1
Non-Parametric Statistics ( Distribution-Free Statistics )
Parameter- Numerical Value that describe the characteristics of a population, represented
by Greek letters, ,,, …etc.
Statistic- Numerical Value that describe the characteristics of a sample, represented by
English letters, …etc.
In parametric tests, such as Z, t and F tests, we have assumed that the variable under
consideration follows normal distribution with parameters and 2. In many cases these
assumptions about the form of the distribution are difficult to justify and; in other cases the
form of the distribution may be known, but not suitable for standard statistical inference
procedures. In either situation statistical procedures can be used to tackle various inference
problems, which only make very weak assumptions abut the distribution of the observations,
with out specifying any parametric form for the distribution. These procedures are referred to
as non-parametric or distribution-free.
Advantages-
(1) Normality assumption not necessary.
(2) Easy to compute and understand.
(3) Less information is used. Sometimes even formal ranking or ordering is not required.
Disadvantages-
(1) A Certain amount of information is ignored.
(2) Not efficient or sharp as parametric tests.
When deciding which statistical procedure should be used to analyze a set of observations,
it is important to consider the scale of measurement of the observations.
There are four principal classifications of measurement level: nominal, ordinal, interval and
ratio.
The nominal scale is the lowest level of measurement, being the case where observations
are categories, with no assumption of ordering between categories.
At ordinal scale of measurement it is possible to rank-order all the categories.
In addition to the ordering property, the interval scale of measurement has the property that
distances between observations are defined in terms of fixed and equal units.
[ For example, observed temperatures are measurements at the interval level, however, the
zero of the measurement depends on the units, i.e. C or F].
If in addition the zero of the scale of measurement is inherently fixed, as in the case of
weight or distances, then the observations are said to be measurements at the ratio scale.
One -Sample Sign Test
To test against appropriate hypothesis.
Ex.1.( Small Sample )
In a given year, the average number of days required by a sample of 15 wholesales of drugs
and drug sundries to convert receivables in to cash were
33.9,35.4,37.3,40.9,27.8,35.5,34.6,41.1,30.0,43.2,33.9,41.3,32.0,37.7,and 35.2 days.
Test =32 against > 32 at =0.01.
Solution. Discard the sample value 32.
2
If the value is >32 , give + sign,
< 32 give – sign.
We have + + + + - + + + - + + + + +
X= number of + signs =12.
X follows Binomial distribution with n=14,p=1/2.
From the table P(X12) =0.0007 < =0.01.
Therefore we reject H0 and conclude that the average number of days is more than 32.
If the sample size n is large ( n30) we use the test statistic
Paired-Sample Sign Test
In problems involving paired data, each pair of sample values can be replaced by a + sign if
the first value is greater than the second, a – sign if the first is less than the second or
discard the two values if they are equal. Then we can proceed as in Ex.1.
EX.2. To determine the effectiveness of a new traffic control system, the number of
accidents that occurred at a random sample of 8 dangerous intersections during the 4 weeks
before and after the installation of the new system were observed with the following results:
Before 9 7 3 1 12 12 5 6
After 5 3 4 11 7 5 5 1
+ + - + + + 0 +
Is the new system more effective at =0.10?
Number of + signs = X=6
X follows binomial distribution with n=7 and p= ½.
P(X6) = 0.063.
Since 0.063 < 0.10, we reject H0.
Conclusion: The new traffic control system is effective.
Test of Randomness.
Runs Test
A run is a sequence of identical occurrences( letters or other kinds of symbols) which is
followed and preceded by different occurrences or by none at all.
A sequence of males and females in a queue.
MMFFFMFF
Number of runs =u=4
Number of occurrences of type I = n1=3(M)
Number of occurrences of type II = n2=4(F)
3
EX.3. The J.P.Morgan Index measures the value of US$ against 15 other currencies and is
reported in the Wall Street Journal. During a recent 18-month period, the value of the dollar
at the end of each month was up(U) or down(D) as shown below;
DUDDDDUUDDUUUUDDDD
Test for randomness at =0.05.
