Which statistical test? Test selection assistant
Answer a few questions about your aim and your data – the assistant names the appropriate standard test, the effect measure to report and the matching calculator on this site. Free, no account, no data entered.
More about this tool
What you can do
- Find the standard test for comparing groups or time points, testing an association, comparing survival, assessing agreement between raters or methods, and evaluating a diagnostic test.
- See which effect measure to report alongside the p value, for example a mean difference with confidence interval, Cohen's d, an odds ratio or a hazard ratio.
- Get the function in R and the menu path in jamovi, SPSS and GraphPad Prism where available.
- Open the matching calculator or tile on pubmed.info directly from the recommendation, e.g. the t-test, Mann-Whitney, chi-square or McNemar calculator, the Kaplan-Meier curve or ICC and kappa.
- Copy the recommendation as text.
How it works
Answer up to five questions; each answer determines the next:
- Aim: compare groups, association, time to event, agreement or diagnostic test.
- Type of outcome: continuous, ordinal or categorical.
- Design: independent groups or paired data (same people measured repeatedly, matched pairs).
- Number of groups or time points: two, or three or more.
- Distribution: approximately normal or large groups, not normal, or unknown – if you do not know, the assistant chooses the cautious rank-based test.
The recommendation covers, among others, Welch's t-test, the paired t-test, one-way and repeated-measures ANOVA, Mann-Whitney, Wilcoxon signed-rank, Kruskal-Wallis, Friedman, chi-square and Fisher's exact test, McNemar, Cochran's Q, Pearson and Spearman correlation, Kaplan-Meier with log-rank test, Cohen's and weighted kappa and the ICC. Short background notes explain level of measurement, dependence, normality and multiple testing. Sources: Altman (Practical Statistics for Medical Research), Bland (An Introduction to Medical Statistics) and du Prel et al. (Dtsch Arztebl Int 2010).
Data and privacy
You do not enter any data, and nothing is transmitted. Only your last answers are remembered in this browser.
Limits: the assistant names standard methods for simple questions. Regression with several explanatory variables, clustered or multilevel data, missing values and meta-analyses are not covered; for these, and for theses, grant applications and publications, discuss the analysis plan with a biostatistician.
Frequently asked questions
t-test or Mann-Whitney U test?
For two independent groups with an approximately normally distributed continuous outcome, or larger groups, the assistant recommends Welch's t-test, which does not require equal variances. For ordinal data, skewed values, outliers or small groups it recommends the Mann-Whitney U test.
Which test for before-and-after measurements?
Paired data need paired methods: the paired t-test or the Wilcoxon signed-rank test for two time points, repeated-measures ANOVA or the Friedman test for three or more, and McNemar or Cochran's Q for yes/no outcomes.
Is a correlation enough to show agreement?
No. A high correlation does not mean that two measurements agree. For continuous measurements the assistant recommends the ICC and a Bland-Altman plot, for categories Cohen's or weighted kappa.
pubmed.info sets no tracking cookies and builds no user profiles. Visits are counted anonymously (cookieless, shortened IP addresses). Ads from Google appear only according to your consent choice.