NON-PARAMETRIC TESTS

Testing (usually called ‘hypothesis testing’ ) play a major role in statistical investigation. In
statistical testing, we are concerned with examining the truth or otherwise, of hypothesis
(assumptions, claims, guesses, etc.) about some feature(s) of one or more populations. Almost
all large and small sample tests such as t, F and 2 are based on the assumptions that the
parent population (from which the sample is drawn) has a specific distribution, such as normal
distribution. The distributions are usually defined through some parameters. Nonparametric
tests do not require such assumption. Hence nonparametric tests are also known as
distribution free tests. The term nonparametric refers to the fact that there are no parameters
involved in the traditional sense of the term parameter used generally. Nonparametric test
statistics utilize some simple aspects of sample data such as the signs of measurements, order
relationships or category frequencies. Therefore, stretching or compressing the scale does not
alter them. As a consequence, the null distribution of the nonparametric test statistic can be
determined without regard to the shape of the parent population distribution.

The inferences drawn from tests based on the parametric tests such t, F and 2 may be
seriously affected when the parent population distributions is not normal. These effects could
be more if when sample size is small. Thus when there is doubt about the distribution of the
parent population, a nonparametric method should be used. In many situations particularly in
social and behavioral sciences observations are difficult or impossible to take on numerical
scales. Nonparametric tests are well suited under such situations.

First step in statistical testing is formulation of a hypothesis. A hypothesis is a statement about

the population. Its plausibility is evaluated on the basis of information obtained by sampling
from the population. A test generally involves two hypotheses. An assertion about the
population in faviour of the ‘existing’ situations is taken as null hypothesis and denoted as H 0.
The negation of the null hypothesis is known as alternative hypothesis and denoted as H 1. H1
plays a decisive role in classifying a test as one-sided or two sided. We first develop a statistic
T(say) on the basis of the sample observations. The statistic T decides whether to reject or
accept the null hypothesis. Usually T follows some distribution. Based on this distribution the
range of T is divided into two groups; the critical region and the region of acceptance. If the
sample point falls in critical region , we reject null hypothesis. The size of the critical region
depends the risk we wish to incur which ultimately gives the significance level of the test. It is
denoted by . It represent the probability of rejecting the null hypothesis when it is triue, also
known as Type I error. Type II error is the probability of rejecting H0 when H1 is true and
denoted by . Commonly used significance levels are 5% and 1% ( = .05 and .01). Finally, a
conclusion is drawn on the basis of the value of T falling or not falling in the critical region.
Some commonly used nonparametric tests are discussed in the sequel.
 Non-parametric Tests

1. Run Test for Randomness

Run test is used for examining whether or not a set of observations constitutes a random
sample from an infinite population. Test for randomness is of major importance because the
assumption of randomness underlies statistical inference. In addition, tests for randomness are
important for time series analysis. Departure from randomness can take many forms.
H0: Sample values come from a random sequence
H1: Sample values come from a non-random sequence

Test statistic: Let r be the number of runs (a Run is a sequence of sign of same kind bounded
by signs of other kind). For finding the number of runs, the observations are listed in their
order of occurrence. Each observation is denoted by a ‘+’ sign if it is more than the previous
observation and by a ‘-‘ sign if it is less than the previous observation. Total number of runs
up (+s) and down (-) is counted. Too few runs indicate that the sequence is not random (has
persistency) and too many runs also indicate that the sequence is not random (is zigzag).

Critical value: Critical value for the test is obtained from the table for a given value of n and
at desired level of significance (). Let this value is rc.
Decision rule: If rc (lower)  r  rc (upper) accept H0. Otherwise reject H0.
Tied values: If an observation is equal to its preceding observation denote it by zero. While
counting the number of runs ignore it and reduce the value of n accordingly.
Large sample sizes: When sample size is greater than 25 the critical value rc can be obtained
using a normal distribution approximation.

