Mann-Whitney U Test: Non-Parametric Analysis for Two Independent Samples

Overview

When comparing two independent groups in psychological research, the Mann-Whitney U test is the non-parametric counterpart to the independent samples t-test. Originally developed independently by Mann and Whitney (1947) and Wilcoxon (1949), it compares the medians of two groups by ranking all scores together and examining the degree of separation between groups — no assumption of normality required.

This unit covers the complete workflow: theoretical foundations, step-by-step hand calculation for small and large samples, SPSS execution, and a direct comparison with the t-test.

Key Terms at a Glance

• Two-sample test: Compares two independent groups on a single variable
• Rank-sum: Ranking all scores together before computing the test statistic
• U statistic: The count of times scores from one group exceed scores from the other group
• Ordinal data: Data that can be ranked but does not meet interval/ratio assumptions

---

2.2 What Are Two-Sample Tests?

Two-sample tests assess whether two independent groups differ on a variable of interest. The null hypothesis takes the form:

H₀: The two populations from which the samples are drawn are identical (X and Y are independent).

When to Use Non-Parametric Two-Sample Tests

The parametric independent-samples t-test requires:

• Sample size ≥ 30 in each group (or normally distributed data)
• Homogeneity of variance
• Interval/ratio level measurement

When these assumptions are violated — small samples, skewed distributions, ordinal data, or gross variance inequality — the Mann-Whitney U test steps in as the appropriate alternative.

Classic Scenarios

Research Question	Groups	Measurement
Do males and females differ in History exam scores?	Male / Female (nominal)	Scores 0–100 (ordinal if n < 30)
Does teaching method affect final exam performance?	Method 1 / Method 2	Exam scores
Do two surgical groups differ in hospital stay?	Standard / Experimental surgery	Days (continuous, but small n)

---

2.3 The Mann-Whitney U Test: Core Logic

The Mann-Whitney U test analyses the degree of separation (or overlap) between an Experimental (E) group and a Control (C) group by counting how many times E scores exceed C scores and vice versa.

$U_1 = N_1 N_2 + \frac{N_1(N_1 + 1)}{2} - \Sigma R_1$

$U_2 = N_1 N_2 + \frac{N_2(N_2 + 1)}{2} - \Sigma R_2$

Where:

• $N_1$, $N_2$ = sample sizes of Group 1 and Group 2
• $\Sigma R_1$, $\Sigma R_2$ = sum of ranks for each group
• $U$ = smaller of $U_1$ and $U_2$
• $U'$ = larger of $U_1$ and $U_2$

The statistic $U_1$ literally counts the number of times a score from Group 1 exceeds a score from Group 2. If there is no difference between groups, both counts should be approximately equal.

Directional Interpretation

Hypotheses	Description
H₀ (Null)	The two sets of scores come from the same population — no systematic difference
H₁ (Non-directional)	The two sets differ systematically (two-tailed test)
H₁ (Directional)	Group E scores are specifically higher (or lower) than Group C scores (one-tailed test)

---

2.4 Assumptions and Background of the U Test

The Mann-Whitney U test, while "assumption-lite" compared to the t-test, is not assumption-free. Key sources (Conover, 1980; Siegel & Castellan, 1988) specify:

1. Random sampling: Each sample randomly selected from its population
2. Independence: The two samples are independent of each other
3. Continuous variable: The original variable is (ideally) a continuous random variable, subsequently ranked
4. Identical distribution shape: The two underlying populations have the same shape (though not necessarily normal)

Homogeneity of Variance

Maxwell and Delaney (1990) noted that identical distribution shapes implies equal dispersion — meaning the Mann-Whitney test also implicitly assumes homogeneity of variance, though this assumption is less often acknowledged. The test is somewhat less sensitive to variance violations than the t-test.

Why Rank Data? Handling Outliers

A key advantage of ranking: by converting raw scores to ranks, outliers lose their disproportionate influence. A score of 2,000 in a dataset with typical scores around 50–100 becomes rank 20 — still "high" but not distorting. This makes the Mann-Whitney U test especially valuable in clinical and applied psychology where extreme cases are common.

Historical Note

The test was independently developed by:

• Mann and Whitney (1947) — the U statistic version described here
• Wilcoxon (1949) — the rank-sum version (Wilcoxon-Mann-Whitney test)

Both yield comparable results despite using different computational formulas and tables.

---

2.5 Step-by-Step: Mann-Whitney U Test for Small Samples (n ≤ 20)

Use when: n ≤ 20 in each group AND data is:

• In rank-order format, OR
• Not normally distributed, OR
• Shows obvious variance inequality between groups

Procedure

STEP 1: Rank ALL scores together (both groups combined), assigning rank 1 to the lowest score.

STEP 2: Sum the ranks for Group 1: $\Sigma R_1$

STEP 3: Sum the ranks for Group 2: $\Sigma R_2$

STEP 4: Compute U and U' using the formulas above.

STEP 5: Compare the smaller of U and U' to the critical value from Table H (based on $N_1$, $N_2$, and significance level). The result is significant if the observed value ≤ the table value.

STEP 6: Translate the statistical result back into the context of the research question.

Worked Example: Assessment Centre Ratings

Two assessment teams independently rated officers on a performance scale. We want to know if the two teams gave significantly different ratings.

