Which statistical test should I use for my research?

Start with the question, not a test name

“Which statistical test should I use?” is a useful question, but it comes after a more precise one: what do you want to estimate? A difference between groups, an association, a change over time, a prediction, and a causal effect are different goals. The same data can support more than one analysis, each answering a different question.

This guide is a map for planning and discussing an analysis. It cannot select a test from a title or a questionnaire alone. Your sampling, measurement, missing data, and intended interpretation all matter.

For an association between two questionnaire scores, the Pearson correlation guide separates the coefficient, uncertainty and two-sided test using the basic study.

Five decisions to write down

1. Define the estimand. State the difference, association, risk, or change you want to describe, and the population to which it applies.
2. Identify the outcome. Is it continuous, binary, ordered, a count, or time until an event? A summed questionnaire score may need different treatment from a single ordered item.
3. Describe the predictors. Record the groups or explanatory variables, whether they are categorical or continuous, and which adjustments were planned before seeing the results.
4. Describe dependence. Were different people measured once, the same people measured twice, or participants sampled within classes or clinics? Repeated observations from one unit are correlated and need an analysis that accounts for that structure. Penn State's repeated-measures lesson explains why independence cannot be assumed in that setting.
5. Check whether the design supports the claim. A regression can adjust for measured variables, but it cannot turn an observational association into a randomized causal comparison.

Follow a complete analysis decision

The undergraduate research example asks whether academic self-efficacy and learning persistence are associated. Each of180 students contributes one response, and each construct is represented by the mean of five questionnaire items.

Read the consultation and analysis plan, then inspect the resulting files. The advanced study also has a question about several predictors together, addressed by multiple regression, and a separate group-comparison question. This shows why naming one test for an entire dissertation can hide different analytic goals. Methods discussed in this guide are not a promise that the service supports every design; feasibility and scope depend on the study.

Match the method to the outcome and design

For a continuous outcome, a comparison of means or a linear model may be a starting point. For a binary outcome, logistic regression or a comparison of proportions may be relevant. For repeated observations, the dependence between measurements must be modeled or otherwise accounted for. These are *families of approaches*, not automatic prescriptions. The amount of data, sparse categories, clustering, missingness, and the effect you need to report can change the choice.

Do not decide solely by asking whether a raw variable is “normally distributed.” In a fitted linear model, diagnostics concern the unexplained variation and model structure. NIST's guide to checking regression assumptions discusses residual plots and the patterns they can reveal. A diagnostic concern calls for investigation, not an automatic switch to a named alternative.

A short handoff checklist

• What is the primary research question and what result would answer it?
• What does each row represent, and can one person or group contribute multiple rows?
• How were participants selected and, if relevant, assigned to groups?
• How is the outcome measured and coded? Are any scores calculated from multiple items?
• Which comparisons were planned, and how many outcomes will be tested?
• Where are values missing, and is the reason known?
• What effect size and uncertainty interval would be useful to report?

Keep a copy of the original data, a codebook, and the questionnaire or measurement protocol with this checklist. If you are still collecting data, a statistical consultation can identify questions to clarify before analysis. If the data are ready, it can help define the analysis and its scope. You can also review the data-analysis service or see how the process works. Consultation is not yet open; check service availability.

For a difference between two groups, follow the worked independent samples t-test guide, including design, Welch inference and reporting.

See the complete research examples

Follow the research questions through the input files, analysis plan and outputs. Each example includes the consultation report, itemized illustrative price and generated files.