
S.No Sample Frequency Probability
. Mean ( X )
1 5 1 1/10
2 6 1 1/10
3 7 2 2/10
4 8 2 2/10
5 9 2 2/10
6 10 1 1/10
7 11 1 1/10
Here, we are given that
= 200, σ = 4, n = 10
2) Since parent population is normal so sampling distribution of sample
means is also normal. Therefore, the mean of this
distribution is given by
E X 200
254
and variance Sampling Distributions
2
2 4 16
Var X 1.6
n 10 10
3) Here, we are given that
= 2550, n = 100, s = 54
First of all, we find the sampling distribution of sample mean. Since sample
size is large (n = 100 > 30) therefore, by the central limit theorem, the
sampling distribution of sample mean follows normal distribution. Therefore,
the mean of this distribution is given by
E X 2550
and variance
s 2 (54) 2 2916
Var ( X ) 29.16
n 100 100
4) Here, we are given that
1 = 68, σ1 = 2.3, n1 = 35
2 = 65, σ2 = 2.5, n2 = 50
To find the mean and standard error, first of all we find the sampling
distribution of difference of two sample means. Let X and Y denote the
mean height of male and female workers of hospital, respectively. Since n1
and n2 are large (n1, n2 > 30) therefore, by the central limit theorem, the
sampling distribution of (X Y) follows normal distribution with mean
E X Y 1 2 68 65 3
and variance
2 2
12 22 2.3 2.5
Var X Y
n1 n 2 35 50
0.1511 0.1250 0.2761
Thus standard Error = Var ( X Y )
0.2761 0.525
5) Here, we are given that
X 7000
N =10000, X = 7000 P 0.70 & n = 100
N 10000 255
Statistical Analysis First of all, we find the sampling distribution of sample proportion. Here, the
sample proportion is not given and n is large so we can assume that the
conditions of normality hold. So the sampling distribution is approximately
normally distributed with mean
E (p) P 0.70
and variance
PQ 0.70 0.30
Var(p) 0.0021 Q 1 P
n 100
Thus, standard error is
standard error = (p) Var(p) 0.0021 0.0458
6) Here, we are given that
25 20
P1 0.25, P2 0.20, n 1 250, n 2 200
100 100
Let p1 and p2be the sample proportions of alcohol drinkers in two cities A and
B respectively. Here the sample proportions are not given and n1 and n2 are
large n1 , n 2 30 so we can assume that conditions of normality hold. So the
sampling distribution of difference of proportions is approximately normally
distributed with mean
E(p1 p 2 ) P1 P2 0.25 0.20 0.05
and variance
P1Q1 P2 Q 2 0.25 0.75 0.20 0.80
Varp1 p 2
n1 n2 250 200
0.75 0.80 1.55
0.00155
1000 1000 1000
Hence, Standard Error (p1-p2) =0.0394
7) We know that the mean and variance of chi-square distribution with n
degrees of freedomare
Mean = n and Variance = 2n
In our case, n = 10, therefore,
Mean = 10 and Variance = 20
8) Here, we are given that
1 2 / 2 2 3
f 2 e ; 0 2
96
256
We have the probability density function of 2 distribution as: Sampling Distributions
1 2
f ( 2 ) e /2
( 2 ) ( / 2)1 ; 0 2
2 / 2
2
Where v = n = 1
by comparision we have
1 3 4 8
2 2
Thus, 8 n 1 8 n 9
Mean = n = 9 and Variance = 2n = 18.
9) Refer Sub-Section 9.6.1.
10) Here, we are given that
60, n 15, X 65 and s 12
We know that the t-variate is
X
t
s/ n
Therefore, we have
65 60 5
t 1.14
12 / 15 4.39
11) We know that the mean and variance of t-distribution with n degrees
of freedomare
n
Mean = 0 and Variance ; n2
n 2
In our case, n = 8, therefore,
8 8
Mean = 0 and Variance 0.8
(80 2) 10
12) Refer Sub-Section 9.6.2
13) Here, we are given that
n 1 15, n 2 10, 1 65, 2 45, s1 60, s 2 50
Thevalue of F-variate can be calculated as follows:
s12 / 12
F
s 22 / 22
where, s12 & s 22 are the values of sample variances.
257
Statistical Analysis Therefore, we have
2 2
F
60 / 65
0.85
0.69
2 2
50 / 45 1.23
14) Refer Sub-Section 9.6.3.
15) We know that the mean and variance of F-distributionwith 1 and 2
degrees of freedom are
2
Mean for 2 2
2 2
and
2 22 1 2 2
Variance 2
for 2 4.
1 2 2 2 4
In our case, 1 5 and 2 12, therefore,
2 12
Mean 1 .2
2 2 10
2(12) 2 (5 12 2) 30 144
Variance 10.8
5(12 2) 2 (12 4) 40 100
16) Same as Sub-section 9.6.3.
258
Statistical Analysis-I
UNIT 10 STATISTICAL ANALYSIS-I
Structure
10.1 Introduction
10.2 Objectives
10.3 Hypothesis
10.3.1 Null and Alternative Hypothesis
10.3.2 Simple and Composite Hypothesis
10.4 Some Basic Concepts
10.4.1 Critical Regions (Region of Rejection)
10.4.2 Type-I and Type-II Error
10.4.3 Level of Significance
10.4.4 Degree of Freedom
10.5 Testing of Hypothesis
10.5.1 Procedure of Testing of Hypothesis
10.5.2 One-tail and Two-tail Test
10.6 Large Sample Tests
10.6.1 Test for the Significance of Population Mean
10.6.2 Test for Equality of Two Population Means
10.6.3 Test for the Significance of Population Proportion
10.6.4 Test for Equality of Two Population Proportions
10.7 Let Us Sum Up
10.8 Key Words
10.9 Suggested Further Reading/References
10.10 Answers to Check Your Progress
10.1 INTRODUCTION
In the previous unit, we have discussed about the very important concept of
‘Statistical Inference’ that is known as sampling distribution. We also
discussed some basic concepts like population sample, parameter, statistic
estimator and estimates. We have discussed that the probability distribution
of a sample statistic is known as sampling distribution of it. In previous unit
we have also discussed the sampling distribution of the statistics mean and
proportion as well as some of the exact sampling distributions, i.e., t, chi-
square and F – distribution along with its properties and applications. In this
unit we shall discuss the concept and meaning of testing of hypothesis. It
provides us significant conclusions on the assumption(s) made for any
259
Statistical Analysis population parameter on the basis of sample based estimated values. The
assumptions which are made for testing purpose are known as statistical
hypotheses.
In testing a hypothesis, we take decision about trueness of statement made for
a population parameter on the basis of observations. For example, a doctor
may be interested to know whether the new medicine is really effective for
controlling blood pressure, a manager may be interested to know whether one
brand of electric bulb is better than the other, a psychologist may be willing
to know whether the IQ of students studying in an Open University is up to
the standard of the students of IIT. Such statistical decisions can be
acceptable/unacceptable only if these are proved/disproved with the help of
sample data taken from the concerned population.
Generally, the calculated value of the sample statistic differs from the
assumed value of the population parameter since the statistic is based on a
part of the population and not on the entire population. Now, a question arises
whether this difference is actually significant or it appears only due to
fluctuations of samples. A little difference sometimes may occur due to a
main cause or due to sources of sampling errors.
It, therefore, seems that some theoretical backgrounds should be developed
for the testing of correctness/incorrectness of the assumed statement
(hypothesis) while the statements are to be verified on the basis of some
sample values (statistic) which themselves may vary from sample to sample
and, thus, may yield some kind of errors. Whatever procedure is applied for
this purpose is generally known as testing of hypothesis.
10.2 OBJECTIVES
After studying this unit, you should be able to
define a hypothesis, null hypothesis, alternative hypothesis, simple and
composite hypothesis;
define and explain the type-I and type-II errors;
explore the concept of critical regions, level of significance and degree of
freedom;
describe the procedure of testing a hypothesis; and
perform the test of hypothesis for large samples for population mean and
difference between two population mean.
10.3 HYPOTHESIS
In our daily life, each one of us comes across with the problems of testing
some kinds of assumptions which we may have in our mind. For example, a
housewife while cooking rice, makes an assumption in her mind that if only
few pieces of rice are checked and found well-cooked then the whole amount
260 of rice may be thought of fully cooked. Thus, she verifies and tests her
believe on the basis of few observations and not with the whole amount of Statistical Analysis-I
rice.; a manufacturer of some goods may think that advertisement of his
product will increase the sale and, therefore, may decide the trueness of his
thinking on the basis records of market evidences; a student, on the basis of
records of passed-out students, may have a belief that few subjects might help
him to pass out a bachelor degree course, etc. Each of these are examples
where any statement (assumption or belief) about some system needs to be
verified, that is, proved or disproved, on the basis of some data/information.
The above discussions and examples now help us to define a statistical
hypothesis. It may be defined in a number of ways: (i) A statistical
hypothesis is an assumption about a population or about its parameter. This
assumption may or may not be true; (ii) A tentative theory about the natural
world; a concept that is not yet verified but that if tried would explain certain
facts or phenomenon; (iii) It is a supposition, a proposition or principle which
is supposed or taken for granted, in order to draw a conclusion or inference
for proof of the point in question.
A statistical hypothesis stands in contrast to a simple assertion. If one makes
some statement and believes it to be true without its verification then it is an
assertion and not a hypothesis. For example, the statement “It is cloudy
weather outside” is an assertion, because the person believes it to be true and
he wants other persons also to believe it; whereas the statement “I think it is
cloudy weather outside” is a hypothesis since he/she involves in his/her
statement the possibility to prove or disprove the statement by others. In fact,
the intention is to determine at a later stage the truth or falsity or probability
of his/her statement.
In many cases, it is necessary to assume some theoretical probability
distributions for testing the hypothesis, particularly when there is no direct
knowledge of the population from which the observations are taken. For
example, a factory owner produces a product of weight exactly 10 g. The
producer finds that the weight of product has been reduced a little at present.
The owner is worried whether weight reduction is due to unavoidable reasons
or due to poor quality of raw material? He may then frame the hypothesis of
no change in weight and use a suitable theoretical probability distribution for
the variations in weight.
10.3.1 Null and Alternative Hypothesis
A statistical hypothesis is an unproved, well-structured, assumed statement
about the population. Null hypothesis is a preliminary part of it, and it is
denoted by symbol H0. It is relating to basic ignorant thought of “no
difference, no change or no effect” between two or more occasions or
situations or places or values. Null means “nothing” and null hypothesis is
that nothing is happening, nothing is present and no changes are observed,
etc., and in this sense, it is a hypothesis which might be falsified on the basis
of the observed data.
261
Statistical Analysis For example, statements like, “effect of new injection is same as the old one
(no difference)”; “the efficiency of new technology is same as the earlier
technology (no change)”; “the two mathematics teachers are equally good (no
one is better than the other)”, etc. Mathematically, if 1 and 2 are values of
similar parameters of two different theoretical populations, then our null
hypothesis will be H0: 1= 2, that is, there is no difference in their values.
Similarly, if 0 is the specified value of the parameter, we should frame our
null hypothesis H0: = 0, that is, there is no change in the given value. Thus,
null hypothesis represents what we would believe by default, before seeing
any evidence. It generally negates any change, any happening, etc. It is
important to mention here that in all types of problems, it is only the null
hypothesis which is either rejected (not accepted) or not rejected (accepted).
According to Prof. R.A. Fisher who coined the term “null hypothesis”, a null
hypothesis is tested for its possible rejection under the assumption that it is
true. Formulation of null hypothesis is a vital step in testing statistical
significance. It should be exact, that is, free of vagueness and ambiguity
Every statistical null hypothesis always occurs in conjunction with another
hypothesis known as “alternative hypothesis” generally denoted by H1. It is a
well-structured assumed statement which complements the null hypothesis
and is stated in such a way that they are mutually exclusive, that is, if one is
true, the other must be false and vice versa. Therefore, when null hypothesis
is rejected, it is equivalent to accept alternative hypothesis.
Some examples of H0 with corresponding H1 are shown below:
(i) If H0: = 0 then H1 may be either 0 or > 0 or < 0
(ii) If H0: 1 = 2 then H1 may be either 1 2, 1 > 2 or 1 < 2
The acceptance of H1: 0 provides us an assurance of significant departure
of the current parametric value to the past /standard value. Likewise,
acceptance of H1: > 0 (or H1: < 0) provides us an assurance of departure
of current value from the past /standard value 0 toward higher (or lower)
level respectively.
10.3.2 Simple and Composite Hypothesis
Since a population is characterized by its parameters, therefore, when value
of parameter(s) is/are known, the population is treated as known (or specified
completely). A hypothesis which completely specifies parameter(s) of a
theoretical population (probability distribution) is called a simple hypothesis.
If a population possesses the only parameter then the statement H0: = 0 is
a simple hypothesis because acceptance of it specifies the population. In
contrast, a hypothesis which does not specify the distribution completely is
known as composite hypothesis.
As an example, let x1, x2,…, xn be a random sample of size n taken from a
normal population (normal probability distribution) with mean and
262 variance σ2. Then a hypothesis H0: = 20, given that σ2 = 5 is a simple
hypothesis since then the normal distribution is completely specified. Statistical Analysis-I
Acceptance of it would decide the value of the parameter . But a hypothesis
like, H0: > 250 is a composite hypothesis as it does not exactly specify the
value of ; it may be any value greater than 250. If both null and alternative
hypotheses are taken together, we may have following cases of combinations
of simple and composite hypotheses:
Case I: Simple null versus Simple alternative Hypothesis, like,
H0: = 0 ; H1: = 1
For instance,
H0: weight of tomato is 130 gms (i.e. H0: = 130gms)
H1: weight of tomato is 200 gms (i.e. H1: = 200gms)
The acceptance of H0 or H1, completely specify the population parameter
(tomato weight).
Case II: Simple null versus Composite alternative Hypothesis, such as
H0: = 0 ; H1: > 0 or H1: < 0 or H1: 0.
For example,
H0: weight of tomato is 130 gm. (i.e. H0: = 130gm).
H1: weight of tomato is above to 130 gm. (i.e. H1: > 130gm).
The acceptance of H0 specifies the population completely, but acceptance of
H1 does not, because a value greater to 130 will not be a unique value.
CHECK YOUR PROGRESS 1
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
1. Define Null Hypothesis and Alternative Hypothesis.
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
2. Describe the simple and composite structure of Null hypothesis and
Alternative hypothesis.
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
263
Statistical Analysis
10.4 SOME BASIC CONCEPTS
A null hypothesis is always tested by the sample data. Let x1, x2,…,xn be a
random sample drawn from a population having unknown parameter . The
collection of all possible values of x1, x2, …, xn is a set called “sample space”
S and a particular choice of x1, x2, …, xn represents a point in that space.
10.4.1 Critical Region (Region of Rejection)
As per criterion of classical testing of hypotheses concepts, in fact, the entire
sample space is partitioned into two disjoint sub-spaces, say
and S , such that if computed value of the test statistic falls in ω ,
we reject the null hypothesis against the alternative hypothesis and if it falls
in ω we do not reject the null hypothesis in favour of alternative hypothesis.
In that sense, the region ω is called a “rejection region” or more popularly a
“critical region”. The is known as “acceptance region”.
As an example, a student appears in total 10 papers, two of each in English,
Physics, Chemistry, Mathematics, and Computer Science. His scores in these
papers are respectively x1, x2, x3…x9, x10 out of maximum marks 100 for each
paper. It is a sample data with sample size n=10.
n
We define statistic t n x i as a sum of all these scores. The range of tn is
i 1
0 t n 1000. For obtaining the distinction award in a course program, a
student needs to have total score more than 750 which is a rule. Suppose any
one student is randomly selected from that course program and we frame
hypotheses as:
H0: selected student is not a distinction award holder
H1: selected student is a distinction award holder.
Now, according to null hypotheses, we have two regions by partitioning the
space [0- 1000]: (i) acceptance region [0-750] and (ii) rejection region [751-
1000]) as shown in figure 3.1.
Figure 10.1: Acceptance and critical regions for Distinction Award
On the basis of sample of size n and statistic tn the computed value may fall
in distinction award region or not, depending how sample observations are.
10
For proving H0, we use test statistic t 10 x i (sum of scores of 10 papers)
i 1
264 and conclude: reject H0 if t10 >750, accept H0 if t10 ≤ 750. Clearly, it
indicates that a basic structure of the procedure of testing of hypothesis needs Statistical Analysis-I
two regions.
The region of rejection has a pre-fixed area denoted by , corresponding to a
cut-off value in a probability distribution of test statistic. The prefixed
probability area, that is, is also called size of test or level of significance or
probability of type I error.
Figure 10.2: Acceptance and Critical regions for Statistic tn.
Symbolically, if w be critical region and tn be the value of statistic based on
sample of size n, then
Pt n ω / H0 Pt n ω, when H0 is true α
Pt n ω / H1 Pt n ω when H1 true β
10.4.2 Type-I and Type-II Error
In previous section, we have discussed a rule that if the value of test statistic
falls in the critical region we reject the null hypothesis, otherwise we do not
reject it. Since the test statistic itself is generated from the sample, which has
been selected at random from the population, it might be very far from the
actual value of the parameter concerned if the sample is not a good
representative of the population or might be very close to the parameter value
if the sample is a good representative. Therefore, in taking decision for
accepting or rejecting null hypothesis on the basis of the value of the
statistics, we may commit some kinds of errors.
For example, an engineer on the basis of a sample infers that a packet of
screws is sub-standard whereas actually it is not or may infer the packet to be
good whereas actually it is sub-standard. Similarly, a person may be declared
a patient of cancer and hence kept under high potency medicines whereas he
is not and vice versa. In both the cases, the error is because of inappropriate
samples of screws and diagnostic tests respectively which have been taken.
Accordingly, the errors which may arise in taking decisions are termed as
Type – I and Type – II errors which are depicted in the following table:
Decision H0 True H1 True
Reject H0 Type - I Error Correct Decision
Accept H0 Correct Decision Type -II Error
265
Statistical Analysis Type-I Error:
From the table we see that the decision relating to rejection of H0, when, in
fact, it is true is called type-I error. The probability of committing type -I
error is called “size of the test” or “level of significance” and is generally
denoted by Thus,
= P [Rejecting H0 when H0 true]
= P [Rejecting H0 | H0 true]
= P [ x ω / H0 ] where x stands for the value of the
statistic under study. Obviously then we have
1- = 1-P [Rejecting H0 | H0 true]
=P [Accepting H0| H0 true] = P [correct decision].
(1-) is probability of correct decision and it correlates to the concept of
100(1-)% confidence interval used in the estimation. Theoretically, the test
procedures are so constructed that the risk of rejecting H0 when it is true is
small.
Type -II error:
The decision relating to acceptance of H0 when it is false, (that is H1 true) is
called type-II error. The probability of committing type-II error is denoted by
. Thus,
= P [Accepting H0 when H0 false]
= P [Accepting H0 when H1 true]
= P [Accepting H0 | H1 true] = P [ x ω / H1 ].
Hence 1- = 1-P [Accepting H0 | H1 is true]
= P [Rejecting H0 | H1 is true] = P [correct decision].
Therefore, (1-), that is, probability of not committing a type-II error is the
probability of correct decision, which is called power of the test.