For large samples, we use the test statistic
Ex.4.
The following is an arrangement of 30 men(M) and 18 women(W) lined up to purchase
tickets for a rock concert:
MWMWMMMWMWMMMWWMMMMWWMWMMMWMMMWWWMWMMMWMWMMMM
WWM
Test for randomness at =0.05.
H0: Arrangement is random.
H1: Arrangement is not random.
Test statistic:
U=27
Conclusion:
There is not enough evidence to suggest that the arrangement is not random.
4
Test of Randomness
Runs Above and Below the Median.
Any sample consisting of numerical measurements or observations can be tested for
randomness, by using the letters a and b to denote values falling above ,a and below, b the
median of the sample.
EX.5. On 24 successive trips between two cities, a bus carried
24,19,32,28,21,23,26,17,20,28,30,24,13,35,26,21,19,29,27,18,26,14,21,and 23 passengers.
At =0.01, test whether it is reasonable to treat these data if they constitute a random
sample.
The median number of passengers is 23.5.
The arrangement of values becomes
abaabbabbaaabaabb aababbb
No. of runs =u=14, n1 =12, n2= 12.
H0: Arrangement is random.
H1: Arrangement is not random.
At =0.01, reject H0 if u 6 or u>20.
Since 6<u=14<20, we accept H0.
There is no real evidence to indicate that the data do not constitute a random sample.
Rank Sums: The Signed-Rank Test
OR Wilcoxon Signed-rank Test
The sign test utilizes the signs of the differences between the observations and in the one-
sample case, or the signs of the differences between the pairs of observations.
In this test, we rank the differences discarding their signs, assigning rank 1 to the smallest
numerical difference, rank n to the largest numerical difference. Zeros are discarded.
If two or more differences are numerically equal, we assign each one of the mean of the
ranks which they jointly occupy.
Then we base the test on
T+= the sum of the ranks of positive values
T- = the sum of the ranks of negative values,
Or T = Min{ T+ , T- }
As shown in the following table.
H1 Reject H0 if Accept H0 or reserve
decision if
<0 T+ T2 T+> T2
>0 T- T2 T-> T2
0 T T T> T
Where T values are found in table XIV.
Using Ex.1, we have
Obs. value x- 0 (0 =32) Ranks
33.9 1.9 1.5
35.4 3.4 6
5
37.3 5.3 9
40.9 8.9 11
27.8 -4.2 8
35.5 3.5 7
34.6 2.6 4
41.1 9.1 12
30 -2 3
43.2 11.2 14
33.9 1.9 1.5
41.3 9.3 13
32 0 -
37.7 5.7 10
35.2 3.2 5
___________________________________________________-
T+ = 94 T- = 11
H0: =32
H1: >32
Decision rule : Reject H0 if T- 16
Since T- =11<16, reject H0.
We conclude that the average number of days is greater than 32.
Large Sample Test.
If n is 15 or more, the distribution of T+ and T-
Can be approximated closely with normal curves. Then we use
Here we assume that each difference is as likely to be positive as negative.
The decision rule is the same as any other Z-test.
Ex.6.Following are the numbers of new loans made by two branches of a finance company,
respectively, over a 16-day period . At 0.05, test whether, on the average, the same number
of loans is made by both branches, against the alternative that branch 1 makes more loans
than branch 2.
Branch Difference Rank
1 2
8 8 0 -
9 6 3 4.5
2 15 -13 15
6 4 2 2.5
15 7 8 10
13 7 6 8
10 15 -5 7
10 12 -2 2.5
6 13 -7 9
7 6 1 1
18 9 9 11
11 14 -3 4.5
10 6 4 6
3 15 -12 14
6
12 2 10 12
14 3 11 13
T+ = 68 T- = 52
H0: 1 =2
H1: 1>2
Test statistic:
At =0.05, reject H0 if Z > 1.645.
Since 0.45 < 1.645, we accept H0.
