14.

4 Mann-Whitney U Test
In Section 14.2 and 14.3, we have discussed the paired sign test and Wilcoxon
matched-pair signed-rank test respectively which are used when the
observations are dependent, that is, observations are paired. When we are
interested in testing of difference in means of two independent populations then
we use two sample t-test. To use the t-test however, it is necessary to make a
set of assumptions (as described in Unit 11 of this course). In particular, it is
necessary that two independent samples be randomly drawn from normal
populations having equal variances and the data be measured in at least of an
interval scale. But in studies of consumer behaviour, marketing research,
experiment of psychology, etc. generally the data are collected in ordinal scale
and the form of the population is not known. Since the parametric t-test could
not be used in such situation, an appropriate non-parametric technique is
needed. In such a circumstance a very simple non-parametric test known as
Mann-Whitney U test may be used.
This test was developed jointly by Mann, Whitney and Wilcoxon. Wilcoxon
considered only the case of equal sample sizes while Mann and Whitney seem
to have been the first to treat the case of unequal sample sizes. Therefore, it is
sometimes called the Mann-Whitney U test and sometimes the Wilcoxon ranks
sum test. Thus, this test may also be viewed as a substitute for the parametric t-
test for the difference between two population means. When assumptions of the
two-sample t-test are fulfilled then this test is slightly weaker than t-test.
Assumptions
This test work under the following assumptions:

(i) The two samples are randomly and independently drawn from their
respective populations.

(ii) The variable under study is continuous.

(iii) The measurement scale is at least ordinal.

(iv) The distributions of two populations differ only with respect to location
parameter.

53
Non-Parametric Tests Let us discuss the general procedure of this test:
Let us suppose that we have two independent random samples X1 , X 2 ,..., X n1
and Y1 , Y2 ,..., Yn 2 drawn from two populations having medians  1 and  2
respectively. Here, we want to test the hypothesis about the medians of two
populations so we can take the null and alternative hypotheses as
H 0 : µ 1 = µ 2 and H 1: µ 1  µ 2 for two-tailed test 
H 0 :  1   2 and H1:  1   2 
or  for one-tailed test 
H 0 :  1   2 and H1:  1   2 

We can also form the null and alternative hypotheses in the form of distribution
functions. If F1(x) and F2(x) are the distribution functions of the first and
second populations respectively and we want to test that the two independent
samples come from the populations that are identical with respect to location,
that is, median.
Therefore, we can take the null and alternative hypotheses as
H1 : F1 (x)  F2 (x) for at least one value of x 
 for two-tailed test 
H1 : F1 (x)  F2 (X) for at least one value of x 
H 0 : F1 (x)  F2 (x) for at least one value of x 
or  for right-tailed test 
H 1 : F1 (x)  F2 (X) for at least one value of x 
H 0 : F1 (x)  F2 (x) for at least one value of x 
or  for left-tailed test 
H 1 : F1 (x)  F2 (X) for at least one value of x 
After setting null and alternative hypotheses, this test involves following steps:
Step 1: First of all, we combine the observations of two samples.
Step 2: After that, ranking all these combined observations from smallest to
largest, that is, the rank 1 is given to the smallest of the combined
observations, rank 2 is given to the second smallest and so on up to
the largest observation. If several values are same (tied), we assign
each the average of ranks they would have received if there were no
repetition.
Step 3: If null hypothesis is true then we can expect that the sum of ranks of
the two samples are equal. Now for convenience, we consider the
smaller sized sample and calculate the sum of ranks of the
observations of this sample. Let S be the sum of the ranks assigned to
the sample observations of smaller sized sample then for testing the
null hypothesis the test statistic is given by
n1 n1  1
U S ; if n1 is small
2
n2  n2  1
U  S ; if n2 is small
2
Step 4: Obtain critical value(s) of test statistic U at given level of significance
under the condition that null hypothesis is true. Table VII in the
Appendix at the end of this block provides the lower and upper
critical values for a given combination of n1 and n2 at α level of
significance for one-tailed and two-tailed test.
54
Step 5: Decision rule: Two-Sample Tests