Being probabilities of type-I and type-II errors, both α and β should be as
small as possible, but with a given set of data, both types of errors cannot be
minimized simultaneously. Therefore, customarily, the test procedures are
developed with an aim to minimize β with a fixed assumed value of α.
10.4.3 Level of Significance
We know that probability of Type-I error is known as level of significance of
a test which is denoted by α. It is also called the size of a test or critical
region. As discussed above, level of significance is a fixed assumed value
which is decided before starting the test procedure. The most commonly used
values of α are 0.05 (5%) and 0.01 (1%). By selecting α = 5%, we mean that
if the null hypothesis is true, there is a 5% chance of rejecting it because of
266
random causes, or equivalently, the concluding statement about H0” is true Statistical Analysis-I
only with 95% assurance. Similar conclusions can be drawn for α = 1%.
Figure 10.3: Critical region area at 5% level of significance.
Consider the figure 10.3, showing probability curve of any probability
distribution, where entire area underlying the curve is obviously equal to
100% out of which let 5% area (shaded area) lies at the end of one side (right
side) of the curve. Let this pre-fixed area constitute the critical region for a
test, that is, here = 5%. Now, as soon as we decide (pre – fix) the value of α
for testing our hypothesis, the cut-off point on x axis (which is shown in the
figure as x0) can easily be obtained on the basis of the used probability
distribution. This value (termed as “critical value”) is used in decision
making for the acceptance/rejection of H0. The calculated statistic tn =f(x1…
xn) may result in a value either less than or greater than the critical value x 0
So, we can make decisions rules, using tn and x0, as follows:
Reject H0 (or equivalently accept H1) if tn > x0; otherwise
accept H0 (or equivalently accept H1) if tn<x0.
This makes it clear that main focus in a testing problem is to decide the
critical value x0 on the basis of the probability distribution of the statistic tn.
This is important to mention here that for standard testing problems, that is,
with known null and alternative hypotheses, the statistic to be used for testing
purpose, is also pre-decided and its sampling distribution is also known.
Some well-known distributions used for most of the testing problems are Z,
χ2, t and F distributions.
10.4.4 Degree of Freedom (df)
The term “degree of freedom” (df) is related with the concept of
independency of sample observations. We have seen that the entire decision
in testing of a hypothesis is based on the sample observations x1, x2, x3, …,xn.
If all these observations are independent among themselves then the
conclusion drawn about H0 is said to be having n degree of freedom.
Suppose for 50 observations x1, x2, …, x50, a restriction (condition) is
50
imposed as x
i 1
i 1000 that is, there sum is fixed and it is 1000.
267
Statistical Analysis If it is like so, then actually, there are now only 49 values which are
independent, because any value (say, the last one) can be obtained using the
equation
49
x 50 1000 x i
i 1
It implies that by putting one condition on the sample values, we lose one
degree of freedom and in general if for a sample of size n, there are k
restrictions ( k<n), the degrees of freedom will be (n-k).
CHECK YOUR PROGRESS 2
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
3. Define a Statistic.
……………………………………………………………………………….
……………………………………………………………………………….
……….………………………………………………………………………
………………….……………………………………………………………
4. Describe the critical region with an example.
……………………………………………………………………………….
……………………………………………………………………………….
……….………………………………………………………………………
………………….……………………………………………………………
5. Describe Type I and Type II errors.
……………………………………………………………………………….
……………………………………………………………………………….
……….………………………………………………………………………
………………….……………………………………………………………
6. Describe the level of significance with example.
……………………………………………………………………………….
……………………………………………………………………………….
……….………………………………………………………………………
………………….……………………………………………………………
268
7. Describe the degree of freedom with example. Statistical Analysis-I
……………………………………………………………………………….
……………………………………………………………………………….
……….………………………………………………………………………
………………….……………………………………………………………
10.5 TESTING OF HYPOTHESIS
So far, we have discussed the basic concepts of the terms which are used in
determining the inference of the Population based on the sample results. The
purpose of this type of inference is to determine whether a certain assumption
or hypothesis about a parameter can be justified by statistical evidence and
also to make valid decisions about the population parameters based on the
analysis of samples. For example, a quality control manager must decide
whether the machines producing a product are working properly on the basis
of samples taken from the lot.
In order to make these statistical decisions, we make certain assumptions
about the population parameters to be tested. Then a sample is taken to
estimate the values of these population parameters. If the estimates favor the
hypothesis, then we do not reject the hypothesis as being correct. Since we
are testing for our assumptions or hypothesis being correct or not, this field of
decision making is known as Testing of Hypothesis. Now we will discuss the
step-by-step procedure of testing a hypothesis in details.
10.5.1 Procedure of Testing of Hypothesis
A hypothesis testing procedure should be designed in a number of steps. If x1,
x2, …, xn is a random sample drawn from a population having unknown
parameter , then steps though which it should pass are:
Step I: The first step in any test is to state the relevant null and alternative
hypotheses to be tested. For example, for testing the significance of
a given value of say, 0, these may be formulated as H0: = 0,
H1: ≠ 0 [or for comparing two values 1 and 2 we may test in
comparative sense, H0: 1 = 2 against H1: 1 2].
Step II: Establish a criterion for acceptance/rejection of null hypothesis, i.e.,
decides the level of significance at which we want to test our
hypothesis. Generally, it is taken as 1% or 5% (α = 0.05 or 0.01).
Step III: Define a statistic tn on sample observations x1, x2, x3,…,xn, and call
1
it a test-statistic. For example, tn= mean = x x i , tn = Max
n
(x1, x2,…,xn) = Highest sample observation, t n x i2 , etc.
Depending upon the objective of the test, other test statistics may be
proportion, difference between means, variance, correlation 269
Statistical Analysis coefficient, etc. In general, a test statistic is taken preferably in the
standard format like Z(for large samples), t (for small samples), 2
and F as:
Statistic - Parameter' s Assumed Value
tn
Standard Error of Sample Statistc
Do numerical computation of tn on the sample data.
Step IV: Obtain the probability distribution of tn, under the null hypothesis
from the assumptions. Generally, with the objectives and the
defined test statistic, the distributions are in the standard form like
Z, 2, t or F or any other well-known distribution in literature.
Step V: Using the distribution and value of , obtain the critical value (or
cut-off value) in that probability distribution.
Step VI: Take decision about H0, in light of step I and II, as
(i) Reject H0 if modulus of calculated value of tn, i.e., t n
critical value (or tabulated value) at level of significance.
(ii) Do not reject H0 if modulus of calculated value of tn, i.e.,
t n critical value (or tabulated value) at level of
significance.
When we take H0: = 0 and H1: > 0 in step-I then decision in
step VI is different and we reject H0: = 0 if tn > critical value. No
modulus sign is used for this case for tn. Similarly, if H1: < 0, we
reject H0: = 0 when tn< - critical value at α level of significance.
10.5.2 One-Tail and Two-Tail Tests
Figure 3 showed that the critical value lies on the right side of the curve, but
always it is not the case. In any test, the critical region may lie either at one
side (at the left tail or at the right tail of the curve) or at both sides, depending
upon how the alternative hypothesis H1 is formulated. Accordingly, we have
one-tailed tests (right or left-tailed tests) or two-tailed test. If it is only at one
side, the size is α, and if is on both the sides, then, the size will be α/2 on both
the sides. Accordingly, the test is said to be one-tailed test or two-tailed test
respectively.
For example, consider a test for testing hypotheses:
(i) H0: = 0, H1: > 0 and (ii) H0: = 0, H1: < 0.
Obviously in (i), H0 will be rejected (H1 will be accepted) when >0; this
means that critical region lies on the upper side (right side) of the
distribution, hence it would be a one-tailed test (right-tailed test). The same
argument shows that (ii) would be a one-tailed test (left-tailed test). So, both
(i) and (ii) are one-tailed tests.
270
Figure 10.4 show the situation of critical values and critical regions for right- Statistical Analysis-I
tailed test.
Figure 10.4: One-tail (right-tail) critical region of size
In contrary, when the alternative hypothesis is like H1: 0 where may be
lower than 0 or may be higher than 0, we have the situation of two-tail test.
Here the critical region may be on the left tail or may be on the right tail
according as < 0 or > 0 and we have two critical values; one on left tail
and another on right tail. Since the level of significance is α, it must be
divided into two such that both sides have critical region equal to α/2 (see the
Figure 3.5 below).
Figure 10.5: Two-tail critical regions of size/2 at each end.
We summarize the three cases related to above discussion as follows:
Null hypothesis Alterative Types of critical region /
hypothesis test
H0 : θ θ0 H1 : θ θ 0 Two tailed test having
critical regions on both
sides.
H0 : θ θ0 H1 : θ θ 0 One tailed test having
critical region on right
hand side.
H0 : θ θ0 H1 : θ θ 0 One tailed test having
critical region on left hand
side.
Likewise, if the null hypothesis is formulated for testing the equality of two
parameters as H0: 1 = 2; we may have alternative hypothesis either H1: 1 >
2 or H1: 1 < 2. Then, in view of the discussions made above and from the
above table, we can see that the test would be right-tailed test when H1: 1 >
2 and left-tailed test when H1: 1 < 2 with level of significance α in each
271
Statistical Analysis case. Further, if in conjunction with H0: 1 = 2, we have H1: 1 ≠ 2, then it
would be a two-tailed test with level of significance α/2 on both sides.
CHECK YOUR PROGRESS 3
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
8. Explain the design of testing of hypothesis.
……………………………………………………………………………….
……………………………………………………………………………….
……….………………………………………………………………………
………………….……………………………………………………………
9. Describe the One tail and Two tail tests with example.
……………………………………………………………………………….
……………………………………………………………………………….
……….………………………………………………………………………
………………….……………………………………………………………
10.6 LARGE SAMPLES TESTS
After discussing some of the preliminaries of testing of hypotheses, now we
shall discuss some test procedures in which our sample is assumed to be of
quite large size though finite.
In Statistics, it has been theoretically proved that if sample taken from any
population is of large size, whatever be the parent distribution, the sampling
distribution of the statistic based upon this sample can be approximated by a
normal distribution. Therefore, one can apply normal distribution-based test
procedures to obtain the cut off points at the pre-defined level of significance
in such cases. Accordingly, such tests are called large sample tests.
Suppose x1, x2,…, xn is a random sample of size n selected from a population
having unknown parameter . Generally, for assuring n to be large enough,
we assume n 30. Let with this assumption, we require a test procedure for
the significance of parameter . In this Section, we shall describe here a
number of tests which are as follows:
10.6.1 Test for The Significance of Population Mean
Our aim here is to test whether the population mean, denoted by µ, assumes a
specific value µ0 or not. Thus, the null hypothesis would be H0: µ = µ0 and
the alternative hypothesis H1 may be any of these:
272 H 1 : μ μ 0 or H 1 : μ μ 0 or H 1 : μ μ 0
If x1, x2,…,xn be a random sample of size n 30 taken from a population Statistical Analysis-I
with mean and finite variance σ2 then by central limit theorem the sample
mean is asymptotically normally distributed with mean and variance σ2/n
irrespective of parent population be normal or non-normal.
Therefore, for mean x = (1/n) Σxi , i=1,2,3,…. N
2
E(x) Var ( x ) V ( x )
n
The test statistic Z which is to be used for testing H0:
x E(x) x μ0
Z ~ N(0, 1) under H0 (assuming σ2 known)
Vx σ/ n
If σ is unknown, then we replace σ by its sample estimate
1 n
s ( x i x ) 2 and then we compute the Z-statistic (since n large)
n 1 i 1
as:
x μ0
Z ~ N(0, 1) under H0 (for large n)
s/ n
The test is one-tail if either H1: µ > µ0 or H1: µ < µ0 and two-tail if
H1 : μ 1 μ 2 as the case may be. The tabulated value zα is available in the
Table 3.1 for one-tail and two-tail tests at different value of α as follows:
Table 10.1
Nature of Test Level of Significance and Critical (Tabulated)
Value
α=0.01 α=0.05 α=0.10
Right Tail Test zα=2.33 zα=1.645 zα=1.28
Right Tail Test zα= -2.33 zα= -1.645 zα= -1.28
Two Tail Test zα=2.58 zα=1.96 zα=1.645
For prefixed level of significance α, the calculated value Z, obtained by
above formula is compared with the tabulated zα as mentioned below:
One Tail Test: When H0: µ = µ0 and either H1: µ > µ0 or H1: µ < µ 0
If Z z and H1: µ > µ0, (right tail test) the null hypothesis is rejected and
significant evidence of departure from standard value is confirmed.
Otherwise, for Z z the H0 is not rejected with conclusion that no
significant evidence exists against the null hypothesis. On the other hand, if
Z z and H1: µ < µ0 (left tail test), we reject H0, otherwise do not reject
for Z z .
273
Statistical Analysis Two Tail Test: When H0: µ = µ0 and H1: µ µ0
If Z z α / 2 the null hypothesis will not be rejected and when Z z α / 2 the
null hypothesis will be rejected.
Example 1: A sample of 100 male students is found to have a mean height of
72 inches. Can it reasonably be regarded as a sample from a large population
with mean height 70 inches and standard deviation 4 inches?
Solution: We are given that
n = 100, x = 72 inches, = 70 inches and σ = 4 inches
We wish to test the null hypothesis:
H0: = 70 against H1: 70
It is two-tail test and the test statistic is Z, since n is more than 30. We have
x μ0 72 70 2
Z 5 .0
σ/ n 4 / 100 0 .4
Since calculated value of test statistic Z is greater than 3, we reject our null
hypothesis H0 at every level of significance (i.e. at 5%, 1% both). We
conclude that sample is not from population with mean 70 and standard
deviation 4.
Example 2: A manufacturer of ball point pens claims that a certain pen he
manufactures has a mean writing-life of 500 pages. A purchasing agent
selects a sample of 100 pens and put them for the test. The mean writing-life
of the sample found 490 pages with standard deviation 50 pages. Should the
purchasing agent reject the manufacturer’s claim at 1% level of
significance?
Solution: Given that
= 500, n = 100, x = 490 and s = 50
We wish to test the null hypothesis
H0: = 500 against H1: < 500
Since population standard deviation (σ) is unknown, so we take estimate of σ
as sample standard deviation (s) in place of σ.
It is left-tail test and the test statistic Z is
x μ 0 490 500 10
Z 2.0
s/ n 50 / 100 5
The tabulated value at 1% level of significance is zα = z0.01 = -2.33 for left
tail test. Since calculated value of Z is greater than the tabulated value, so we
do not reject the null hypothesis and conclude that the purchasing agent will
be agree with the manufacture’s claim at 1% level of significance.
274
CHECK YOUR PROGRESS 4 Statistical Analysis-I
Note: i) Check your answers with those given at the end of the unit.
10) A sample of 900 bolts has a mean length 3.4cm. Is the sample
regarded to be taken from a large population of bolts with mean 3.25
cm. and standard deviation 2.61 cm. at 5% level of significance?
11) A big company uses thousands of CFL lights every year. The brand
that the company has been using in the past has average life 1200
hours. A new brand is offered to the company at a price lower than
they are paying for the old brand. Consequently, a sample of 100 new
brand CLF light is tested and yields an average life of 1180 hours
with standard deviation 90 hours. Should the company accept to the
new brand?
10.6.2 Test for Equality of Two Population Means
Let there be two populations say population-I and population-II. The
population-I has mean μ 1 and variance σ 12 whereas the population-II has mean
μ 2 and variance σ 22 .
For testing the hypothesis that two-population means are same, we draw
independent random samples of size n1 and n2 respectively from population-I
and population-II. Let x1i (i =1, 2, …, n1) and x2i (i=1, 2,…, n2) be the
observations in samples selected from population-I and population-II
respectively and x 1 and x 2 be the corresponding sample means. In order to
test the null hypothesis, we assume that both sample sizes are large enough to
ensure that sampling distribution of the test statistic to be used is
asymptotically normal distribution. As we have discussed in previous unit
that the sampling distribution of difference of two large sample means
follows asymptotic normal distribution with mean (µ1-µ2) and variance
(σ 12 / n 1 σ 22 / n 2 ). The scheme of selecting the samples from both the
populations is shown in the following fig 10.6:
Population I (may not normal) Population II (may not normal)
µ1, σ 12 µ2 , σ 22
x 11 , x 12 , x 13 ,..., x 1n 1 x 21 , x 22 , x 23 ,..., x 2n 2
Mean x 1 Mean x2
Normal N μ 1 , σ12 /n 1
Normal N μ 2 , σ 22 /n 2
Normal Population of mean x 1 Normal Population of mean x 2
Fig 10.6: Two independent samples from two different populations having
normal distribution of means
275
Statistical Analysis Thus, the Test statistic
(x 1 x 2 ) E(x 1 x 2 )
Z ~ N(0, 1) .
V(x 1 x 2 )
We assume that population variances σ12, σ22 are known to us. Thus we have
E ( x 1 x 2 ) E ( x1 ) E ( x 2 ) μ1 μ 2
and
σ12 σ 22
Var ( x 1 x 2 ) Var ( x 1 ) Var ( x 2 )
n1 n 2
Our null hypothesis here is:
H0: µ1 = µ2 (no difference in means)
against the alternative hypotheses
H 1 : μ 1 μ 2 or H 1 : μ 1 μ 2 or H 1 : μ 1 μ 2
Now from the above, we get the test statistic Z as
( x1 x 2 ) (μ1 μ 2 )
Z ~ N(0, 1)
σ12 σ 22
n1 n 2
Since under H0, µ1 = µ2, so µ1 - µ2= 0, and hence
x1 x 2
Z ~ N(0, 1) under H 0
σ12 σ 22
n1 n 2
For prefixed level of significance, α, the calculated Z, by the above formula,
is compared with the tabulated zα. The tabulated value of zα may be
determined from the Table 3.1 for either one-tail or two-tail, as the case may
be.
For One Tailed Test: When H0: µ1 = µ2 and H1: µ1> µ2 or H1: µ1< µ2
For H1: µ1> µ2 or H1: µ1< µ2 (right tail test), if Z z the null hypothesis is
not rejected and we conclude that, there is no significant evidence against the
equality of means of two populations. So, both means will be treated as same
at the α-level of significance otherwise if Z z H0 is rejected at α-level of
significance and the difference between the two means µ1and µ2 is
established and concluded that one mean is greater than the other.
For H1: µ1< µ2 (left tail test), if Z z the null hypothesis is rejected at α
level of significance, otherwise for Z z , H0 is not rejected.
276
For Two Tail Test: When H0: θ1 = θ2 and H1: θ1 θ2 Statistical Analysis-I
For H1: θ1 θ2 and if Z z α / 2 , the null hypothesis is not rejected and for
Z z α / 2 the null hypothesis is rejected at α-level of significance.
Case-I: When Variances are Equal and Known
If 12 22 2 i.e. populations have same variances, i.e., 2 (known) then
for testing H 0 : μ1 μ 2 , the test statistic Z reduces to
x1 x 2
Z ~ N[0, 1]
1 1
σ
n1 n 2
Case-II: When Variances are Equal and Unknown
If 12 , 22 are unknown these two could be estimated by their respective
unbiased estimators s12 , s 22 respectively, where
1 n1 1 n2
s12 x1i x1 2 ; s 22 x 2i x 2 2
n 1 1 i 1 n 2 1 i 1
When σ12 = σ22= σ2 (unknown), then the estimate of σ2 is:
2 n 1 1s12 n 2 1s 22
σ̂
n1 n 2 2
x1 x 2
and Z ~ N0, 1
2 1 1
σ̂
n1 n 2
Case-III: When Variances are Unequal and Unknown
If 12 22 and 1 , 2 (both are unknown), then if sample sizes n1, n2 are
large enough, we use s12 and s22 respectively for σ12 and σ22. Then the Z
statistic will be:
x 1 x 2
Z ~ N 0, 1
s /n s /n
2
1 1
2
2 2
Example 3: In two samples of women from Punjab and Tamil Nadu, the
mean height of 1000 and 2000 women are 67.5 and 68.0 inches respectively.