The critical values for two-sided test at 5% level of significance are

rc (lower) =  - 1.96  ; rc (upper) =  + 1.96 

For one-sided tests, these are

rc (left tailed) =  - 1.65  , if r  rc , reject H0
rc (right tailed) =  + 1.65  , if r  rc , reject H0
2n 1
where   16n  29
and  
3 90
Example 1: Data on value of imports of selected agricultural production inputs from U.K. by
a county (in million dollars) during recent 12 years is given below: Is the sequence random?
5.2 5.5 3.8 2.5 8.3 2.1 1.7 10.0 10.0 6.9 7.5 10.6

H0: Sequence is random. H1: Sequence is not random.

5.2 5.5 3.8 2.5 8.3 2.1 1.7 10.0 10.0 6.9 7.5 10.6
+ - - + - - + 0 - + +

Herte n = 11, the number of runs r = 7. Critical nvalues for  = 5% (two sided test) from the
table are rc (lower) = 4 and rc (upper) = 10. Since rc (lower)  r  rc (upper), i.e., observed
r lies between 4 and 10 the H0 is accepted. The sequence is random.

III-58
 Non-parametric Tests

2. WALD-WOLFOWITZ Two-Sample Run Test

Wald –wolfowitz run test is used to examine whether two random samples come from
populations having same distribution. This test can detect differences in averages or spread or
any other important aspect between the two populations. This test is efficient when each
sample size is moderately large (greater than or equal to 10).
H0: Two sample come from populations having same distribution
H1: Two sample come from populations having different distributions

Test statistic: Let r denotes the number of runs. To obtain r, list the n 1+ n2 observations from
two samples in order of magnitude. Denote observations from one sample by x’s and other by
y’s. Count the number of runs.

Critical Value: Difference in location results in few runs and difference in spread also result
in few number of runs. Consequently, critical region for this test is always one-sided. The
critical value to decide whether or not the number of runs are few, is obtained from the table.
The table gives critical value rc for n1 (size of sample 1) and n2 (size of sample 2) at 5% level
of significance.
Decision rule: If r  rc reject H0.
Tie: In case x and y observations have same value place the observation x(y) first if run of x(y)
observation is continuing.
Large sample sizes: For sample sizes larger than 20 critical value rc is given below.
rc =  - 1.96  at 5% level of significance
2n1n2 (2n1n2  n1 
where   1 2n1n2 and   2
 n2 ) (n1  n2 ) (n1 
n1 
n2
Example 2: To determine if a new hybrid seeding produces a bushier flowering plant,
following data was collected. Examine if the data indicate that new hybrid produces larger
shrubs than the current variety?
Shrubs Girth (in inches)
Hybrid x 31.8 32.8 39.2 36.0 30.0 34.5 37.4
Current y 35.5 27.6 21.3 24.8 36.7 30.0
variety
H0: x and y populations are identical
H1: There is some difference in girth of x and y shrubs.

Consider the combined ordered data.

21.3 24.8 27.6 30.0 30.0 31.8 32.8 34.5 35.5 36.0 36.7 37.4 39.2
y y y y x x x x y x y x x

Test statistic r =6 (total number of runs). For n1 = 7 and n2 = 6, critical value rc at 5% level of
significance is 3. Since r > rc, we accept H0 that x and y have identical distribution.

III-59
 Non-parametric Tests

3. Median Test for Two Samples

To test whether or not two samples come from same population median test is used. It is more
efficient than the run test but each sample should be of size 10 at least. In this case, the
hypothesis to be tested is
H0 : Two samples come from populations having same distribution.
H1 : Two samples come from populations having different distribution.
Test Statistic: χ2 (Chi-square). To test the value of test statistics two samples of sizes n 1 and
n2 are combined. Median M of the combined sample of size n = n1 + n2 is obtained. Number of
observations below and above the median M for each sample is determined. This is then
analyzed as a 2 × 2 contingency table in the manner given below.
Number of observations
Sample 1 Sample 2 Total
Above Median a b a+b
Below Median c d c+d
a+c= n1 b+d = n2 n = a+b+ c+d

Test Statistic: χ2 = (ad  bc)2 (a  b  c  d )

(a  c)(b  d)(a  b)(c  d)
Decision rule: if χ ≥ χc2 reject H 0 otherwise accept it.
2

Tie: ties are ignored and n is adjusted accordingly.

