The Mann-Whitney U Test, Explained

After reading this you will know when to reach for the Mann-Whitney U test instead of a t-test, how to compute U by hand from ranks, and how to read the p-value and the rank-biserial effect size.

What the test does

The Mann-Whitney U test asks a simple question: if you draw one value from group A and one value from group B at random, how often does the A value come out larger? If the two groups are identical, that should happen about half the time. If A values tend to sit above B values, the test detects it.

It works on ranks, not on the raw numbers. That makes it robust to skew and to outliers, and it means you never assume the data are normally distributed. A single extreme value that would wreck a mean barely moves a rank.

Consider five recovery times in days for treatment A: 12, 15, 11, 18, 14, and five for treatment B: 20, 22, 17, 25, 19. Every A value is smaller than every B value. The test should return the strongest possible signal, and below you will see it does exactly that.

When to use it, and when not to

Use the Mann-Whitney U test when you have two independent samples and you want to compare their distributions, but you cannot trust the normality assumption behind a two-sample t-test. Small samples, visibly skewed data, or ordinal outcomes such as Likert ratings all point to this test.

Do not use it when your two samples are paired (the same subject measured twice, or matched pairs). For that, use the Wilcoxon signed-rank test. For three or more independent groups, use the Kruskal-Wallis H test, which is the direct generalization. If your data really are roughly normal, the two-sample t-test has slightly more power and is the conventional choice.

The formula and the intuition

Pool both samples, sort all the values, and assign ranks from 1 to n_A + n_B. Sum the ranks that belong to group A; call that R_A. The U statistic for A is:

Here n_A is the size of sample A, n_B is the size of sample B, and R_A is the sum of A's ranks. The subtracted term removes the minimum possible rank sum, so U_A counts the number of pairs (one A value, one B value) in which the A value wins.

The two U values are linked by a fixed identity, so you only need one:

The reported statistic is usually the smaller of the two. For the p-value with larger samples, U is standardized. Under the null hypothesis its mean and standard deviation are:

Then z = (U - \mu_U) / \sigma_U feeds a normal approximation. For small samples the calculator uses the exact distribution of U instead, which is more accurate when ties are absent.

A worked example with the demo data

Effect size: rank-biserial correlation

A p-value tells you whether the difference is detectable, not how large it is. The rank-biserial correlation fills that gap:

Here U is the value tied to group A, and the product n_A n_B is the total number of A-versus-B pairs. The measure runs from -1 to +1. A value of 0 means A wins exactly half the pairings. A value near \pm 1 means one group dominates.

For the demo data using U_A = 1: r = 1 - \frac{2 \cdot 1}{25} = 1 - 0.08 = 0.92. That is a very large effect, which matches the fact that A loses 24 of 25 pairings.

Reading the results

Three numbers matter. Read them in order.

U

The count of pairings won by one group, from 0 to n_A n_B. Values near the midpoint n_A n_B / 2 mean no difference. For the demo, the midpoint is 12.5 and the observed U is 1, far from it.

p-value

The probability of a U this extreme or more if the distributions were identical. Compare it to your chosen threshold, commonly 0.05. The demo gives 0.016, so you reject the null.

Rank-biserial r

The size of the difference on a -1 to +1 scale. Rough guides: 0.1 small, 0.3 medium, 0.5 large. The demo's 0.92 is very large.

Report all three. A p-value alone can be significant with a trivial effect if the samples are large, or nonsignificant with a big effect if they are tiny.

Common mistakes

Watch for ties. When two values are equal they share the average of the ranks they would occupy. Many ties inflate the variance and make the normal approximation less accurate; the calculator applies a tie correction, but heavy ties still weaken the test.

Do not read a significant result as "the medians differ" unless the two distributions have similar shape. With different spreads the test can flag a difference in variability rather than in center.

Finally, the test needs both samples to be genuine independent draws. If group membership is confounded with time, batch, or operator, the U statistic will pick up that nuisance rather than the effect you care about.

Related tools on this site

If your data look roughly normal, compare the result against the t-test calculator. Check that assumption first with the normality test battery and the Levene's test for equal variances. For more than two groups use the Kruskal-Wallis H test or, for means, one-way ANOVA. If you have a test statistic in hand and only need the tail probability, the p-value calculator converts z, t, χ² or F values directly. For categorical outcomes, see the chi-square test of independence and Fisher's exact test.

Related tools

Chi-Square Test of Independence

Test association between two categorical variables from a contingency table.

One-Way ANOVA Calculator

Compare the means of three or more groups with an F-test.

P-value Calculator

Convert z, t, χ² or F statistics into p-values.

t-Test Calculator

One-sample, two-sample (pooled/Welch) and paired t-tests from raw data.

A/B Test Significance Calculator

Two-proportion z-test with p-values, uplift and CI of the difference.

Chi-Square Goodness-of-Fit Test

Test whether observed counts match an expected distribution.