The Wilcoxon Two Sample Test

Example:
#A = #B = , W = , p <=

Characteristics:
A most usefull test to see whether the values in two samples differ in size. It resembles the &B=> Median-Test in scope, but it is much more sensitive. In fact, for large numbers it is almost as sensitive as the &B=> Two Sample Student t-test. For small numbers with unknown distributions this test is even more sensitive than the Student t-test.
As it is only on rare occasions that we do know that values are Normal distributed, this test is to be preferred over the Student t-test.

H0:
The populations from which the two samples are taken have identical median values. To be complete, the two populations have identical distributions.

Assumptions:
None realy.

Scale:
Ordinal.

Procedure:
Rank order all N = m + n values from both samples (m and n) combined. Sum the ranks of the smallest sample (Wsmallest). This value is used to determine the level of significance.

Level of Significance:
Look up the level of significance in a table using Wsmallest, m and n.
Calculating the exact level of significance is based on calculating all possible permutations of ranks over both samples. This is computationally demanding if n and m are larger than 7.

Approximation:
If m>10 and n>10,
Z = ( Wsmallest - 0.5 - m * ( m + n + 1 ) / 2 ) / sqrt( m * n * ( m + n + 1 ) / 12 )
is approximately Normal distributed.
(Use Wsmallest - 0.5 if Wsmallest > N*(N+1)/4, else use Wsmallest + 0.5)

Remarks:

Recently (summer 2006), a user of this web-site has discovered a bug in the calculations of the normal approximation. If the sum of the ranks of the sample with fewer observations was greater than the sum of the ranks of the sample with more observations, the script calculates the p-value using the smallest sum but then uses the smaller sample number as m in the normal approximation. This has been corrected as of November 2006.

In this example, exact probabilities are calculated for m <= 10 or n <= 10. If both are larger than 7 this can take more time than is available within this system (the number of calculations grows as N!/(m!*n!), with N!=N*(N-1)*(N-2)*...*1). Therefore, if it is anticipated that the calculations take too much time, the Normal approximation is used (ie, too many permutations to check). However, the resulting values are unreliable and this will be indicated with a *. You are advised to check the level of significance in a table.
For m > 10 and n > 10 the Normal approximation is used. A perl script of the test is available here. A minimalist Windows version (with dosperl interpreter) is available here (<500 kB).

Return to: Statistics