Note: This test can be extended to k samples. Number of observations below and above the
combined median M from a 2 × k contingency table.

Example 3: Perform a median test on the problem of example 1 for the testing that the two
samples come from same population.
H0 : x and y populations are identical.
H1 : There is some difference in girth of x and y shrubs. Seventh

value 32.8 is the median of combined ordered sequence.

Number of observations
x y Total
Above M 4 2 6
Below M 2 4 6
6 6 12
12(16  4) 4
χ2 2   1.33 .
= 3
6.6.6.6
Since χ2 =1.33 < χc2 =3.84, H0 is accepted. It is concluded that two samples come from the
same population. There is no significance difference in the girth of hybrid and current variety
of shrub.
Note: This example is simple to demonstrate test procedure. In real situation n should be at
least 20 and each cell frequency at least 5.
III-60
 Non-parametric Tests

4. Sign Test for Matched Pairs

In many situations, comparison of effect of two treatments is of interest but observations
occur in pairs. Thus the two samples are not truly random. Because of such pair-wise
dependence ordinary two sample tests are not appropriate. In such situations when one
member of the pair is associated with the treatment A and the other with treatment B,
nonparametric sign test has wide applicability. It can be applied even when qualitative data
are available. As the name suggests it is based on the signs of the response differences D i. If
ith pairs of observations is denoted by (x i, yi) where x is the effect of treatment A and y to B
then Di = xi – yi. The hypothesis to be tested is
H0 : No difference in the effect of treatments A and B.
H1 : A is better than B.

Test Statistic: Let S be the number of ‘-’ signs.

Critical value: Critical value Sc corresponding to n the number of pairs is given in Table 3.
Significance level is given by α1 as critical region is one sided (left tailed).
Decision rule: If S ≤ Sc reject H0, other wise accept H0.
Tie: In case two values of a pair are equal, reject that pair and reduce the number of
observations accordingly.
Note: In case alternative H1 is that there is some difference in effect of A and B, S represents
either the number of negative signs or the number of positive signs whichever turn out to be
smaller. A critical region is two sided and significant level is given by α2 for finding Sc.

Example 4: In a market study, two brands of lemonade were compared. Each of 50 judges
tasted two samples, one of brand A and one of brand B with the following results. 35
preferred brand A, 10 preferred B, and 5 could not tell the difference. Thus n = 45 and S = 10.
Assuming α1 = 5%, critical value Sc = 16 from Table 3. Since S < S c, we reject H0 of no
difference in favour of the alternative H1 that the brand A is preferred.

5. WILCOXON Signed Rank Test for Matched Pairs

In situations where there is some kind of pairing between observations in the two samples
ordinary two sample tests are not appropriate. Signed rank tests are useful in such situations.
When observations are measured data, signed rank test is more efficient than sign test as it
takes account of the magnitude of the observed differences, if the difference between the
response of the two treatments A and B is to be tested the test hypothesis is
H0 : No difference in the effect of treatments A and B.
H1 : Treatment A is better than B.

Test Statistic: T represents the sum of ranks with negative signs. For calculating T, obtain
the differences Di = xi – yi where xi’s are response of treatment A and y i’s of treatment B.
Rank the absolute values of differences. Smallest give rank 1. Ties are assigned average
ranks. Assign to each rank sign of observed difference. Obtain the sum of negative ranks.

III-61
 Non-parametric Tests

Critical value: Tc is given in Table 4 for n no. of pairs. Significance level is given by α 1 as
critical region is one sided.
Decision rule: T ≤ Tc reject H0, other wise accept it.
Tie: Discard the pair for which difference = 0 and reduce n accordingly. Equal differences are
assigned average ranks.

