Brunner-Munzel Test
Also known as: Brunner-Munzel Testi, generalized Wilcoxon test, nonparametric Behrens-Fisher test, probabilistic index test
The Brunner-Munzel test is a nonparametric two-sample hypothesis test that estimates the probabilistic superiority index P(X < Y) — the probability that a randomly selected observation from one group exceeds a randomly selected observation from the other. Introduced by Brunner and Munzel in 2000 as a solution to the nonparametric Behrens-Fisher problem, it remains valid even when the two groups have unequal variances or differently shaped distributions, making it a robust alternative to the Mann-Whitney U test in heteroscedastic settings.
Key highlights
- Does not assume equal variances, making it valid in heteroscedastic two-sample settings where Mann-Whitney U can produce inflated error rates.
- Produces a directly interpretable effect size — the probabilistic superiority index — rather than requiring a separate calculation.
- Works with both continuous and ordinal data without requiring any distributional shape assumption.
- Computationally straightforward and available in major statistical computing environments.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use the Brunner-Munzel test to compare two independent groups on a continuous or ordinal outcome when the assumption of equal variances cannot be justified — the situation Brunner and Munzel called the nonparametric Behrens-Fisher problem. Core assumptions are: (1) the two samples are independent, (2) the measurement scale is at least ordinal, and (3) no normality is required. It is a direct upgrade over the Mann-Whitney U test in settings where the distribution shapes or spreads differ between groups. For very small samples (n < 10 per group) a permutation variant provides better control of Type I error. The test is not suitable for paired or repeated-measures data; use the Wilcoxon signed-rank test instead.
Strengths & limitations
- Does not assume equal variances, making it valid in heteroscedastic two-sample settings where Mann-Whitney U can produce inflated error rates.
- Produces a directly interpretable effect size — the probabilistic superiority index — rather than requiring a separate calculation.
- Works with both continuous and ordinal data without requiring any distributional shape assumption.
- Computationally straightforward and available in major statistical computing environments.
- Requires at least approximately 10 observations per group for the large-sample t-approximation to be reliable; for smaller samples the permutation version should be used.
- Less familiar than Mann-Whitney U in many applied fields, which can complicate peer review and reporting.
- Like all nonparametric rank-based tests, it is slightly less powerful than a parametric t-test when normality and equal variance truly hold.
- Not directly applicable to more than two groups; for multi-group comparisons the Kruskal-Wallis test or its heteroscedastic extensions are needed.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does the Brunner-Munzel test differ from the Mann-Whitney U test?
Both tests are based on ranks and compare two independent groups, but Mann-Whitney U assumes that the two groups have the same distributional shape (including equal variance) under the null hypothesis. If that shape assumption is violated, the Mann-Whitney U test can produce incorrect Type I error rates. The Brunner-Munzel test uses separate within-group variance estimates, so it remains valid when the groups have different spreads — analogous to the way Welch's t-test improves on Student's t-test for unequal variances.
What does the probabilistic superiority index mean?
The index P-hat estimates P(X < Y), the probability that a randomly drawn observation from Group 2 is larger than a randomly drawn observation from Group 1. A value of 0.5 means no stochastic dominance; 0.7 means there is a 70 % chance that a Group 2 value exceeds a Group 1 value. It is a scale-free effect size that is easy to communicate to non-statisticians.
When should I use the permutation version?
The asymptotic (t-approximation) version is reliable when each group has at least 10 observations. With smaller samples the approximation is inaccurate and the permutation-based version proposed by Neubert and Brunner (2007) should be used instead, as it controls the Type I error rate more accurately in small-sample settings.
Can the test be used for ordinal data such as Likert scales?
Yes. Because the test only requires that the measurement scale be ordinal — it never computes differences or sums of raw values — it is entirely appropriate for Likert-scale and other ordinal responses. This is one of its practical advantages over tests that require interval-level measurement.
Sources
- 1.Brunner, E. & Munzel, U. (2000). The Nonparametric Behrens-Fisher Problem: Asymptotic Theory and a Small-Sample Approximation. Biometrical Journal, 42(1), 17–25.
- 2.Neubert, K. & Brunner, E. (2007). A studentized permutation test for the nonparametric Behrens-Fisher problem. Computational Statistics & Data Analysis, 51(10), 5192–5204.
- 3.Brunner, E., Bathke, A. C., & Konietschke, F. (2019). Rank and Pseudo-Rank Procedures for Independent Observations in Factorial Designs. Springer.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). Brunner-Munzel Test. ScholarGate. https://scholargate.app/statistics/brunner-munzel-test