Based on the relationship between the Mann-Whitney Test and the Wilcoxon Rank-Sum Test, we can modify the exact test described in Wilcoxon Rank-Sum Exact Test to provide an exact test for Mann-Whitney.

The one-tail Wilcoxon Rank-Sum exact test is as follows:

p-value = PERMDIST(*W*, *n*1, *n*2, TRUE)

T-crit = PERMINV(*α*, *n*1, *n*2)

This turns into the one-tail Mann-Whitney exact test is as follows:

p-value = PERM2DIST(COMBIN(MIN(*n*1, *n*2)+1,2)+*U*, *n*1, *n*2, TRUE)

U-crit = PERM2INV(*α*, *n*1, *n*2)-COMBIN(MIN(*n*1, *n*2)+1, 2)

The two-tail Wilcoxon Rank-Sum exact test is as follows:

p-value = 2 * PERMDIST(*W*, *n*1, *n*2, TRUE)

T-crit = PERMINV(*α*/2, *n*1, *n*2)

This turns into the two-tail Mann-Whitney exact test is as follows:

p-value = 2 * PERM2DIST(COMBIN(MIN(*n*1, *n*2)+1,2)+*U*, *n*1, *n*2, TRUE)

U-crit = PERM2INV(*α*/2, *n*1, *n*2)-COMBIN(MIN(*n*1, *n*2)+1, 2)

The p-values for the two-tailed test are correct as long as the p-value for the one-tailed test is at most .5.

**Real Statistics Functions**: The Real Statistics Resource Pack contains the following functions which implement the process described above.

**MANNDIST**(*x*, *n*1, *n*2, *tails*) = value of the Mann-Whitney distribution at *x* based on and elements, where *tails* = 1 (default) or 2; i.e. the p-value as defined above

**MANNINV**(*p*, *n*1, *n*2, *tails*) = inverse of the Mann-Whitney distribution at *p*; i.e. the least value of such that MANNDIST(*x*, *n*1, *n*2, *tails*) ≥ *p*, where *tails* = 1 (default) or 2; i.e. U-crit as defined above

**MANN_EXACT**(R1, R2, *tails*) = p-value of the Mann-Whitney exact test on the data in ranges R1 and R2, where *tails* = 1 or 2 (default)

Thus for the two-tailed test for Example 1 of Mann-Whitney Test, we have

MANNDIST(H11,H5,I5,2) = .117928

MANNINV(.05,H5,I5,2) = 34

MANN_EXACT(A6:A17,B6:B16,2) = .117928

**Observation**: These functions are quite computationally intensive. Depending on the power of your computer they compute quite quickly for values of *n = n*1 + *n*2 up to about 24 or 25, a little higher when *n*1 is much higher than *n*2. For larger values of *n* these functions are probably too slow to be of practical use for most situations.

Dear Charles,

could you please explain me why you multiply the PERM2DIST by two to obtain the p-value for the two-tail Mann-Whitney exact test?

–> p-value = 2 * PERM2DIST(COMBIN(MIN(n1, n2)+1,2)+U, n1, n2, TRUE)

Does that have something to do with the test being two-tailed? I’m currently writing my thesis and in my text book I have the following data: μ = 12; standard deviation = 4,59; the calculated U-value = 11. In the text book they then calculate the normal distribution of these values and multiply the obtained value by 2 to get the p-value (0,828), but unfortunately they don’t explain why they do it. If you draw a quick sketch of the distribution, you see that the U is left of the μ and if you want to calculate the area from -infinity to U, I wouldn’t multiply it by two – could you please explain me why you do that? Thanks a lot.

Best regards from Austria,

Sonja

Sonja,

For a symmetric distribution the two-tailed p-value is double the one-tailed p-value.

Charles

Thanks a lot, Charles!