To take the decision about the null hypothesis, the calculated value of
test statistic U (computed in Step 3) is compared with the critical
(tabulated) value (obtained in Step 4) at a given level of significance
(α) under the condition that null hypothesis is true. Since test may be
one-tailed or two-tailed so following cases arise:
For two-tailed test: When H0 :  1   2 and H1 :  1   2
For two-tailed test, we see critical values at α/2 for α level of
significance. If calculated value of test statistic U is either less than or
equal to the lower critical value (U L,  / 2 ) or greater than or equal to
the upper critical value (U U,  / 2 ) , that is, U  U L,  / 2 or U  U U,  / 2 then
we reject the null hypothesis at α level of significance. However, if
computed U lies between these critical values, that is,
U L,  / 2  U  U U,  / 2 , then we do not reject the null hypothesis at α%
level of significance.
For one-tailed test:
For one-tailed test, we see the critical value at α for α level of significance.
Case I: When H 0 :  1   2 and H1 :  1   2 (right-tailed test)
If calculated value of test statistic U is greater than or equal to the
upper critical value (U U,  ), that is, U  U U,  then we reject the null
hypothesis at α level of significance. However, if computed U is less
than upper critical value (U U,  ) , that is, U  U U,  then we do not
reject the null hypothesis at α% level of significance.
Case II: When H 0 :  1   2 and H1 :  1   2 (left-tailed test)
If calculated value of test statistic U is less than or equal to the lower
critical value (U L,  ), that is, U  U L,  then we reject the null
hypothesis at α level of significance. However, if computed U is
greater than lower critical value (U L,  ), that is, U  U L,  then we do
not reject the null hypothesis at α% level of significance.
For large (n1 or n2 >20):
When either n1 or n2 exceeds 20, the statistic U is approximately
normally distributed with mean
n 1n 2
E U   … (7)
2
and variance
n1n 2  n1  n 2  1
Var  U   … (8)
12
The proof of mean and variance of test statistic(T) is beyond the
scope of this course.
Therefore in this case, we use normal test (Z-test) (described in Unit
10 of Block 3 of this course.). The test statistic of Z-test is given by
U  E U  U  E U 
Z  ~ N  0, 1 
SE  U  Var  U 
55
Non-Parametric Tests n1 n 2
U
 2 ~ N  0,1  Using equations  … (9)
n 1 n 2  n 1  n 2  1  (7) and (8) 
12
After that, we calculate the value of test statistic Z and compare it
with the critical value(s) given in Table 10.1 at prefixed level of
significance α. Take the decision about the null hypothesis as
described in Section 10.2 of Unit 10 of this course.
Let us do some examples to become more user friendly with the test explained
above:
Example 5: A Statistics professor taught two special sections of a basic course
in which students in each section were considered outstanding. He used a
“traditional” method of instruction (T) in one section and an “experimental”
method of instruction (E) in the other. At the end of the semester, he ranked the
students based on their performance from 1 (worst) to 20 (best).
T 1 2 3 5 8 10 12 13 14 15
E 4 6 7 9 11 16 17 18 19 20

Test whether there is any evidence of a difference in performances based on the

two methods at 5% level of significance.
Solution: It is case of two independent populations and the assumption of
normality of both the populations is not given also the given data are in the
form of ranks so we cannot use two-sample t-test in this case. Also sample
sizes are small so we cannot use the Z-test. Therefore, we go for Mann-
Whitney U test.
Here, we want to test that performance of the students based on the two
methods is different. If  1 and  2 denote the average (median) ranks of the
students who taught by traditional” and “experimental” methods respectively
so our claim is  1   2 and its complement is  1   2. Since complement
contains the equality sign so we can take the complement as the null hypothesis
and the claim as the alternative hypothesis. Thus,
H0 : 1   2
H 1 :  1   2  both methods are different 
Since the alternative hypothesis is two-tailed so the test is two-tailed test.
Here, combined ranks are given. Also n1 = n2 so we take any sum of ranks(S)
of first sample or second sample. We consider first sample. Thus,
S = 1 + 2 + 3 + 5 + 8 + 10 + 12 + 13 + 14 + 15 = 83
Here, we consider the first sample so the test statistic is given by
n1 n1  1 10  (10  1)
U S  83   28
2 2
Since test is two-tailed so the lower and upper critical values of test statistic for
two-tailed test corresponding n1  10 and n2  10 at 5% level of significance are
U L,  / 2  U L, 0.025  79 and U U,  / 2  U U, 0.025  131.
Since calculated value of test statistic U (= 28) is less than critical values
(= 79 and 131) so we reject the null hypothesis and support the alternative
56
hypothesis i.e. we support the claim at 5% level of significance. Two-Sample Tests
Thus, we conclude that the samples fail to provide us sufficient evidence
against the claim so we can say that performances based on two methods are
different.
Example 6: The following data show the median amount to be spent on a
birthday gift by students who are unemployed and employed:
Unemployed 200 350 500 250 150 275
Employed 550 500 1000 450 700 200 500 600 1000

