Skip to main content

Parametric vs. Non-Parametric Two-Sample Tests

Overview​

One of the most practical decisions in psychological research is: should I use a parametric test (t-test) or a non-parametric test (Mann-Whitney U or Wilcoxon)? This unit synthesises the comparison drawn in the IGNOU MPC-006 curriculum, situating it within the broader debate in statistical methodology.

2.13 Power Efficiency: The Central Trade-Off​

The power efficiency of a non-parametric test is often defined relative to its parametric equivalent. For the Mann-Whitney U and Wilcoxon tests:

When the assumptions of the t-test are fully met, the non-parametric tests require slightly larger sample sizes to achieve the same statistical power. However, when t-test assumptions are violated, non-parametric tests often outperform parametric alternatives.

Asymptotic Relative Efficiency (ARE)​

For normally distributed data, the Mann-Whitney U test has an ARE of approximately 0.955 compared to the t-test — meaning it requires only about 5% more observations to achieve equivalent power. This is remarkably high for a "less powerful" test.

For data from non-normal distributions, the ARE can exceed 1.0 — meaning the non-parametric test is actually more powerful.

Full Comparison Table​

FeatureIndependent t-testMann-Whitney UDependent t-testWilcoxon MPSR
DesignTwo independent groupsTwo independent groupsMatched/repeated measuresMatched/repeated measures
Data levelInterval/ratioOrdinal or aboveInterval/ratioOrdinal or above
Normality requiredYesNoYesNo (symmetric differences)
Sample sizeIdeally ≥ 30/groupAny (table for ≤20, Z for >20)Ideally ≥ 30 pairsAny (table for ≤25, Z for >25)
Sensitive to outliersYes (highly)No (ranks neutralise outliers)YesModerate
What it testsMean differenceMedian / rank-sum differenceMean difference (paired)Median difference score
Power efficiency100% (baseline)~95.5% under normality100% (baseline)~95.5% under normality
ComputationModerateModerate (laborious for large n)ModerateModerate
SPSS menuAnalyze → Compare MeansAnalyze → Non-Param → 2 IndependentAnalyze → Compare MeansAnalyze → Non-Param → 2 Related

The Two Camps of Statistical Opinion​

The Non-Parametric Camp​

Proponents argue non-parametric tests:

  • Are simpler to compute (especially with small samples)
  • Have fewer assumptions (no normality, no homogeneity of variance)
  • Can be applied more widely to ordinal, skewed, or small-sample data
  • Reduce outlier influence via the ranking procedure
  • Are appropriate by default for Likert-scale and clinical data common in psychology

The Parametric Camp​

Advocates counter that parametric tests:

  • Are robust to mild violations of assumptions (especially with n > 30)
  • Maintain higher power efficiency under ideal conditions
  • Are directly interpretable in terms of means (a more intuitive parameter)
  • Benefit from the Central Limit Theorem in large samples

Decision Framework: When to Use Each Test​

Practical Rule (Siegel & Castellan Recommendation)​

Use the t-test unless:

  1. Data is in rank-order format, OR
  2. Sample is small AND distribution is clearly non-normal, OR
  3. Sample is small AND variances are obviously unequal

When any of these conditions holds, switch to the appropriate non-parametric test.

Additional Consideration: Pressed for Time​

The IGNOU MPC-006 text notes: "If you are particularly pressed for time or have a large number of analyses to do, there is particularly nothing inappropriate about using non-parametric statistics, even in cases where t-tests might have been used."

This pragmatic stance reflects the reality that for most psychological data — ordinal scales, small clinical samples, convenience samples — non-parametric tests are often the safer choice by default.

When Parametric Assumptions Fail in Psychology​

Indian and international psychology research routinely encounters situations where non-parametric tests are mandatory:

SituationWhy Non-Parametric?Preferred Test
Comparing 12 patients before/after therapyn < 25, distribution unknownWilcoxon MPSR
Assessing Likert-scale attitude differences between two cultural groupsOrdinal data, often skewedMann-Whitney U
Ranking job candidates assessed by two different panelsData inherently ordinal (ranks)Mann-Whitney U
Pilot study with n = 8 per groupToo small for CLT, normality unverifiableMann-Whitney U
Measuring recovery time in two surgical groups with outliersOutliers distort t-testMann-Whitney U

Self-Assessment Questions​

Level 1 — Recall​

  1. What is the parametric equivalent of the Mann-Whitney U test? What is the parametric equivalent of the Wilcoxon MPSR test?

  2. What is the approximate Asymptotic Relative Efficiency of the Mann-Whitney U test relative to the t-test for normally distributed data?

  3. List three conditions under which the IGNOU text recommends choosing a non-parametric test over a t-test.

Level 2 — Application​

  1. A researcher studying the effect of meditation on stress (n = 15 participants, pre–post design) finds that the difference scores are clearly skewed. Should they use: (a) Paired t-test, (b) Mann-Whitney U, or (c) Wilcoxon MPSR? Explain your choice.

  2. Across five studies, all using the same two-group design with n = 25 per group, three studies report normal distributions and two report moderate skewness. Design a consistent analytical strategy that is valid across all five studies.

Level 3 — Critical Thinking​

  1. A senior researcher insists: "With a sample of n = 50 per group, you can always use the t-test regardless of distribution shape." Is this correct? Under what conditions might it still be advisable to use the Mann-Whitney U test with large samples?

  2. Compare the statement "non-parametric tests are assumption-free" with the actual assumption list for Mann-Whitney U and Wilcoxon MPSR. What is misleading about the common claim that these tests "require no assumptions"?

🧠 Memory Aid​

"MIND the Pair":

  • Mann-Whitney → Mismatch (two independent groups with no pairing)
  • Independent samples → t-test is the parametric twin
  • Non-normal or small n → switch from t to Mann-Whitney
  • Dependent/paired → Wilcoxon MPSR is the non-parametric choice
  • Paired t-test → Wilcoxon is the "pair"-analogous non-parametric test

Power shortcut: Non-parametric ≈ t-test in power unless distributions are perfectly normal — and in psychology, they rarely are.

📚 External Resources​

Foundational Reading​

Tutorials and Guides​

Video Resources​

Research Articles​

  • 📄 Siegel, S., & Castellan, N. J. (1988). Non-Parametric Statistics for the Behavioral Sciences (2nd ed.). McGraw-Hill. (Core textbook referenced throughout IGNOU MPC-006)
  • 📄 Zimmerman, D. W. (1998). Invalidation of parametric and nonparametric statistical tests by concurrent violation of two assumptions. The Journal of Experimental Education, 67(1), 55–68.
  • 📄 de Winter, J. C., & Dodou, D. (2010). Five-point Likert items: t test versus Mann-Whitney-Wilcoxon. Practical Assessment, Research & Evaluation, 15(1), 11.

Source PDFs: