The Mann-Whitney U test: a complete guide

The Mann-Whitney U test pools both groups, ranks every observation from lowest to highest, and then asks whether one group's ranks cluster systematically higher than the other's. The raw values are used only to establish the ordering; after that they are discarded.

That single design decision explains most of the test's behaviour. Because a score of 41 and a score of 4,100 both simply rank tenth, an extreme value cannot distort the result the way it would distort a mean. The cost is that genuine information about the size of the gaps is thrown away, which is why the test has slightly less power than a t-test when the t-test's assumptions do hold.

This also explains a property that surprises people: Mann-Whitney is not simply a weaker t-test. When the data are genuinely normal, it retains about 95% of the t-test's power — a small price. When the data are heavily skewed or contain outliers, it can be substantially more powerful than the t-test, because a single extreme value inflates the t-test's estimate of variability and shrinks the statistic. The choice is therefore not a trade of rigour for safety. On the kind of data that make people reach for it, the rank-based test is frequently the better-performing option as well as the more defensible one.

A common misconception is that you should test for normality first and let the result choose the test. In practice this is unreliable: with small samples the normality test has little power to detect real departures, and with large samples it flags trivial ones while the t-test is already robust. Choose on the measurement level and the visible shape of the distribution, and say so in your methods section.

Sample size is worth a note here too. The exact test becomes computationally heavy as groups grow, which is why software switches to the normal approximation above roughly 20 per group. The approximation is good at that size and better beyond it, so the switch costs nothing in practice — but it does mean the statistic you report changes form, from a U compared against a critical value to a z with an associated p. Both are the same test; only the presentation differs.

That last row is the one most write-ups skip, and it changes what your significant result is entitled to say.

Strictly, Mann-Whitney tests whether a randomly chosen observation from one group is likely to exceed a randomly chosen observation from the other — a property called stochastic dominance. Only when the two distributions have the same shape and spread does this reduce to a statement about medians.

Two groups of five participants complete a task. Group A scores 12, 15, 19, 24, 31. Group B scores 17, 22, 28, 35, 41.

Step 1 — pool and rank

Step 2 — sum the ranks

Rₐ = 1 + 2 + 4 + 6 + 8 = 21 and Rₖ = 3 + 5 + 7 + 9 + 10 = 34

As a check, the two must total N(N+1)/2 = 10 × 11 / 2 = 55. Here 21 + 34 = 55.

Step 3 — compute U for each group

Uₐ = nₐnₖ + nₐ(nₐ+1)/2 − Rₐ = 25 + 15 − 21 = 19

Uₖ = nₐnₖ + nₖ(nₖ+1)/2 − Rₖ = 25 + 15 − 34 = 6

A second check: Uₐ + Uₖ must equal nₐnₖ = 25. Here 19 + 6 = 25.

Step 4 — the decision

The test statistic is the smaller of the two, so U = 6. For nₐ = nₖ = 5 at α = .05 two-tailed, the critical value is 2. Since U = 6 is greater than 2, we do not reject the null: p = .222 (exact).

A word on why the test is so widely used in health and education research specifically. Outcome measures in those fields are frequently bounded scales, symptom counts or ratings — data that are ordinal by construction, bounded at both ends, and rarely symmetric. The t-test's assumptions are a poor fit for that shape of data, and transforming it to force normality tends to make the results harder to interpret rather than easier. A rank-based test sidesteps the problem entirely: it makes no distributional assumption at all, and the conclusion it supports translates naturally into plain language.

Handling ties

When two or more observations share a value, each receives the average of the ranks they would have occupied. Three observations competing for ranks 4, 5 and 6 all receive rank 5.

Ties reduce the variance of the rank sums, so the standard error must be adjusted. Every statistical package applies this correction automatically, but the hand formula does not — with many ties, a hand calculation will be slightly wrong. Likert data typically produce a great many ties, which is a good reason to let software do the arithmetic.

Floor and ceiling effects deserve a mention, because they arise constantly with rating scales. If a large share of respondents choose the same option — everyone selecting 5 on a 5-point item — the ranks become heavily tied and the test loses power sharply, since ranking cannot separate observations that are identical. When the effect is that pronounced, no rank-based test will rescue the analysis. The honest response is to report the distribution directly and note the ceiling, rather than presenting a significance test the data cannot support.

Exact test or normal approximation?

With larger samples, U is converted to a z score:

z = (U − μ₢) / σ₢, where μ₢ = nₐnₖ/2 and σ₢ = √[nₐnₖ(N+1)/12]

Five per group demonstrates the mechanics but has almost no power. Here is the same test on a realistic sample, using the normal approximation your software will apply.

Two groups of 40 rate a service on a 7-point scale. The rank sums come out as Rₐ = 1,384 and Rₖ = 1,856, which correctly total N(N+1)/2 = 80 × 81 / 2 = 3,240.

Step 1 — U

Uₐ = nₐnₖ + nₐ(nₐ+1)/2 − Rₐ = 1,600 + 820 − 1,384 = 1,036

Uₖ = 1,600 + 820 − 1,856 = 564. These sum to 1,600, as they must.

Step 2 — the normal approximation

μ₢ = nₐnₖ/2 = 800

σ₢ = √[nₐnₖ(N+1)/12] = √[1,600 × 81 / 12] = √10,800 = 103.92

z = (564 − 800) / 103.92 = −2.271, giving a two-tailed p = .023.

Step 3 — effect size

r = |z| / √N = 2.271 / √80 = 2.271 / 8.944 = 0.254 — a small to medium effect.

Mann-Whitney is one member of a family, each matched to a design. Choosing the wrong member is a common and easily avoided error.

The pattern is consistent: each rank-based test replaces the raw values with their ranks and asks the same structural question. They are all somewhat less powerful than their parametric counterparts when the parametric assumptions genuinely hold, and considerably more trustworthy when they do not.

The usual effect size is the rank-biserial correlation, most simply computed from the z score:

r = z / √N, where N is the total sample size across both groups

An alternative that is easier to explain to a non-technical reader is the common language effect size: the probability that a randomly chosen member of one group scores higher than a randomly chosen member of the other. It is computed directly from U as U / (nₐnₖ). In our worked example, 19 / 25 = 0.76 — a participant from group B outscores one from group A about 76% of the time.

1. Claiming a difference in medians without checking distribution shape

Only valid when the two distributions have similar shapes. Otherwise the claim is about one group tending to score higher.

2. Reporting means alongside a Mann-Whitney

Report medians and interquartile ranges — the statistics that match the test you chose.

3. Using it on paired data

Mann-Whitney requires independent groups. For before-and-after measurements use the Wilcoxon signed-rank test.

4. Choosing the test based on a normality test alone

Normality tests are underpowered in small samples and oversensitive in large ones. Decide on measurement level and distribution shape.

5. Running it on three or more groups pairwise

That inflates the false positive rate. Use Kruskal-Wallis, then follow up with corrected pairwise comparisons.

6. Omitting the effect size

A p-value alone says nothing about magnitude. Report the rank-biserial correlation or the common language effect size.