Can the samples be regarded as drawn from the same population having
standard deviation 2.5 inches? In other words, can the mean height of Punjab
and Tamil Nadu women be regarded as same?
Solution: Given n1= 1000, n2 = 2000, x1 67.5, x2 68.0, σ 2.5 inches
We wish to test the null hypothesis that the samples are drawn from the same
population, that is, mean height of Punjab and Tamil Nadu women are same.
H0: 1 = 2 against H1: 1 2 277
Statistical Analysis It is two tail test and test statistic Z is
x1 x 2 67.5 68.0
Z 5.16
1 1 1 1
2.5
n1 n 2 1000 2000
The calculated Z 5.16 tabulated zα/2=1.96 at5% level of significance
which implies rejection of hypothesis H0.
We conclude that women of Punjab and Tamil Nadu differ significantly in
terms of their mean height.
Example 4: IGNOU conducts both face-to-face and distance-mode classes
for particular courses indented both to be identical. One sample of 50
students yields examination results mean and SD as:
x 1 80 .4 , s1 12 .8
and other sample of 100 distance-mode students yields mean and SD
examination results as
x 2 78 .3 , s 2 20 .5
Are the two means of education methods statistically equal at 5% level?
Solution: Given that
n 1 50 , x 1 80.4, s 1 12.8;
n 2 100 , x 2 78.3, s 2 20.5
We wish to test the null hypothesis that the two means of teaching methods
are statistically equal, that is, no difference in both type of teaching methods,
or
H0: 1 = 2 against H1 : μ1 μ 2
Since population standard deviations σ1 and σ2 are unknown, so we use
estimates of σ1 and σ2 as sample standard deviations s1 and s2 respectively.
It is two tail test and the test statistic Z is
x1 x 2 80.4 78.3 2 .1
Z 0.77
2 2
s 12 s 22
12.8
20.5 3.28 4.20
n1 n 2 50 100
The calculated Z 0.77 tabulated z α/2 1.96 so we accept H0 at 5% level of
significance. We conclude that both methods of education, i.e., face-to-face
and distance-mode are statistically equal. No significant difference observed
in sample data.
278
CHECK YOUR PROGRESS 5 Statistical Analysis-I
Note: i) Check your answers with those given at the end of the unit.
12) Two brands of electric bulbs are quoted at the same price. A buyer
tested a random sample of 200 bulbs of each brand and found the
following information:
Mean life (hrs.) S.D. (hrs.)
Brand A 1300 41
Brand B 1280 46
Is there significant difference in the quality(life) of two brands of
electric bulb at 1% level of significance?
13) A marketing company has undertaken an advertisement campaign for
a popular breakfast food and claimed that it is an improved product. A
survey is conducted to find out the monthly demand of 100 consumers
before and after the campaign. We have following information.
Mean S.D.
Before 120 25
After 138 36
Analyze the above data and test whether the campaign was successful?
10.6.3 Test for the Significance of Population Proportion
Sometime we come across the populations divided into two groups or
categories and instead of testing the significance of population mean, we may
be interested to test the significance of population proportion, for example,
proportion of car-owners, smokers, cancer patients, etc., in the concerned
population.
In general, let N be the size of a finite population and ‘A’ be an attribute.
Suppose there are M people (or items) among N (M < N) who possess
attribute A (or in favor of A).
Let in a random sample of size n < N is drawn from the same population N
which contains x people possess attribute A, remaining (n−x) do not possess
it where (0 x n) . The scheme is shown in fig. 10.7.
279
Statistical Analysis POPULATION of Size N
M units
(N –M) units
Possessing A
Sample n (n < N)
x
n -x
Fig. 10.7: Sample from population possessing attribute A (having x
units in favor of A)
Let us denote population and sample proportions, respectively by
M x
P and p
N n
It is shown in the previous unit that E(p) = P. Suppose our problem here is to
test the significance of Population proportion, that is, whether in the
population the proportion is equal to a given value (say P0) or not, for
instance we want to know whether the proportion of cancer patients is 18%;
proportion of literate persons is 27%, etc. Obviously, then the null hypothesis
will be H0: P = P0 and it can be tested, as usual, by selecting a sample of
prefixed size from the population.
From the theory of sampling distribution, we observe that for sufficiently
large N, the distribution of p is asymptotically normal with mean P and
PQ
variance . Hence a test statistic for testing H0: P = P0 and alternative
n
hypothesis either H1 : P P0 or H1: P > P0 or H1: P < P0 the Z statistic can be
written as
p E(p) pP
Z ~ N 0, 1
V(p) PQ
n
Two-Tail Test: If Z Z / 2 we reject H0 against H1 : P P0 at α level of
significance in a two-tail test and conclude that sample data does not support
to the claimed population proportion P0. When Z Z / 2 we do not reject H 0
inferring that claimed or past population proportion still stands as revealed by
the sample data.
One-Tail Test: When to test H0: P = P0 against H1: P > P0, (right-tail test) we
280 reject H0 if Z >Z. It is equivalent to conclude that in the population
proportion possessing A has been significantly improved compare to earlier Statistical Analysis-I
P0. In contrary to this, when H0: P = P0 against H1: P < P0 (left-tail test) is
tested, we reject H0 if Z <-Z at level of significance. It infers there might
be significant reduction in the proportion in the present scenario.
Example 5: A machine produces a large number of items out of which 30%
are found to be defective. In a random sample of 100 items, 40 are found to
be defective. Is there any evidence of more deterioration of quality at 5%
level of significance.
Solution: The company manager knows that his machine produces 30%
defective items. The “defects” is an attribute which some objects of
population possess, so P0=0.30, and he is worry whether this has increased
presently?
It is given that sample proportion p = 40/100 = 0.40
The null hypothesis would be
H0: P = P0 = 0.30 against H1: P > 0.30
It is right-tail test and the test statistic is obtained as
p P0 0.40 0.30 0.10
Z 2.18
P0 Q 0 /n 0.30 0.70/100 0.0458
The calculated Z= 2.18 is greater than the tabulated value z = 1.64 at =
0.05, Hence, we reject null hypothesis H0 and conclude that there is
significant deterioration of quality of items.
Example 6: A dice is thrown 900 times and draw of a 2 or a 5 is observed
450 times. Show that the die cannot be regarded as as unbiased at 5% level of
significance.
Solution: If getting a 2 or 5 is a success, then in usual notions, we have from
the questions
n = 9000, x = number of successes = 450, p = 450/900 = 1/2
Here, H0: Dice is unbiased, i.e., Probability of getting a 2 or 5, P = P0 = 1/3,
and hence
1 1
H 0 : P0 against H 1 : P0
3 3
It is two tail test and test statistic Z is
p P0 (1 / 2) (1 / 3) (1 / 6)
Z
10.6
P0 Q 0 (1 / 3) (2 / 3) 2
n 900 8100
Calculated Z 10.6 tabulated z α/2 1.96, so H 0 is rejected and we infer
that the dice is significantly biased at 5% level of significance. 281
Statistical Analysis CHECK YOUR PROGRESS 6
Note: i) Check your answers with those given at the end of the unit.
14) In a sample of 100 M.Sc. Economics 1st year students of IGNOU, it
was seen that 54 came from Science background and the rest are from
other. Can we assume that students from both backgrounds (i.e.
Science and others) have equally performed in the examination at 1%
level of significance?
15) Out of 20 patients who are given a particular injection, 18 survived.
Test the hypothesis that the survival rate is 80% or more at 5% level
of significance?
10.6.4 Test for Equality of Two Population Proportions
Suppose there are two populations, each having persons (or items) possessing
attribute A. We are interested to test whether proportions possessing A in
both the population are same?
For example, percentage of failed students in two different Sections in a
school are same or populations of smokers in two different cities are same or
one city exhibit more number smokers than that in another city. We may use
notations P1 and P2, respectively to denote proportions of units in two
populations possessing any attribute A. Obviously, the Null and alternative
hypothesis will be:
H0: Failed percentage in both sections are same (proportion of Smokers in
both the cities are same), i.e., P1 = P2
H1: P1 ≠ P2
Let samples of sizes n1 and n2 be the samples selected independently from the
Population-I and Population-II (scheme of selection of samples is depicted in
figure 10.7).
Population I Population II
P1=X1/N1 P2= X2/N2
Sample n1 Sample n2
p1= x1/n1 p2= x2/n2
Fig 10.6.1: Two independent samples from different populations
possessing attribute A
Let n1 has x1 units and n2 has x2 units possessing the attribute A (say A =
282 smoking-habit) in samples. Therefore, from the theory of sampling
distribution of (p1 - p2), following are the expected values, variances and Statistical Analysis-I
standard errors of statistic respectively:
p1 = (x1 / n1) ; p2 = (x2 / n2)
E (p1) = P1 ; E (p2) = P2 E (p1 - p2) = P1- P2
P1Q1 PQ PQ P Q
V (p1) = V (p2) = 2 2 ; V(p1 p 2 ) 1 1 2 2
n1 n2 n1 n2
P 1 Q1 P2 Q2 P1 Q1 P 2 Q 2
SE(p1 ) ; SE(p 2 ) ; SE(p1 p 2 )
n1 n2 n1 n2
Under the assumptions of large population and sample sizes due to central
limit theorem, the term (p1-p2) is distributed asymptotically normal with mean
E(p1-p2) and variance V(p1-p2) and the Z-statistic is:
Z
p1 p 2 Ep1 p 2 ~ N [ 0, 1]
Vp1 p 2
Under H0: P1= P2 = P (given value), and H1: P1≠ P2, the test statistic Z is
p1 p 2
Z ~ N [ 0, 1]
1 1 where Q = 1-P.
PQ
n1 n 2
For H1: P1 P2 and if Z z α / 2 , the null hypothesis is not rejected and for
Z z α / 2 the null hypothesis is rejected at α-level of significance.
If P is unknown, its unbiased sample-based estimate p̂ is used to replace P in
Z as
p1 p 2 n 1 p1 n 2 p 2
Z where p̂ ;
q̂ 1 p̂
1 1 n1 n 2
p̂ q̂
n1 n 2
In such case for alterative hypothesis H1: P1> P2 the test statistic Z is used
(instead of Z ) and H0: P1=P2 is rejected when Z z (right-tail test).
Otherwise, if H1: P1< P2 the H0: P1=P2 is rejected when Z z (left-tail
test).
Example 7: In a random sample of 100 persons from town A, 55 are found to
be high consumers of wheat. In another sample of 90 from town B, 45 are
found to be high consumers of wheat. Do these data reveal a significant
difference between town A and town B, so far as the proportions of high
wheat consumers are concerns (at α= 0.05).
Solution: Given x 1 55 , x 2 45 , n 1 100 , n 2 90 283
Statistical Analysis Let us set up the hypothesis that the proportions of consumers of wheat in the
two towns say P1 and P2, are same.
i.e., H0: = P1 = P2 = P against H1 : P1 P2
From the data the corresponding sample proportions are obtained as:
x1 55 x 2 45
p1 0.55 and p2 0.50
n 1 100 n 2 90
Since, here population proportions P1 and P2 are unknown, we obtain the
estimate of combined proportion (p) of high wheat consumers in the two
towns as:
n 1 p1 n 2 p 2 x 1 x 2 55 45 10 10 9
p̂ q̂ 1 .
n1 n 2 n 1 n 2 100 90 19 19 19
As per assumption of H1, it is two tail test. The value of test statistic Z as
p1 p 2 p1 p 2
Z
S.E.p1 p 2 1 1
p̂q̂
n1 n 2
0.55 0.50 0.05
0.666
10 9 1 1 0.075
19 . 19 100 90
The calculated Z = 0.666 is smaller than the tabulated z / 2 z0.025 1.96 at
5% level of significance, so H0 is not rejected against H1 and accordingly we
conclude that towns A and B do not differ significantly so far as the high
wheat consumption is concerned.
Example 8: A machine produced 80 defective articles in a batch of 400.
After overhauling it produced 45 defectives in a batch of 300. Has the
machine improved due to overhauling? (Take = 0.01).
Solution: Here the population is produced articles and let attribute A stands
for defective articles. Let P1 and P2, respectively proportions of defective
articles before and after overhauling. Since null hypothesis always negate the
things, we assume that the proportions of defective articles are same before
and after overhauling, that is, H0: P1 = P2 = P (says) against H1: P1 > P2
(which means that proportion of defective articles would be smaller than
earlier before overhauling).
The sample defective proportions before and after overhauling are:
x1 80 x2 45
p1 0.20 and p2 0.15
n 1 400 n 2 300
Since P unknown, the pooled estimate of proportion is
284
n 1 p1 n 2 p 2 80 45 5 5 23 Statistical Analysis-I
p̂ and q̂ 1 p̂ 1 .
n1 n 2 400 300 28 28 28
It is right-tail test. The calculated value of test statistic Z is obtained as:
p1 p 2 p1 p 2
Z
S.E. p1 p 2 1 1
p̂q̂
n 1 n 2
0.20 0.15 0.050
1.852
5 23 1 1 0.027
28 28 400 300
The tabulated (or critical) value at 1% level of significance is zα = 2.33 for
one right-tail test. Hence, the calculated Z = 1.852 < tabulated zα = 2.33
implying that H0 is not rejected. Thus, we conclude that overhauling did not
improve the situation.
CHECK YOUR PROGRESS 7
Note: i) Check your answers with those given at the end of the unit.
16) The proportion of literates between groups of people of two districts
A and B are tested. Out of the 100 persons selected at random from
each district, 50 from district A and 40 from district B are found
literates. Test whether the observed difference in sample proportion of
literates is statistically significant at 1% level of significance?
10.7 LET US SUM UP
In this unit we have discussed:
1. Statistical hypothesis, Null hypothesis, Alternative hypothesis , Simple &
Composite hypothesis;
2. Type-I and Type-II errors, Critical region, One tailed and two tailed test;
3. Procedure of testing of hypothesis;
4. Level of significance and degree of freedom;
5. Design of a test of significance;
6. Large sample tests for Population Mean and difference between two
Population Means; and
7. Large sample tests for Population Proportion and difference between two
Population Proportions.
285
Statistical Analysis
10.8 KEY WORDS
Null Hypothesis: A statistical hypothesis that usually asserts that nothing
special is happening with respect to some characteristic of the underlying
population.
One-tailed (or directional) Test: The rejection region is located in just one
tail of the sampling distribution.
Two-tailed (or nondirectional) Test: Rejection regions are located in both
tails of the sampling distribution.
10.9 SUGGESTED FURTHER READING/
REFERENCES
Witte, R., & Witte, J. (2017). Statistics. Hoboken, NJ: John Wiley & Sons.
10.10 ANSWERS TO CHECK YOUR PROGRESS
1) Please refer to section 10.3.1 and 10.3.2.
2) Please refer to section 10.3.3.
3) Please refer to section 10.4.
4) Please refer to section 10.4.
5) Please refer to section 10.5.
6) Please refer to section 10.6.1.
7) Please refer to section 10.6.2.
8) Please refer to section 10.7.
9) Please refer to section 10.8.
10) We first set up our null hypothesis
H0: = 3.25 cm. against H1: μ 3.25
It is two tail test using one sample.
Given, x 3.4 cm, n 900 cm, μ 3.25 cm and σ 2.61 cm
3.40 3.25
Z 1.73
2.61 / 900
Since ׀Z<׀1.96, we conclude that H0 is not rejected at 5% level of
significance.
11) We given that
µ0 = 1200, n = 100, x 1180 , s = 90
To test the null hypothesis regarding mean
286
H0: = 1200 against H1: < 1200 Statistical Analysis-I
The sample is showing lower mean value which is a doubt. It is one
tail left test. Then the test statistic Z is
x μ0 1180 1200 20
Z 2.22
s/ n 90 / 100 9 .0
Since calculated Z > zα = -2.33, so we do not reject H0 at 5% level of
significance.
12) Given that
n1 200, x1 1300, σ1 41
n 2 200, x 2 1280, σ 2 46
We wish to test the null hypothesis
H0: 1 = 2 against H1 : μ1 μ 2 (i.e., mean life of both brands differ)
It is two tail test and test statistic is
x1 x 2 1300 1280 20
Z
σ 2
σ 2
41 2
46 2 4.36 = 4.59
1 2
n1 n 2 200 200
Calculated Z 4.59 tabulate d z α/2 2.58, we reject H 0 at 1% level
of significance.
13) Given that
n1 100, x1 120, s1 25
n 2 100, x 2 138, s 2 36
We wish to test the null hypothesis
H0: 1 = 2 against H1 : μ1 μ 2
It is one tail (left) test and the test statistic Z is
x1 x 2 120 138 18
Z
s2
1 s 2
2 25 2 2
36 4.38 = -4.11
n1 n 2 100 100
Since calculated Z = -4.11 < tabulated - zα = -2.33, so we reject H0 at
5% level of significance.
14) Here, n = 100 and x = Number of students from science background
54
p proportion of students from Science background 0.54
100
We wish to test the hypothesis: H0: P = P0 = 0.5 against H1: P0 ≠ 0.5
It is two tail test and test statistics Z is 287
Statistical Analysis p P0 0.54 0.50 0.04
Z 0.80
P0 Q 0 /n 0.5 0.5 / 100 0.05
The calculated value ׀Z = ׀0.80 is smaller than the tabulated value z =
2.58 at 1% level of significance then we do not reject null hypothesis.
15) Here, the null hypothesis is
H 0 P P0 against H1 : P P0
80 20 x 18
P0 0.80 ; Q0 0.20 p 0 .9
100 100 n 20
Also, by using one tail right test, test statistic Z is
p P0 0.1 0.1
Z 1.12
P0 Q 0 /n 0.80.2 / 20 0.089
The calculated Z= 1.12is smaller than the tabulated zα = 1.645 at 5%
level of significance and we do not reject the null hypothesis.
16) Let p1 and p2 stand for proportions of literates in districts A and B
respectively. We have to test H0: P1 = P2 = P against H1 : P1 P2 .
Given
x1 50 40
p1 0.50 and p2 0.40
n 1 100 100
n 1 p1 n 2 p 2 50 40
Also p̂ 0.45 q̂ 1 p̂ 0.55
n1 n 2 100 100
It is two tail test and, we calculate the Z statistic
p1 p 2 p1 p 2
Z
S.E. p1 p 2 1 1
p̂q̂
n 1 n 2
0.50 0.40 0.10
2.04
1 1 0.049
0.45 0.55 100 100
The calculated value Z 2.04 2.58 (critical value in two tail test at
1% level of significance), then the null hypothesis is not rejected.
