Chapter – 11

Non parametric tests

11.0. Objectives:
Having gone through the lesson the learner will be able to understand
1. Types and application of non parametric statistics

11.1. Non-Parametric tests

The parametric and non-parametric statistical tests are commonly employed

in behavioural researches. A parametric statistical test is one, which specifies certain

conditions about the parameter of the population from which a sample is taken. Such

statistical tests are considered to be more powerful than non-parametric tests and

should be used if their basic requirements or assumptions are met. These

assumptions are based upon the nature of the population distribution as well as

upon the type of measurement scales used in quantifying the data. A non-

parametric statistical test is one, which does not specify any conditions about the

parameter of the population from which the sample is drawn. Since these statistical

tests do not make any specified and precise assumption about the form of the

distribution of the population, these are also known as distribution free statistics. The

non-parametric statistics do not specify any rigid conditions like parametric statistical

tests although certain assumptions are associated with them. Non-parametric tests,

the variables under study should be continuous and the observations should be

independent. Some of the important non-parametric tests are Chi Square, Sign test,

Median test, Sign rank test, U’ test.

Three are several advantages and disadvantages of non-parametric statistical

tests in comparison to parametric statistical tests. Some of the important advantages

are as follows.

124
 1. Simplicity and facilitation in derivation: Most of the non-parametric statistics

can be derived by using simple computational formulas. This advantage does not lie

with most of the parametric statistics, the derivat'lon of which requires an advanced

knowledge mathematics.

2. Wider scope of application: Since non-parametric statistics as compared to

parametric statistics are based upon fewer and less rigid and elaborate assumptions

regarding the form of population distribution, they can be easily applied to much

wider situations.

3. Speed of application: When the sample size is small, calculation of non-

parametric statistics is faster than parametric statistics.

4. Susceptibility to violation of assumptions: In case of non-parametric

statistics assumptions are fewer and less elaborate than in the case of parametric

statistics. Therefore, assumptions of non-parametric statistics are less susceptible to

violation. Not only this, these violations are easier to check and can be readily and

economically taken care of with the non-parametric statistics.

5. Type of measurement required: Non-parametric statistics require

measurement based upon a nominal scale and ordinal scale whereas parametric

statistics require measurement based upon the interval scale and/or ratio scale. As

treatments associated with either nominal scale or ordinal scale are easier than

treatments associated with either interval scale or ratio scale, the parametric

statistics have a better case for applicability than the non-parametric statistics.

6. Impact of sample size: When sample size is 10 or less than 10, non-

parametric statistics are easier, quicker and more efficient than the parametric

statistics. If the assumptions of parametric statistics are violated for such small

cases, the result is likely to get badly affected. Therefore, for this sample size, non-

125
parametric statistics are always superior to the parametric statistics. The reader

should note that as the sample size increases, non-parametric statistics become

time-consuming, laborious and less efficient than the parametric statistics.

7. Statistical efficiency: Non-parametric tests are often more convenient than

the parametric tests. If the data are such that they meet all assumptions of non-

parametric statistics but not of parametric statistics, non-parametric statistics have

statistical efficiency equal to parametric statistics. If both parametric and non-

parametric statistics are applied to the data, which fulfil all assumptions of

parametric tests, the distribution-free statistics become more eficient with a small

sample size but they become less and less efficient as sample size increases.

Despite the above advantages certain disadvantages are associated with

them. The main ones are as given below:

1. According to Moses (1952), the non-parametric statistics have lower

statistical efficiency than parametric statistics when sample size is large, preferably

above 30.

2. If all assumptions of parametric statistics are fulfilled, Siegel (1956) and

Siegel & Castellan (1988) consider the use of non-parametric statistics as simply

"wasteful of data".

3. It is also said that the probability tables for testing the significance of non-

parametric statistics are widely scattered in different publications, which, for a

behavioural scientist, is difficult to locate and interpret.

11.1. Chi-square (X2) test

The Chi-suare is one of the most important non-parametric statistics, which is

used for several purposes. It is a non-parametric statistic because it involves no

assumption regarding the normalcy of distribution or homogeneity of the variances.

126
The chi-square test is used when the data are expressed in terms of frequencies of

proportions or percentages. The chi-square applies only to discrete data. However,

any continuous data can be reduced to the categories in such a ways that they can

be treated as discrete data and then, the application of chi-square is justified. Chi-

square is not stable when computed from a table in which any observed frequency is

less than 5, unless a correction for continuity called Yates correction is made.

The additive property of Chi-square: When several X 2 s have been computed from

independent experiments (i.e., from tables based upon different samples), these

may be summed to give a new chi-square with df = the sum of the separate dfs. The

fact that chi-squares may be added to provide an overall test of a hypothesis is

important in many experimental studies. Combining the data from several

experiments will often yield a conclusive result, when separate experiments, taken

alone, provide only indications.

