When to Use Non-Parametric Tests

Non-parametric tests are the alternatives you reach for when your data breaks the assumptions that t-tests, ANOVA, and Pearson correlation depend on. Instead of assuming a normal distribution and working with means, they work with the ranks of your data, which makes them valid for skewed distributions, ordinal variables, small samples, and data with outliers. This guide explains when parametric assumptions fail, how rank-based tests get around them, which non-parametric test replaces which parametric one, and how to choose and report the right test.

Every major test earlier in this cluster has a non-parametric twin. If you know the t-test, ANOVA, and correlation, this guide mostly teaches you the fallback for each, and the judgment about when to switch.

What Parametric Tests Assume

Parametric tests such as the t-test, ANOVA, and Pearson correlation are called parametric because they assume the data follows a distribution with specific parameters, usually the normal distribution defined by its mean and standard deviation. Those assumptions are what let the tests work with means and produce their p-values. When the assumptions hold, parametric tests are the right choice, and they're slightly more powerful than the alternatives. When the assumptions fail, the p-values can be wrong, and that's when non-parametric tests earn their place.

Four situations commonly break parametric assumptions.

• Non-normal distributions. The data is clearly skewed rather than symmetric. Income, reaction times, and wealth are classic examples that pile up at one end with a long tail. When the distribution is far from normal, a test built on normality can mislead.
• Ordinal data. The variable is a rank or rating rather than a true measurement. Likert-scale responses, satisfaction ratings, and severity grades have order but no guarantee that the gap between 1 and 2 equals the gap between 4 and 5. Means aren't strictly meaningful on such data, but ranks are.
• Small samples. With very few observations, you can't verify normality, and the central limit theorem can't rescue you. When n is small and you have no external reason to assume normality, a rank-based test is the safer choice.
• Outliers. One or two extreme values can distort a mean and inflate a standard deviation, dragging a parametric test with them. Because rank-based tests use order rather than magnitude, a wild outlier becomes just the highest rank rather than a number that warps the whole calculation.

Checking these is largely visual, which is why plotting comes before testing. A histogram or box plot reveals skew and outliers at a glance, as the visualization guide covers. Formal normality tests exist, but on small samples they lack the power to detect non-normality, and on large samples they flag trivial departures, so the plot is usually more informative than the test.

How Rank-Based Tests Work

The core idea is simple: replace each value with its rank in the combined data, then test the ranks instead of the raw numbers. The smallest value becomes rank 1, the next becomes rank 2, and so on. All the tests below run on those ranks.

This single move is what gives non-parametric tests their two signature properties. Because ranks don't care about the distance between values, a distribution's shape stops mattering, so there's no normality assumption. And because the largest value is just the highest rank no matter how extreme it is, a single outlier can't distort the result. The cost is that ranking discards the magnitude information in the data, which is the source of the small power disadvantage discussed later. What it keeps is the order, which for many research questions is exactly what matters.

The Non-Parametric Equivalents

Every common parametric test has a rank-based counterpart that answers the same question without the normality assumption. This table is the heart of the guide: find the parametric test you would have used, and read across to its non-parametric equivalent.

Mann-Whitney U: two independent groups

The Mann-Whitney U test (also called the Wilcoxon rank-sum test) is the non-parametric answer to the independent-samples t-test. It compares two independent groups by ranking all observations together and checking whether one group's ranks tend to be higher than the other's. Use it when you'd have run an independent t-test but the data is skewed, ordinal, or has outliers.

Wilcoxon signed-rank: paired measurements

The Wilcoxon signed-rank test replaces the paired-samples t-test. It's for before-and-after designs, crossover studies, or matched pairs, where each subject serves as its own control. It ranks the sizes of the differences between the paired measurements, so it tests whether the differences tend to be positive or negative without assuming those differences are normally distributed.

Kruskal-Wallis: three or more groups

The Kruskal-Wallis H test is the non-parametric equivalent of one-way ANOVA, extending the Mann-Whitney logic to three or more independent groups. Like ANOVA, a significant result tells you at least one group differs from the others, but not which. The follow-up is discussed below.

Spearman: association between variables

Spearman rank correlation replaces Pearson correlation. It correlates the ranks of two variables rather than their values, which makes it appropriate for ordinal data or for monotonic relationships that aren't strictly linear. The choice between Spearman and Pearson, and how to report each, is covered in detail in the guide to using correlation coefficients in research papers, with the mechanics of the coefficient itself in the guide to what the correlation coefficient measures.

A Decision Framework

Choosing between a parametric test and its non-parametric equivalent comes down to a short sequence of questions.

1. What's your data type? If the outcome is ordinal (ranks, ratings, Likert items), go non-parametric. If it's continuous, keep going.
2. Plot the distribution. Make a histogram or box plot. If the data is roughly symmetric with no severe outliers, a parametric test is fine. If it's clearly skewed or has influential outliers, favor the non-parametric version.
3. How big is the sample? With a large sample, the central limit theorem makes parametric tests robust to moderate non-normality, so mild skew is not disqualifying. With a small sample, you can't verify normality, so a rank-based test is safer unless you have strong external grounds to assume normality.
4. Match the design to the test. Two independent groups point to Mann-Whitney; paired measurements to Wilcoxon signed-rank; three or more groups to Kruskal-Wallis; an association to Spearman.

When the parametric assumptions genuinely hold, prefer the parametric test for its slightly greater power. When they don't, switch without hesitation, since a valid non-parametric result beats an invalid parametric one every time.