Examine that the median amount to be spent on a birthday gift by students who
are unemployed is lower than for students who are employed by Mann-
Whitney U test at 1% level of significance.
Solution: It is case of two independent populations and the assumption of
normality of both the populations is not given so we cannot use t-test in this
case. Also sample sizes are small so we cannot use the Z-test. Therefore, we go
for Mann-Whitney U test.
Here, we want to test that the median amount to be spent on a birthday gift by
students who are unemployed is lower than for students who are employed. If
 1 and  2 denote the average (median) amount to be spent on a birthday gift by
students who are unemployed and employed respectively so our claim is
 1   2 and its complement is  1   2 . Since complement contains the equality
sign so we can take the complement as the null hypothesis and the claim as the
alternative hypothesis. Thus,
H0 :  1   2

H1 :  1   2  median amount to be spent by students who are 

 unemployed is lower than the students who are employed 
Since the alternative hypothesis is left-tailed so the test is left-tailed test.
To perform the Mann-Whitney U test, we combine the observations of two
samples and for convenient, we arrange the data in ascending order. After that
we rank all these combined observations from smallest to largest as:
Unemployed Ranks Unemployed Ranks
150 1 200 2.5
200 2.5 450 7
250 4 500 9
275 5 500 9
350 6 550 11
500 9 600 12
700 13
1000 14.5
1000 14.5
Total 27.5

Since n1 (= 6) is small than n2 (= 9) so we take sum of the ranks(S) assigned to

the sample of unemployed students. Thus, S = 27.5.

57
Non-Parametric Tests Here, we consider the first sample so the test statistic is given by

n1  n1  1
U S
2
6   6  1
 27.5   6.5
2
To decide about the null hypothesis, the calculated value of test statistic U is
compared with the lower critical (tabulated) value at 1% level of significance.
Since test is left-tailed so the lower critical value of test statistic for left-tailed
corresponding to n1= 6 and n2 = 9 at 1% level of significance is
U L,   U L, 0.01  29.

Since calculated value of test statistic U (= 6.5) is less than lower critical value
(= 29) so we reject the null hypothesis and support the alternative hypothesis
i.e. we support the claim at 1% level of significance.
Thus, we conclude that the samples fail to provide us sufficient evidence
against the claim so we may assume that the median amount to be spent on a
birthday gift by students who are unemployed is lower than for students who
are employed.
Now, you can do following exercises in same manner.
E7) Write one difference between two-sample t-test and Mann-Whitney U
test.
E8) Write one difference between Wilcoxon matched-pair singed-rank test
and Mann-Whitney U test.
E9) The senior class in a particular high school had 25 boys. Twelve boys
lived in villages and other thirteen lived in a town. A test was conduct
to see that village boys in general were physically fit than the town
boys. Each boy in the class was given a physical fitness test in which a
low score indicates poor physical condition. The scores of the village
boys (V) and the town boys(T) are as follows:

Village 15.7 8.2 6.5 7.2 9.0 4.5 10.6

Boys(V) 12.4 16.2 12.9 11.4 5.6

Town 12.7 3.2 11.8 7.9 5.6 6.7 12.6

Boys(T) 7.9 2.7 6.1 3.6 6.5 2.8

Test whether the village boys are more fit than town boys at 5% level of
significance.
E10) The following data represent lifetime (in hours) of batteries for two
different brands A and B:
Brand A 40 30 55 40 40 35 30 40 50 45 40 35

Brand B 45 60 50 60 35 50 55 60 50 50 40 55

Examine, is the average life of two brands is same at 1% level of

significance?