The formula for calculating X2 is given below

(fo - fe)2
X2 =
Σ fe

Where X2 = chi-square, fo=obtained or observed frequency; and fe = expected

frequency. df = (r-1)(c-1) where r is number of rows and c is number of columns.
Example: The items in an attitude scale are answered by underlying one of the

following phrases: Strongly approve, approve, indifferent, disapprove, and strongly

disapprove. The distribution of answers to an item marked by 100 subjects is shown

in table-11.1. Do these answers diverge significantly from the distribution to be

expected if there are no preferences in the group.

127
Table -11.1:
Strongly Strongly
approve Approve Indifferent Disapprove disapprove
Observed (fo) 23 18 24 17 18 100
Expected (fe) 20 20 20 20 20 100
(fo - fe) 3 2 4 3 2
(fo - fe)2 9 4 16 9 4
(fo - fe)2 .45 .20 .80 .45 .20
fe

(fo - fe)2
X2 =
Σ fe
X2 = 2.10 df = 4 P lies between .70 and .80 Which shows that the chi-

square obtained is not significant.

11.2. Sign test

The sign test gets its name from the fact that it uses plus and minus signs

rather than quantitative measures as its data. It is particularly useful for research in

which quantitative measurement is impossible or infeasible, but in which it is

possible to rank with respect to each other the two members of each pair.

The sing test is applicable to the case of two related samples when the

experimenter wishes to establish that two conditions are different. The only

assumption underlying this test is that the variable under consideration has a

continuous distribution.

128
Example: Sign test applied to data consisting of 10 pairs of scores obtained under
two conditions of spelling (1) words in context (2) words spelled separately.
(1) (2) (3=1-2)
15 12 + Signs
18 15 + + 7
9 10 - - 2
15 16 - 0 1
18 18 0 10
12 10 +
15 12 +
16 13 +
14 12 +
22 19 +
Tables are available which give the number of signs necessary for

significance at different levels, when N is small. When the number of pairs is as

large as 20, X2 test may be applied.

11.3. Wilcoxon Sign rank test

The test we have just discussed, the sign test, utilizes information simply

about the direction of the differences within pairs. If the relative magnitude as well as

the direction of the differences is considered, a more powerful test can be made.

The Wilcoxon matched pairs signed ranks test does just that: it gives more weight to

a pair which shown a large difference between the two conditions than to a pair

which shows a small difference.

The Wilcoxon test is a most useful test for the behavioural scientist. With

behavioural data, it is not uncommon that the researcher can (a) tell which member

of a pair is “greater than” which i.e., tell the sign of the difference between any pair,

and (b) rank the differences in order of absolute size. That is, he can make the

129
judgement of “greater than” between any pair’s two performance, and also can make

that judgement between any two difference scores arising from any two pairs. With

such information, the experimenter may use the Wilcoxon test.

Method: Let d = the difference score for any matched pair, representing the

difference between the pair’s scores under the two treatments. Each pair has one d.

To use the Wilcoxon test, rank all the d’s without regard to sign: give rank of 1 to the

smallest d, the rank 2 to the next smallest, etc. when one ranks scores without

respect to sign a d of –1 is given a lower rank than a d of either –2 or +2.

Then to each rank affix the sing of the difference. That is, indicate which

ranks arose from negative d’s and which ranks arose from positive d’s.

Occasionally the two scores of any pair are equal. That is, no difference

between the two treatments is observed for that pair, so that d=0. Such pairs are

dropped from the analysis. This is the same practice that we follow with the sign test.

N= the number of matched pairs minus the number of Pairs whose d=0.

Another sort of tie can occur. Two or more d’s can be of the same size. We

assign such tied cases the same rank. The rank assigned is the average of the

ranks which would have been assigned if the d’s had differed slightly. Thus three

pairs might yield d’s of –1, -1, and +1. Each pair would be assigned the rank of 2, for

(1+2+3)/3= 2. Then the next d in order would receive the rank of 4, because ranks

1,2 and 3 have already been used.

Let T= the smaller sum of like signed ranks. That it, T is either the sum of the

positive ranks or the sum of the negative ranks, whichever sum is smaller. Table of

signed ranks gives various values of T and their associated levels of significance.

That is, if an observed T is equal to or less than the value in the body of signed rank

130
table under a particular significance level for the observed value of N, the null

hypothesis may then be rejected at the level of significance.

When N is larger than 25, signed rank table of significance cannot be used.

However, it can be shown that such cases the sum of the ranks, T, is practically

normally distributed and the significance is obtained by converting T into z value

and a z value of above 1.96 indicates significance below 0.05 level.

N(N+1)
T-
4
z=
N(N+1)(2N+1)
24

Example: Social perceptiveness score of twins in nursery school and at home.

Social Social Rank with
Pair perceptiveness perceptiveness d Rank less
score of twin in score of twin at of d frequent
nursery school home sign
A 82 63 19 7
B 69 42 27 8
C 73 74 -1 -1 1
D 43 37 6 4
E 58 51 7 5
F 56 43 13 6
G 76 80 -4 -3 3
H 82 79 3 2
T=4

11.4. Median test

The median test is used to see if two groups (not necessarily of the same

size) come from the same population or from populations having the same median.

