The rank-based answer to the two-sample t-test
When data are ordinal, skewed, or too few to trust means, Mann-Whitney replaces values with their ranks. Pool both groups, rank everything from smallest up, and ask whether one group's ranks pile up at the top. U counts, for each observation in one group, how many observations in the other group it beats, the most literal "which group tends to be bigger" statistic there is.
Two groups of five (78, 82, 85, 90, 88 against 72, 75, 80, 77, 74): group 1's rank sum is 39 of a possible 40-ish spread, U₁ = 24, U₂ = 1. The normal approximation gives z ≈ −2.30, two-tailed p ≈ 0.0216. Group 1's values dominate group 2's almost completely, and the test noticed without ever trusting a mean.
What it tests, precisely
Strictly, Mann-Whitney tests whether a random value from one group tends to exceed a random value from the other. It becomes a test of medians only when the two distributions share a shape. Report it as the tendency test it is, and small samples are welcome, with the caveat that exact tables beat the normal approximation under roughly n = 8 per group.
Reaching for Mann-Whitney as a "safe" default on clean, plentiful data. When the t-test's assumptions genuinely hold, ranks throw away information and power with it.
Frequently asked questions
Which U do I report?
Either. They sum to n1·n2, and papers report U for the group of interest or the smaller U. This page shows both, with the smaller flagged as the test statistic.
How are ties handled?
Tied values share the average of their ranks, and the standard error carries a tie correction. Both happen automatically here.