Mann-Whitney U Test vs t-Test: When to Use Which (Worked Examples)

Use the Mann-Whitney U test (also called the Wilcoxon rank-sum test) when comparing two independent groups and you cannot safely assume normality — typically because the sample size is small, the data is ordinal (Likert ratings, ranks), or there are clear outliers that distort the mean. The t-test assumes the sampling distribution of the difference in means is approximately normal, which holds for any continuous data when sample sizes are large (n ≥ 30 per group) by the central limit theorem. When samples are small AND the population distribution is heavily skewed or contains outliers, the t-test’s p-value can be misleading; Mann-Whitney works on ranks and is robust to those issues. The price: when the t-test’s assumptions actually hold, Mann-Whitney is roughly 95% as efficient (about 5% less power), so reach for it only when you have a real reason to doubt normality.

Combine both groups into one list and rank from smallest to largest. Assign tied values the average of the ranks they span. Sum the ranks for group 1, call it R1. Compute U1 = R1 - n1(n1+1)/2, where n1 is the size of group 1. Compute U2 similarly for group 2; U1 + U2 = n1 × n2 always. Use the smaller of U1 and U2 as the test statistic U, look it up in a Mann-Whitney table for small samples, or use the normal approximation for n1, n2 ≥ 10: z = (U - n1×n2/2) / sqrt(n1×n2×(n1+n2+1)/12). The intuition: if the groups are exchangeable, the ranks shuffle randomly between them and U falls near the middle. If group 1 is systematically larger, ranks pile up in group 1 and U deviates.

Eight subjects per condition. Reaction times (ms) for condition A: 245, 252, 261, 273, 287, 295, 312, 480. Condition B: 268, 271, 274, 282, 291, 298, 305, 318. Notice the outlier in A (480) that would distort the mean. Combine and rank from smallest: A245=1, A252=2, A261=3, B268=4, B271=5, A273=6, B274=7, B282=8, A287=9, B291=10, A295=11, B298=12, B305=13, A312=14, B318=15, A480=16. R(A) = 1+2+3+6+9+11+14+16 = 62. U(A) = 62 - 8×9/2 = 62 - 36 = 26. U(B) = 8×8 - 26 = 38. Smaller U = 26. Normal approximation z = (26 - 32) / sqrt(64×17/12) = -6 / sqrt(90.67) = -6 / 9.52 = -0.63. Two-tailed p ≈ 0.53. Fail to reject — no evidence the conditions differ. A naive t-test on these same data, distorted by the 480 ms outlier, would have given a much smaller p-value driven by the outlier rather than a real difference. This is exactly the situation Mann-Whitney was designed for.

You compare satisfaction ratings (1-7 Likert) for two product designs. Group A (n=10): 5, 6, 4, 7, 5, 6, 5, 6, 7, 6. Group B (n=10): 3, 4, 5, 4, 3, 5, 4, 5, 4, 5. Likert data is ordinal — the distance between "5" and "6" is not guaranteed to equal the distance between "3" and "4" — so a t-test on the raw scores assumes more than the measurement actually justifies. Combine and rank with ties getting average ranks. After ranking and summing R(A) = 142.5, R(B) = 67.5. U(A) = 142.5 - 55 = 87.5. U(B) = 100 - 87.5 = 12.5. Smaller U = 12.5. z = (12.5 - 50) / sqrt(100 × 21 / 12) = -37.5 / sqrt(175) = -37.5 / 13.23 = -2.83. Two-tailed p ≈ 0.0046. Reject — group A rates the product significantly higher. The interpretation is in terms of stochastic dominance: a random observation from group A is more likely to be higher than a random observation from group B, not "the means differ by X units."

When the t-test’s assumptions actually hold (normal populations, similar variances), Mann-Whitney has asymptotic relative efficiency of 0.955 — meaning you need about 5% more observations to achieve the same power as the t-test. That cost is small, and many practitioners use Mann-Whitney as the default for two-group comparisons in fields where data is reliably non-normal (clinical trials with skewed lab values, ecology counts, satisfaction surveys). When distributions are heavy-tailed or contain outliers, Mann-Whitney can be substantially MORE powerful than the t-test because the t-test’s power is throttled by the outliers inflating the standard deviation. Here is the surprising part most courses leave out: against truly contaminated or fat-tailed distributions, Mann-Whitney can be 1.5x to 3x more powerful than the t-test. It is not always the conservative choice.

Two independent groups, continuous or ordinal outcome. (1) Is the outcome ordinal (Likert, ranks, ordered categories)? → Mann-Whitney. (2) Is n ≥ 30 per group and the data continuous? → t-test is robust to non-normality by CLT. (3) Is n < 30 per group? Check normality with a histogram or Shapiro-Wilk test. If normality is plausible, use t-test. If not, Mann-Whitney. (4) Are there outliers? Decide whether they are data-entry errors (drop and use t-test) or genuine observations (use Mann-Whitney to keep them without distortion). (5) Are variances roughly equal? If not, Welch’s t-test (or Mann-Whitney) handles unequal variances better than the pooled t-test.

Snap a photo of the two columns of data and StatsIQ runs both tests automatically, presenting the t-test result, Welch’s t-test result, and Mann-Whitney U result side by side along with normality diagnostics (Shapiro-Wilk, Q-Q plot, histogram) so you can defend your choice. The app calls out outliers, suggests transformations, and explains which test is most defensible given your sample size and data shape.