We conclude that there is no difference in the average number of loans made by the two
branches or reserve judgement.
Rank Sums: The U- Test
( Mann-Whitney U Test or Wilcoxon Test)
To determine the locations of two populations differ.
[ Non-parametric alternative to the two sample t-test.]
H0: The two population locations are the same.
H0: 1 =2
H1: 1>2 OR 1<2 OR 1 2
Ex.7.
Suppose we wish to determine whether the observations are drawn from the same
populations at =0.05.
I Rank II Rank
18 3.5 20 5
28 8 21 6
23 7 14 1.5
29 9 18 3.5
32 10 14 1.5
_________ ________
T1= 37.5 T2 = 17.5
We begin the U test by arranging the data jointly, as if they comprise one sample.
Then we use the two related statistics
7
Or the statistic U, which is the smaller of the two.
An alternative approach is to use just T = Min{ T1,T2 }, and use the corresponding table to
get the critical values.
In our problem,
H0: 1 =2
H1: 1 2
Test statistic: T= 17.5
Decision rule: Reject H0
If T TL or T TU
Where TL and TU are the table values.
At =0.05{two sided test] n1 = n2= 5 TL =18 TU=37.
Since T=17.5 < TL, we reject H0.
The locations of two populations are not the same.
If we are testing
H0: 1 =2
H1: 1>2
The decision rule is reject H0 if T TU, and if H1: 1<2, reject H0 if T TL .
Large Sample Test
If n1 ,n2 or both exceed 10, the sampling distribution of T is approximately normal and we use
the test statistic
and the test procedure is the same as any other Z-test.
Reject H0 if Z>Z/2 for two-sided test
Z>Z for one-sided ( greater tail ) test,
Z<-Z for one-sided ( lesser tail) test.
8
The H-Test or Kruskal-Walis Test
1/ Comparison of 2 or more populations.
2/ Rank or quantitative data but non-normal.
3/The samples are independent.
H0: The locations of all populations are the same.
H1:At least two population locations are not the same.
Test Statistic:
Where Ti is the sum of the ranks assigned to the ni values of the ith sample, where all
samples are ranked jointly from low to high.
EX.8
A company’s trainees are randomly assigned to groups which are taught a certain industrial
inspection procedure by three different methods. The following are the scores they obtained
in a test at the end of the training period. Are the three methods equally effective at
=0.05. { Use H-test].
A Ri B Ri C Ri
94 17 85 12 89 15
87 14 82 10 67 2
91 16 79 8 72 4.5
74 6 84 11 76 7
86 13 61 1 69 3
97 18 72 4.5
80 9
H0: The locations of all populations are the same.
H1: At least two i ‘s differ.
Test Statistic:
Critical Value: At =0.05 , CV = .
Reject H0 if H >
Since H > 5.991, we reject H0.
Conclusion: The three training methods are not equally effective and at least one is more
effective than the others.
Non-parametric Measure and Test of Correlation
Spearman Rank Correlation
In some cases, the assumptions under regression could not be met or one or both variables
may be ordinal scaled. Then we cannot employ the Pearson correlation or regression
techniques to analyse the relationship between the two variables.
9
Spearman rank correlation coefficient in the sample rs( estimate of the population parameter
s) is defined as
If there are ties in giving ranks, use the average rank to be given to those values.
Ex.( N0. 17.23)
No. of years on Rank No. of Units Rank di
assembly line (X) manufactured (Y)
15.1 6 110 4 -2 4
7.0 8 105 6 2 4
18.6 3 115 2 1 1
23.7 1 127 1 0 0
11.5 7 98 8 -1 1
16.4 4 103 7 -3 9
6.3 9 87 9 0 0
15.4 5 108 5 0 0
19.9 2 112 3 -1 1
Total 20
To test H0:s =0 against a suitable alternative, we use table 9 of the textbook.
For n=9, =0.05, reject H0 if rs >0.600 for one-sided case
rs > 0.683 for two-sided case.
For n> 30, we use