58
 Two-Sample Tests
14.5 KOLMOGOROV-SMIRNOV TWO-SAMPLE
TEST
In Mann-Whitney U test, we have tested the hypothesis that the two
independent samples come from the populations that are identical with respect
to location whereas Kolmogorov-Smirnov two-sample test is sensitive to
differences of all types that may exist between two distributions, that is,
location, dispersion, skewness, etc. Therefore, it is referred as a general test.
This test was developed by Smirnov. This test also carries the name of
Kolmogorov because of its similarity to the one-sample test developed by
Kolmogorov. In Kolmogorov-Smirnov one sample test, the observed (sample
or empirical) cumulative distribution function is compared with the
hypothesized cumulative distribution function whereas in two-sample case the
comparison is made between the empirical cumulative distributions functions
of the two samples.
Assumptions
The assumptions necessary for this test are:
(i) The two samples are randomly and independently drawn from their
respective populations.
(ii) The variable under study is continuous.
(iii) The measurement scale is at least ordinal.
Let us discuss the general procedure of this test:
Let us suppose that we have two independent random samples X1 , X 2 ,..., X n1
and Y1 , Y2 ,..., Yn 2 drawn from first and second populations with distribution
functions F1(x) and F2(x) respectively. Also let S1(x) and S2(x) be the sample or
empirical cumulative distribution functions of samples drawn from first and
second populations respectively.
Generally, we want to test whether independent random samples come from
populations having the same distribution functions in all respect or the
distribution functions of two populations differ with respect to location,
dispersion, skewness, etc. So we can take the null and alternative hypotheses as
H 0 : F1 (x)  F2 (x) for all values of x
H1 : F1 (x)  F2 (x) for at least one value of x
After setting null and alternative hypotheses, the procedure of Kolmogorov -
Smirnov two-sample test summarise in the following steps:
Step 1: This test is based on the comparison of the empirical (sample)
cumulative distribution functions, therefore, first of all we compute
sample cumulative distribution functions S1(x) and S2(x) from the
sample data as the proportion of the number of sample observations
less than or equal to some number x to the number of observations,
that is,
The number of sample observations less than or equal to x
S1  x  
n1
and
59
Non-Parametric Tests The number of sample observations less than or equal to x
S2  x  
n2
Step 2: After finding the empirical cumulative distribution functions S1(x)
and S2(x) for all possible values of x, we find the deviation between
the empirical cumulative distribution functions for all x. That is,
S1 (x)  S2 (x) for all x
Step 3: If the two samples have been drawn from identical populations then
S1(x) and S2(x) should be fairly close for all value of x. Therefore, we
find the point at which the two functions show the maximum
deviation. So we take the test statistic which calculate the greatest
vertical deviation between S1(x) and S2(x), that is,
D  sup S1 (x)  S2 (x)
x

which is read as “D equals the supreme over all x, of the absolute

value of the difference S1 (x)  S2 (x) ”
Step 4: Obtain critical value of test statistic at α % level of significance under
the condition that null hypothesis is true. Table V and VI in
Appendix provides the critical values of test statistic for equal and
unequal sample sizes respectively at different level of significance.
Step 5: Decision Rule:
To take the decision about the null hypothesis, the test statistic
(calculated in Step 3) is compared with the critical (tabulated) value
(obtained in Step 4) for a given level of significance (α).
If computed value of D is greater than or equal to critical value (D  ) ,
that is, D  D then we reject the null hypothesis at α level of
significance, otherwise we do not reject H0.
For large sample (n1 = n2 = n > 40 and n1 >16 for unequal):
For a sample size n1 = n2 = n > 40 and n1 >16 for unequal sample
sizes, the critical value of test statistic at given level of significance is
approximated by the formula given in last row of Table V & VI in
the Appendix.
Now, it is time to do some examples based on above test.
Example 7: The following data represent lifetime (in hours) of batteries for
two different brands A and B:
Brand A 40 30 55 40 40 35 30 40 50 45 40 35
Brand B 45 60 50 60 35 50 55 60 50 50 40 55

Examine by the Kolmogorov-Smirnov two-sample test, is the average life of

two brands same at 5% level of significance?

