If you have ever opened a statistics textbook, looked at the long list of tests — t-test, ANOVA, Mann-Whitney U, Kruskal-Wallis — and thought, “Which one is mine?”, you are not alone. Nearly every student reaches this fork in the road, and the question behind it is the one this guide answers: parametric vs nonparametric tests — what is the difference, and how do you choose?
The short answer is that the choice depends on your data, not on your preference. Parametric tests assume your data follow a particular distribution (usually the normal distribution); nonparametric tests make far fewer assumptions. Pick the wrong family and your results can be misleading. Pick the right one and your analysis is both honest and convincing. This guide explains both families in plain English, compares them side by side, and gives you a simple five-step method for choosing correctly every time.
What Is a Parametric Test?
A parametric test is a statistical test that makes assumptions about the parameters of the population your data come from — most importantly, its mean and standard deviation. The word “parametric” literally means “based on parameters”.
The best-known assumption is normality: parametric tests generally assume that your data are approximately normally distributed — the familiar bell-shaped curve. They also usually assume that your data are measured on an interval or ratio scale (think test scores, reaction times, or height in centimetres) and that the groups you compare have roughly equal variances, an assumption called homogeneity of variance.
The tests you meet most often in this family are the independent-samples t-test, the paired-samples t-test, one-way ANOVA, Pearson’s correlation, and linear regression. When their assumptions are met, parametric tests are the more powerful option: they are more likely to detect a real effect when one exists, and they let you draw conclusions about population means, which is usually exactly what a researcher wants to report.
The catch is the “when their assumptions are met” part. If your data are badly skewed, full of outliers, or measured on an ordinal scale, a parametric test can give you a p-value that looks convincing but means very little. That is why the first skill in choosing a test is learning to check the assumptions — not just learning the test names.
What Is a Nonparametric Test?
A nonparametric test is a “distribution-free” test: it does not assume your data follow a normal distribution or any other specific shape. Instead of working with the raw scores, most nonparametric tests work with the ranks of your data — they sort your scores from smallest to largest and analyse the order rather than the exact values.
Because they make so few assumptions, nonparametric tests are the safe choice when your data are skewed, your sample is small, or your measurements are ordinal — for example, Likert-scale ratings, rankings, or categories with a natural order. They are also the right choice when outliers would distort a mean: one extreme score can pull a group mean a long way, but it can only move a rank by one place.
The most common nonparametric tests are the Mann-Whitney U test (the rank-based alternative to the independent t-test), the Wilcoxon signed-rank test (the alternative to the paired t-test), the Kruskal-Wallis test (the alternative to one-way ANOVA), the Friedman test (the alternative to repeated-measures ANOVA), Spearman’s rank correlation (the alternative to Pearson’s r), and the chi-square test.
There is a trade-off. When your data genuinely are normal, a nonparametric test has slightly less statistical power than its parametric equivalent — it is a little less likely to detect a real effect. Think of it this way: parametric tests use all the information in your exact scores, while rank-based tests deliberately throw away some of that information to buy robustness. When the assumptions hold, that information is worth keeping; when they do not, it is worth trading away.
Parametric vs Nonparametric Tests: The Key Differences
| Feature | Parametric tests | Nonparametric tests |
|---|---|---|
| Assumption about distribution | Assume data follow a known distribution, usually normal | Make no assumption about the shape of the distribution |
| Type of data | Interval or ratio data | Nominal or ordinal data (also work with interval or ratio data) |
| What it compares | Means (and variances) | Medians, distributions, or ranks |
| Statistical power | Higher when assumptions are met | Slightly lower when data are normal |
| Sample size | Works best with moderate to large samples | Works well with small samples |
| Outliers | Sensitive — one extreme score can distort the mean | Resistant — ranks limit the influence of any single score |
| Typical examples | t-test, ANOVA, Pearson correlation, regression | Mann-Whitney U, Wilcoxon, Kruskal-Wallis, Spearman |
The Matching Pairs: Every Parametric Test Has a Nonparametric Partner
| Research question | Parametric test | Nonparametric equivalent |
|---|---|---|
| Compare two independent groups | Independent-samples t-test | Mann-Whitney U test |
| Compare two related groups (before/after) | Paired-samples t-test | Wilcoxon signed-rank test |
| Compare three or more independent groups | One-way ANOVA | Kruskal-Wallis test |
| Compare three or more related groups | Repeated-measures ANOVA | Friedman test |
| Measure association between two variables | Pearson correlation (r) | Spearman rank correlation (rho) |
| Test association between categories | — (no parametric equivalent) | Chi-square test of independence |
Notice the pattern: the two tests in each row answer the same research question — they just make different assumptions about the data. Once you know the question you are asking, choosing between parametric and nonparametric is usually just a matter of checking the assumptions.
How to Choose: A Five-Step Method
Step 1: What type of data do you have?
Start with your measurement scale. If your data are nominal (categories with no order, such as eye colour) or ordinal (ordered categories, such as Likert ratings or rankings), the decision is made for you: use a nonparametric test. Parametric tests need interval or ratio data — numbers where the gaps between values mean something consistent, like test scores or reaction times.
Step 2: Is your data approximately normally distributed?
If your data are interval or ratio, look at their shape. A histogram or a Q-Q plot will show you at a glance whether the scores pile up in a bell shape or lean heavily to one side. For a formal check, the Shapiro-Wilk test is the standard choice for smaller samples. A word of caution: with very large samples, normality tests can flag trivial departures as “significant”, so always look at the plot as well as the p-value.