In the median test the null hypothesis is that there is no difference between the two

sets of scores because they have been taken from the same population.

131
 To perform the median test, we first determine the median score for the

combined group (i.e. the median for all scores in both samples). Then we

dichotomize both sets of scores at that combined median and cast these data in a

2x2 table and calculate the chi-square. Above Vs below the common median

constitutes one category and group-I Vs group-II the other category. However when

N1 and N2 are small the chi-square test is not accurate and the exact method of

computing probabilities should be used.

Group-I Group-II Total

No. of scores above combined median A B A+B

No. of scores below combined median C D C+D

Total A+C B+D N = n 1 + n2

Example: A clinical psychologist wants to investigate the effects of a tranquilizing

drug upon hand tremor. Fourteen psychiatric patients are given the drug, and 18

other patients matched for age and sex are given placebo (i.e., a harmless dose).

Since the medication is in pill form the patients do not know whether they are getting

the drug or not. The first group is the experimental, the second the control group.

Tremor is measured by a steadiness tester. The following table gives the

scores of the two groups: a + sign indicates a score above the common median, a –

sign a score below the common median. Does the drug increase steadiness as

shown by lower scores in the experimental group? As we are concerned only if the

drug reduces tremor, this is a one tailed test.

Table: Median test applied to experimental and control groups. Plus signs indicate

scores above the median, minus signs scores below the common median.

132
 N=14 Sign N=18 Sign
Experimental control
53 + 48 -
39 - 65 +
63 + 66 +
36 - 38 -
47 - 36 -
58 + 45 -
44 - 59 +
38 - 53 +
59 + 58 +
36 - 42 -
42 - 70 +
43 - 71 +
46 - 65 +
46 - 46 -
55 +
61 +
62 +
53 +
Common median = 49.5
The common median is 49.5. in the experimental group 4 scores are above

and 10 below the common median instead of the 7 above and 7 below to be

expected by chance. In the control group, 12 scores are above and 6 below the

common median instead of the expected 9 in each category. These frequencies are

entered in the following table and X2 is computed by formula with correction for

continuity.

Below Above Total

median median
Experimental 10 4 14
Control 6 12 18
16 16 32
The X2 is, when corrected = 3.17. a X 2 of 3.17 with 1 degree of freedom yields
a p which lies at .08, about midway between .05 and .10.
X2 is generally applicable in the median test. However, when N 1 and N2 are
small the X2 test is not accurate and the exact method of computing probabilities
should be used.

133
11.5. Mann Whitney U test
When at least ordinal measurement has been achieved, the Mann Whitney U

test may be used to test whether two independent groups have been drawn from the

same population. This is one of the most powerful of the nonparametric tests, and it

is a most useful alternative to the parametric t test when the researcher wishes to

avoid the t test’s assumptions, or when the measurement in the research is weaker

than interval scaling.

Let n1 = the number of cases in the smaller of two independent groups, and n 2

= the number of cases in the larger. To apply the U test, we first combine the

observations or scores from both groups, and rank these in order of increasing size.

In this ranking, algebraic size is considered, i.e, the lowest ranks are assigned to the

largest negative numbers, if any. Then U is computed by the formula

U = n1n2 + N1(n1+1) - R1
2
And

U = n1n2 + N2(n2+1) - R2
2

Where R1= sum of the ranks assigned to group whose sample size is n1.

R2= sum of the ranks assigned to group whose sample size is n2.

The above formulas yield different U’s. It is the smaller of these that we want.

The larger value is U’. The significance of the U is obtained by using the tables of

significance. Two types of tables are available where neither n1 or 22 is larger than

8, and when n2 is between 9 and 20. When n 2 > 20 the significance is determined

134
by converting U into z scores. A z score of 1.96 or above indicates significance

beyond 0.05 level.

n1n2
U-
2
z=
(n1)(n2)(n1+n2+1)
12

For example, we might have used this method in finding the value of U for the data

given in the example for small samples above. The E and C scores fore that

example are given again in the following table, with their ranks.

E score Rank C score Rank

78 7 110 9
64 4 70 5
75 6 53 3
45 1 51 2
82 8
R2=26 R1=19

For those data, R1=19 and R2= 26, and it will be remembered that n 1=4 and n2=5. By
applying formulas

U = n1n2 + n2(n2+1) - R2 = (4)(5)+ 5(5+1) - 26 =9

2 2

U = n1n2 + n1(n1+1) - R1 = (4)(5)+ 4(4+1) - 19 = 11

2 2
Of the two U’s 9 and 11 the former (9) is smaller and is the actual U and the latter
(11) is larger and is U’.

135
11.6. Unit end questions
1. Differentiate parametric tests with non parametric tests
2. Describe Chi-square test
3. Differentiate the conditions under which Wilcoxon sing rank test and Manwitly
U test are used.
11.7. Source
1. Sidney Siegel , Nonparametric statistics for the behavioural sciences,. New
Delhi: Mc.Graw-Hill.
2. Singh A.K. (1997) Tests measurements and Research Methods in
Behavioural science Patna: Bharati Bhavan Publishers and Distributors.

136