Written and maintained by CASRAI Editorial Board
Last updated
SPSS runs the Mann-Whitney U test from two different places in the menu, and they don’t show you the same thing by default. The older route prints one compact table with every statistic on it; the newer route opens with a one-line summary that hides the actual test statistics until you double-click into it. Neither is wrong, but if you don’t know which one you’re looking at, it’s easy to report the wrong number or miss the mean-rank comparison the test is actually built on. This guide covers both menu paths, how to read the mean-rank and test-statistic output each one produces, and when Mann-Whitney U is the right test to run instead of an independent-samples t-test in the first place.
Two menu paths to the same test
Both routes run the identical Mann-Whitney U calculation; they differ only in interface and default output.
- Analyze > Nonparametric Tests > Legacy Dialogs > 2 Independent Samples — the older procedure, and still the faster one for a single, specific Mann-Whitney U test. It opens a plain dialog, runs immediately, and prints the full Ranks and Test Statistics tables in the output viewer with no extra clicks.
- Analyze > Nonparametric Tests > Independent Samples — the newer “one-click” procedure, built around an automated field-type check and a Hypothesis Test Summary view. For two groups and a continuous or ordinal test field, it defaults to running Mann-Whitney U automatically, but the summary view it opens with shows only the decision (reject or retain the null), not the U statistic itself — you have to double-click the summary row to see the underlying numbers.
Both are valid, current SPSS procedures — the Legacy Dialogs path was not removed when the newer one was added, and each covers a case the other doesn’t handle as cleanly: Legacy Dialogs is quicker when you already know exactly which test you want and need the numbers on screen immediately; the newer dialog is more useful when you’re screening several fields for the right nonparametric test and want SPSS to pick one based on the data’s characteristics.
Running it via Legacy Dialogs
Go to Analyze > Nonparametric Tests > Legacy Dialogs > 2 Independent Samples. Move your outcome variable into Test Variable List and your two-group categorical variable into Grouping Variable, then click Define Groups and enter the two group codes exactly as they’re coded in your data — SPSS will not run the test until both codes are supplied, the same requirement as the independent-samples t-test dialog. Under Test Type, Mann-Whitney U is checked by default alongside three other rank-based tests (Kolmogorov-Smirnov Z, Moses extreme reaction, Wald-Wolfowitz runs); leave the other three unchecked unless you specifically need them, since each adds its own table to the output. If your sample is small or has many tied ranks, click Exact and select Exact (or Monte Carlo for a very large tied dataset) rather than relying on the default asymptotic approximation — see the reporting section below for why this matters.
Running it via the Independent Samples dialog
Go to Analyze > Nonparametric Tests > Independent Samples. On the Fields tab, move your outcome into Test Fields and your grouping variable into Groups. On the Settings tab, either leave Automatically compare distributions across groups selected (SPSS will choose Mann-Whitney U for two groups by default) or choose Customize tests and tick Mann-Whitney U (2 samples) directly if you want to guarantee it runs regardless of what SPSS’s automatic choice would be. The Settings tab also has its own Test Options exact-test control, in the same place as the Legacy Exact button. Running it produces the Hypothesis Test Summary view first; double-click the row (or right-click and choose View Test Statistics) to open the detail view.
Reading the mean-rank output
Both routes produce a Ranks table (Legacy Dialogs shows it directly; the newer dialog’s detail view shows the equivalent under a different layout) with three columns per group: N, Mean Rank, and Sum of Ranks. This is the table that actually explains the result — the test statistics below it are a formalization of what the mean ranks already show.
- SPSS pools both groups into one combined ranking of every observation, low to high, before computing anything — ranks are never computed within a group separately.
- The group with the higher mean rank is the group whose values tend to sit higher in the combined ordering. If the two mean ranks are close, no amount of statistical significance changes that the two groups are barely distinguishable in practice; if they’re far apart, that’s the effect the test statistic is quantifying.