288
Statistical Analysis-II
UNIT 11 STATISTICAL ANALYSIS-II
Structure
11.1 Introduction
11.2 Objectives
11.3 Procedure for Small Sample Test
11.3.1 Test for Population Mean
11.3.2 Test for Difference of Two Population Means
11.3.3 Paired t-test
11.4 Chi-Square Test
11.4.1 Test for Goodness of Fit
11.4.2 Test for Independence of Attributes
11.5 F-Test
11.5.1 Test for equality of Two Variances
11.6 Let Us Sum Up
11.7 Key Words
11.8 Suggested Further Reading/References
11.9 Answers to Check Your Progress
11.1 INTRODUCTION
The entire large sample theory, as applied to testing of hypothesis in the
previous unit, is based on the assumption that sample size being large enough
(generally more than 30) the sampling distribution of the test statistic can
always be approximated by the normal distribution owing to the results of
central limit theorem.
You might have observed there that due to this reason, the test statistics used
everywhere was Z, where Z stands for the normal distribution with mean zero
and variance one. However, if the sample size n is small (practically less
than 30), the sampling distribution of many test statistics are far from the
normal, and as such, normal distribution is not the basis of those tests. For
small samples, therefore, the assumption of normality of the distributions of
test statistic is not valid and exact sampling distribution of the test statistic
has to be known. Fortunately, in the literature of sampling distributions,
many such distributions are obtained which can be applied for small sample
problems of testing of hypotheses. In this Unit, we will study various tests of
significance based on these statistics.
The small sample tests for testing the significance of population mean and the
test for the equality of two population means are very frequent in real life
problems. We shall discuss here the relevant testing procedures, namely,
t- tests, derived on the basis of t- distribution. 289
Statistical Analysis In testing of hypothesis or even in the theory of estimation, we generally
assume that the random variable under study follows a particular probability
distribution, such as, normal, binomial, Poisson distribution, etc., but such
assumptions need to be verified as to whether our assumptions are true or not.
We have already discussed and have also seen that the genesis of the problem
of testing the hypotheses lies in this fact.
Although, in many testing problems, we frequently deal with certain
variables (that is, quantitative characteristics), but in many real-world
situations of business and other areas, the collected data are qualitative by
nature (that is, attributes), classified into different categories or groups
according to one or more attributes. Such types of data are known as
“categorical data”. For example, the number of persons in a sample may be
categorized into different categories of their age, income, job, etc. Then a
question arises “how the inference problems, particularly, the problems of
testing of hypothesis, arising out of categorical data can be tackled?” In this
unit, we shall see that chi- square test, one of the sampling distributions, is
helpful for such problems. We shall discuss here two most widely used tests
of chi-square.
Sometimes, it is required to test whether the variances of two different
populations are equal or not; similar to the situation of testing the equality of
population means discussed in the previous unit. We shall discuss this
problem and will see how the F-distribution can be applied for this purpose.
11.2 OBJECTIVES
After studying this unit, you should be able to:
define the meaning of small sample procedures of testing the
hypothesis;
describe the small sample test procedures for single population mean
and difference of two population means on the basis of t-distribution;
describe the concept and procedure of paired t-test;
explain the chi-square test for the goodness of fit and independence
of attributes; and
describe the test for the equality of two variances using F-
distribution.
11.3 PROCEDURE FOR SMALL SAMPLE TESTS
Throughout in this unit unless stated otherwise, we shall assume that the
random sample x1, x2,…,xn is a sample of size n (where n < 30) selected from
a population having parameter ( which is unknown and to be tested). For
developing a test procedure for the parameter θ based upon the small sample,
we follow some steps, which are more or less similar as already mentioned in
290
the previous unit under Section 3.4. For readiness of the material, we briefly Statistical Analysis-II
present here the steps:
Step I: Formulation of null hypothesis H0, to be tested against the
alternative hypothesis H1. For example, H0: = 0 against
alternative H1: 0 or H0: = 0 against H1: >0 (or <0), etc.
Step II: Decision about the level of significance, α. Generally, we take =
5% (or 1%) as the case may be desired.
Step III: Depending upon the nature of the null and alternative hypotheses,
selection of an appropriate test statistic for testing H0, and
computation of the test statistic on the basis of the random sample
of size n.
Step IV: Depending upon the assumption of the null hypothesis, selection
of an appropriate sampling distribution which the test statistic has
to follow. The sampling distribution to be used for deciding the
critical region (either one-tailed or two-tailed, as the case may be)
is always known as soon as we choose our test statistic and
formulate our null and alternative hypotheses.
Step V: Using the distribution and value of , obtain the critical value (or
cut-off value) in that sampling distribution.
Step VI: Take the decision for rejecting or not rejecting H0, in the light of
step I and II, in the same way as mentioned in Step VI of Section
3.4, which is as follows:
(i) Case-I: When we take H1: 0 then we reject H0 if modulus of
calculated value of test statistic is greater than or equal to the critical
value (or tabulated value) at level of significance. We do not reject H0
if modulus of calculated value of test statistic is less than the critical
value (or tabulated value) at level of significance.
(ii) Case-II: When we take H1: >0 or H1: <0 then step VI is different
and we reject H0: =0 if test statistic is greater than the critical value.
No modulus sign is used for this case for the test statistic. Similarly, if
H1: <0, we reject H0: =0 when the test statistic is less than the
negative of the critical value at α level of significance.
CHECK YOUR PROGRESS 1
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
1. Describe the procedure of testing of hypothesis for the small sample.
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
291
Statistical Analysis 11.3.1 Test for Population Mean
In some cases, we attempt to test the hypothesis regarding the significance of
a population mean when the sample size is small (n<30) and population
variance σ2 is unknown.
For example, for Boarding schools, one can assume that the average score in
mathematics paper is 55 (that is, µ0= 55; μ0 being the assumed population
mean) or average weight of an Indian army person is 60 kg (that is, µ0 = 60).
Our aim here would be to test whether the assumed mean in these populations
is same as given or not, where populations are respectively of school
children and of army persons.
To test the significance of given population mean, as usual, we assume that a
random sample x1, x2,…, xn of size n from the given population (assuming
normal) has been drawn. For testing purpose, we set up the hypotheses H0
(null) and H1(alternative) as
H 0 : 0 against H1 : 0 or H1 : 0 or H1 : 0
Let X be the statistic computed on the basis of the random sample x1, x2,…,
xn; E(X) and V(X) respectively be the mean and variance of X, then under
the null hypothesis, the test- statistic t is given by
X E(X) x μ0
t ~ t (n 1)d.f.
VX s/ n
x i 1
and s 2 x i x 2 is an
where, E(X) is sample mean x
n n -1
unbiased estimate of the unknown population variance σ2. It is well known
that statistic t follows a t- distribution with (n-1) degrees of freedom (df).
Substituting the values of x, μ 0 , s and n in the expression of t, we find the
calculated value of t. Now we search the critical value (or cut-off value or
tabulated value) of the test statistic t from the t-table at (n-1) df at α level of
significance. Comparing these two values we take the decision as:
(A) One -Tail Test:
Case I: When H 0 : μ μ 0 against H1 : μ μ 0 (right-tail test), H0 is rejected
if calculated value of t > tabulated t at (α, (n-1) df).
Case II: When the null hypothesis is H0: µ = µ0 against H1: µ < µ0 (left-tail
test), H0 is rejected if calculated t< −tabulated t at (α, (n-1) df).
(B) Two-Tail Test: When H 0 : μ μ 0 against H1 : μ μ 0 , H0 is rejected if
calculated value of t > tabulated t at (α/2, (n-1) df).
Example 1: A hand wash manufacturing company was selling a particular
type of brand-hand wash through a large number of retail shops. Before an
292 advertising campaign, the mean sale per weak per shop was 60 dozen. After
the heavy campaign to judge the impact, a random sample of 16 shops was Statistical Analysis-II
taken and the mean sale was found to be 64 dozen with standard deviation 6.
Can you consider the advertisement campaign effective?
Solution: Given that
x 64, μ 0 60, s 6 and n 16
For testing H0: =0 (=60) against H1:> 60 dozens (right- tail test) the test
statistic t is:
x μ 0 64 60 4
t 2.67
s/ n 6/ 16 1.5
The critical value (or tabulated value) of t for (n-1) = 15 df at 5% level of
significance [(one- tail right- hand test)] is t15,0.05 = 2.602 (obtained from the
t- table). Since the calculated value is greater than tabulated value, we reject
H0 at 5% level of significance, that is, H1is accepted and we conclude that
advertising campaign is effective for increasing the bath-soap sale in the
market.
Example 2: The mean Share price of Real estate companies is Rs.68. After a
month, these Share prices of all companies have been changed. A sample of
10 Pharma companies was taken and their share prices were noted as: 70, 76,
75, 69, 70, 72, 68, 65, 75, 72. Test, whether mean share price is still the
same?
Solution: Given sample size n = 10 (< 30), so it is a small sample case and
we wish to test the null hypothesis that the population mean is still 70.0
besides all changes, that is,
H0: =0 =68 against H1:≠ 0 = 68
The test statistic is
x μ0
t
s/ n
For calculating the value of t, we prepare the following table:
S. No. Sample value Deviations d2
(x) d= (x -a), with a=70
1 70 0 0
2 76 6 36
3 75 5 25
4 69 -1 1
5 70 0 0
6 72 2 4
293
Statistical Analysis 7 68 -2 4
8 65 -5 25
9 75 4 16
10 72 2 4
Total 11 115
The assumed value is 70, so we get
x 70
d 70 11 70.1
n 10
2 1 2
d
2
1 2
11 1 121
s d 115 115 11.43
n -1 n 10 1 10 9 10
s 11.43 3.38
It is a two-tail test and the test statistic is
x μ0 70.1 68 2.10
t 1.962
s/ n 3.38/ 10 1.07
The tabulated value of t for n -1 = 9 df and at 5% level of significance, that is,
t9,0.025 is 2.262.
Since the calculated value of t is smaller than the tabulated value, we do not
reject H0 at 5% level of significance and, therefore, conclude that mean prices
of the Shares in the given month is still 68 besides many ups and downs.
CHECK YOUR PROGRESS 2
Note: i) Check your answers with those given at the end of the unit.
2) Describe the procedure of small sample test for single population mean.
3) An automobile tyre manufacturer claims that the average life of a
particular category of his tyre is 18000 kms when used under normal
driving conditions. A random sample of 16 tyres was tested. The mean
and S.D. of life in the sample were 20,000 and 60,000 kms.,
respectively. Assuming the life of the tyre in kms to be normally
distributed, decide whether the manufacture’s claims are true at 1%
level of significance.
4) A random sample of weights in kg of10 students of IGNOU are given
as 48, 50, 62, 75, 80, 60, 70, 56, 52, 78. Can we say that mean of the
distribution of weights of all students of IGNOU from which the above
sample was taken is equal to 60 kg.
294
11.3.2 Tests for Equality of Two Population Means Statistical Analysis-II
Suppose we have to compare means of two groups when sample sizes taken
from these two groups are small. For example, performance of English-
medium and Hindi-medium school children in mathematics paper are to be
compared or performance of two brand-promotion campaigns of a product in
two different markets are to be compared.
Let us assume that two independent random samples x1, x 2 , ..., x n1 and
y1 , y2 , ..., yn2 of sizes n1 and n2 respectively are drawn from two normal
populations N(1,σ12) and N(2,σ22). Further, suppose the variances of the
both populations are unknown but are equal, i.e., σ12 = σ22 = σ2 (say) where σ2
is an unknown value.
In this situation, we want to test the null hypothesis
H 0 : 1 2 against H1 : 1 2 or H1 : 1 2 or H1 : 1 2
We calculate the statistic t as (assuming H0 true):
xy
t ~ t ( n 1n 22 )
1 1
s
n1 n 2
Here, x and y are the means of first and second sample respectively and
2 2
s2
n 1 1s12 n 2 1) s 22 where, s12
x x and s 22
y y
n1 n 2 2 n1 1 n 2 1
After substituting values of x, y, s, n1 and n 2 in the expression of t, we obtain
the calculated value of t. Then we look for critical value (or cut-off value or
tabulated value) of test statistic t from the t-table. On comparing these two
values, we take the decision as below:
[A] One-Tail Test:
Case I: When H 0 : μ 1 μ 2 against H1 : μ 1 μ 2 (right-tail test), H0 is
rejected if calculated value of t > tabulated t at (α, (n1 + n2-2) df).
Case II: When H 0 : μ 1 μ 2 against H1 : μ 1 μ 2 (left-tail test), H0 is rejected
when if calculated value of t < −tabulated t at (α, (n1 + n2-2) df).
[B] Two-Tail Test: When H 0 : μ 1 μ 2 against H1 : μ 1 μ 2 , H0 is rejected if
calculated value of t > tabulated t at (α/2, (n1 + n2-2) df).
Example 3: In a random sample of 10 pigs fed by diet A, the increase in
weight in pounds in a certain period were 12, 8, 14, 16, 13, 12, 8, 14, 10, 9
lbs. In another random sample of 10 pigs fed by diet B, the increase in
weights in the same period were14, 13, 12, 15, 16, 14, 18, 17, 21, 15 lbs. Test 295
Statistical Analysis whether diets A and B differ significantly regarding their effect on increase in
weight.
Solution: Obviously here the hypothesis of no difference in mean weights of
pigs under two diets will be tested, that is,
H0: 1 = 2 against H1: 1 ≠ 2.
In order to find the value of t-statistic, we prepare the following table:
Diet A Diet B
x d1 = (x-a); with d12 y d2 = (y-b); with d22
a=12 b=16
12 0 0 14 -2 4
8 -4 16 13 -3 9
14 2 4 12 -4 16
16 4 16 15 -1 1
13 1 1 16 0 0
12 0 0 14 -2 4
8 -4 16 18 2 4
14 2 4 17 1 1
10 -2 4 21 5 25
9 -3 9 15 -1 1
2 2
d 1 4 d 1 70 d 2 5 d 2 65
We calculate the test statistic t:
xy
t
1 1
s
n1 n 2
Here, a=12, b= 16 are assumed means
xa
d 1
12
4
11.6, yb
d 2
16
5
15.5
n1 10 n2 10
2 1
d 12
d 1
2
2
d 2
2
s d 2
n 1 n 2 2 n1 n2
2 2
1 4 5 68.4 62.5
70 65 7.27
10 10 2 10 10 18
s 7.27 2.70
296
Putting these values in test statistic t, we get Statistical Analysis-II
11.6 15.5 3.90 3.90
t 3.23
1 1 2.70 0.45 1.21
2.70
10 10
Also, calculated t 3.23 and tabulated t n1 n 2 2,α/2 t 18 , 0.025 2.101
The calculated value of t is greater than the tabulated value so H0 is rejected
at 5% level of significance and we conclude that diets A and B differ
significantly in terms of increase in weights of pigs. Definitely any one diet
performs better to other significantly.
Example 4: The means of two random samples of sizes 10 and 8 are 210.40
and 208.42 respectively. The sum of squares of the deviations from means is
26.94 and 24.50, respectively. Can samples be considered to have been drawn
from the normal populations having equal mean.
Solution: Given that
n1 10, n 2 8, x 210.40, y 208.92,
n1 n2
2 2
x i x 26.94,
i 1
y
i 1
i y 24.50,
So
1 1
s2
n1 n 2 2
x i
2
x y i y
2
16
51.44 3.215
s 3.215 1.79
We wish to test the null hypothesis that both the samples are drawn from the
normal populations having the same mean. Let μ1 and μ2 be means of the first
and second populations respectively, then the null hypothesis will be
H0: 1 = 2 against H1: 1 ≠2.
It is a two- tail test and we calculate the test-statistic t as
xy 210.40 208.92 1.48 1.48
t 1.74
1 1 1 1 1.79 0.47 0.85
s 1.79
n1 n 2 10 8
Also, tabulated value t n 1 n 2 2 , α/2 t 16 , 0.25 2.12
The calculated value of t is smaller than the tabulated value so we do not
reject H0 at 5% level of significance. We, therefore, conclude that both
samples are from the normal populations having same means.
297
Statistical Analysis CHECK YOUR PROGRESS 3
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
5) Describe the procedure of small sample test for difference between
two population means.
6) Two different types drugs A and B were tried on some patients for
increasing their weights. The 6 persons were given drug A and other 7
persons were given drug B. The increase in weight (in ponds) is given
below:
Drug A 5 8 7 10 9 6 -
Drug B 9 10 15 12 14 8 12
Do the two drugs differ significantly with regard to their mean weight
increment?
7) To test the effect of fertilizer on wheat production, 26 plots of land
with equal areas were chosen. Half of these plots were treated with
fertilizer and the other half were untreated. Other conditions were the
same. The mean yield of wheat on the untreated plots was 4.6 quintals
with a standard deviation of 0.5 quintals, while the mean yield of the
treated plots was 5.0 quintals with standard deviations of 0.3 quintal.
Can we conclude that there is significant improvement in wheat
production because of the effect of fertilizer at 1% level of
significance?
11.3.3 Paired t-Test
In the sub-section 11.3.2, for testing the difference of two population means,
we assumed that the populations are different from which the samples are
drawn independently. Sometimes, we may come across with the problems of
testing the equality of means within the same population when two samples
at different times or occasions are taken on same units.
For instance, if we wish to test the effectiveness of a new diet marketed for
reducing weight, the weights of a small group of individuals (say, 15
individuals) may be recorded before and after administering the diet which
will give rise to two samples of same size. Here the observations may be
grouped into pairs for each individual. The hypothesis to be tested is again
same as considered in sub-section 4.2.2, that is, there is no effect of
administering the diet, or equivalently, that the average weights before and
after administering the diet are same. The sample values taken before and
after administering the diet (denoted by xi and yi respectively for i = 1, 2, …,
n) may be put in the following manner for n individuals:
Sample I: x1, x2, x3, … xn
298 Sample II: y1, y2, y3, … yn.
Whenever we face such type of experiments for testing the equality of two Statistical Analysis-II
population means, we apply t-test but due to paired observations, the test is
called a paired t- test. The procedure of testing under paired t-test is described
below:
Using the notations of sub-section 4.2.2, we wish to test
H0: 1 = 2 or D 1 2 0
for which the test statistic t is
d
t ~ t n 1d.f.
sd / n
1
where d i x i y i i 1, 2, ..., n and similarly , d d i x y
n
2
s
1
d d 2
1 2
d i
di
2
d i
n 1 n 1 n
We need to calculate only Σdi , Σdi2 for obtaining the value of t. Then we
obtain the critical value (or cut-off value or tabulated value) of test statistic t
from the t-table.
The methods of taking decision about the rejection of H0 or not rejecting H0,
on the basis of calculated and tabulated values of t-statistic, are exactly
similar as mentioned in sub-section 4.2.2 for H1: μ1 > μ2 ; μ1 < μ2 and μ1 ≠ μ2
except that here degrees of freedom (df) is (n – 1) instead of (n1 + n2-2).
Note: Sometimes data are recorded in the form of increments/decrements
only, like; weight of child increased by 2 kg, marks increased by 3%,
temperature reduced to 30C ,etc. Then actually for n objects, this set of data
show [d1 , d2 ,d3 , .., dn]. Accordingly, our hypothesis will be
H 0 : D 0 against H1 : D 0 or H1 : D 0
The sample with larger mean may be treated as the first sample and hence is
denoted by x. Therefore, this has to be checked before attempting the paired-
t test that whether values (xi, yi) are given or increments/ decrements are
given in the data set.
