Mann-Whitney U Test Assumptions

Search “Mann-Whitney U test assumptions” and a surprising number of results tell you the test has none — that it is the “assumption-free” alternative to the independent-samples t-test. That is an overcorrection. The Mann-Whitney U test drops the t-test’s normality assumption, but it does not drop every assumption, and the ones it keeps decide whether you can defend your conclusion when a reviewer pushes back on it.

The Mann-Whitney U test (also called the Wilcoxon rank-sum test; Mann, H. B., & Whitney, D. R., 1947, “On a Test of Whether one of Two Random Variables is Stochastically Larger than the Other,” The Annals of Mathematical Statistics, 18(1), 50–60) compares two independent groups by ranking all observations together and comparing the rank sums. Its formal null hypothesis, per SciPy’s own documentation, is that “the distribution underlying sample x is the same as the distribution underlying sample y” — a statement about entire distributions, not means or medians specifically. This guide covers what the test actually requires, what it doesn’t, and the one assumption — equal distribution shape — that determines whether you’re allowed to describe a significant result as “a difference in medians” at all.

The two assumptions the test genuinely needs

Strip away the folklore and two requirements remain, both about how the data were collected rather than how they’re distributed:

Assumption	What it means in practice	What happens if it’s violated
Independence	Observations are independent within each group (one participant’s value doesn’t influence another’s) and the two groups are independent of each other (not the same people measured twice, and not matched/paired cases).	The test is invalid, not just less powerful. A repeated-measures or matched-pairs design that gets fed into a Mann-Whitney U instead of the Wilcoxon signed-rank test understates the correlation between measurements and can produce a misleading p-value.
Ordinal or continuous measurement	The outcome variable can be meaningfully ranked — a Likert-type item, a count, a skewed continuous measure, or an interval/ratio variable used ordinally. See levels of measurement.	On genuinely nominal/categorical data (blood type, department name), ranking is meaningless and the test doesn’t apply — use a chi-square or Fisher’s exact test instead.

That’s the complete list of requirements for the test to be valid. Notice what isn’t on it: no assumption about the shape of either distribution, no assumption of equal variances, and no minimum sample size beyond what you need for adequate power. This is exactly why the test exists — it’s the tool reached for when the t-test’s normality assumption in the independent-samples t-test can’t be met.

What it doesn’t require: normality

The Mann-Whitney U test does not assume either group is normally distributed, and it does not require a normality test as a gatekeeping step before you’re “allowed” to run it. This is precisely the point of a rank-based nonparametric test: by converting raw values to ranks, it removes distributional shape from the calculation entirely. If you’re used to routing every analysis through a Shapiro-Wilk or Kolmogorov-Smirnov test first, Mann-Whitney is the one common case where that step is not what determines validity — it determines whether the t-test’s assumption holds, not whether Mann-Whitney’s does.

That said, pre-testing normality to decide which test to run at all is its own known problem. Rochon, Gondan & Kieser (2012, BMC Medical Research Methodology 12:81) simulated exactly this two-stage procedure — a Shapiro-Wilk pre-test choosing between the t-test and Mann-Whitney U — and found that while it “seemed to satisfactorily maintain the nominal significance level” overall, the conditional error rate for skewed (exponential) data was “considerably larger than alpha = .05” specifically among the pairs the pre-test had accepted as normal, while running the unconditional t-test with no pre-test at all was alpha-robust. Their finding at small samples reinforces the same caution: Shapiro-Wilk “lacks power to detect deviations from normality” exactly when you’d most want it to catch a problem. The practical takeaway is to choose your test from the study design and measurement level, not from a pre-test’s verdict on a small sample.

The assumption most guides skip: equal shape, for a median comparison

This is the part of the “assumption-free” myth that causes real reporting errors. The test’s null hypothesis, again, is that the two distributions are identical — not that the two medians are equal. SciPy’s own mannwhitneyu documentation is explicit about this: the two-sided alternative is stated as “the distributions are not equal,” and only “under a more restrictive set of assumptions” can “the alternative hypotheses be expressed in terms of the locations of the distributions.”

That restrictive set of assumptions is similar shape and spread between the two groups. When both distributions have roughly the same shape (both skewed the same direction, similar variance, no wildly different outliers), a significant Mann-Whitney result can be reported as a difference in medians — the ranking evidence and the median comparison point the same direction. When the shapes clearly differ, a significant result only licenses the more general claim of stochastic dominance: one group’s values tend to be systematically higher (or lower) than the other’s, without a specific claim about medians. Reporting “the medians differed significantly” when the shapes are visibly different overstates what the test showed — check the two distributions (a pair of histograms or box plots) before choosing which sentence to write, the same caveat that applies to a significant Kruskal-Wallis result across more than two groups.

