Writing 7 min read

Choosing the Right Statistical Test: A Practical Guide

Selecting the wrong statistical test is one of the most common errors in published research. Peer reviewers catch it, editors reject papers for it, and it undermines the validity of your conclusions. Yet most training covers specific tests in isolation rather than teaching you how to choose between them.

The choice of statistical test depends on four factors: the type of outcome variable, the number of groups, whether the groups are independent or paired, and whether your data meets normality assumptions. Work through these in order and the right test becomes clear.

Step 1 — What type of outcome variable do you have?

Continuous

Measurements on a scale with meaningful intervals — weight, absorbance, blood pressure, enzyme activity. This is the most common type in lab research and opens the widest range of statistical tests.

Categorical / Binary

Outcomes that fall into categories — dead/alive, infected/not infected, responder/non-responder. Use chi-square or Fisher's exact test.

Ordinal

Ranked data where intervals are not equal — pain scale, Likert scale, cancer stage. Use non-parametric tests (Mann-Whitney, Kruskal-Wallis).

Step 2 — How many groups?

Two groups: t-test (if normal) or Mann-Whitney U (if not). Three or more: ANOVA (if normal) or Kruskal-Wallis (if not), followed by a post-hoc test. Never run multiple t-tests — with 5 comparisons at p < 0.05, the probability of at least one false positive is 23%.

Step 3 — Independent or paired?

Are the same subjects measured more than once (before/after, repeated time points, matched pairs)? Use paired tests: paired t-test, Wilcoxon signed-rank, or repeated-measures ANOVA. Different subjects in each group? Use independent tests.

Step 4 — Does your data meet normality assumptions?

Parametric tests (t-test, ANOVA) assume approximately normal data. Check with a histogram or Q-Q plot. For small samples (n < 15) where normality is uncertain, use the non-parametric equivalent to be conservative. With large samples (n > 30), parametric tests are robust to non-normality due to the central limit theorem.

Quick reference table

Two independent groups, continuous, normal → Independent t-test

Non-normal: Mann-Whitney U

Two paired groups, continuous, normal → Paired t-test

Non-normal: Wilcoxon signed-rank

Three+ independent groups, continuous, normal → One-way ANOVA + post-hoc

Non-normal: Kruskal-Wallis + Dunn's

Three+ paired groups → Repeated-measures ANOVA

Non-normal: Friedman test

Categorical outcome → Chi-square

Small expected counts: Fisher's exact

Correlation between two continuous variables → Pearson r (normal) or Spearman ρ (non-normal)

💡 Always report effect sizes (Cohen's d, eta-squared, r) alongside p-values. A statistically significant result with a tiny effect size may not be clinically or practically meaningful.

Use the Statistical Test Selector

Answer 5 questions about your data and get an instant recommendation with the exact test, assumptions to check, and software commands for SPSS, R, and GraphPad.

Try it free →