Which Statistical Test Should I Use? A Decision Guide With Table

Reviewed by Prof. Dr. Alexander Holtermann · Updated 16 October 2026 · 6 min read

Your research question, the level of measurement of your outcome, the number of groups and whether samples are paired decide which statistical test fits.

Key takeaways

Four questions determine the test: Do you compare groups or look at relationships? What level of measurement is the outcome? How many groups or variables? Are the samples independent or paired?
Comparing two independent groups on a metric outcome: Welch’s t-test; ordinal or non-normal small samples: Mann-Whitney U test.
Three or more groups: one-way ANOVA or Kruskal-Wallis test. Same people measured twice: paired t-test or Wilcoxon signed-rank test.
Relationship between two metric variables: Pearson correlation or linear regression; ordinal: Spearman. Two categorical variables: chi-square test. Binary outcome with predictors: logistic regression.
Choose the test before collecting data, based on your hypotheses, and justify the choice in your methods section.

Four questions that lead to the right test

Choosing a statistical test feels overwhelming because there are so many. In practice, a small number of tests covers almost all student theses, and four questions lead you to the right one.

1. What is your research question? Do you want to compare groups (“Do A and B differ?”) or examine a relationship (“Is X related to Y?”) or predict an outcome?
2. What is the level of measurement of the outcome (dependent variable)? Nominal, ordinal or metric (interval/ratio)?
3. How many groups or variables? Two groups, three or more, one predictor or several?
4. Are the measurements independent or paired? Different people in each group, or the same people measured repeatedly or matched pairs?

Decision table: comparing groups

For group comparisons, find the row for your outcome and the column for your design. Parametric tests come first; the rank-based alternative is in parentheses.

Outcome	Two independent groups	Two paired measurements	3+ independent groups	3+ repeated measurements
Metric, roughly normal	Welch’s t-test	Paired t-test	One-way ANOVA (Welch’s ANOVA)	Repeated-measures ANOVA
Ordinal or non-normal metric	Mann-Whitney U test	Wilcoxon signed-rank test	Kruskal-Wallis test	Friedman test
Nominal (categories)	Chi-square test (Fisher’s exact test)	McNemar’s test	Chi-square test	Cochran’s Q test

Decision table: relationships and prediction

If you are interested in how variables are related, or want to predict one from others, the level of measurement of both variables decides.

Variables	Test or model	Effect size
Two metric variables, linear	Pearson correlation	r
Two ordinal variables, or metric with outliers	Spearman’s rank correlation (Kendall’s tau)	r_s
Two nominal variables	Chi-square test of independence	φ, Cramér’s V
Metric outcome, one or more predictors	Linear regression (simple or multiple)	R², β
Binary outcome (yes/no), one or more predictors	Logistic regression	Odds ratio
Metric outcome, group factor plus metric covariate	ANCOVA or regression with dummy variables	partial η²

Worked examples from student research

These typical thesis questions show how the four questions translate into a test.

Research question	Outcome	Design	Test
Does exam anxiety differ between bachelor’s and master’s students?	Scale score (metric)	2 independent groups	Welch’s t-test
Does a relaxation training reduce exam anxiety?	Scale score (metric)	Same people before/after	Paired t-test
Do three teaching formats lead to different quiz scores?	Points (metric)	3 independent groups	One-way ANOVA + Tukey HSD
Is study group membership related to passing?	Passed yes/no (nominal)	2 × 2 table	Chi-square test
Are study hours related to exam scores?	Two metric variables	Relationship	Pearson correlation, linear regression
Does satisfaction (single 5-point item) differ between online and in-person students?	Ordinal	2 independent groups	Mann-Whitney U test
Which factors predict dropping out of a course?	Dropped out yes/no	Several predictors	Logistic regression

Parametric or non-parametric?

Parametric tests (t-test, ANOVA, Pearson) make assumptions about distributions, mainly approximate normality, and in return are slightly more powerful and provide means and confidence intervals. Non-parametric (rank-based) tests need fewer assumptions.

Choose parametric when the outcome is metric and either roughly normal or the groups have about 30 or more cases each.
Choose rank-based when the outcome is ordinal, or metric with small groups and clear skewness or outliers.
Do not run both and report the one you like better. Decide in advance, or report the second as a robustness check.
Check assumptions with plots and descriptive statistics, not only with formal tests; see the article on normality testing.

Common mistakes when choosing a test

These mistakes cost points in theses and are easy to avoid.

Running several t-tests instead of one ANOVA for three or more groups (inflates the Type I error).
Treating paired data as independent, for example comparing before and after scores with an independent t-test; you lose power and violate the independence assumption.
Using a t-test for percentages or yes/no answers instead of a chi-square test.
Using linear regression for a binary outcome instead of logistic regression.
Computing Pearson correlations between single ordinal items without justification.
Choosing the test after looking at the results.

Planning: test choice and sample size go together

Once you know the test, plan the sample size with a power analysis. For the most common tests, the power analysis tool gives the required n for α = .05 and power .80: Welch’s t-test with a medium effect (d = 0.5) needs 64 per group; a paired t-test 34 people; a correlation of r = .30 needs 85 people; a 2 × 2 chi-square test with w = .30 needs 88. Write the planned test, α, the expected effect size and the resulting sample size into your methods section or preregistration.

Frequently asked questions

How do I know which statistical test to use?

Answer four questions: comparison or relationship, level of measurement of the outcome, number of groups or variables, and independent or paired samples. The decision tables in this article map the answers to a test.

What test do I use to compare two groups?

For a metric outcome, Welch’s t-test (independent) or the paired t-test (same people). For ordinal or clearly non-normal data in small samples, the Mann-Whitney U test or the Wilcoxon signed-rank test. For categories, the chi-square test.

What test do I use for Likert scale data?

For a single Likert item, rank-based tests such as Mann-Whitney U or Spearman correlation. For a scale score averaged from several items, parametric tests such as the t-test, ANOVA or Pearson correlation are common practice.

When do I use a chi-square test instead of a t-test?

When the outcome is categorical (e.g., passed/failed) rather than a metric score. A t-test compares means; a chi-square test compares frequencies.

Can I use a t-test with small samples?

Yes, if the values are roughly normally distributed. With small samples and clear skewness or outliers, the Mann-Whitney U test is safer. In either case, small samples mean low power.

Does SPSS, R or JASP choose the right test for me?

No. Statistics programs run whatever test you select. You choose the test based on your research question, design and level of measurement; the software then calculates it. Programs such as JASP help by showing assumption checks next to the results.

Sources and further reading

Howell, D. C. (2013). Statistical Methods for Psychology (8th ed.). Wadsworth Cengage Learning.
Agresti, A., & Finlay, B. (2018). Statistical Methods for the Social Sciences (5th ed.). Pearson.
Gravetter, F. J., & Wallnau, L. B. (2017). Statistics for the Behavioral Sciences (10th ed.). Cengage.

Your university’s examination regulations and style guide always take precedence.