Worked example

The dataset below is illustrative, not from a real study, chosen specifically to show why the shape check matters. All statistics were computed directly from these two lists of values (average-rank handling for ties, continuity-corrected normal approximation) rather than looked up or estimated.

Clinic A wait times, minutes (n = 6): 12, 15, 14, 22, 13, 41
Clinic B wait times, minutes (n = 7): 18, 21, 19, 25, 23, 20, 29

Value	Group	Rank
12	A	1
13	A	2
14	A	3
15	A	4
18	B	5
19	B	6
20	B	7
21	B	8
22	A	9
23	B	10
25	B	11
29	B	12
41	A	13

Group A’s rank sum is R_A = 1+2+3+4+9+13 = 32, giving U_A = R_A − n₁(n₁+1)/2 = 32 − 21 = 11. Group B’s U_B = 31, and U_A + U_B = 42 = n₁×n₂, confirming the arithmetic. With no ties, the exact variance is n₁n₂(N+1)/12 = 6×7×14/12 = 49, so σ_U = 7. The continuity-corrected z is (11 − 21 + 0.5) / 7 = −1.357, giving a two-tailed p = .175 — not significant at α = .05. The rank-biserial effect size is r = 1 − 2U_A/(n₁n₂) = 1 − 22/42 = .48, a moderate effect despite the non-significant p-value at this small sample size.

Now the shape check: Clinic A’s median is 14.5, Clinic B’s is 21 — consistent with U_A being the smaller rank sum. But Clinic A’s values are tightly clustered from 12–22 except for one 41-minute outlier stretching its distribution rightward, while Clinic B’s seven values sit in a tighter, more symmetric 18–29 band. The shapes are not comparable. Even if this result had reached significance, the honest sentence would describe Clinic A’s wait times as stochastically lower than Clinic B’s — not “Clinic A’s median wait time was significantly lower,” which implies a shape symmetry this data doesn’t have.

Checking the assumptions before you run the test

Independence is a design question, not something a statistical check can rescue after the fact — confirm it from how the data were collected (two genuinely separate groups, not paired or repeated measurements) before analysis, not from the output.
Level of measurement is confirmed the same way: can every value in the combined dataset be placed in a meaningful rank order? If not, Mann-Whitney doesn’t apply regardless of sample size.
Shape similarity is the one assumption worth checking visually after the fact, because it changes how you’re allowed to word the conclusion, not whether the test runs. Overlay a histogram or box plot of both groups before deciding between “difference in medians” and “stochastic dominance” language.

How to report it

A complete APA-style report includes the test statistic, the approximation used, the p-value, both groups’ medians (and IQR or range), and an effect size — for example: “Wait times differed significantly between Clinic A (Mdn = 14.5) and Clinic B (Mdn = 21.0), U = 11, z = −1.36, p = .175, r = .48.” Report the exact p-value rather than the normal-approximation p-value when either group is small (under roughly n = 20) or ties are heavy — both SPSS and Stata can produce the exact test on request, and the software-specific guides linked here cover exactly where to find that option in each package’s output.

Frequently asked questions

Is the Mann-Whitney U test really assumption-free?

No. It drops the normality assumption the independent-samples t-test needs, but it still requires independent observations and data that can be meaningfully ranked (ordinal or continuous). “Fewer assumptions than the t-test” is accurate; “no assumptions” is not.

Does the Mann-Whitney U test require a normal distribution?

No, and this is the entire reason to use it — it’s reached for specifically when a group’s distribution is too skewed, too small, or too far from normal for a t-test to be trustworthy.

Does Mann-Whitney U require equal variances between groups?

Not for the test’s validity. Unequal variance is one common way two distributions end up with different shapes, though, which affects whether you can describe a significant result as a median difference (see above) even though it doesn’t affect whether the test itself is valid.

What happens if I run Mann-Whitney U on paired or repeated-measures data?

The independence assumption is violated and the p-value can’t be trusted, because the test’s variance calculation assumes no correlation between the two groups’ values. Use the Wilcoxon signed-rank test for paired data instead.

Can I use Mann-Whitney U on ordinal Likert-scale data?

Yes — ordinal data is exactly the measurement level the test is built for, since it only uses rank order and never assumes equal spacing between scale points the way a t-test’s mean does.

What’s the difference between Mann-Whitney U and the Kruskal-Wallis test?

Mann-Whitney U compares exactly two independent groups; Kruskal-Wallis extends the same rank-based logic to three or more groups and reduces algebraically to Mann-Whitney U when there are only two.

Mann-Whitney U Test Assumptions: What It Actually Requires