- Tied values (the same score appearing in both groups) are each given the average of the ranks they’d otherwise occupy. A dataset with many ties — common with short ordinal scales like a 5-point Likert item — needs the tie-corrected variance SPSS applies automatically to the asymptotic test, and is also the clearest case for requesting an exact test instead of trusting the asymptotic p-value.
Reading the test statistics
The Legacy Dialogs Test Statistics table prints, in one place: Mann-Whitney U, Wilcoxon W, Z, and Asymp. Sig. (2-tailed) — plus an Exact Sig. [2*(1-tailed Sig.)] row if you requested the exact test. The newer dialog’s detail view reports the same underlying test but under its own generic nonparametric-procedure column headers: Test Statistic (the U value), Standard Error, Standardized Test Statistic (the Z value), and Asymptotic Sig. (2-sided test) — the numbers match Legacy Dialogs’ output for the same data, only the table’s column names differ.
- Wilcoxon W is the sum of ranks for whichever group has fewer cases (or the first-listed group if the two sample sizes are equal) — it’s printed because the Mann-Whitney U test and the Wilcoxon rank-sum test are the same calculation under two different historical names; SPSS reports both statistics from one procedure rather than making you choose. It is not the same test as the Wilcoxon signed-rank test, which is the paired-samples analogue, not the two-independent-groups one.
- Asymp. Sig. is a normal approximation to the true sampling distribution of U, and it degrades with small samples or heavy tied ranks — the same limitation the asymptotic approximation has for the Wilcoxon signed-rank test. As a rough guide, if either group has fewer than about 20 cases, or the data have substantial ties, request the exact test and report Exact Sig. instead of the asymptotic value.
- Running Mann-Whitney U on exactly two groups is mathematically the same calculation as running a Kruskal-Wallis test with k = 2 — the two procedures reduce to each other algebraically, which is why Kruskal-Wallis is described as the extension of Mann-Whitney to three or more groups, not the other way around.
Mann-Whitney U vs. an independent-samples t-test: which one applies
The independent-samples t-test compares group means directly and assumes the outcome is at least approximately normally distributed within each group. Mann-Whitney U compares group mean ranks and makes no such assumption. The choice is about what your data can support, not a preference:
| Situation | What’s actually the problem | Use instead |
|---|---|---|
| Outcome is ordinal (Likert-type item, pain/severity scale, rank/grade) rather than interval or ratio | A t-test’s mean assumes equal distances between scale points, which an ordinal scale doesn’t guarantee | Mann-Whitney U test |
| Outcome is continuous but clearly non-normal in a small sample, and a transform doesn’t fix it | The t-test’s small-sample validity depends on approximate normality; check it directly rather than assume it | Mann-Whitney U test |
| One or two extreme outliers are present, sample is small | A t-test compares means, which outliers pull disproportionately; a rank-based test is far less sensitive to their exact magnitude | Mann-Whitney U test |
| Group variances differ sharply and the outcome is otherwise reasonably normal | Homogeneity of variance is the actual issue, not normality — switching to a rank-based test changes what’s being tested, not just how | Welch’s t-test (the “Equal variances not assumed” row of the independent-samples t-test output), not Mann-Whitney U |
| Sample is reasonably large (each group roughly 30+) and only mildly non-normal | The Central Limit Theorem means the sampling distribution of the mean is close enough to normal for the t-test to remain valid | Independent-samples t-test — Mann-Whitney U isn’t required, though it remains a valid alternative |
None of this is really a Mann-Whitney-specific decision — it’s the same normality/measurement-level logic that applies across all the parametric-vs-nonparametric choices covered in the Kruskal-Wallis guide for three or more groups.
Assumptions of the Mann-Whitney U test
- Independent observations, both within and across the two groups — each case contributes to exactly one group and one rank. Repeated measures or matched pairs need the Wilcoxon signed-rank test instead.
- The outcome is at least ordinal — measured on a scale where “higher” and “lower” are meaningful, even if the distances between values aren’t. See levels of measurement for the full nominal/ordinal/interval/ratio distinction.
- Exactly two independent groups. For three or more, run a Kruskal-Wallis test instead.
