Statistical Tests Compared: Which Test to Use

Choosing the right statistical test comes down to three questions: what kind of outcome you are measuring, how many groups you are comparing, and whether those groups are independent or paired. This guide lays out the common tests against those questions, and links to the calculators where we have them.

The quick reference table

Your outcome Two independent groups Three or more groups A relationship
A number / scale
e.g. score, height
Independent t-test One-way ANOVA Pearson correlation / regression
A rank / ordinal
e.g. 1–5 rating
Mann-Whitney U Kruskal-Wallis Spearman correlation
A category
e.g. pass/fail
Chi-square test Chi-square test Chi-square of association

If your groups are paired

When the same people are measured twice (before and after, say), the choice changes: use a paired t-test for numeric outcomes, a Wilcoxon signed-rank test for ordinal ones, and McNemar’s test for categorical ones. For three or more repeated measures, use a repeated-measures ANOVA or the Friedman test.

Two things people get wrong

  • Using a t-test on clearly non-normal, small samples — a Mann-Whitney U is safer there.
  • Reading significance as importance — always report an effect size alongside the p-value, because a trivial effect can be significant in a large sample.

Not sure your choice fits your design? Try our Which Statistical Test tool, or send us your variables and we’ll confirm the right test — and run it if you like.

Want a statistician to confirm and run it?

Send us your research question and variables for a fixed-price quote.

Ask a statistician

Frequently asked questions

What if I have more than one outcome or predictor?
Then you likely need a regression model rather than a single test. Multiple predictors, control variables or nested data all point to modelling — send us the design and we’ll advise.
How do I know if my data are normally distributed?
Check with a histogram and a normality test, and consider your sample size. For small, clearly skewed samples, use the non-parametric option (Mann-Whitney, Kruskal-Wallis, Spearman).