Note: In good practice, if we talk about increments then we take di = (yi- xi),
and alternative hypothesis H1 : D 0 ( one sided right-tail test). For
decrements we take di = (xi-yi) and same alternative H1 : D 0
Example 5: A group of 12 children was tested to find out how many digits
they would repeat from memory after hearing them once. They were given
practice session for this test. Next week they were retested. The results
obtained were as follows:
299
Statistical Analysis
1 1
Child no. 1 2 3 4 5 6 7 8 9 10
1 2
Recall
6 4 5 7 6 4 3 7 8 4 6 5
Before
Recall
6 6 4 7 6 5 5 9 9 7 8 7
After
Can the memory practice session improve the performance of children?
Solution: We set up the null hypothesis that the children have no
improvement after the practice session, that is, no effect due to practice or
both populations before and after practice have same mean,
So, we have H0: 1 = 2 against H1: 1<2 (one sided left- tail test)
The 1 and 2 stand for the average before and after the memory practice
session of test respectively. In this example, same 12 children are tested,
hence, for testing H0, we use paired t-test. For computing t value, we do the
following calculations:
Child Digit Recall d = (x-y) d2
No. Before (x) After (y)
1 6 6 0 0
2 4 6 -2 4
3 5 7 -2 4
4 7 7 0 0
5 6 8 -2 4
6 4 5 -1 1
7 3 5 -2 4
8 7 9 -2 4
9 8 9 -1 1
10 4 7 -3 9
11 6 8 -2 4
12 5 7 -2 4
2
d 19 d 35
d
d 19
1.58
n 12
1 1 di
2
d i d n 1 d i n
2 2 2
s
d
n 1
300
2 Statistical Analysis-II
1
35
19 1
4.92 0.45
11 12 11
s d 0.45 0.67
Substituting the values, we have test statistic t as
1.58 1.58
t 8.17
0.67/ 12 0.1934
The critical (tabulated) value of t for 11 df at 5% level of significance is
t11,0.05 = -1.796 (one-sided left-test). Since calculated value of t is smaller than
the tabulated value, so H0 is rejected at 5% level of significance and we
conclude that the children have improvement after the memory practice
session.
CHECK YOUR PROGRESS 4
Note: i) Check your answers with those given at the end of the unit.
8) Describe the procedure of paired t-test for small sample.
9) To verify whether a course “Post Graduate Certificate in Climate
Change (PGCCC)” improved performance, a similar test was given to
10 participants both before and after the course. The original marks
out of 100 (before course) recorded in a alphabetical order of the
participants are 42, 46, 50, 36, 44, 60, 62, 43, 70 and 53. After the
course the marks are in the same order 45, 46, 60, 42, 60, 72, 63, 43,
80 and 65. Test whether the course PGCCC was useful?
11.4 CHI-SQUARE TESTS
In the previous units, we discussed parametric tests in which we first assumed
the distribution of the parent population and then performed a test about some
parameter(s) of the population(s) such as mean, variance, proportion, etc.
Generally, the parametric tests are based on the assumption that the parent
population is a normal population and, thus, involves one or more parameters
for its complete specification. In this section, we shall discuss about some
χ 2 (chi-square) tests. Some chi square tests are generally known as non-
parametric tests, that is, we do not presume that the given set of data have
come from any specific distribution (particularly from a normal distribution).
Thus, there is no question of involving a parameter in the test.
11.4.1 Test for Goodness of Fit
The chi-square ( χ 2 ) test for goodness of fit was given by Karl Pearson in
1900, which is the oldest non-parametric method of testing of hypothesis. In
this test, given a set of observed frequencies for values of a discrete variable,
we test whether the data follows a specific probability distribution or not.
301
Statistical Analysis Under the assumption of a particular distribution for the given data (as for
example, binomial, Poisson, etc.), which is considered to be the null
hypothesis to be tested; expected frequencies for the given variable values are
obtained which would be expected to occur if the data follows the assumed
distribution. This test is known as “goodness of fit test” since the aim is to
observe how close are the given values of the variable to the frequencies
yielded by the assumed probability distribution, or, in other words, we judge
the fitness of the data to the assumed theoretical distribution. This is a non-
parametric test since we do not assume any specific probability distribution
beforehand for the data.
Assumptions for the test:
1. Sample observations are independent.
2. The measurement scale is at least nominal.
3. The observations may be classified into non-overlapping categories.
Let a random sample of size N be drawn from a population with unknown
distribution of k characteristics and the data categorized into k groups or
classes. Also, let O1, O2, …, Ok be the observed frequencies (as given in the
data set) and E1, E2, …,Ek be the corresponding expected frequencies as are
expected if the data follows the assumed distribution. We generally put a
linear constraint of equality on the observed and expected frequencies which
is O i E i .
To perform the chi- square goodness of fit test the steps are as follows:
Step 1: First of all, we form the null and alternative hypothesis. Null
hypothesis is that the given data set follows the given probability distribution
or the pattern specified in the question. Alternative hypothesis contradicts this
assumption.
Step 2: For computing expected frequencies corresponding to given values of
the variable, for the assumed distribution (mentioned in the null hypothesis),
we compute those required sample statistic(s) using the sampled data which
are necessary to obtain an estimate of the corresponding population
parameter(s) and assume that it is equal to the theoretical value of the
corresponding parameter. This, in fact, provides us an estimated value of the
parameter which is generally not known in most of the cases. However, if
value(s) of parameters are given, they can be used directly for the purpose
and there is no need to estimate it on the basis of the given sample data. For
instance, let mean (and/or variance) of the theoretical distribution is needed
for calculating the expected frequencies, we compute mean (and/or variance)
of the sample for this purpose.
Step 3: In the next step, we find the probability that an observation falls
(belongs) to a particular category (to a particular value of the variable) using
the assumed probability distribution.
302
Step 4: The calculated probabilities are then used to find the corresponding Statistical Analysis-II
expected frequencies using the result
Ei Npi ; for all i =1, 2, …, k
Step 5: For testing the null hypothesis, compute the test statistic
2
k
Oi Ei 2 ~ 2k r 1
i 1 Ei
The 2 statistic follows chi-square distribution with (k –r –1) degrees of
freedom where, r is the number of parameters estimated from the sample.
Step 6: Now obtain the tabulated value of χ2 at the specified level of
significance and (k-r-1) degrees of freedom from the table and compare it
with the calculated value and take the decision as follows:
If computed value of test statistic is greater than tabulated value with
k r 1 degrees of freedom and at α level of significance, we may reject
the null hypothesis, otherwise we shall not reject it.
Note: Table 5 in Appendix provides the critical value of chi-square
distribution at α level of significance.
Example 6: The following data are collected during a test to determine
consumer preference among five leading brands of soaps:
Brand
A B C D E
Preferred
Number of
194 205 204 196 202
Customers
Test the hypothesis that the preference is uniform over the five brands at the
5% level of significance.
Solution: Here, we are interest to test the null hypothesis that the preference
of customers over five brands is uniform, that is, all brands are equally
preferred by customers. Therefore, we have
H0: The preference of customers over the five brands of bath soap is uniform
H1: The preference of customers over the five brands of bath soap is not
uniform
Here, equivalently we can say that a uniform distribution is assumed for the
data., that is, the proportions of customers for each brand is same. In other
1
words, p1 p2 p3 p4 p5 p
5
where pi ( i = 1, 2,…, 5) is the proportion of customer for the ith brand.
For testing the null hypothesis, the test statistic χ2 as mentioned above.
303
Statistical Analysis The theoretical or expected number of customers or frequency for each brand
is obtained by multiplying the appropriate probability by total number of
customers, that is, sample size N. Therefore,
1
E1 E 2 E 3 E 4 E 5 Np 1000 200
5
For calculating the value of test statistic, we prepare the following table:
Brand Observed Expected (O−E) (O−E)2 O E 2
Frequency Frequency E
(O) (E)
A 194 200 -6 36 0.18
B 205 200 5 25 0.13
C 204 200 4 16 0.08
D 196 200 −4 16 0.08
E 202 200 1 1 0.01
Total 1000 1000 0.48
Therefore, the calculated value of the test statistic is 0.48.
k
χ2
O i E i 2 0.48
i 1 Ei
The tabulated value of chi-square with k 1 5 1 4 degrees of freedom
and 5% level of significance is 7.78. Since computed value of test statistic is
less than tabulated value 2 at 5% level of significance, therefore, we do not
reject the null hypothesis at and conclude that the preference of customers
over the five brands of bath soap is uniform.
CHECK YOUR PROGRESS 5
Note: i) Check your answers with those given at the end of the unit.
10) Describe the chi-square goodness of fit test.
11) The following table gives the numbers of aircraft accidents that
occurred during the various days of the week:
Days Mon Tue Wed Thu Fri Sat Sun
Number of Accidents 14 15 8 20 11 9 14
Test whether the accidents are uniformly distributed over the week.
11.4.2 Test for Independence of Two Attributes
In the analysis of research data, the investigator often needs to decide
whether two variables or attributes are related (associated). For example, a
sociologist may wish to know whether level of formal education is associated
304
with income, a quality-control engineer may be interested to know whether Statistical Analysis-II
the number of defective items produced on the various production lines is
independent of the day of the week, a doctor may wish to know whether a
drug is effective in curing colds, etc. If there is no association between two
variables, we say that they are independent. In other words, we can say that
two variables are independent if the distribution of one is not depending on
the distribution of other. Similarly, in case we are given two attributes instead
of two variables and we want to test the independence of two attributes, we
may use chi-square test for independence. This test will indicate only whether
or not any association exists between attributes. The null hypothesis in such a
test is always that the attributes are independent against the alternative
hypothesis that they are not independent. Symbolically, we write
H0: The two attributes are independent;
H1: The attributes are not independent.
To conduct the test, a sample is drawn from the population and the observed
frequencies are cross-classified according to the two criteria so that each
observation belongs to one and only one level of each criterion. The cross-
classification can be conveniently displayed by mean of a table called a
contingency table. Therefore, a contingency table is an arrangement of data
into a two-way classification. One of the classifications is entered in rows and
the other in columns.
Assumptions:
1. The sample of N observation is random sample.
2. Each observation in the sample may be classified according to two
criteria so that each observation belongs to one and only one level of each
criterion.
Let the N observations are classified into the following type of contingency
table where we have p levels of the attribute A and q levels of the attribute B.
Here Oij ( i = 1, 2,…, p and j = 1, 2,…, q) represents the observed frequency
of the cell (i, j).
Table
B B1 B2 … Bj … Bq Total of
rows
A
A1 O11 O12 … O1j … O1q R1
A2 O21 O22 … O2j … O2q R2
. . . . . .
. . . . . .
. . . . . .
Ai Oi1 Oi2 … Oij … Oiq Ri
. . . . . .
. . . . . .
. . . . . .
Ap Op1 Op2 … Opj … Opq Rp
Total of C1 C2 … Cj … Cq N
columns 305
Statistical Analysis To calculate the expected frequencies corresponding to each Oij, we use the
multiplication law of probability. According to this law if two events are
independent then the probability of their joint occurrence is equal to the
product of their individual probabilities. Therefore, if two criteria of
classification of observations in the contingency table are independent, the
joint probability for a cell is equal to the product of the two corresponding
marginal probabilities. Since, row totals are showing the marginal
probabilities P(Ai); i = 1, 2,…,p and column totals are showing probabilities
P(Bj); j = 1, 2,…, q; the probability in the cell (i., j) is equal to P(Ai)P(Bj).
Clearly, we know that P(Ai) = Ri/N and P(Bj) = Cj/N.
Thus, we have
Ri Cj
P AiB j .
N N
To obtain expected frequency Eij for each cell, we multiply this estimated
probability by the total sample size. Thus,
Ri Cj
Eij N. .
N N
which reduces to
R i C j Sum of i th row Sum of jth column
E ij
N Total sample size
Thus, it is evident that calculation of expected frequencies is quite simple.
Once each Eij is calculated, for testing the null hypothesis H0, we use the
same χ2- statistics, which is
2
p q
O ij Eij
2
~ 2p 1q 1
i 1 j1 Eij
Here the test statistic follows as a chi-square distribution with (p−1) (q−1)
degrees of freedom.
The decision about rejecting or not rejecting the null hypothesis of
independence of the two attributes against the alternative hypothesis is
exactly same as mentioned in the sub-section 11.4.1 above.
Example 7: Calculate the expected frequencies for the following data
presuming the two attributes condition of child and condition of home and
check whether these are independent.
Condition of Child Condition of Home
Clean Dirty
Clear 75 45
Fairly clean 85 15
306 Dirty 40 40
Solution: Here, we want to test the null hypothesis that condition of home Statistical Analysis-II
and condition of the child are independent, that is,
H0: Condition of child is independent of condition of home against
H1: Condition of child depends upon the condition of home.
For the calculation of χ2 statistic, we do the following computations for
expected frequencies:
Condition of Condition of Home Total
Child
Clean Dirty
Clear 75 45 120
Fairly clean 85 15 100
Dirty 40 40 80
Total 200 100 300
Under H0, Eij’s are obtained as
R 1 C1 120 200 R C 2 120 100
E11 80; E 12 1 40;
N 300 N 300
R 2 C1 100 200 R C 2 100 100
E 21 66.67 67; E 22 2 33.33 33;
N 300 N 300
R 3 C1 80 200 R C 2 80 100
E 31 53.33 53; E 32 3 26.67 27;
N 300 N 300
For calculating the value of χ2 test statistic, we prepare the following table:
Observed Expected (O – E) (O – E)2 O E 2
Frequency Frequency E
(Oij) (Eij)
75 80 −5 25 0.3125
45 40 5 25 0.6250
85 67 18 324 4.8350
15 33 −18 324 9.8180
40 53 −14 169 3.1880
40 27 14 169 6.2590
Total 25.0375
Therefore,
2
χ
p q O ij E ij
2
25.0375
i 1 j1 E ij
307
Statistical Analysis The degrees of freedom will be (p –1)(q –1) = (3 – 1)(2 – 1) = 2.
The tabulated value of chi square at 2 df and 5% level of significance is 5.99.
Since calculated value of test statistic is greater than tabulated value, we
reject our null hypothesis and conclude that condition of home does affect the
condition of child.
CHECK YOUR PROGRESS 6
Note: i) Check your answers with those given at the end of the unit.
12) Discuss chi-square test for testing independence of attributes.
13) 1500 families were selected at random in a city to test the belief that
high income families usually send their children to public schools and
low-income families often send their children to government schools.
The following results were obtained in the study conducted.
Income Public School Government Total
School
Low 300 600 900
High 400 200 600
Total 700 800 1500
Use chi-square test at 1% level to state whether the two attributes are
independent.
11.5 F-TEST
11.5.1 TEST FOR EQUALITY OF TWO VARIANCES
While applying t-distribution for testing H0: 1 = 2, the basic assumption
was: the population variances of the two populations are equal. But, in fact, it
is a very rare case in practice. Therefore, sometimes it is necessary to test the
equality of population variances first. The F distribution is used for
ascertaining the equality of two population variances. We shall describe here
the F-statistic for this purpose which follows the F distribution.
Let x 1 , x 2 , ..., x n1 be a sample of size n1 from a normal population with
mean 1 and variance σ12.Similarly, let y1 , y 2 , ..., y n 2 be a sample of size n2
from another normal population with mean 2 and variance σ22. We will
show how the two samples are used for defining F-statistic. According to the
objective of the test, the null hypothesis will be
H 0 : σ12 σ 22
against the alternative hypothesis
H1 : σ12 σ 22 or H1 : σ12 σ 22 or H1 : σ12 σ 22
308
To test H0, we calculate the test statistic Statistical Analysis-II
s12
F 2 ~ F(n1 1, n 2 1) ; where
s2
1 1
s12 x x 2 and s 22 y y 2 .
n1 1 n 2 1
The sampling distribution of F-statistic so defined, follows a F-distribution
with ν1 n 1 1 and ν 2 n 2 1 degrees of freedom. As a norm, larger
variance is taken in the numerator of F-ratio and the degree of freedom
corresponding to this variance is taken as ν 1 .The sample with larger s2 is
treated as the first sample s12 with population variance σ12.
Using the calculated of F-statistic, we take the decision about H0 as:
[A] For One-Tail Test: (i) When H 0 : σ12 σ 22 with H1 :σ12 > σ 22
H0 is rejected when calculated value of F Fn1 1,n 2 1 α , the tabulated
value, at α% level of significance and ν1 and ν2 df, otherwise, H0 is not
rejected.
(ii) When H 0 : σ12 σ 22 with H1 : σ12 σ 22
H0 is rejected when calculated value of F Fn1 1,n 2 1 1 α , not rejected
otherwise.
[B] For Two -Tail Test: When H 0 : σ12 σ 22 and H1 : σ12 σ 22
H0 is rejected when calculated value of F Fn11,n 2 1 α/2 or
F Fn1 1,n 2 1 1 α/2 , otherwise not rejected.
Example 8: Two random samples gave the following results:
Size Mean Sum of Squares of deviation
from the Mean
Sample I 9 59 26
Sample II 11 60 32
Test whether both samples are from the same normal population?
Solution: Since we have to test whether both the samples are from same
normal population, therefore, we will test two things separately:
(i) The equality of two population means,
(ii) The equality of two population variances.
That is H 0 : 1 2 first and then test H 0 : 12 22 309
Statistical Analysis The equality of two means will be tested using t-test whereas equality of two
variances will be tested using F-test. But since t-test is based on the prior
assumption that both population variances are same, therefore, first we apply
F-test and later the t-test (when F-test accepts equality of variance
hypothesis).
Given that n 1 9, x 59, x x 2
26
n 2 11, y 60, y y 2
32
we get
1 1 1
s12 x x 2 26 3.25 s 22 y y
2
n1 1 9 1 n 2 1
1
36 3.60
11 1
s2
For H 0 : 12 22 , H1 : 12 22 (right-tail test), the test statistic is: F 12
s 2
2 2 s 22
Since s s therefore we take reverse F
2 1 ~ Fn 2 1,n1 1 = (3.60/3.25) =1.1
s12
The tabulated values at 1% and 5% level of significance are F10, 80(.01) = 3.34
and F10, 8(0.05) = 5.81 (one sided left-tail test) respectively. Since calculated F
is smaller than tabulated F at both level of significance; H0 is not rejected
and we conclude that the variances of two populations are same. Now,
applying t-test, the equality of two population means can be tested as
discussed earlier in 4.2.2.
Example 9: The following data relate to the number of items produced in a
shift by two workers A and B for some days:
A 26 37 40 35 30 30 40 26 30 35 45
B 19 22 24 27 24 18 20 19 25
Can it be inferred that A is more stable (or consistent) worker compared to
B?
Solution: We wish to test the null hypothesis that worker A and B are
equally efficient in terms of stability of item production. Let σ12 and σ22 be
the variances for the two workers A and B respectively. Then our hypothesis
will be
H 0 : σ12 σ 22 against H1 : σ12 σ 22
For calculating s12 and s 22 , we first find two means which are found to be 34
and 22 respectively. Then we prepare following table:
310
Items x x x x 2 Items y y y y 2 Statistical Analysis-II
produced produced by
by A x 34 B (variable y 22
(variable x) y)
26 -8 64 19 −3 9
37 3 9 22 0 0
40 6 36 24 2 4
35 1 1 27 5 25
30 -4 16 24 2 4
30 -4 16 18 −4 16
40 6 36 20 −2 4
26 -8 64 19 −3 9
30 -4 16 25 3 9
35 1 1
45 11 121
Total 374 0 380 198 0 80
The test statistic is
s12
F
s 22
where s12 and s 22 are as same as given above.
