CHAPTER 6

Non-parametric Statistics

Most of the hypothesis-testing procedures discussed in the previous chapters are based
on the assumption that the random samples are selected from normal populations.
These testing procedures are referred to as parametric methods. In this chap-
ter, we consider a number of alternative test procedures, called non-parametric or
distribution-free methods. These often assume no knowledge whatsoever about the
distributions of the underlying populations, except perhaps that they are continuous.

We discuss two non-parametric tests in this chapter. The first, called the Wilcoxon
signed-rank test, is a test for a population mean, analogous to the one-sample t-test.
The second one is the Wilcoxon rank-sum test or Mann-Whitney test which is
analogous to the two-sample t-test.

6.1 Wilcoxon signed-rank test

The Wilcoxon signed-rank test applies is used on symmetric continuous distribution.

Under this condition, we can test the null hypothesis µ̃ = µ̃0 , where µ̃ is the population
median. Note that for symmetric distributions the median and the mean are approxi-

6-1
6.1. Wilcoxon signed-rank test 6. Non-parametric Statistics

mately equal and as result, some textbooks represents the null hypothesis in terms of
the population. The procedure for testing the null hypothesis is as follows:

i. First subtract µ̃0 from each sample value, discarding all differences equal to zero.

ii. Rank the remaining differences without regard to sign. A rank of 1 is assigned
to the smallest absolute difference (i.e., without sign), a rank of 2 to the next
smallest, and so on.

iii. When the absolute value of two or more differences is the same, assign to each
the average of the ranks that would have been assigned if the differences were
distinguishable. For example, if the fifth and sixth smallest differences are equal
in absolute value, each is assigned a rank of 5.5.

iv. If the hypothesis µ̃ = µ̃0 is true, the total of the ranks corresponding to the
positive differences should nearly equal the total of the ranks corresponding to
the negative differences.

Let us represent these totals by w+ and w− , respectively. We designate the smaller

of w+ and w− by w. Then the null hypothesis µ̃ = µ̃0 can be rejected in favor of the
alternative µ̃ < µ̃0 only if w+ is small and w− is large. Likewise, the alternative µ̃ > µ̃0
can be accepted only if w+ is large and w− is small. For a two-sided alternative, we
may reject H0 in favor of H1 if either w+ or w− , is sufficiently small.

Example 6.1.1. The nickel content, in parts per thousand by weight, is measured
for six welds. The results are 9.3, 0.9, 9.0, 21.7, 11.5 and 13.9. Use the Wilcoxon
signed-rank test to the test H0 : µ̃ ≤ 5 against H0 : µ̃ > 5.

Solution 6.1.1. The table of the differences and signed ranks is computed as follows

6-2
6. Non-parametric Statistics 6.1. Wilcoxon signed-rank test

x x−5 Signed-rank

9.0 4.0 1
0.9 -4.1 -2
9.3 4.3 3
11.5 6.5 4
13.9 8.9 5
21.7 16.7 6

Since the alternative is that µ̃ > 5, we compute w+ as follows

w+ = 1 + 3 + 4 + 5 + 6

= 19

Consulting the Wilcoxon signed-rank test table of critical values, we find that

p = P(W+ > 19)

= 0.0469

Hence we reject the null hypothesis and conclude that the population median is greater
than 5.

Note that since the sample size is 6 in the above example, summing up all the
absolute ranks we get, w+ + w− = 1 + 2 + . . . + 6 = 21. Therefore, since we only had
one negative signed-rank, 2, we could have found w+ = 21 − 2 = 19. It is the case that
for sample size n,

6-3
6.1. Wilcoxon signed-rank test 6. Non-parametric Statistics

w+ + w− = 1 + 2 + . . . + n
n(n + 1)
=
2

and therefore in cases where there are many more positive ranks than the negative
ranks, its easier to compute w− . Then compute w+ = n(n + 1)/2 − w− .

Example 6.1.2. Lets consider the data in Example 6.1.1 to test H0 : µ̃ = 9 against
H0 : µ̃ ̸= 5.

Solution 6.1.2. The table of the differences and signed ranks is computed as follows

x x−9 Signed-rank

9.0 0.0 –
0.9 -8.1 -4
9.3 0.3 1
11.5 2.5 2
13.9 4.9 3
21.7 12.7 5

From the above table the value of the test static is w+ = 11. The sample size is now
5 since we discard differences of zeros. From the Wilcoxon singed-rank test table, we
have that p-value when w+ = 12 is given by 2(0.1562) = 0.3124. Thus, p = P(W+ >
11) > 0.3124 and therefore we fail to reject the null hypothesis.

6-4
6. Non-parametric Statistics 6.2. Wilcoxon rank-sum test

Large-sample approximation

When n ≥ 15, the sampling distribution of W+ (or W− ) approaches the normal distri-
bution with mean and variance given by

n(n + 1) 2 n(n + 1)(2n + 1)

µW + = and σW +
= (6.1)
4 24

Therefore, when n exceeds the largest value in Wilcoxon signed-rank test table, the
statistic, Z = (W+ − µW+ )/σW+ can be used to determine the critical region for the
test.

6.2 Wilcoxon rank-sum test

The Wilcoxon rank-sum test, also known as the Mann-Whitney test, is used to test
the difference in population means in a certain cases where the populations are not
normal. There are however two assumptions required for this test:

i. the populations should be continuous;

ii. their density functions should be identical.

Let X1 , X2 , . . . , Xm be a random sample from one population with mean, µX and

Y1 , Y2 , . . . , Yn be a random sample from another with mean, µY . The test is performed
by ordering the m + n observations obtained from the two samples, and assigning ranks
1, 2, . . . , m+n. The test statistic, denoted by W , is the sum of the ranks corresponding
to X1 , . . . , Xm .

Example 6.2.1. Resistances, mΩ, are measured for five wires of one type and six
wires of another type. The results are as follows:

6-5
6.2. Wilcoxon rank-sum test 6. Non-parametric Statistics

X: 36 28 29 20 38
Y: 34 41 35 47 49 46

Use the Wilcoxon rank-sum test to test H0 : µX ≥ µY versus H1 : µX < µY .

Solution 6.2.1. Lets order the 11 values and assign them ranks as follows

Value Rank Sample Value Rank Sample

20 1 X 38 7 X
28 2 X 41 8 Y
29 3 X 46 9 Y
34 4 Y 47 10 Y
35 5 Y 49 11 Y
36 6 X

Then w = 1 + 2 + 3 + 6 + 7 = 19. From the Wilcoxon rank-sum test table of critical

values we have that p = P(W > 19) = 0.026. Therefore we reject the null hypothesis.

6-6