Example 5: Blood pressure reading of ten patients before and after medication for reducing
the blood pressure are as follows. Test the null hypothesis of no effect against the alternative
that medication is effective.
Patient 1 2 3 4 5 6 7 8 9 10
Before x 86 84 78 90 92 77 89 90 90 86
treatment
After y 80 80 92 79 92 82 88 89 92 83
treatment
Differences 6 4 -14 11 0 -5 1 1 -2 3
Rank 7 5 9 8 Discard 6 1.5 1.5 3 4
Sign + + - + Discard - + + - +

Rank sum of negative differences = 3+6+9 = 18. therefore value of test statistic T= 18. for n =
9 and α1 =5 % Tc = 8 from table 4. Since T>Tc null hypothesis of no effect of medication is
accepted.

6. KOLMOGOROV-SMIRNOV Test
In situations where there is unequal number of observations in two samples Kolmogorov-
Smirnov test is appropriate. This test is used to test whether there is any significance
difference between two treatments A and B (say). The test hypothesis is
H0 : No difference in the effect of treatments A and B.
H1 : There is some difference in the effect of treatments A and B.

Test Statistic: The test statistic is Dm,n = sup F (x)  G (x) , F and G are the sample
m n
empirical distributions of sample observations of two samples respectively with respective
sample sizes m and n. F(xi) is calculated as the average number of sample observations of the
first sample that are less than x i. Similarly G(xi) is calculated. Dm,n is largest value of the
absolute difference between F(x) and G(x).
Critical value: Tabulated value of Dm,n is available for different values of m, n and for
different level of significance. is given in Table 4 for n no. of pairs. Significance level is
given by α1 as critical region is one sided.
Decision rule: If the calculated value of Dm,n is greater than the Tabulated value of Dm,n, H0 is
rejected otherwise it is accepted.

III-62
 Non-parametric Tests

Example 6: The following data represent the lifetimes (hours) of batteries for different
brands:
Brand A 40 30 40 45 55 30
Brand B 50 50 45 55 60 40
Are these brands different with respect to average life?
We first calculate the sample empirical distributions of two samples:
x F6 (x) G6 (x) F6 (x)  G6
(x)
30 2/6 0 2/6
40 4/6 1/6 3/6
45 5/6 2/6 3/6
50 5/6 4/6 1/6
55 1 5/6 1/6
60 1 1 0

D6,6 = sup F6 (x)  G6 = 3/6, from Table the critical value for m = n = 6 at level  = .05 is
(x)
4/6. Since the calculated value of Dm,n is not greater than the Tabulated value, H0 is not
rejected and it is concluded that the average length of life for two brands is the same.

Some Selected references

Bhattacharya, G.K. and Johnson, R.A. Statistics concepts and Methods. New York, John
Wiley and Sons. pp 505-521.
Neave, H.R. and Worthington, P.L. Distribution free tests. London Unwin Hyman, pp 161-
164,328,337-341.
Neter, J.W.W. and Whitemore, G.A. applied Statistics. London, Allyn and Bacon Inc. pp 360-
388.
Ostle, B. Statistics in Reasersch. Ames. Iowa, USA. The Iowa State University. pp 466-473.

PRACTICAL EXERCISES
1. Maximum level of a lake each year for a period of 20 years is given below. It is desired to
test (a) whether the sequence is generated by a random process, or (b) the process contains
a trend. The presence of trend will have significant environmental policy implications.
Year Level Year Level
(Above 190 (Above 190
meters) meters)
1 6.6 11 6.0
2 6.5 12 5.8
3 6.4 13 5.9
4 6.5 14 5.6
5 6.4 15 5.5
6 6.4 16 5.3
7 6.3 17 5.1
8 6.2 18 5.3
9 6.1 19 5.4
10 5.9 20 5.2

III-63
 Non-parametric Tests

2. Seasonal rainfall at two meteorological observations of a district is given below.

Examine by using Run test and Median test whether the rainfall of two observations
can be considered as same.
Seasonal rainfall(cm)
observations
Year A B
1985 25.34 24.31
1986 49.35 45.13
1987 39.62 42.83
1988 42.90 46.94
1989 57.66 57.50
1990 24.89 30.70
1991 50.63 48.37
1992 38.47 38.45
1993 43.25 44.00
1994 50.83 50.00
1995 22.02