Using the values obtained in the table, we have
1
s12 x x 2 1 380 38 s 22 1 y y 2 1 80 10
n1 1 10 n 2 1 8
Therefore, for one sided right tailed test
38
F 3.8
10
The tabulated value of F10, 8 (0.01) = 3.34 at 1% level of significance. Since the
calculated value of F is greater than the tabulated value, we reject the null
hypothesis and conclude that the worker A is more stable than worker B.
CHECK YOUR PROGRESS 7
Note: i) Check your answers with those given at the end of the unit.
14) Two sources raw materials are under consideration by a bulb
manufacturing company. Both sources seem to have similar 311
Statistical Analysis characteristics but the company is not sure about their respective
uniformity. A sample of 12 lots from source A yields a variance 125
and a sample of 10 lots from source B yields a variance of 112. Is it
likely the variance of source A significantly differ to the variance of
source B at significance level = 0.01?
15) Two samples are drawn from two different normal populations.
Sample I 60 65 71 74 76 82 85 87 - -
Sample II 61 66 67 85 78 63 85 86 88 91
Test the equality of two population variances at 1% level of significance.
11.6 LET US SUM UP
In this unit, we discussed some testing procedures for different kinds of
hypotheses based on the assumption that the sample(s) collected were small
in sizes so that the assumption of normality of the data were not fulfilled. The
tests described are as follows:
1. Small sample tests for testing the hypotheses using t-tests for (i)
significance of population mean and (ii) equality of two population
means;
2. Paired t-test when the two samples are selected from the same population
on same number of units with the aim to test the equality of means;
3. Chi-square tests for (i) testing the goodness of fit of a theoretical
probability distribution on a given set of data and (ii) testing the
independence of two attributes with different levels of classification;
4. Test based on F distribution for testing the equality of two population
variances.
11.7 KEY WORDS
Degrees of freedom (df): The number of values free to vary, given one or
more mathematical restrictions.
11.8 SUGGESTED FURTHER READING/
REFERENCES
Witte, R., & Witte, J. (2017). Statistics. Hoboken, NJ: John Wiley & Sons.
11.9 ANSWERS TO CHECK YOUR PROGRESS
1) Please refer to section 11.3
2) Please refer to section 11.4
312
3) We are given that Statistical Analysis-II
0 = 18,000, n = 16, x 20,000 , s = 6,000
and we wish to test the null hypothesis
H0: = 18,000 against H1: >18,000
The test statistic is
x μ 20000 18000 2000
t 1.33
s/ n 6000 / 16 1500
The critical (tabulated) value of t for (n-1) = 15d.f. at 1% level of
significance [(one tail (right)] is t15,0.01 = 2.602. Since the calculated
value is smaller than tabulated value, we do not reject H0 at 1% level
of significance.
4) Given that past value of mean weight is 60 kg. which is like an old
information. We wish to test the null hypothesis that
H 0 : μ μ 0 60 against H1 : μ μ 0 60
The test statistic is:
x μ0
t ~ t (n 1)
s/ n
From the data one can calculate: Mean = (630/10) = 63
1
s2 x x 2 1252 139.11 s 139.11 11.79
n 1 9
The test statistic is
t
63 60 3 0.80
11.79 / 10 37.26
The tabulated value of t for n – 1 = 10–1 = 9 degree of freedom for
5% level or t9,0.025 is 2.262 (two tail test). Since the calculate value of t
is smaller than the tabulated value, we do not reject H0 at 5% level of
significance.
5) Please refer to section 11.5
6) Setting the hypothesis that there is no difference between diets A and
B, i.e., H0: 1 = 2 against H1: 1 ≠2,
we calculate the test statistic:
xy
t
1 1
s
n1 n 2
From the data we may have: 313
Statistical Analysis
x 8
d 1
8
3
7.5 y b
d 2 12 4 11.43
n1 6 , n2 7
2 1
d12
d
1
2
2
d
2
2
s d 2
n1 n 2 2 n1 n2
1
19
32 39 42 17.5 36.71 4.93
6 7 2 6 7 11
s 4.93 2.22
Putting the values in test statistic
7.5 11.43 3.93 3.93
t 3.17
1 1 2.22 0.56 1.24
2.22
6 7
Calculated t 3.17 and tabulated t 11 , 0.025 2.201 (two tail at 5 %)
The calculated value is greater than the tabulated value so H0 is
rejected at 5% level of significance (two tail test).
7) Given that:
n1 13, x 4.6, s1 0.5,
n 2 13, y 5.0, s2 0.3
1 1
s2
n1 n 2 2
x i
2
x y i y
2
n1 n 2 2
n 1s12 n 2 s 22
1 1
13 13 2
13 0.5 13 0.3 4.42 0.18
2 2
24
s 0.18 0.43
We want to test the null hypothesis that the there is no significant
improvement in wheat production due to fertilizer, that is
H0: 1 = 2 against H1: 1 ≠2
We calculate the test statistic
xy 4.6 5.0 0.4
t 2.35
1 1 1 1 0.17
s 0.43
n1 n 2 13 13
Tabulated value t24, 0.005=-2.792 at α=1% . The calculated value is
314 greater than the tabulated value, so we do not reject H0.
8) Please refer to section 11.6 Statistical Analysis-II
9) We set up the null hypothesis
H0: 1 = 2 against H1: 1<2 (one-tail left)
So, to test the given H0, we use paired t-test for which the test statistic
is:
d
t ~ t (n 1)d.f.
sd / n
From the data one can calculate:
d
d 70
7.0
n 10
2 1 2
d 1
2
702 1 300 33.33
s
d d 790
n 1 n 9 10 9
s d 33.33 5.77
Putting the values, we have test statistic
7 .0
t 3.83
5.77/ 10
The critical (tabulated) value of t for 9 df at 5% level of significance
is t9,0.05 = -1.833 (one tail left test). Since calculated value of t is
smaller than the tabulated value, so H0 is rejected at 5% level of
significance.
10) Please refer to section 11.4.
11) Here, we are interest to test that the null hypothesis
H0: The accidents are uniformly distributed over the week
against the alternative hypothesis
H1: The accidents are not uniformly distributed over the week
For testing the null hypothesis, the test statistic is
2
k
Oi Ei 2 ~ 2k 1
i 1 Ei
Since the uniform distribution is one in which all outcomes
considered to have an equal or uniform probability. Therefore, the
probability that the accident occurs in each day is same. Thus,
1
p1 p 2 p3 p 4 p5 p
7
315
Statistical Analysis The theoretical or expected frequency for each day is obtained by
multiplying the appropriate probability by the total number of
accidents, that is, sample size N. Therefore,
1
E1 E 2 E 3 E 4 E5 Np 91 13
7
Therefore, the test statistic is
2
k
Oi Ei 2 7.6923
i 1 Ei
The tabulated value of chi-square with k 1 7 1 6 degrees of
freedom and at 1% level of significance is 16.81. Since computed
value of test statistic is less than tabulated value at 1% level of
significance therefore, we do not reject the null hypothesis.
12) Please refer to section 11.5.
13) Here, we want to test the null hypothesis
H0: Two attributes are independent
against the alternative hypothesis
H1: Two attributes are not independent
For testing the null hypothesis, that test statistic is
2
p q
O ij Eij
2
~ 2p 1q 1
i 1 j1 Eij
Now, under H0, the expected frequencies can be obtained as:
Ri Cj Sum of i th row Sum of jth column
Eij =
N Total sample size
Therefore,
R 1 C1 900 735 R C2 900 765
E11 441, E12 1 459,
N 1500 N 1500
R 2 C1 600 735 R C2 600 765
E 21 294, E 22 2 306
N 1500 N 1500
Therefore,
2
p q
O ij Eij
2
220.98
i 1 j1 Eij
The degrees of freedom will be (p –1)(q –1) = (2 – 1)(2 – 1) =1. The
tabulated value of chi square 1df and α=1% is 6.63. Since calculated
value of χ 2 is greater than tabulated value at 1% level of significance
therefore, we reject our null hypothesis.
316
14) Given that n 1 12, s12 125 ; n 2 10, s 22 112 Statistical Analysis-II
We wish to test the null hypothesis that the variances of source A
and B are same, that is
H 0 : σ12 σ 22 H 1 : σ12 σ 22
against
which means one source is larger to other. The test statistic is
s12
F
s 22
Therefore,
125
F 1.116
112
The tabulated value of F11, 9, (0.01) = 3.10(one tail right) at 1% level of
significance. Since the calculated value of F is smaller than the
tabulated value of F, therefore, we do not reject the null hypothesis.
15) We wish to test the null hypothesis
H 0 : σ 22 σ12 H 1 : σ 22 σ12
against
The test statistic is
s12
F 2
s2
From the data we have obtained:
x
x 600 75 , y
y 770 77
n1 8 n2 10
1
s12 x x 2 1 636 90.86
n1 1 7
1
s 22 y y2 1 1200 133.33
n 2 1 9
Since s 22 s12 therefore we take reverse of the F- statistic
s 22 133.33
F 1.47
s12 90.86
Now the tabulated value of F9,7 (0.05) = 5.21 (one tail right at 5% level).
Since the calculated value of F is smaller than the tabulated value of
F, therefore, we do not reject the null hypothesis.
317
Statistical Analysis
UNIT 12 ANALYSIS OF VARIANCE TESTS
Structure
12.1 Introduction
12.2 Objectives
12.3 Analysis of Variance (ANOVA)
12.3.1 Significance ofAnalysis of Variance
12.3.2 Degrees of Freedom
12.3.3 Uses of ANOVA
12.4 One-way Analysis of Variance (ANOVA)
12.4.1 Basic Assumptions of One-Way ANOVA
12.4.2 Test of Hypothesis in One-Way ANOVA
12.5 Two-way Analysis of Variance (ANOVA)
12.5.1 Basic Assumptions of Two-way ANOVA
12.5.2 Test of Hypothesis in Two-way ANOVA
12.6 Let Us Sum Up
12.7 Key Words
12.8 Suggested Further Reading/References
12.9 Answers to Check Your Progress
12.1 INTRODUCTION
In Unit 4 of this block, we tested the equality of means of two independent
groups by using t-test. Sometimes situations may arise where testing of more
than two means is required. As an example, in crop-cutting experiments it
may be required to test whether under similar conditions the average yield of
some crop in a number of fields is same or not. For obvious reasons, in such
cases, t-test cannot be applied.Generally, for such situations, the technique of
Analysis of Variance (ANOVA) is used, in which the testing of equality of
several means is doe by dividing the population variability into different
components. The usual F-test is used to test the equality of means of several
groups.
As its name suggests, the analysis of variance focuses on variability. It
involves the calculation of several measures of variability, when the total
variability of the population is divided into many components, like,
variability within the smaller sub-groups, variability between the smaller sub-
groups, etc. In other words, ANOVA is a technique which split up the total
variation of data which may be attributed to various “sources” or “causes” of
variation. There may be variation between variables and also within different
levels of variables. In this way, Analysis of Variance is used to test the
318
homogeneity of several population means by comparing the variances Analysis of Variance
Tests
between the sample and within the sample.
In this unit, we shall discuss the one-way as well as two-way Analysis of
Variance. One-way Analysis of Variance is a technique where only one
independent variable at different levels is considered which affects the
response variable whereas in two-way Analysis of Variance technique, we
will consider two variables at different levels which affect the response
variables.
12.2 OBJECTIVES
After studying this unit, you should be able to:
understand the Analysis of Variance technique;
describe various types of assumptions underlying the Analysis of
Variance technique and applications of it;
define various types of linear models used in Analysis of Variance
technique;
understand how to test the hypothesis under One-way Analysis of
Variance; and
explain the method of performing Two-way ANOVA Test.
12.3 ANALYSIS OF VARIANCE (ANOVA)
The statistical technique known as "Analysis of Variance"(abbreviated as
ANOVA), was propounded by Professor R.A. Fisher in 1920’s, in which he
developed the method of testing equality of means of several sub-
populations by dividing the total variability into different components.
Variation is inherent in nature, so analysis of variance means examining the
variation present in data or parts of data. In other words, analysis of variance
means to find out the cause of variation in the data. The total variation in any
set of numerical data of an experiment is due to number of causes which may
beAssignable causes and Chance causes.
The variation in the data due to assignable causes can be detected, measured
and controlled whereas the variation due to chance causes is not in the control
of human being and cannot be traced or find out separately.
According to Professor R.A. Fisher, Analysis of Variance (ANOVA) is the
"separation of variance ascribable to one group of causes from the variance
ascribable to other group". So, by this technique, the total variation present in
the data is divided into two components of variation; one due to assignable
causes (Between the Groups Variability) and the other due to chance causes
(Within Group Variability). Analysis of variance technique can be classified
as Parametric ANOVA and Non-Parametric ANOVA.The topic of this unit is
related to parametric ANOVA.
319
Statistical Analysis Parametric ANOVA can be classified as simply ANOVA if only one
response variable is considered. If more than one response variables are
under consideration than it is called multivariate analysis of variance
(MANOVA).
If we consider, only one independent variable which affects the
response/dependent variable then it is called One-way ANOVA. If the
independent variables/explanatory variables are more than one, say n, then it
is called n-way ANOVA. If n is equal to two than the ANOVA is called
Two-way classified ANOVA.
12.3.1 Significance of Analysis of Variance
One obvious question may arise that why do we call it Analysis of Variance,
even though we are testing for the equality of means? Why do not we simply
call it the Analysis of Means? How do we propose test for means by analysis
the variances? As a matter of fact, in order to determine if means of several
populations are equal, we do consider the measure of variance, 2 .
The estimate of population variance 2 is computed by two different
estimates of it each one by a different method. One approach is to compute an
estimator of 2 in such a manner that even if the population means are not
equal it will have no effect on the value of this estimator. This means that the
differences in the values of the population means does not affect the value of
2 as calculated by a given method. This estimator of 2 is the average of
the variances found within each of the samples. For example, if we take 10
samples of size n, then each sample will have a mean and a variance. Then
the mean of these 10 variances would be considered as an unbiased estimator
of 2 , the population variance, and its value remains appropriate irrespective
of whether the population means are equal or not. This is really done by
pooling all the sample variances to estimate a common population variance,
which is the average of all sample variances. This common variance is known
as variance within sample or 2within
The second approach to calculate the estimate of 2 is based upon the Central
Limit Theorem and is valid only under the null hypothesis assumption that all
the population means are equal. This means that in fact, if there are no
differences among the population means, then the computed value of 2 by
the second approach should not differ significantly from the computed value
of 2 by the first approach. Hence, if these two values of 2 are
approximately the same, then we can decide to accept the null hypothesis of
equality of means.
The second approach results in the following computation:
Based upon the Central Limit Theorem, we have previously found that the
standard error of the sample means is calculated by:
X
320 n
or, the variance would be: Analysis of Variance
Tests
2 2
x 2
n or, 2 n x
Thus, by knowing the square of the standard error of the mean x 2 , we
could multiply it by n and obtain a precise estimate of 2 . This approach of
estimating 2 , is known as 2between . Now, if all population means are equal
then, 2between value should be approximately the same as 2between value. A
significant difference between these two values would lead us to conclude
that this difference is the result of difference between the population means.
But, how do we know that any difference between these two values is
significant or not? How do we know whether this difference, if any, is simply
due to random sampling error or due to actual differences among the
population means?
R. A. Fisher developed a Fisher test or F-test to answer the above question.
He determined that the difference between the 2between and 2within could be
expressed as a ratio to be designated as the F-value, so that
2between
F
2within
In the above, case, if the population means are exactly the same, then 2between
will be equal to the 2within and the value of F will be equal to 1.
However, because of sampling errors and other variations, some disparity
between these two values will be there, even when the null hypothesis is true,
meaning that all population means are equal. The extent of disparity between
the two variances and consequently, the value of F, will influence our
decision on whether to accept or reject the null hypothesis. It is logical to
conclude that if the population means are not equal then their sample means
will also vary greatly from one another, resulting in a large value of 2between ,
and hence a larger value of F ( 2within is based only on sample variances and
not on sample means and hence is not affected by differences in sample
means). Accordingly, larger the value of F statistic, the more likely the
decision is to reject the null hypothesis. But how large the value of F be so as
to reject the null hypothesis? The answer is that the computed value of F
must be larger than the critical value of F, given in the table for a given level
of significance and calculated number of degrees of freedom. (The F
distribution is a family of curves, so that there are different curves for
different degrees of freedom.)
12.3.2 Degrees of Freedom
We have talked about the F-distribution being a family of curves, each curve
reflecting the degrees of freedom relative to both 2between and 2within . This
321
Statistical Analysis means that the degrees of freedom are associated both with the numerator as
well as with the denominator of the F-ratio.
Since the variance between samples 2between comes from many samples and if
there are k number of samples, then the degrees of freedom, associated with
the numerator would be (k – 1).
The denominator is the mean variance of the variances of k samples and since
each variance in each sample is associated with the size of the sample (n),
then the degrees of freedom associated with each sample would be (n – 1).
Hence, the total degrees of freedom would be the sum of degrees of freedom
of k sample or df = k(n – 1), when each sample is of size n.
12.3.3 Uses of ANOVA
The following are some of the uses of ANOVA:
1. To test the homogeneity of several means (say, k groups) that is, H0:
μ1 = μ2 = . . . = μk
If H0 is rejected then we can say that there is a significant difference
between these k groups or there is a significant effect of these k
independent variables.
2. To test the relationship between two variables
This test provides evidence that dependent variable Y and independent
variable X are related in their movements. If Y do not relate with Xthen
we expect H0: μ1 = μ2 = … = μk, which is the null hypothesis for testing
the absence of relationship.
3. Test for Linearity of Regression
If in 2, some relationship is established, the next step is to find the
appropriate regression function. At the first stage we try to find out
whether the linear regression fits the observed data, that is the null
hypothesis will be
H0: μi=+Xi
with the sample model Yij = μi + eij, when α and β are the parameters of
the model.
4. Test for Polynomial Regression
If a linear relationship is not established, we proceed to test the
hypothesis
H0: μi = α+β1 Xi + β2 Xi 2 +...+ βk Xi k
That is, the relationship between X and Y can be explained by a
polynomial of degree k.
5. Some Other Uses of ANOVA
Test of Homogeneity of a Group of Regression Coefficients.
Test for Equality of Regression Equations from p Groups.
322 Test for Multiple Linear Regression Model.
CHECK YOUR PROGRESS 1 Analysis of Variance
Tests
Note: i) Use the space given below for your answers.
ii) Check your answers with those given at the end of the unit.
1.Describe the Analysis of Variance and differentiate between One-way
and Two-way ANOVA.
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
2. Explain briefly the uses of Analysis of Variance.
……………………………………………………………………………….
…………………………………………………………………………………
……….………………………………………………………………………
………………….……………………………………………………………
12.4 ONE-WAY ANALYSIS OF VARIANCE
(ANOVA)
One-factor analysis of variance (or one-way analysis of variance) is a special
case of ANOVA.Whereas two sample t-test is used to decide whether two
groups (two levels) of a factor have the same mean; one-way ANOVA
generalizes this problem to k levels (greater than two) of a factor.
In the following, subscript i refers to the ith level of the factor and subscript j
refers to the jth observations within a level of factor. For example, y23 refers to
third observation in the second level of a factor.
The observations on different levels of a factor can be exhibited as shown
below:
Level of Observations Totals Means
a factor
1 y11y12............. y1n y1. y 1.