Solution: Here, we want to test that the average life of batteries of two brands
is same. If F1(x) and F2(x) are cumulative distribution functions of the life of
batteries of brand A and brand B respectively then our claim is F1(x) = F2(x)
and its complement is F1(x) ≠ F 2(x). Since the claim contains the equality sign
so we can take claim as the null hypothesis and complement as the alternative
hypothesis. Thus,
60
 H 0 : F1 (x)  F2 (x) for all values of x Two-Sample Tests

H1 : F1 (x)  F2 (x) for at least one value of x

So the test statistic is given by
D  sup S1  x   S2  x 
x
where, S1(x) and S2(x) are the empirical cumulative distribution functions of
the lives of the batteries of brand A and brand B drawn in the samples
respectively which are defined as:
number of batteries of brand A whose average life  x
S1  x  
n1
number of batteries of brand B whose average life  x
S2  x  
n2
Calculation for S1  x   S2  x  :
Average Frequency C.F. SA(x) Frequency C.F. SB(x) S1  x   S2  x 
Life (Brand A) Brand A (Brand B) Brand B

30 2 2 2/12 0 0 0 2/12
35 2 4 4/12 1 1 1/12 3/12
40 5 9 9/12 1 2 2/12 7/12
45 1 10 10/12 1 3 3/12 7/12
50 1 11 11/12 4 7 7/12 4/12
55 1 12 12/12 2 9 9/12 3/12
60 0 12 12/12 3 12 12/12 0
Total n1 = 12 n2 = 12
From the above calculation, we have
7
D  sup S1  x   S2  x    0.58
x 12
The critical value of test statistic for equal sample sizes n1 = n2 = 12 at 5%
level of significance is 6/12 = 0.5.
Since calculated value of test statistic (= 0.58) is greater than critical value
(= 0.5) so we reject the null hypothesis i.e. we reject the claim at 5% level of
significance.
Thus, we conclude that the samples provide us sufficient evidence against the
claim so the average life of batteries of two brands is different.
Example 8: The following are the marks in Statistics of B.Sc. students taken
randomly from two colleges A and B:
Marks 0-10 10-20 20-30 30-40 40-50 50-60 60-70 70-80 80-90 90-100
College (A) 2 2 4 6 3 3 4 8 7 5
College (B) 1 1 2 5 7 3 3 2 6 6

Apply Kolmogorov-Smirnov test to examine that the distribution of marks in

college A and college B is same at 1% level of significance.

Solution: Here, we wish to test that the distribution of marks in Statistics of

B.Sc. students in college A and college B is same. If F1(x) and F2(x) are
distribution functions of the marks in college A and college B respectively then

61
Non-Parametric Tests our claim is F1(x) = F2(x) and its complement is F1(x) ≠ F2(x). Since the claim
contains the equality sign so we can take claim as the null hypothesis and
complement as the alternative hypothesis. Thus,
H 0 : F1 (x)  F2 (x) for all values of x

H1 : F1 (x)  F2 (x) for at least one value of x

So the test statistic is given by
D  sup S1  x   S2  x 
x

where, S1(x) and S2(x) are the empirical distribution functions of samples
drawn from college A and college B respectively which are defined as:
number of students of college A whose marks  x
S1  x  
n1
number of students of college B whose marks  x
S2  x  
n2

Calculation for S1  x   S2  x  :
Marks Frequency C.F. SA(x) Frequency C. F. SB(x) S A x   S B x 
(College A) (College A) (College B) (College B)

0-10 2 2 0.0455 1 1 0.0278 0.0177

10-20 2 4 0.0909 1 2 0.0556 0.0354
20-30 4 8 0.1818 2 4 0.1111 0.0707
30-40 6 14 0.3182 5 9 0.2500 0.0682
40-50 3 17 0.3864 7 16 0.4444 0.0581
50-60 3 20 0.4545 3 19 0.5278 0.0732
60-70 4 24 0.5455 3 22 0.6111 0.0657
70-80 8 32 0.7273 2 24 0.6667 0.0606
80-90 7 39 0.8864 6 30 0.8333 0.0530
90-100 5 44 1 6 36 1 0
Total n1 = 44 n2 = 36

From the above calculation, we have

D  sup S1  x   S2  x 
x

 0.0732
Since n1 = 44 > 16 so we can calculate the critical value of test statistic for
unequal sample sizes n1 = 44 and n2 = 36 at 1% level of significance by the
formula
n1  n 2 44  36
D n,  1.63  1.63  0.366
n 1n 2 44  36
Since calculated value of test statistic D (= 0.0732) is less than critical value
(= 0.366) so we do not reject the null hypothesis i.e. we support the claim at
1% level of significance.

62
Thus, we conclude that the samples fail to provide us sufficient evidence Two-Sample Tests
against the claim so we may assume that the distribution of marks in Statistics
of B.Sc. students in college A and college B are same.
Now, you can try the following exercises.

 


63