Step 3: How big is your sample?
With a small sample — roughly fewer than 30 observations per group — it is hard to judge normality at all, and a single outlier carries more weight. When samples are small and the distribution is doubtful, a nonparametric test is the safer default. With larger samples, the central limit theorem helps parametric tests stay valid even when the raw data are somewhat skewed.
Step 4: Check for outliers and equal variances
Scan for extreme scores that do not belong to the same pattern as the rest. If a few wild values are genuine data (not typing errors), a rank-based test will stop them from dominating your result. Also check that your groups have roughly similar spread; wildly unequal variances are another reason to switch families.
Step 5: Match the test to your question
Finally, confirm the test answers the question you actually asked. Comparing means? That is parametric territory. Comparing medians or overall distributions? Nonparametric. And remember that switching families changes what you report: with a Mann-Whitney U test you report medians and a rank-based effect size, not group means.
Worked Example: Choosing a Test for a Real Study
Imagine a student teacher comparing two teaching methods. She collects exam scores (out of 100) for two groups of 22 students each — one group taught with lectures, the other with problem-based learning. Before choosing a test, she works through the five steps.
First, her data are interval data (exam scores), so either family is possible. But when she plots the scores, the lecture group shows a long tail of very low marks, and the Shapiro-Wilk test returns a p-value of .01 — clear evidence against normality. Three students scored below 30 while the rest clustered between 65 and 80.
An independent-samples t-test would compare the means — and those three low scores would drag the lecture group’s mean down, making the teaching methods look more different than they really are. So she runs a Mann-Whitney U test instead, which compares the ranks of the scores rather than their exact values. The result shows the problem-based group scoring systematically higher (median 74 versus 66), and the conclusion is honest about what was actually compared: the overall distribution of scores, not a mean distorted by outliers.
Had the data been roughly normal with no extreme scores, the t-test would have been the better choice: more power, and a conclusion stated in terms of mean scores, which most readers find easier to interpret. The point of the example is not that one family is “better” — it is that the data, not habit, should make the decision.
Common Mistakes to Avoid
- Running a t-test on Likert-scale data without a second thought. A 1-to-5 rating is ordinal, not interval. The gap between “agree” and “strongly agree” is not necessarily the same as the gap between “neutral” and “agree”, so a test that assumes equal intervals is on shaky ground. A Mann-Whitney U or Kruskal-Wallis test is the honest choice.
- Testing normality on the wrong variable. For a paired t-test, it is the differences between the paired scores that should be roughly normal, not the raw scores themselves. For ANOVA, it is the residuals. Checking the wrong thing gives you false confidence.
- Switching tests until you get p < .05. Running the t-test, disliking the result, then trying the Mann-Whitney U (or the reverse) is test-shopping. Decide on the test from your data and your assumptions before you look at the p-value — and say so in your write-up.
- Believing nonparametric means “no assumptions at all”. Nonparametric tests drop the normality assumption, but they still assume your observations are independent and, for most rank tests, that the groups have similarly shaped distributions. Independence is never optional.
- Reporting means after running a rank-based test. If you used a Mann-Whitney U test, your conclusion is about ranks and medians. Reporting group means alongside it confuses your reader and misstates what the test actually compared.
Frequently Asked Questions
Can I use a parametric test if my data are not normal?
Sometimes. Mild departures from normality with a large sample are usually fine, because parametric tests are fairly robust. But severe skew, a small sample, or heavy outliers are genuine reasons to switch to a nonparametric test. When in doubt, run the nonparametric version — a slightly conservative answer you can defend beats a flashy one you cannot.
Which type of test is more powerful?
When the assumptions are met, parametric tests are more powerful: they are more likely to detect a real effect. But when the assumptions are violated, the balance flips — a nonparametric test can actually be more powerful than a parametric test applied to badly skewed data. Power belongs to the test whose assumptions fit the data.
Can I use nonparametric tests with large samples?
Yes. Nothing stops you, and with very large samples the power difference between the two families becomes tiny. In practice, though, large clean samples usually satisfy the parametric assumptions, so researchers tend to use the parametric test and report means. Either choice is defensible if you can explain it.
Is the chi-square test parametric or nonparametric?
Nonparametric. The chi-square test works with frequencies in categories, not with means or variances, and it assumes nothing about the shape of a distribution. That is why it appears in the nonparametric column even though it has no parametric partner.
What should I do if my sample is very small?
With fewer than about 10 to 15 observations per group, checking normality is almost impossible, so a nonparametric test is the safer default. Just as important, do not lean only on the p-value: with tiny samples, report effect sizes and confidence intervals as well, and be honest about how much (or how little) the data can actually show.
Key Takeaways
- Parametric tests assume your data follow a known distribution (usually normal) and compare means; nonparametric tests make almost no distributional assumptions and compare ranks, medians, or distributions.
- Every common parametric test has a nonparametric partner that answers the same research question — t-test to Mann-Whitney U, ANOVA to Kruskal-Wallis, Pearson to Spearman.
- Choose by checking your data type, the shape of your distribution, your sample size, and outliers — in that order — and decide before you look at the p-value.
- When the assumptions hold, parametric tests give you more power; when they do not, the nonparametric test is the honest (and sometimes more powerful) choice.