Statistical Test Selector
Answer a few questions about your research objective, variables, and study design to identify statistical tests that may fit your analysis.
Find an appropriate statistical test
The recommendation updates as you describe your study.
Research objective
Outcome variable
Predictor or grouping variable
Number of groups or measurements
Study design
Distribution and assumptions
Agreement design
Your recommendation will appear here
Start by choosing your research objective.
How the statistical test selector works
A statistical test should follow the research question and data structure—not the other way around.
The selector first identifies whether your goal is comparison, association, prediction, agreement, survival analysis, or a one-sample distributional question. It then considers the measurement level of the outcome, the predictor or grouping variable, the number of groups, and whether observations are independent or related.
The resulting recommendation includes a commonly used primary method, possible alternatives, and issues to verify before analysis.
Research objective
Different questions require different families of statistical procedures.
Variable type
Continuous, ordinal, categorical, count, and survival outcomes are modeled differently.
Data dependence
Independent groups require different tests from matched or repeated observations.
Assumptions
Model form, residuals, outliers, sample size, and variance structure matter.
A practical statistical test guide
Common tests organized by question and data structure.
| Research question | Typical design | Common test or model |
|---|---|---|
| Compare one continuous sample with a reference value | One sample | One-sample t testWilcoxon signed-rank |
| Compare two independent groups | Continuous outcome | Welch's t testMann–Whitney U |
| Compare two related measurements | Pre–post or matched pairs | Paired t testWilcoxon signed-rank |
| Compare three or more independent groups | Continuous or ordinal outcome | One-way ANOVAWelch ANOVAKruskal–Wallis |
| Compare three or more related measurements | Repeated observations | Repeated-measures ANOVAFriedman testMixed model |
| Assess association between categorical variables | Contingency table | Chi-square testFisher's exact test |
| Assess association between numeric or ordinal variables | Two variables | Pearson correlationSpearman correlationKendall's tau |
| Predict a continuous outcome | One or more predictors | Linear regressionRobust regressionMixed-effects model |
| Predict a binary or categorical outcome | One or more predictors | Binary logisticOrdinal logisticMultinomial logistic |
| Analyze count or rate data | Event counts or exposure-adjusted rates | Poisson regressionNegative binomial regression |
| Analyze time until an event | Censoring may occur | Kaplan–MeierLog-rank testCox regression |
| Measure reliability or agreement | Raters, items, or methods | Cohen's kappaFleiss' kappaICCBland–Altman |
Before choosing a test
- Define the estimand. State exactly what difference, association, probability, or effect you want to estimate.
- Identify the unit of analysis. Participants, classrooms, schools, repeated visits, and observations are not interchangeable.
- Respect dependence. Repeated, matched, clustered, or nested observations usually need methods that model their correlation.
- Inspect the data. Check coding, missingness, impossible values, outliers, distributions, and group sizes.
- Plan effect sizes. A p value alone does not show magnitude or practical relevance.
When a simple test is not enough
A basic test selector cannot fully handle factorial experiments, multilevel sampling, longitudinal trajectories, latent variables, survey weights, propensity scores, mediation, moderation, multiple imputation, compositional data, network data, Bayesian analysis, or machine-learning validation.
For clustered or repeatedly measured data, mixed-effects models or generalized estimating equations may be more appropriate than treating every observation as independent. For observational causal questions, statistical adjustment alone does not guarantee causal identification.
Frequently asked questions
The appropriate test depends on your research objective, outcome variable, predictor or grouping variable, number of groups or measurements, whether observations are independent or paired, and whether important assumptions are reasonably satisfied.
No. It provides an educational starting point, not a definitive analysis plan. Complex sampling, multilevel data, repeated observations, missing data, small samples, multiple outcomes, and causal questions may require specialist advice.
Parametric tests model particular distributional features and usually make assumptions about errors or residuals. Nonparametric or rank-based tests use fewer distributional assumptions but still have design requirements and may test a different statistical hypothesis.
Normality is generally most relevant to model residuals or within-group distributions, not merely the raw pooled outcome. Graphs, sample size, outliers, design balance, and robustness should be considered alongside formal tests.
Use it to compare the mean of a continuous outcome between two independent groups when the observations are independent and the model assumptions are acceptable. Welch's t test is often preferred when group variances or sample sizes differ.
Use it for two related measurements, such as pretest and posttest scores from the same participants or matched pairs, when the pairwise differences are approximately normal and serious outliers are absent.
Use ANOVA to compare a continuous outcome across three or more groups or conditions. The exact form depends on whether groups are independent, measurements are repeated, covariates are included, or the design contains multiple factors.
A chi-square test of independence is commonly used to assess association between two categorical variables in independent observations when expected cell counts are adequate. Fisher's exact test is an alternative for sparse small tables.
Pearson correlation describes linear association between continuous variables. Spearman correlation describes monotonic rank association and may be more suitable for ordinal variables, strong outliers, or clearly non-linear monotonic relationships.
The model is largely determined by the outcome: linear regression for continuous outcomes, binary logistic regression for two-category outcomes, ordinal logistic regression for ordered categories, multinomial logistic regression for unordered categories, and count models such as Poisson or negative binomial regression for counts.
Report descriptive statistics, the test statistic, degrees of freedom when applicable, exact p value, confidence interval, an appropriate effect size, assumption checks, missing-data handling, and enough methodological detail for reproducibility.
No. Statistical significance does not establish practical importance, causality, measurement quality, or replicability. Interpret p values together with effect sizes, confidence intervals, study design, and substantive context.