Assessment Team 1 Ratings	Assessment Team 2 Ratings
N₁ = 12 officers	N₂ = 13 officers
ΣR₁ = 148	ΣR₂ = 177

Computing U: $U = (12)(13) + \frac{12(13)}{2} - 148 = 156 + 78 - 148 = 86$

Computing U': $U' = (12)(13) + \frac{13(14)}{2} - 177 = 156 + 91 - 177 = 70$

Decision: The smaller value is U' = 70. For N₁ = 12, N₂ = 13, α = 0.05 (two-tailed), the critical value from Table H is 41.

Since 70 > 41, we fail to reject H₀. There is no significant difference in ratings between the two assessment teams.

Significance Rule for Small Samples

For the Mann-Whitney U test (small sample): The result is significant if U ≤ critical table value. This is the opposite direction from t-tests — smaller U = greater separation between groups.

---

2.6 Step-by-Step: Mann-Whitney U Test for Large Samples (n > 20)

When both sample sizes exceed approximately 20, the sampling distribution of U approaches normality, allowing conversion to a Z-score:

$Z = \frac{U - \frac{N_1 N_2}{2}}{\sqrt{\frac{N_1 N_2 (N_1 + N_2 + 1)}{12}}}$

Significance Criteria for Z

Test Type	Significance at α = .05
Two-tailed	Z > 1.96
One-tailed	Z > 1.64

Why Restrict to Small Samples?

For large samples, the ranking procedure becomes laborious, and violations of parametric assumptions become less consequential (the Central Limit Theorem mitigates normality concerns). Thus, Mann-Whitney U is most distinctively useful for small-to-moderate sample sizes where parametric tests are unreliable.

---

2.7 Computing Mann-Whitney U in SPSS

The SPSS workflow for non-parametric two-sample tests:

Analyse → Non-Parametric Tests → Legacy Dialogs → 2 Independent Samples

Detailed Steps

1. Analyse → Non-Parametric Tests → 2 Independent Samples
2. Move the dependent variable (e.g., test scores) into the Test Variable List
3. Move the grouping variable (e.g., gender, treatment group) into the Grouping Variable box
4. Click Define Groups → enter the codes for each group (e.g., 1 = male, 2 = female) → Continue
5. Under Test Type, ensure Mann-Whitney U is checked
6. For exact probabilities: click Exact → Exact → Continue
7. Click OK

Reading SPSS Output

SPSS provides:

• Mean Rank for each group
• Sum of Ranks for each group
• U statistic and W statistic (Wilcoxon's version)
• Z-score (for large samples)
• Asymptotic significance (2-tailed) p-value
• Exact significance p-value (if requested)

---

Self-Assessment Questions

Level 1 — Recall

1. What is the non-parametric equivalent of the independent samples t-test?

2. What is the null hypothesis in the Mann-Whitney U test?

3. For a small sample Mann-Whitney U test with α = .05 (two-tailed), you obtain U = 15 and U' = 41. The critical table value is 23. What is your decision?

Level 2 — Application

4. A researcher compares anxiety scores of yoga practitioners (n = 8) vs. non-practitioners (n = 9). Scores are ordinal (Likert scale 1–5) and distributions are skewed. Which test is appropriate, and why? What would constitute a significant result?

5. After combining and ranking scores from two groups, you find ΣR₁ = 95 (N₁ = 10) and ΣR₂ = 120 (N₂ = 11). Compute U and U'. Which value do you compare to the critical table?

Level 3 — Critical Thinking

6. A colleague argues: "Ranking data loses information, so non-parametric tests are always inferior to parametric tests." Evaluate this claim. Under what conditions would you choose the Mann-Whitney U test even when a t-test could technically be run?

7. If two groups have identical medians but very different variances, would the Mann-Whitney U test reliably detect no group difference? What does this tell us about the test's implicit assumptions?

---

🧠 Memory Aid

"MANN ranks, U counts"

• Merge all scores into one list
• Assign ranks (1 = lowest)
• Note the rank sums per group
• Now compute U and U' with the formulas

U = N₁N₂ + [N₁(N₁+1)/2] - ΣR₁

Think of it as: "Total possible comparisons, plus correction, minus what was actually ranked to Group 1."

---

📚 External Resources

Foundational Reading

• 🌐 Wikipedia: Mann-Whitney U test — Comprehensive coverage of the test's history, assumptions, and formula derivation
• 🌐 Wikipedia: Non-parametric statistics — Broader context for distribution-free methods in psychology

Tutorials and Guides

• 📖 Laerd Statistics: Mann-Whitney U Test — Step-by-step guide with SPSS screenshots
• 📖 Statistics How To: Mann Whitney U Test — Clear worked examples with interpretation
• 📖 Real Statistics: Mann-Whitney Test — Excel and manual calculation approach

Video Lectures

• 🎥 Statistics 101: Mann-Whitney U — YouTube — Visual, conceptual walkthrough
• 🎥 SPSS: Mann-Whitney U Test — YouTube — Full SPSS tutorial with output interpretation

Research Articles

• 📄 Fay, M. P., & Proschan, M. A. (2010). Wilcoxon-Mann-Whitney or t-test? On assumptions for hypothesis tests and multiple interpretations of decision rules. Statistics Surveys, 4, 1–39. Link
• 📄 Nachar, N. (2008). The Mann-Whitney U: A test for assessing whether two independent samples come from the same distribution. Tutorials in Quantitative Methods for Psychology, 4(1), 13–20. Link
• 📄 Hart, A. (2001). Mann-Whitney test is not just a test of medians: differences in spread can be important. BMJ, 323(7309), 391–393.

---

Source PDFs:

• 📄 Block-4/Unit-2.pdf — Pages 21–33
• 📚 MPC-006 Statistics in Psychology