- Does not require normality — that’s the entire reason to use it in place of a t-test. It also does not require equal variances in the strict sense a t-test does.
- Similar distribution shape matters for how you word the result. When the two groups’ distributions have similar shape and spread, a significant result can be described as a difference in medians. When the shapes clearly differ, the technically correct interpretation is that one group tends to produce systematically higher (or lower) values than the other — a difference in stochastic dominance, not specifically a median difference — the same caveat that applies to interpreting a significant Kruskal-Wallis result.
Worked example: reading the output and reporting it
Illustrative example (not real data): a researcher compares a 10-point ordinal satisfaction rating between two independent groups, 6 cases in Group A and 7 in Group B (N = 13). The outcome is ordinal and the sample is small, so Mann-Whitney U is used in place of an independent-samples t-test.
| Group | n | Median | Sum of ranks | Mean rank |
|---|---|---|---|---|
| Group A | 6 | 4.5 | 25.5 | 4.25 |
| Group B | 7 | 8.0 | 65.5 | 9.36 |
Group B’s mean rank (9.36) sits well above Group A’s (4.25) in the pooled ranking of all 13 scores. Working through the standard U formula on these rank sums gives U = 4.5, with a normal-approximation z = -2.29 and Asymp. Sig. (2-tailed) = .022. Because the sample is small and several scores tie between the two groups, the exact test is the more defensible number to report here rather than the asymptotic approximation, but the asymptotic figures are used below for illustration.
The effect size most commonly reported alongside Mann-Whitney U is r = z / √N: here, 2.29 / √13 ≈ .63, a large effect by the same small/medium/large (.10/.30/.50) thresholds Cohen proposed for r in other contexts. The write-up:
“A Mann-Whitney U test indicated that satisfaction ratings were significantly higher in Group B (Mdn = 8, n = 7) than in Group A (Mdn = 4.5, n = 6), U = 4.5, z = -2.29, p = .022, r = .63.”
Note what the sentence reports and doesn’t: the U statistic, the z and p from the asymptotic test (or Exact Sig. in place of p if the exact test was requested, which for a dataset this small and this tied it should be), the effect size, and both groups’ medians and sample sizes — not group means, which the test never used.
Frequently asked questions
Is the Mann-Whitney U test the same as the Wilcoxon rank-sum test?
Yes — they are the same test under two different historical names, which is why SPSS’s Legacy Dialogs output prints both a Mann-Whitney U value and a Wilcoxon W value from the same procedure. This is a different test from the Wilcoxon signed-rank test, which applies to paired, not independent, samples.
Can I use Mann-Whitney U with more than two groups?
Not directly — for three or more independent groups, run a Kruskal-Wallis test. Kruskal-Wallis with exactly two groups reduces algebraically to Mann-Whitney U, so the two are consistent, not competing, tests.
Does a significant Mann-Whitney U result mean the medians are different?
Only when the two groups’ distributions have similar shape and spread. If the shapes clearly differ, the honest interpretation is that one group’s values tend to be systematically higher or lower than the other’s, not specifically that the medians differ.
Why does SPSS give both an asymptotic and an exact significance value?
The asymptotic value is a normal approximation that becomes less reliable with small samples or many tied ranks. The exact test calculates the true p-value from the actual permutation distribution instead of approximating it, and is the safer choice whenever either group is small (roughly under 20) or the data have substantial ties.
What effect size should I report with Mann-Whitney U?
The most common choice is r = z / √N, interpreted on the same rough .10/.30/.50 small/medium/large scale used for correlation coefficients. Rank-biserial correlation is a less common but valid alternative some journals prefer. See the effect size entry and the effect size guide for the broader context of why an effect size belongs alongside every significance test, not just parametric ones.
Do I need to check normality before running Mann-Whitney U?
No — not requiring a normality check is the point of choosing it. If you’re deciding between Mann-Whitney U and an independent-samples t-test, that decision is exactly where checking normality matters; once you’ve decided Mann-Whitney U is the right test, no further normality test is needed to run it.