3. An experiment was performed to determine if self fertilized and cross fertilized plants
have different growth rates. Pairs of plants one self and other cross fertilized were
planted in 15 pots. Their heights were measured after specified period of time.
(a) perform the sign test ot determine whether there is any difference in the growth
rates of self fertilized and cross fertilized plants.
(b) Perform Wilcoxon signed rank test to determine of crossed plants jave a higher
growth rate.

Plant height (cms) Plant height (cms)

Pair Crossed Self fertilized Pair Crossed Self fertilized
fertilized fertilized
1 45.5 40.0 9 41.2 41.2
2 40.0 42.3 10 42.7 42.0
3 42.8 41.2 11 43.3 42.0
4 41.6 41.3 12 41.0 40.7
5 37.9 36.7 13 46.0 43.5
6 42.5 38.0 14 39.2 40.6
7 44.1 39.8 15 44.3 42.5
8 40.7 38.9

III-64
 Non-parametric Tests

Table 1: Critical values for runs up and down test

α1 = 5 % α1 = 2.5 % α1 = 1 % α1 = 0.5 %
α2 = 10 % α2 = 5 % α2 = 2 % α2 = 1 %
n Lower Upper Lower Upper Lower Upper Lower Upper
3 - - - - - - - -
4 - - - - - - - -
5 1 - 1 - - - - -
6 1 - 1 - 1 - 1 -
7 2 - 2 - 1 - 1 -
8 2 - 2 - 2 - 1 -
9 3 8 3 - 3 - 2 -
10 3 9 3 - 3 - 2 -
11 4 10 4 10 3 - 3 -
12 4 11 4 11 4 - 3 -
13 5 12 5 12 4 12 4 -
14 6 12 5 13 5 13 4 13
15 6 13 6 14 5 14 4 14
16 7 14 6 14 6 15 5 15
17 7 15 7 15 6 16 6 16
18 8 15 7 16 7 16 6 17
19 8 16 8 17 7 17 7 18
20 9 17 8 17 8 18 7 18
21 10 18 9 18 8 19 8 19
22 10 18 10 19 9 20 8 20
23 1 19 10 20 10 20 9 21
24 1 20 11 20 10 21 10 22
25 12 21 11 21 11 22 10 22

α1 : Significance level for one sided test

α2 : Significance level for two sided test

Source: Distribution Free Tests by H.R. Neave and P.L. Worthington. London, Unwin
Hyman.

III-65
 Non-parametric Tests

Table 2: Critical values for the two sample run test.

n1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
n2
2 2 2 2 2 2 2 2 2 2
3 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3
4 2 2 2 3 3 3 3 3 3 3 3 4 4 4 4 4
5 2 2 3 3 3 3 3 4 4 4 4 4 4 4 5 5 5
6 2 2 3 3 3 3 4 4 4 4 5 5 5 5 5 5 6 6
7 2 2 3 3 3 4 4 5 5 5 5 5 6 6 6 6 6 6
8 2 3 3 3 4 4 5 5 5 6 6 6 6 6 7 7 7 7
9 2 3 3 4 4 5 5 5 6 6 6 7 7 7 7 8 8 8
10 2 3 3 4 5 5 5 6 6 7 7 7 7 8 8 8 8 9
11 2 3 4 4 5 5 6 6 7 7 7 8 8 8 9 9 9 9
12 2 2 3 4 4 5 6 6 7 7 7 8 8 8 9 9 9 10 10
13 2 2 3 4 5 5 6 6 7 7 8 8 9 9 9 10 10 10 10
14 2 2 3 4 5 5 6 7 7 8 8 9 9 9 10 10 10 11 11
15 2 2 3 4 5 6 6 7 7 8 8 9 9 10 10 11 11 11 12
16 2 3 4 4 5 6 6 7 8 8 9 9 10 10 11 11 11 12 12
17 2 3 4 4 5 6 7 7 8 9 9 10 10 11 11 11 12 12 13
18 2 3 4 5 5 6 7 8 8 9 9 10 10 11 11 12 12 13 13
19 2 3 4 5 6 6 7 8 8 9 10 10 11 11 12 12 13 13 13
20 2 3 4 5 6 6 7 8 9 9 10 10 11 12 12 13 13 13 14

