MerQuR+Statistics Glossary

Statistics Glossary → Brunner-Munzel test

Brunner-Munzel testconcept

A rank test for two independent groups of whether the probability that a randomly chosen value from one group is smaller than one from the other differs from 0.5, without requiring equal variances or equal distribution shapes.

The Mann-Whitney U test also works with ranks, but it tests the null hypothesis that the two distributions are identical; if the groups differ in spread or shape, it can give a significant result even when the mean ranks are equal, and its Type I error rate is distorted. The Brunner-Munzel test targets the relative effect directly: p̂ = P(X < Y) + ½·P(X = Y). The null hypothesis is p = 0.5, meaning that neither group tends to produce systematically larger values. The statistic estimates the variance separately for each group and is referred to a t distribution with Welch-type degrees of freedom, making it something like a rank-based counterpart of Welch's t-test.

p̂ is also an interpretable effect size (the probability of superiority): 0.70 means that a randomly chosen value from the second group exceeds one from the first with 70% probability. With very small samples (fewer than about 10 per group) the t approximation can weaken and a permutation version is recommended. It should be remembered that the test concerns this probability, not the difference in medians.

An everyday example

You are comparing the delivery times of two courier firms. One always delivers in 2-3 days; the other sometimes takes 1 day and sometimes 6. Because the spreads differ so much, Mann-Whitney can mislead. Brunner-Munzel asks whether a random parcel arrives later with the second firm than with the first, and a value such as p̂ = 0.55 may show that the two firms are in fact quite close.

In MerQur

In the Mann-Whitney U test, the 'Extra outputs: Brunner-Munzel' checkbox adds the test and the estimate p̂ = P(X < Y) + ½P(X = Y).

Analyses where this concept appears

Mann-Whitney U

Related concepts

Rank-biserial correlationWelch correction