Mann-Whitney U Test: A Complete Guide

What Is the Mann-Whitney U Test?

Henry Mann and Donald Ransom Whitney published the test in 1947 as a generalization of Frank Wilcoxon's 1945 rank-sum procedure. The two tests are mathematically equivalent — running a Wilcoxon rank-sum test and a Mann-Whitney U Test on the same data produces identical p-values. Many software packages use the names interchangeably, which explains why you may see "Wilcoxon rank-sum test" in R output when you call wilcox.test().

The test does not require the data to be normally distributed. Instead of computing means, it computes a statistic U that counts how many times an observation from group 1 exceeds an observation from group 2. When both groups are identically distributed (H₀ true), we expect U to be approximately n₁n₂ / 2. A very large or very small U signals that one group consistently outranks the other.

Assumptions of the Mann-Whitney U Test

The Mann-Whitney U Test has four assumptions. Meeting them ensures the p-value is valid and the result is interpretable.

The Mann-Whitney U Test Formula

Two formulas calculate U — one for each group. You then take the smaller of the two as the test statistic.

The rank sum W is computed by pooling both samples, ranking all N = n₁ + n₂ values from 1 to N (with ties receiving averaged ranks), then summing only the ranks belonging to each group. The U statistic directly counts the number of times a group 1 observation exceeds a group 2 observation — a cleaner interpretation than the formula might suggest.

Large-Sample Normal Approximation

When both sample sizes exceed about 20, the U statistic follows an approximately normal distribution. Convert U to a z-score using these formulas, then look up the p-value in the standard normal table.

7-Step Calculation Procedure

The following steps apply to any Mann-Whitney U Test, whether done by hand or with software. The RANK Framework below captures the key phases as a memory aid.

Worked Examples

Both examples below show every calculation step explicitly. Example 1 uses a healthcare scenario with small samples where an exact U is compared to the critical value. Example 2 uses a larger educational dataset requiring the normal approximation.

Example 1 — Pain Scores: Treatment vs. Control

Example 2 — Student Exam Scores: Two Teaching Methods

Effect Size for the Mann-Whitney U Test

A statistically significant Mann-Whitney U Test tells you the groups differ — but not by how much. Always report an effect size. Two measures are appropriate: rank-biserial correlation (r) and Cliff's delta (δ).

Rank-Biserial Correlation (r)

The rank-biserial correlation r ranges from −1 to +1. An r of +0.60 means that in 80% of all possible pairings, the group 1 observation exceeds the group 2 observation. An r of 0 means the groups perform identically by rank. The benchmarks (small = 0.1, medium = 0.3, large = 0.5) come from Cohen's (1988) guidelines adapted for nonparametric tests.

Cliff's Delta (δ)

Cliff's delta is a more direct measure of how often values in group 1 exceed values in group 2. It equals 2r (when using the rank-biserial formula), ranging from −1 (group 2 always higher) to +1 (group 1 always higher). Values near 0 mean nearly equal group performance. Many researchers prefer δ because it has a clearer probabilistic interpretation: δ = 0.40 means that a randomly chosen observation from group 1 exceeds a randomly chosen observation from group 2 by a margin of 40 percentage points more often than not.

How to Report the Mann-Whitney U Test in APA Format

APA 7th edition (2020) requires you to report the U statistic, sample sizes, p-value, and an effect size. The median for each group should appear in the text or a table. Always report exact p-values (e.g., p = .023) rather than threshold statements (e.g., p < .05) unless the p-value is below .001.

APA Reporting Examples

Running the Mann-Whitney U Test in Software

Mann-Whitney U Test in R

R's built-in wilcox.test() function performs the Wilcoxon rank-sum test, which is identical to the Mann-Whitney U Test. R reports the W statistic (= U₁ in the formula above).

# Define the two groups
group_a <- c(2, 3, 4, 4, 5, 6, 7, 8)
group_b <- c(4, 5, 6, 7, 7, 8, 9, 10)

# Two-tailed Mann-Whitney U Test with exact p-value
result <- wilcox.test(group_a, group_b,
                     alternative = "two.sided",
                     exact = TRUE,
                     conf.int = TRUE)
print(result)

# Compute rank-biserial correlation (effect size)
# install.packages("rstatix") if needed
library(rstatix)
wilcox_effsize(data, score ~ group, ci = TRUE)

# Output includes: W (= U₁), p-value, confidence interval

Mann-Whitney U Test in Python (SciPy)

Python's SciPy library provides scipy.stats.mannwhitneyu(), which directly computes the U statistic and p-value with both exact and asymptotic methods.

from scipy import stats
import numpy as np

group_a = [2, 3, 4, 4, 5, 6, 7, 8]
group_b = [4, 5, 6, 7, 7, 8, 9, 10]

# Mann-Whitney U Test (two-tailed, exact p-value)
U_stat, p_value = stats.mannwhitneyu(
    group_a, group_b,
    alternative='two-sided',
    method='exact'      # use 'asymptotic' for large n
)

print(f"U = {U_stat}, p = {p_value:.4f}")

# Effect size: rank-biserial correlation
n1, n2 = len(group_a), len(group_b)
r_effect = 1 - (2 * U_stat) / (n1 * n2)
print(f"r = {r_effect:.3f}")

Mann-Whitney U Test in SPSS

/* SPSS: Nonparametric Tests → 2 Independent Samples */
NPAR TESTS
  /M-W= score BY group (1 2)
  /STATISTICS DESCRIPTIVES
  /MISSING ANALYSIS.

/* Or via menus: */
/* Analyze → Nonparametric Tests → Legacy Dialogs → */
/* 2 Independent Samples → Mann-Whitney U        */

Mann-Whitney U vs. Other Tests

Mann-Whitney U Test vs. Independent Samples t-Test

When data are normally distributed, the t-test is slightly more powerful. The efficiency loss from using Mann-Whitney on normal data is only about 5% — a modest price for protection against non-normality violations. When data are skewed or contain outliers, Mann-Whitney is the appropriate choice. Check normality with a Shapiro-Wilk test for small samples (n < 50) or assess using a Q-Q plot.

Mann-Whitney U vs. Wilcoxon Signed-Rank Test

These two tests are frequently confused despite serving different purposes. The Mann-Whitney U Test applies to two independent groups — completely separate participants in each group. The Wilcoxon signed-rank test applies to two related groups — matched pairs or the same participants measured twice (before and after). Using the wrong test is a methodological error. If participants are matched or measurements are repeated, the Wilcoxon signed-rank test is correct; if they are separate and unrelated, use Mann-Whitney.

Mann-Whitney U vs. Kruskal-Wallis Test

The Kruskal-Wallis test is the nonparametric extension of one-way ANOVA — it compares three or more independent groups. The Mann-Whitney U Test handles exactly two groups. If a Kruskal-Wallis test is significant, Mann-Whitney U Tests are used as post-hoc pairwise comparisons (with Bonferroni correction applied to control the family-wise error rate).

When to Use the Mann-Whitney U Test: Decision Guide

Real-World Applications of the Mann-Whitney U Test

The Mann-Whitney U Test appears across scientific disciplines precisely because ordinal data and non-normal distributions are the rule rather than the exception in real research.

Common Mistakes and Misconceptions

Entity and Formula Glossary

Mann-Whitney U Test Calculator

Enter your two data groups below — separate values with commas. The calculator assigns ranks, computes U, runs both exact (small n) and asymptotic (large n) p-values, and generates the APA-formatted result.

Frequently Asked Questions