2 y21 y22............. y2n y2. y 2.
. . . .
. . . .
. . . .
k yk1 yk2............. ykn yk. y k.
A linear mathematical model for one-way classified data is used which can
be written as
yij= μi+ eij wherei = 1, 2, . . ., k; j = 1, 2, . . ., n.
Here, yij is continuous dependent or response variable, whereas μi is discrete
independent variable, also called an explanatory variable. Total number of
observations is 323
Statistical Analysis k
n1 + n2 + … +nk = n i N
i 1
In fact, this model decomposes the responses into a mean for each level of a
factor and error term, that is,
Response = A mean for each level of a factor + error term
The analysis of variance provides estimates for each level mean. These
estimated level means are the predicted values of the model and the
difference between the response variable and the estimated/predicted level
means are the residuals.
That is,
y ij i eij implying that eij y ij i
The above model can also be re-written as yij= μ + (μi-μ) +eij
or yij= μ +αi + eij, i= 1, 2, ..., k and j= 1, 2, ..., ni
where, αi= μi– μ.
12.4.1 Assumptions
The following are the basic assumptionsof one-way ANOVA:
1. Dependent variable measured on interval scale;
2. k samples are independently and randomly drawn from the population;
3. Population can be assumed reasonably to have a normal distribution;
4. k samples have approximately equal variance;
5. Various effects are additive in nature; and
6. eij are independently and identically distributed (i.i.d.) normal variables
with mean zero and variance σe2.
12.4.2 Procedure of One-Way ANOVA Test
Now, when we discuss the step-by-step computation procedure for one way
analysis of variance for k independent samples, the first step of the procedure
is to make the null and alternative assumptions.
Null Hypothesis
We want to test the equality of the population means, that is, to test the
homogeneity of effect of different levels of a factor. Hence, the null
hypothesis is given by
H0: μ1 = μ2 = . . . = μk
against the alternative hypothesis
H1: μ1 ≠ μ2 ≠ . . . ≠ μk (or some μi's are not equal)
Determining Level of Significance:
The next is to decide a criterion for acceptance or rejection of null hypothesis
324 i.e., cut-off value and level of significance , at which we have to test our
null hypothesis. As discussed in Unit 3 and 4, level of significance is a fixed Analysis of Variance
Tests
assumed value which is decided before starting the test procedure. The most
commonly used values of α are 0.05 (5%) and 0.01 (1%). Let us take it as
= 5% (or 1%) if not given anywhere or follow whatever given for α.
Computation of Ratio of Variations F:
Since here F ratio contains only two elements, which are the variances
between the samples and within the samples respectively, as discussed
before, let us recapitulate the calculation of these variances.
If all the means of samples were exactly equal and all samples were exactly
representative of their respective populations so that all the sample means
were exactly equal to each other and to the population mean, then there will
be no variance. However, this can never be the case. We always have
variation both between samples and within samples, even if we take these
samples randomly and from the same population. This variation is known as
the total variation.
2
The total variation designated by X X , where X represents individual
observations for all samples and X is the grand mean of all sample means and
equals µ, the population mean, is also known as the total sum of squares or
SST, and is simply the sum of squared differences between each observation
and the overall mean. This total variation represents the contribution of two
elements. These elements are:
a) Variance between samples. The variance between samples may be due
to the effect of different treatments, meaning that the population means
may be affected by the factor under consideration, thus making the
population means actually different, and some variance may be due to the
inter-sample variability. This variance is also known as the sum of
squares between samples. Let this sum of squares be designed as SSB.
Then SSB is calculated by the following steps:
Step-I: Take k samples of size n each and calculate the mean of each
sample, designated as X1 , X 2 , X 3 ,... , X k .
Step-II: Calculate the grand mean X of the distribution of these sample
means, so that
k
X
i 1
i
X
k
Step-III:Take the difference between the means of the various samples
and the grand mean, which can be denoted as
( X1 X ), ( X 2 X ), ( X 3 X ), ... , ( X k X )
Step-IV: Square these deviations or differences individually, multiply
each of these squared deviations by its respective sample size
and sum up all these products, so as to find 325
k
n X
Statistical Analysis 2
i i X , where n i size of the ith sample.
i 1
This will be the value of the SSB.
However, sometimes individual observations in each k samples are not
available, but means of the samples are given, then the step 1 can be skipped
and we start the computation procedure from step 2. Divide SSB by the
degrees of freedom, which are (k – 1), where k is the number of samples and
this would give us the value of 2between , so that:
SSB
2between
(k 1)
This is also known as mean square between samples or MSB.
b) Variance within samples. Even though each observation in a given
sample comes from the same population and is subjected to the same
treatment, some chance variation can still occur. The variance may be due
to sampling errors or other natural causes. The variance or sum of squares
is calculated through the following steps:
Step-I: Calculate the mean value of each sample, that is,
X1 , X 2 , X 3 ,..., X k .
Step-II: Take one sample at a time and take the deviation of each item in
the sample from its mean. Do this for all the samples, so that we
would have a difference between each value in each sample and
their respective means for all values in all samples.
Step-III:Square these differences and take a total sum of all these
squared differences (or deviations). This sum is also known as
SSW or sum of squares within samples.
Step-IV: Divide this SSW by the corresponding degrees of freedom. The
degrees of freedom are obtained by subtracting the total number
of samples from the total number of items.
Thus if N is the total number of items or observations, and k is
the number of samples, then, df = (N – k) which are the degrees
of freedom within samples (If all samples are of equal size n,
then df = k(n – 1), since (n – 1) are the degrees of freedom for
each sample and there are k samples).
This figure SSW/df is also known as 2within , or MSW (mean of sum of
squares within samples).
Now the value of F can be computed as:
2between SSB / df SSB /(k 1) MSB
F 2
within SSW / df SSW /( N k ) MSW
326
Construction of ANOVA Table Analysis of Variance
Tests
After the various calculations for SSB, SSW and the degrees of freedom have
been, made, these figures can be presented in a simple table called Analysis
of Variance table or simply ANOVA table, as follows;
ANOVA Table
Source of Sum of Degrees of Mean Squares Variance
Variation Squares Freedom Ratio F
Treatment SSB (k -1 ) SSB MSB
MSB F
(k 1) MSW
Within SSW (N – k) SSW
MSW
(N k)
Total SST
Then, the variance ratio is
MSB
F
MSW
This calculated value of F is then compared with the critical value of F,
obtained from the F-table at respective d.f. and a decision is made about the
validity of null hypothesis. The critical value of F from the table for α level of
significance and degrees of freedom as follows:
df(numerator) = (k – 1)
df (denominator) = (N – k)
This value of F from the table is compared with calculated value of F and if
calculated value of F is less than the critical value of F, we cannot reject the
null hypothesis.
Now after discussing the procedure of One-way ANOVA test let us practice
to solve some examples.
Example 1: To test whether all professors teach the same material in
different sections of the introductory statistics class or not, 4 sections of the
same course were selected and a common test was administered to 5 students
selected at random from each section. The scores for each student from each
section were noted and are given below. We want to test for any differences
in learning, as reflected in the average scores for each section.
Student # Section 1 Section 2 Section 3 Section 4
Scores (X1) Scores (X2) Scores (X3) Scores (X4)
1 8 12 10 12
2 10 12 13 15
3 12 10 11 13
4 10 8 12
10
5 5 13 14
10
Totals X 1 45 X 2 55 X 3 60 X 4 60
Means X1 9 X 2 11 X 3 12 X 4 12
327
Statistical Analysis Solution: The method is as follows:
1) State the null hypothesis. We are assuming that there is no significant
difference among the average scores of students from these 4 sections and
hence all professors are teaching the same material with the same
effectiveness, i.e.
H 0 : 1 2 3 4 .
H1 : All means are not equal or at least two means differ from each other
2) Establish a level of significance. Let α = 0.05.
3) Calculate the variance between the samples, as follows:
a) The mean of each sample is,
X1 9, X 2 11, X 3 12, X 4 12
b) The grand mean or X is:
X
X 9 11 12 12 11
n 4
c) Calculate the value of SSB.
SSB n X X
2
5 (9 11) 2 5 (11 11) 2 5 (12 11) 2 5 (12 11) 2
= 20 + 0 + 5 + 5 = 30
d) The variance between samples 2between or MSB is given by:
MSB
SSB
30 (30) 10
df (k 1) 3
4) Calculate the variance within samples, as follows:
The find the sum of squares within samples (SSW) we square each
deviation between the individual value of each sample and its mean, for
all samples and then sum these squares deviations, as follows:
X X
2 2 2 2 2
i Xi 1 X1 X 2 X 2 X 3 X 3 X 4 X 4
We have the mean of each sample as
X1 9, X 2 11, X 3 12, X 4 12
Thus
X
2
i Xi (8 9) 2 (10 9) 2 (12 9) 2 (10 9) 2 (5 9) 2
(12 11) 2 (12 11) 2 (10 11) 2 (8 11) 2 (13 11) 2
(10 12) 2 (13 12) 2 (11 12) 2 (12 12) 2 (14 12) 2
(12 12) 2 (15 12) 2 (13 12) 2 (10 12) 2 (10 12) 2
= 1 + 1 + 9 + 1 + 16 +1 + 1 + 1 + 9 + 4+4 + 1 + 1 + 0 + 4
+0 + 9 + 1 + 4 + 4
= 28+16+10+18=72
328 Then SSW = 28 + 16 + 10 + 18 = 72
Now, the variance within samples, 2within , or MSW is given by Analysis of Variance
Tests
SSW SSW 72 72
MSW 4.5
df ( N K ) 20 4 16
MSB 10
Then the F-ratio 2.22 .
MSW 4.5
We can construct an ANOVA table for the problem solved above as follows:
ANOVA Table
Source of Sum of Degrees of Mean Square Variance
Variation Squares Freedom Ratio F
Treatment SSB = 30 (k – 1) = 3 SSB MSB
MSB
(k 1) MSW
30 = 2.22
MSB 10
3
Within (or SSW = 72 (N – k) = 16 SSW
MSW
error) (N k)
72
MSW 4.5
16
Total SST = 102
Now, we check for the critical value of F from the table for α = 0.05 and
degrees of freedom as follows:
df(numerator) = ( k – 1) = (4 – 1) = 3
df (denominator) = (N – k) = (20 – 4) = 16
This value of F from the table is given as 3.24. Now, since our calculated
value of F = 2.22 is less than the critical value of F = 3.24, we cannot reject
the null hypothesis.
Example 2: A department store chain is considering building a new store at
one of the three locations. An important factor in making such a decision is
the household income in these areas. If the average income per household is
similar then they can pick any one of these locations. A random survey of
various households in each location is undertaken and their annual combined
income is recorded. This data is tabulated as follows:
329
Statistical Analysis Annual Household Income ($ 1, 000s)
Area (1) Area (2) Area (3)
70 100 60
72 110 65
75 108 57
80 112 84
83 113 84
- 120 70
- 100 -
Total 380 763 420
Test if the average income per household in all these localities can be
considered as the same at α = 0.01.
Solution: If μi denotes the mean of the ith area ( i= 1, 2, 3) then the null
hypothesis is:
H 0 : 1 2 3
against H1 : At least two means are not equal.
The null hypothesis can be tested by computing the F-ratio for the data given
and then comparing it with the critical ratio of F from the table.
As before, let us first calculate the values of SSB and SSW.
380 763 420
Here: X1 76 X 2 109 X3 70
5 7 6
76 109 70 255
so that, X 85.
3 3
Then, SSB 5(76 85) 2 7(109 85) 2 6(70 85) 2
405 4032 1350 = 5787
n1 n2 n k1
SSW (X i1 X1 ) 2 (X i 2 X 2 ) 2 (X ik X k ) 2
i 1 i i 1
where n i 5 n1 7 n3 6 k 3
5 7 6
SSW (X i1 X1 ) 2 (X i 2 X 2 ) 2 (X i 3 X 3 ) 2
i 1 i 1 i 1
(70 76) 2 (72 76) 2 (75 76) 2 (80 76) 2 (83 76) 2
330
(100 109) 2 (110 109) 2 (108 109) 2 (112 109) 2 Analysis of Variance
Tests
(113 109) 2 (120 109) 2 (100 109) 2
(60 70) 2 (65 70) 2 (57 70) 2 (84 70) 2 (84 70) 2 (70 70) 2
= 36 + 16 + 1 + 16 + 49 +81 + 1 + 1 + 9 + 6 + 121 + 81
+100 + 25 +169 + 196 + 196 + 0
= 118+310 + 686
Then, SSW = 118 + 310 + 686 = 1114.
SSB 5787
Now, MSB 2893.5
(k 1) 2
SSW 1114
MSW 74.26
(N k) 15
MSB 2893.5
Then, F 38.96
MSW 74.26
We can construct an ANOVA table for the problem solved above as follows:
ANOVA Table
Source of Sum of Degrees of Mean Square F
Variation Squares Freedom
Treatment SSB =5787 (k – 1) = 2 SSB MSB
MSB F
(k 1) MSW
5787 = 38.96
2893.5
2
Within (or SSW =1114 (N – k) = SSW
MSW
error) 15 (N k)
1114
74.26
15
Total SST =6901
The critical value of F from table for α = 0.01 and df 2 and 15 respectively is
6.36. Since our calculated value of F is higher than the table value of F, we
cannot accept the null hypothesis.
CHECK YOUR PROGRESS 2
Note: i) Check your answers with those given at the end of the unit.
3) There are three sections of an introductory course in Statistics. Each
section is being taught by a different professor. There are some
complaints that at least one of the professors does not cover the 331
Statistical Analysis necessary material. To make sure that all the students receive the same
level of material in a similar manner, the chairperson of the department
has prepared a common test to be given to students of the three
sections. A random sample of seven students is selected from each class
and their test scores out of a total of 20 points are tabulated as follows;
Students Section (1) Section (2) Section (3)
1 20 12 16
2 18 11 15
3 18 10 18
4 16 14 16
5 14 15 16
6 18 12 17
7 15 10 14
Total 119 84 112
Do you think that at 95% confidence level, there is significant
difference in the average test scores of students taught by the different
professors?
4) Able Insurance Company wants to test whether three of its salesmen, A,
B, and C, in a territory make similar number of appointments with
prospective customers during a given period of time. A record of
previous four months showed the following results for the number of
appointments made by each salesman for each month.
Salesman
Month (A) (B) (C)
1 8 6 14
2 9 8 12
3 11 10 18
4 12 4 8
Totals 40 28 52
Do you think that at 95% confidence level, there is significant
difference in the average number of appointments made by the
salesmen per month?
332
Analysis of Variance
12.5 TWO-WAY ANALYSIS OF VARIANCE Tests
(ANOVA)
In the previous Section we considered the case where only one predictor/
independent variable/explanatory was categorized at different levels. In this
section, let us consider the case with two categorical predictors, each
categorized at different levels and a continuous response variable. Then it is
called two-way classification and the analysis is called Two-way ANOVA.
In such an ANOVA, generally we have an experiment in which we
simultaneously study the effect of two factors in the same experiment. For
each factor, there will be a number of classes/groups or levels. In the fixed
effect model, there will be only fixed levels of the two factors. We shall first
consider the case of one observation per cell. Let the factors be A and B and
the respective levels be A1,A2,…,Ar and B1, B2, …, Bqs. Let yij be the
observation/response/dependent variable under the ith level of factor A and
jthlevel of factor B. Further, let μ1A, μ2A,…,μrAbe the means of levels
A1,A2,…,Ar and μ1B, μ2B,…,μsB be the means of levels B1, B2, …, Bsin the
population. The observations then can be represented in a table as follows:
TWO-WAY CLASSIFIED DATA
A/B B1 B2 … Bj … Bs TOTAL MEAN
A1 y11 y12 … y1j … y1s y1. 1.
A2 y21 y22 … y2j … y2s y2. 2.
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
Ai yi1 yi2 … yij … yis yi. i.
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
Ar Yr1 Yr2 … Yrj … Yrs Yr. r.
Total y.1 y.2 … y.j … y.s y. .= G
Mean .1 .2 … .j … .s - ..
Mathematical Model
Here, the mathematical model may be written as
yij = μij+ eij whereeij’s are error terms.
12.5.1 Assumptions of Two-Way ANOVA
For the validity of F-test, the following assumptions should be satisfied:
(i) All the observations yij are independent;
(ii) Different effects (effects of levels of factor A, effects of levels of
factor B and error effect) are additive in ‘nature’; 333
Statistical Analysis (iii) eij’s are independently and identically distributed normal variables
with mean zero and variance e2 that is, eij i i d N 0, e2 ; and
(iv) There is no interaction between different levels of factor A and B.
12.5.2 Test of significance in Two-way ANOVA
As we have discussed above that if we are interested in studying the
simultaneous effect of two independent factors on the dependent variable, we
use two-way ANOVA. For example, we may wish to study the simultaneous
effects of five varieties of wheat (first criterion) and four different types of
fertilizers (second criterion) on the yield (dependent variable) or test the
stress level of employees in three different organizations in different regions,
and so on. In such situations, we can also apply two separate one-way
ANOVA for each treatment/factor. However, it is more advantageous to use
two-way ANOVA because the variance can be reduced by introducing the
second factor.
In two-way ANOVA, the total variation in the data is divided into three
components: variation due to the first criterion (factor), variation due to the
second criterion (factor) and variation due to error.
Now let us discuss the testing procedure of two-way ANOVA briefly
mentioning the main steps and formulae as follows:
Step 1:We first formulate the null hypothesis (H0) and alternative hypothesis
(H1). In two-way ANOVA, we can test two hypotheses
simultaneously: one for different levels of factor A and the other for
different levels of factor B. If factor A has r levels, we can set up the
null and alternative hypotheses as follows:
H 0 A : 1 2 ... r
H 1A At least one i j (i j 1,2,..., r )
Similarly, if factor B has s levels, we can set up the null and
alternative hypotheses as follows:
H 0 B : 1 2 ... s
H 1B At least one i j (i j 1,2,..., s)
Step 2:We calculate the Correction factor (CF) and the raw sum of squares
(RSS) using the formulae given below:
G2
Correction Factor (CF) andRaw Sum of Squares
N
r s
RSS y ij2
i 1 j1
where G = sum of all observations, N = the total number of
observations, and y ij observation of the ith level of factor A and the
jth level of factor B.
334
Step 3:We calculate the total sum of squares (TSS), the sum of squares Analysis of Variance
Tests
between rows or sum of squares due to factor A (SSA), the sum of
squares between columns of sum of squares due to factor B (SSB) and
sum of squares due to error (SSE) as follows:
Total Sum of Squares(TSS) RSS Correction Factor (CF)
1 r 2
Sum of Squares due to Factor A (SSA ) y i Correction Factor (CF)
q i 1
1 s 2
Sum of Squares due to Factor B (SSB) y i Correction Factor (CF)
p i 1
Sum of Squares due to Error (SSE) = TSS – SSA – SSB
where y i the sum of the observations of the ithlevel of factor A.
y j the sum of the observations of the jth level of factor B.
