Parametric vs Non-Parametric Tests: How to Choose
When to use a t-test or ANOVA and when to use Mann–Whitney, Wilcoxon or Kruskal–Wallis instead, with a simple decision table for medical research.
One of the most common questions in thesis analysis is whether to use a parametric or a non-parametric test. The choice depends on the type of data, its distribution and the sample size, not on which test gives the smaller p-value.
What is the difference?
Parametric tests assume that the data (or the differences between groups) follow a particular distribution, usually the normal distribution. They compare means. Examples: t-tests, ANOVA, Pearson correlation.
Non-parametric tests make fewer assumptions about the distribution. Most work on the ranks of the data and compare medians or distributions. Examples: Mann–Whitney U, Wilcoxon signed-rank, Kruskal–Wallis, Spearman correlation.
Matching pairs of tests
| Situation | Parametric test | Non-parametric alternative |
|---|---|---|
| Two independent groups | Independent samples t-test | Mann–Whitney U test |
| Same participants measured twice | Paired t-test | Wilcoxon signed-rank test |
| Three or more independent groups | One-way ANOVA | Kruskal–Wallis test |
| Same participants, three or more times | Repeated measures ANOVA | Friedman test |
| Relationship between two numerical variables | Pearson correlation | Spearman rank correlation |
How to decide
- Check the variable type. Ordinal data (such as a pain score from 0 to 10, or Likert responses) are often better handled with non-parametric tests.
- Look at the distribution. Draw a histogram or Q–Q plot for each group. Severe skew or extreme outliers favour non-parametric tests.
- Consider the sample size. With larger samples, t-tests and ANOVA are fairly robust to moderate non-normality. With very small groups, normality is hard to check and non-parametric tests are a safer choice.
- Check equal variances for t-tests and ANOVA (Levene's test). If variances differ, use Welch's t-test or Welch's ANOVA.
Normality tests such as Shapiro–Wilk are useful, but do not rely on them alone. In large samples they flag trivial departures; in small samples they miss important ones. Look at the plots too.
A worked example
You compare length of hospital stay between patients with and without a surgical site infection. Most patients stay 3–5 days, but a few stay over 20 days. The data are strongly skewed, so you report the median (IQR) for each group and use the Mann–Whitney U test.
If instead you compare mean haemoglobin between two groups of 60 women each, and the histograms look roughly symmetrical, an independent t-test with mean (SD) is appropriate.
Common mistakes
- Running both tests and reporting whichever gives a significant result.
- Using a paired test for independent groups, or the reverse.
- Reporting means with a non-parametric test (report medians and IQR instead).
- Running several t-tests for three groups instead of ANOVA, which inflates the chance of a false positive.
Key takeaways
- Parametric tests compare means and assume a distribution; non-parametric tests use ranks.
- Decide from the data type, the shape of the distribution and the sample size.
- Report the summary statistic that matches the test: mean (SD) or median (IQR).
- Choose the test before looking at p-values.
References
- Kirkwood BR, Sterne JAC. Essential Medical Statistics. 2nd ed. Oxford: Blackwell Science; 2003.
- Altman DG. Practical Statistics for Medical Research. London: Chapman & Hall; 1991.
- Mishra P, Pandey CM, Singh U, Gupta A, Sahu C, Keshri A. Descriptive statistics and normality tests for statistical data. Ann Card Anaesth. 2019;22(1):67–72. PubMed
This article is for education. Always follow the requirements of your university, guide, ethics committee and target journal.