Mann-Whitney U Test Calculator

Frank Wilcoxon (1945) introduced two rank-based procedures in a single short paper: one for paired observations (now the Wilcoxon signed-rank test) and one for two independent samples of equal size (the Wilcoxon rank-sum test). Two years later, Mann and Whitney (1947) generalised the rank-sum to unequal sample sizes and reformulated the statistic as a count of pairwise wins, which they called U. The two statistics are algebraically equivalent (U = W − n(n + 1) / 2), but historical accident left us with two names for the same test. R reports W; SPSS, SciPy, and the textbooks of the biostatistical tradition report U.

The Mann-Whitney U is a non-parametric procedure. It does not assume normality of the underlying distributions, only that observations are independent and that the measurement scale is at least ordinal. Under the more restrictive location-shift model (the two distributions have the same shape, differing only by an additive constant), the test addresses the location of the second distribution relative to the first. Without that assumption, the null hypothesis is the more general statement P(X > Y) = 0.5. The choice matters for interpretation: a significant Mann-Whitney with grossly heteroscedastic groups may reflect a spread difference rather than a location shift, in which case the Brunner-Munzel (2000) test, which tests P(X > Y) = 0.5 directly under heteroscedasticity, is preferable.

For tied observations the average-rank correction is essential. Without it, the asymptotic z-statistic is biased toward zero, and the p-value becomes anti-conservative for heavily tied ordinal data (Likert scales, ordinal grades). The calculator applies the ties correction Var(U) = n₁ n₂ (n + 1) / 12 × (1 − Σ(tⱼ³ − tⱼ) / (n³ − n)) automatically and reports both the raw and continuity-corrected z. For small samples (combined N ≤ 30, with n₁ × n₂ ≤ 200), the calculator builds the exact null distribution of U by convolution and returns the exact p-value, the gold standard for sparse data, matching R's exactRankTests::wilcox.exact() and Python's scipy.stats.mannwhitneyu(method = "exact").

Effect sizes for rank-based tests sit on multiple scales. The rank-biserial correlation r = 1 − 2U / (n₁ n₂) is the standardised pair-wins quantity recommended by APA. Cliff's delta (1993) is the same value expressed as P(X > Y) − P(X < Y) and is conventional in clinical epidemiology and ordinal data analysis. The common-language effect size (CLES) of McGraw and Wong (1992) is P(X > Y) + 0.5 P(X = Y), the probability that a randomly chosen value from group A exceeds a randomly chosen value from group B. All three are derived from the same U and are reported together so you can pick the convention your reviewer expects.

The Hodges-Lehmann (1963) median is the recommended companion to the U-test's p-value. For two independent samples it is the median of all pairwise differences xᵢ − yⱼ; for paired data it is the median of the Walsh averages (xᵢ + xⱼ) / 2. Under the location-shift model it estimates the median shift, and its 95% confidence interval, obtained by inverting the Mann-Whitney U (or Wilcoxon signed-rank) test, has nominal coverage even under non-normality. Most online U-test calculators omit it; the calculator computes the point estimate and the CI via the z-quantile index lookup on the sorted outer-product differences (Walsh averages for the paired case).

The Mann-Whitney U feeds into the broader two-group testing surface. For approximately normal data with similar variances, an independent-samples t-test (available in the two-sample t-test calculator ) is slightly more powerful. For ≥ 3 groups the non-parametric extension is the Kruskal-Wallis test; the parametric counterpart is the ANOVA calculator. For 2 × 2 categorical comparisons see the Fisher's exact test calculator. And for a publication-ready APA write-up of any non-parametric analysis with the full Hodges-Lehmann CI by a PhD statistician, the statistical analysis service covers everything end-to-end.