Step 4:We determine the degrees of freedom (df) as
The degrees of freedom (df) for factor A = r – 1
The degrees of freedom (df) for factor B = s – 1
The degrees of freedom (df) for error = (r – 1) (s – 1)
Step 5: We obtain the various mean sums of squares as follows:
SSA
Mean sum of squares due to factor A (MSSA )
r 1
SSB
Mean sum of squares due to factor B (MSSB)
s 1
SSE
Mean sum of squares due to error (MSSE )
(r 1) (s 1)
Step 6: We calculate the value of the test statistics using the formulae given
below:
MSSA
FA
MSSE
MSSB
FB
MSSE
ANOVA TABLE
Sources of DF SS MSS F-Test
Variation
Variation (r-1) SSA SSA MSSA
(MSSA ) FA
Due to r 1 MSSE
335
Statistical Analysis Factor A
Variation (s-1) SSB SSB MSSB
( MSSB) FB
Due to s 1 MSSE
Factor B
Variation (r-1)(s-1) SSE SSE
(MSSE )
Due to (r 1) (s 1)
Error
Total rs-1 TSS
Step 7:We take decisions about the null hypotheses for factor A and factor B
as explained below:
(i) Compare the calculated value of FAwith tabulated value of FAat
the respective df’s. If calculated value is greater than the tabulated
value then reject the hypothesis H0A, otherwise it may be
accepted.
(ii) Compare the calculated value of FBwith tabulated value of FB at
the respective df’s. If calculated value is greater than the tabulated
value then reject the hypothesis H0B, otherwise it may be
accepted.
Now after discussing the procedure of Two-way ANOVA test let us practice
to solve some examples:
Example 3: Future group wishes to enter the frozen shrimp market. They
contract a researcher to investigate various methods of groups of shrimp in
large tanks. The researcher suspects that temperature and salinity are
important factor influencing shrimp yield and conducts a two-way analysis of
variance with their levels of temperature and salinity. That is each
combination of yield for each (for identical gallon tanks) is measured.The
recorded yields are given in the following chart:
Categorical variable Salinity (in pp)
Temperature 700 1400 2100 Total Mean
600 F 3 5 4 12 4
700 F 11 10 12 33 11
800 F 16 21 17 54 18
Total 30 36 33 99 11
Compute the ANOVA table for the model.
Solution: Since in each all there is one observation. So we will use the
model.
yij =μ+ i+βj+ eij
336
where, yij is the yield corresponding to ith temperature and jthsalinity, μ is the Analysis of Variance
Tests
general mean, i is the effect due to ith temperature, βj is the effect due to
jthsalinity and eij ~ i i d N (0, σ2). The hypotheses to be tested are
H0A: α1= α2 = α3against H1A: α1≠ α2 ≠ α3 ≠ 0
H0B: 1=2 = 3against H1B: 1≠ 2 ≠ 3 ≠ 0.
The computations are as follows:
Grand Total (G) = 99
No. of observations(N) = 9
Correction Factor CF = (99 99) / 9 = 1089
Raw Sum of Square (RSS) = 1401
Total Sum of Square (TSS) = RSS - CF= 1401-1089 = 312
Sum of Square due to Temperature (SST) = (12)2/3+(33)2/3+(54)2/3-1089 =
294
Sum of Square due to Salinity (SSS)= (30)2/3+(36)2/3+(33)2/3-1089 = 6
Sum of Square due to Error= TSS – SST – SSS= 312 -294 -6 = 12
ANOVA TABLE
Sources of DF SS MSS F-Test
Variation
Due to 2 294 147 FT=MSST/MSSE=147/3
Temperature =49
Due to Salinity 2 6 3 FS = 3/3 = 1
Due to error 4 12 3
Total 8 312
Since tabulated value of F2,4 at 5% level of significance is 10.649 which is
less than the calculated FT for testing the significant difference in shrimp
yield due to differences in levels of temperature (49). So, H0A isrejected.
Hence, there aredifferences in shrimp yield due to temperature at 5% level of
significance.
Since tabulated value of F2,4 at 5% level of significance is 10.649 which is
greater than the calculated FS for testing the significant difference in shrimp
yield due to difference in the level of salinity (1). So, H0B is not rejected.
Hence there is no any difference in shrimp yield due the salinity level of 5%
level of significance.
CHECK YOUR PROGRESS 3
Note: i) Check your answers with those given at the end of the unit.
5) An experiment was conducted to determine the effect of different
dates of planting and different methods of planting on the field of
337
Statistical Analysis sugar-cane. The data below show the fields of sugar cane for four
different data and the methods of planting:
Dates of Planting
Method of October November February March
Planting
I 7.10 3.69 4.70 1.90
II 10.29 4.79 4.50 2.64
III 8.30 3.58 4.90 1.80
Carry out an analysis of the above data.
6) A researcher wants to test four diets A, B, C, D on growth rate in
mice. These animals are divided into 3 groups according to their
weights. Heaviest 4, next 4 and lightest 4 are put in Block I, Block II
and Block III respectively. Within each block, one of the diets is
given at random to the animals. After 15 days increase in weight is
noted and given in the following table:
Blocks Treatments/Diets
A B C D
I 12 8 6 5
II 15 12 9 6
III 14 10 8 5
Perform a two-way ANOVA to test whether the data indicate any
significant difference between the four diets due to different blocks.
12.6 LET US SUM UP
In this unit, we have discussed:
1. Basic ideas and concepts related to andthe technique of Analysis of
Variance as applied to test the equality of several population means
simultaneously;
2. Basic underlying assumptions of ANOVA, terminologies and
notations which are frequently used in Analysis of Variance;
3. Different applications of ANOVA;
4. Types of data which can be analyzed through ANOVA technique
and the meaning of one -way and two-way classified data as well as
the concept of one-way and two-way ANOVA;
5. Formation of null and alternative hypotheses to be tested and
different steps of One-way ANOVAfor testing the hypothesis under
consideration;
6. Null and alternative hypotheses to be tested and different steps of
two-way ANOVA for testing the hypothesis under consideration;
338
Analysis of Variance
12.7 KEY WORDS Tests
Analysis of variance (ANOVA): An overall test of the null hypothesis for
more than two population means.
Two-factor ANOVA: A more complex type of analysis of variance that tests
whether differences exist among population means categorized by two factors
or independent variables.
12.8 SUGGESTED FURTHER READING/
REFERENCES
Witte, R., & Witte, J. (2017). Statistics. Hoboken, NJ: John Wiley & Sons.
12.9 ANSWERS TO CHECK YOUR PROGRESS
1) Please refer to section 12.6excluding sub-sections 12.3.1 and 12.3.2.
2) Please refer to section 12.3.3.
3) The null hypothesis states that there are no differences among the mean
scores of the three sections(denoted by μi, i = 1, 2, 3), so that we have:
H 0 : 1 2 3
H1 : At least two means differ.
a) As we know, the sum squares between the samples is given by:
k
SSB n i ( X i X ) 2
i 1
In our case, we have 3 samples, therefore,
119 84 112
X1 17, X 2 12, X3 16
7 7 7
45
And hence, X 15
3
Then, SSB 7(17 15) 2 7(12 15) 2 7(16 15) 2 =28+ 63 + 7 = 98.
b) Sum of squares within (SSW) samples is calculated as follows:
nI n2 nk
SSW (X i1 X1 ) 2 (X i 2 X 2 ) 2 (X ik X k ) 2
I1 i 1 i 1
(20 17) (18 17) (18 17) 2 (16 17) 2
2 2
(14 17) 2 (18 17) 2 (15 17) 2
(12 12) 2 (11 12) 2 (10 12) 2 (14 12) 2
(15 12) 2 (12 12) 2 (10 12) 2
(16 16) 2 (15 16) 2 (18 16) 2 (16 16) 2
(16 16) 2 (17 16) 2 (14 16) 2
= 58 339
Statistical Analysis Further, we have N = 21, k = 3 so that k-1= 2 and N-k= 18. Therefore,
calculated
SSB / df 98 / 2
F 15.21
SSW / df 58 / 18
The tabulated F2,18 at 5% level of significance is 3.55. Hence, since
Ftab<Fcal; H0 is rejected implying that mean scores of three sections
are not same.
4) In the usual notations,our null hypothesis is
H0: 1 2 3 , and H1 at least two means are unequal.
40 28 52
Here X1 10 X2 7 X3 13
4 4 4
10 7 13 30
so that, X 10.
3 3
Then, SSB 4(10 10) 2 4(7 10) 2 4(13 10) 2 = 0 + 36 + 36 = 72
Degrees of freedom = df = (k – 1) = ( 3 – 1) = 2.
4 4 4
SSW (X i1 X1 ) 2 (X i 2 X 2 ) 2 (X i 3 X 3 ) 2
i 1 i 1 i 1
(8 10) 2 (9 10) 2 (10 10) 2 (12 10) 2
(6 7) 2 (8 7) 2 (10 7) 2 (4 7) 2
(14 13) 2 (12 13) 2 (18 13) 2 (8 13) 2 = 82
Degrees of freedom = df = (N – k) = (12 – 3) = 9.
Then, the F-ratio is given as:
SSB / df 72 / 2 36
F 3.95
SSW / df 82 / 9 9.1
The F-ratio from the table at 95% confidence level and degrees of
freedom 2 and 9 respectively is given as 4.26.
Since our calculated value of F is less than the tabulated value of F, we
cannot reject the null hypothesis.
5) Here we have H 0 A : There is no differencein the average yield due to
different methods of planting, that is,
H 0 A : 1 2 3 against, H1A: α1≠α2≠α3
H 0 B : There is no difference in the average yield due to different dates
of planting, that is,
H 0 B : 1 2 3 4
H 1 B : 1 2 3 4
340
G= y ij = Total of all observations Analysis of Variance
Tests
=7.10+3.69+4.70+1.90+10.29+4.79+4.58+2.64+8.30+
3.58+4.90+1.80
=58.28
N = No. of observations = 12
G2 58.28 58.28
Correction Factor = CF = 283.0465
N 12
Raw Sum of Square (RSS) =
7.10 2 3.69 2 4.70 2 1.90 2 10 .29 2 4.79 2 4.58 2
2 2 2 2 2
2.64 8.30 3.58 4 .90 1.80 355 .5096
Total Sum of Square (TSS) = RSS – CF = 355.5096-283.0465= 72.4631
Sum of Square due to Dates of Planting (SSD)
D12 D22 D32 D24
CF
3 3 3 3
2 2 2 2
=
25.69 12.06 14.18 6.35
283.0465
3 3 3 3
M12 M 22 M32
Sum of Square to Method of Planting (SSM) CF
4 4 4
2 2 2
=
17.39 22.31 15.58
283.0465
4 4 4
= 286.3412-283.0465 = 3.2947
Sum of Square due to Error (SSE)= TSS-SSD-SSM
= 72.4631-3.2947-65.8917 = 3.2767
65.8917 3.2947
MSSD 21.9639 MSSM 1.6473
3 , 2 ,
3.2767
MSSE 0.5461
6
MSSM 1.6473
For testing H 0 A : FM is 3.02
MSSE 0.5461
MSSD 21.9639
For testing H 0 B : FD is 40.22
MSSE 0.5461
The tabulated value of F2, 6 at 5% level of significance is 5.14 which
is greater than the calculated value of FM (3.02) so H0A is accepted.
So, we conclude that there is no significant difference among the
different methods of planting.
341
Statistical Analysis The tabulated value of F3, 6 at 5% level of significance is 4.76 which is
less than calculated value of FD (40.22). So, we reject the null
hypothesis H 0 B . Hence there is a significant difference among the
dates of planting.
6) Null Hypotheses are:
H01: There is no significant difference between mean effects of diets.
H02: There is no significant difference between mean effects of
different blocks.
Against the alternative hypothesis
H11: There is significant difference between mean effects of diets
H12: There is significant difference between mean effects of different
blocks.
Blocks Treatments/Diets
A B C D Totals
I 12 8 6 5 31
II 15 12 9 6 42
III 14 10 8 5 37
Totals T1.=41 T2.=30 T3.=23 T4.=16 110
Squares of observations
Blocks Treatments/Diets Totals
A B C D
I 144 64 36 25 269
II 225 144 81 36 486
III 196 100 64 25 385
Totals 565 308 181 86 1140
Grand Total = G = y ij = 110
G 2 110
2
Correction Factor (CF) = 1008.3333
N 12
2
Raw Sum of Squares (RSS) = yij 1140
Total Sum Squares (TSS) = RSS - CF = 1140 - 1008.3333 = 131.6667
Sum of Squares due to Treatments/ Diets (SST)
T12. T22. T32. T42.
= CF
3 3 3 3
342
= 41 30 23 16 1008.3333 Analysis of Variance
Tests
= 1681 900 529 256 1008.3333
= 1122 - 1008.3333 = 113.6667
1 2
Sum of squares due to block (SSB) =
4
T.1 T.22 T.32 CF
= 31 42 37 1008.3333
= 1023.5 – 1008.3333 = 15.1667
Sum of Squares due to Errors (SSE) = TSS – SST SSB
= 131.6667 – 113.6667 15.1667= 2.8333
Mean Sum of Squares due to Treatments (MSST)
SST 113.6667
= 37.8889
df 3
Mean Sum of Squares due to Blocks (MSSB)
SSB 15.1667
= 7.58335
df 2
Mean Sum of Squares due to Errors (MSSE)
SSE 2.8333
0.4722
df 6
ANOVA TABLE
Source of SS df MSS F
Variation
Between 113.6667 3 113.6667 37.8889
37.8889 F1 80.2391
Treatments/ 3 0.4722
Diets
Between Blocks 15.1667 2 15.1667 7.58335
7.58335 F2 160.596
2 0.4722
Due to Errors 2.8333 6 2.8333
0.4722
6
Total 131.6667 11
Tabulated F at 5% level of significance with (3, 6) degree of freedom
is 4.76 & tabulated F at 5% level of significance with (2, 6) degree of
freedom is 5.14
Conclusion: Since calculated F1 > Tabulated F1, so we reject H01 and
conclude that there is significant difference between mean effect of
diets.
Also calculated F2 is greater than tabulated F2, so we reject H02 and
conclude that there is significant difference between mean effect of
different blocks. 343
Statistical Analysis
344
You might also like
- Mathematics For Economists by Carl P Simon and Lawrence E Blume 2004 PDF100% (1)Mathematics For Economists by Carl P Simon and Lawrence E Blume 2004 PDF952 pages
- Business Statistics & Analytics For Decision Making Assignment 1 Franklin Babu100% (1)Business Statistics & Analytics For Decision Making Assignment 1 Franklin Babu9 pages
- Unit 1 Measures of Central Tendency: StructureNo ratings yetUnit 1 Measures of Central Tendency: Structure30 pages
- Research Ii: Whole Brain Learning System Outcome-Based EducationNo ratings yetResearch Ii: Whole Brain Learning System Outcome-Based Education16 pages
- Laporan SPSS (1) - Fitri Aulia - 4192411015No ratings yetLaporan SPSS (1) - Fitri Aulia - 41924110158 pages
- Usage of Measures of Central Measures and Tendencies in Real LifeNo ratings yetUsage of Measures of Central Measures and Tendencies in Real Life5 pages
- Measures of Central Tendency and Variability100% (1)Measures of Central Tendency and Variability65 pages
- Measure of Central Tendency Dispersion ANo ratings yetMeasure of Central Tendency Dispersion A8 pages
- MBA_U3_Quantitative Techniques for Business DecisionsNo ratings yetMBA_U3_Quantitative Techniques for Business Decisions18 pages
- Business Research Methods and Statistics Using SPSS (Chapter 7 - Describing and Presenting Your Data)No ratings yetBusiness Research Methods and Statistics Using SPSS (Chapter 7 - Describing and Presenting Your Data)29 pages
- Statistical Analysis_ Descriptive Stat (2)No ratings yetStatistical Analysis_ Descriptive Stat (2)6 pages
- Business Statistics - Session Descriptive StatisticsNo ratings yetBusiness Statistics - Session Descriptive Statistics28 pages
- Measures of Central Tendency: Mean, Median ModeNo ratings yetMeasures of Central Tendency: Mean, Median Mode6 pages
- Measures of Central Tendency & VariabilityNo ratings yetMeasures of Central Tendency & Variability25 pages
- Finals. Fmch. Measure of Central Tendency Shape of The Distribution of DispeNo ratings yetFinals. Fmch. Measure of Central Tendency Shape of The Distribution of Dispe5 pages
- Electromagnetic Induction 1. ACJC - 08H2P1 - Q31No ratings yetElectromagnetic Induction 1. ACJC - 08H2P1 - Q3122 pages
- Mathematics For Economists by Carl P Simon and Lawrence E Blume 2004 PDFNo ratings yetMathematics For Economists by Carl P Simon and Lawrence E Blume 2004 PDF952 pages
- The High-Field Path To Fusion Energy: Learn MoreNo ratings yetThe High-Field Path To Fusion Energy: Learn More14 pages
- Astronomy Rotation Revolution Equinoxes Solstices Aphelion Perihelion Sun Angles Direct/Indirect Sun Migration of Direct Sun Latitudinal ZonesNo ratings yetAstronomy Rotation Revolution Equinoxes Solstices Aphelion Perihelion Sun Angles Direct/Indirect Sun Migration of Direct Sun Latitudinal Zones13 pages
- What Event Is This? (Circle) : A. Winter Solstice B. Autumnal Equinox C. Summer Solstice D. Vernal EquinoxNo ratings yetWhat Event Is This? (Circle) : A. Winter Solstice B. Autumnal Equinox C. Summer Solstice D. Vernal Equinox1 page
- Portugal Golden Visa European CitizenshipNo ratings yetPortugal Golden Visa European Citizenship7 pages
- Esime U.P. Ticomán: XXX X A4 2SE6 Cigüeñal Jorge VegaNo ratings yetEsime U.P. Ticomán: XXX X A4 2SE6 Cigüeñal Jorge Vega1 page
- CHAPTER 8 - Noncurrent Asset Held For SaleNo ratings yetCHAPTER 8 - Noncurrent Asset Held For Sale19 pages
- 3a. T24 Loan and Deposits - Parameter V1.0100% (2)3a. T24 Loan and Deposits - Parameter V1.023 pages
- Sept-24 Price List Bs6.2 (PV) - Xuv 7oo (D)No ratings yetSept-24 Price List Bs6.2 (PV) - Xuv 7oo (D)1 page
- CA IPCC Old Accounts Question Paper Nov 2018 StudycafeNo ratings yetCA IPCC Old Accounts Question Paper Nov 2018 Studycafe15 pages
- Vijayawada, Andhra Pradesh Vijayawada, Andhra PradeshNo ratings yetVijayawada, Andhra Pradesh Vijayawada, Andhra Pradesh16 pages
- Daily 500 Maximum 2500 Minimum 1000 Lead Time 15 Days EOQ 1200 Emergency Lead Time 3 DaysNo ratings yetDaily 500 Maximum 2500 Minimum 1000 Lead Time 15 Days EOQ 1200 Emergency Lead Time 3 Days5 pages
- Grade 11 - Maths II - EM 3rdTT - 2024 Southern ProvinceNo ratings yetGrade 11 - Maths II - EM 3rdTT - 2024 Southern Province5 pages
- RPWORLD - A Leading Mechanical Part Manufacturing SupplierNo ratings yetRPWORLD - A Leading Mechanical Part Manufacturing Supplier22 pages
- Following the Trend: Diversified Managed Futures Trading (Wiley Trading) download pdf100% (1)Following the Trend: Diversified Managed Futures Trading (Wiley Trading) download pdf55 pages
- REVIVE PO #2123 - Bridge and Bracing - 24 ProgressNo ratings yetREVIVE PO #2123 - Bridge and Bracing - 24 Progress1 page