Significance level 5%

Source: Statistics in Research by Borten Ostle. Ames. Iowa USA. Iowa State University
Press.

III-66
 Non-parametric Tests

Table 3: Critical values for the Sign test (Matched pairs)

α1 5% 2.5 % 1% 0.5 % α1 5% 2.5 % 1% 0.5 %

n α2 10 % 5% 2% 1% n α2 10 % 5% 2% 1%
1 - - - - 26 8 7 6 6
2 - - - - 27 8 7 7 6
3 - - - - 28 9 8 7 6
4 - - - - 29 9 8 7 7
5 0 - - - 30 10 9 8 7
6 0 0 - - 31 10 9 8 7
7 0 0 0 - 32 10 9 8 8
8 1 0 0 0 33 11 10 9 8
9 1 1 0 0 34 11 10 9 9
10 1 1 0 0 35 12 11 10 9
11 2 1 1 0 36 12 11 10 9
12 2 2 1 1 37 13 12 10 10
13 3 2 1 1 38 13 12 11 10
14 3 2 2 1 39 13 12 11 11
15 3 3 2 2 40 14 13 12 11
16 4 3 2 2 41 14 13 12 11
17 4 4 3 2 42 15 14 13 12
18 5 4 3 3 43 15 14 13 12
19 5 4 4 3 44 16 15 13 13
20 5 5 4 3 45 16 15 14 13
21 6 5 4 4 46 16 15 14 13
22 6 5 5 4 47 17 16 15 14
23 7 6 5 4 48 17 16 15 14
24 7 6 5 5 49 18 17 15 15
25 7 7 6 5 50 18 17 16 15

α1 : Significance level for one sided test

α2 : Significance level for two sided test

Source: Distribution Free Tests by H.R. Neave and P.L. Worthington. London, Unwin
Hyman.

III-67
 Non-parametric Tests

Table 4 Critical values for the Wilcoxon signed rank test

α1 5% 2.5 1% 0.5 α1 5% 2.5 % 1 % 0.5 %

% %
n α2 10 % 5% 2% 1% n α2 10 % 5% 2% 1%
1 - - - - 26 110 98 84 75
2 - - - - 27 119 107 92 83
3 - - - - 28 130 116 101 91
4 - - - - 29 140 126 110 100
5 0 - - - 30 151 137 120 109
6 2 0 - - 31 163 147 130 118
7 3 2 0 - 32 175 159 140 128
8 5 3 1 0 33 187 170 151 138
9 8 5 3 1 34 200 182 162 148
10 10 8 5 3 35 213 195 173 159
11 13 10 7 5 36 227 208 185 171
12 17 13 9 7 37 241 221 198 182
13 21 17 12 9 38 256 235 211 194
14 25 21 15 12 39 271 239 224 207
15 30 25 19 15 40 286 264 238 220
16 35 29 23 19 41 302 279 252 233
17 41 34 27 23 42 319 294 266 244
18 47 40 32 27 43 336 310 281 261
19 53 46 37 32 44 353 327 296 276
20 60 52 43 37 45 371 343 312 291
21 67 58 49 42 46 389 361 328 307
22 75 65 55 48 47 407 278 345 322
23 83 73 62 54 48 426 296 362 339
24 91 81 69 61 49 446 415 379 355
25 100 89 76 68 50 466 434 397 373

α1 : Significance level for one sided test

α2 : Significance level for two sided test

Source: Distribution Free Tests by H.R. Neave and P.L. Worthington. London, Unwin
Hyman.

III-68