The Power Tradeoff

The usual objection to non-parametric tests is that they're less powerful, meaning likelier to miss a real effect. That's true, but the size of the penalty is routinely overstated, and it runs in both directions.

On genuinely normal data, a parametric test holds only a modest power advantage, on the order of a few percent, because ranking discards some information the parametric test can use. But that comparison only applies when the data actually is normal. On skewed data or data with outliers, the situation reverses: the violated assumptions damage the parametric test more than the loss of magnitude information damages the rank-based test, so the non-parametric version can be the more powerful choice as well as the more valid one. The penalty is real only in the one case, normal data, where you'd have used the parametric test anyway.

The practical implication is that "non-parametric tests are weaker" is not a reason to force a parametric test onto data that violates its assumptions. The power you'd gain is illusory when the assumptions are broken.

After a Significant Kruskal-Wallis: Post-Hoc Tests

Like one-way ANOVA, a significant Kruskal-Wallis result tells you that at least one group differs, but not which pairs. The standard follow-up is pairwise Mann-Whitney tests between the groups, with a correction for running multiple comparisons.

The correction matters for the same reason it matters after ANOVA. Each test carries its own false-positive risk, and running several inflates the overall chance of a spurious result. With three groups you run three pairwise tests, and the familywise error rate climbs from 5 percent to roughly 14 percent if left uncorrected. A Bonferroni correction addresses this by dividing the significance threshold by the number of comparisons, so three comparisons are each tested at about .0167 rather than .05. The Holm correction is a slightly more powerful alternative that achieves the same error control. This is the same multiple-comparisons problem that makes running several t-tests in place of ANOVA unreliable, covered in the guide to ANOVA.

Reporting Non-Parametric Tests

Reporting conventions differ from parametric tests in one important way: because these tests work on ranks, you describe the data with the median and interquartile range rather than the mean and standard deviation, since the median is the matching measure of center for skewed or ordinal data.

• Describe with median and IQR, not mean and SD. The pairing follows from the same logic as choosing the median for skewed data in the first place, covered in the guide to standard deviation and variance.
• Report the test statistic, sample sizes, and exact p-value. Mann-Whitney reports U, Wilcoxon reports W or T, Kruskal-Wallis reports H with its degrees of freedom, and Spearman reports its coefficient.
• State which test and why. Name the non-parametric test and briefly note the reason for choosing it, such as non-normal data or ordinal measurement, so a reviewer sees the choice was deliberate.
• Report an effect size where available. Significance alone doesn't convey magnitude. Rank-biserial correlation for Mann-Whitney, or epsilon-squared for Kruskal-Wallis, gives the reader the size of the effect. The same hypothesis-testing logic of a null and a p-value applies throughout.

Common Mistakes

• Defaulting to non-parametric out of caution. When the parametric assumptions genuinely hold, the parametric test is more powerful. Rank-based tests are the fallback for violated assumptions, not a universal safe choice.
• Running a normality test on a large sample and switching on a trivial result. On large samples, normality tests flag departures too small to matter. A histogram plus the robustness of parametric tests at large n usually settles it better than a significance test.
• Reporting the mean and SD with a non-parametric test. If the data justified a rank-based test, describe it with the median and IQR, which match. Reporting a mean reintroduces the very sensitivity to skew you switched to avoid.
• Skipping the post-hoc correction after Kruskal-Wallis. Uncorrected pairwise comparisons inflate the false-positive rate exactly as uncorrected multiple t-tests do. Apply a Bonferroni or Holm correction.
• Using a rank-based test on nominal data. Mann-Whitney, Wilcoxon, and Kruskal-Wallis all require data that can be meaningfully ranked, meaning at least ordinal. For nominal categories with no order, a chi-square test is the appropriate tool instead.
• Assuming Mann-Whitney only compares medians. It's often described as a median comparison, but it's strictly a test of whether one group's values tend to rank higher than the other's. When the two groups have similarly shaped distributions, that reads cleanly as a difference in medians; when the shapes differ, interpret it as a difference in ranks.

Where This Fits in the Cluster

Non-parametric tests are the safety net under the parametric methods covered elsewhere in this cluster. When your data suits the assumptions, use the parametric versions: the t-test for one or two groups, and ANOVA for three or more. When the assumptions fail, this guide's equivalents keep your analysis valid. For the reasoning behind significance itself, see the guide to hypothesis testing, and for choosing between Pearson and Spearman specifically, the guide to using correlation coefficients in research papers.

When your analysis is written up, the methods and results sections are where a subject-matter editor earns their place, because a reviewer who spots the wrong test for the data, or a mean reported where a median belongs, will question the rest. Editor World's academic editing, dissertation editing, and journal article editing services include review of statistical reporting, and you choose your own editor by field so the person reading your methods knows your discipline's conventions.

Frequently Asked Questions

Page last reviewed: July 2026. Content reviewed by Editor World editorial staff. Editor World, founded in 2010 by Patti Fisher, PhD, graduate of The Ohio State University, provides professional editing and proofreading services for academic researchers, doctoral candidates, faculty, business professionals, and authors worldwide. 100% human editing, no AI at any stage. BBB A+ accredited since 2010 with 5.0/5 Google Reviews and 5.0/5 Facebook Reviews. More than 100 million words edited for over 8,000 clients in 65+ countries. Multiple Gold and Bronze Stevie Award winner. Native English editors from the United States, the United Kingdom, and Canada. Less than 5% of applicants are accepted to the editor panel. Recommended by the Boston University Economics Department, University of San Diego, University of Michigan, UCLA, University of Missouri, and more.