Statistical Methods

How to Choose the Right Statistical Test for Medical Research

Match research objectives, variable types and study design to appropriate statistical tests.

Start with the research question, not the software

A statistical test is chosen to answer an objective. Write down the primary outcome, explanatory variable, number of groups, whether observations are independent or paired, and whether the objective concerns a difference, association, prediction, or estimation. The same dataset may legitimately require different tests for different objectives.

Research question	Outcome	Typical first choice
Is smoking status associated with disease status?	Two categorical variables	Chi-square; Fisher exact for sparse tables
Is mean haemoglobin different between independent groups?	Continuous, two independent groups	Independent-samples t-test; Welch variant if variances differ
Did the same participants improve after treatment?	Continuous, paired observations	Paired t-test, considering the distribution of paired differences
Do three treatment groups differ in a continuous score?	Continuous, independent groups	ANOVA or an appropriate alternative
Are two continuous measures associated?	Continuous/ordinal variables	Pearson or Spearman correlation, depending on assumptions

Check the design and assumptions

For categorical comparisons, inspect cell counts and the underlying sampling design. For continuous outcomes, examine distribution shape, outliers, equal variance assumptions where relevant, and independence. A non-significant normality test is not proof of normality. When the study design involves matched groups, clustering, repeated measures, or confounding, a simple test may be insufficient.

Worked clinical example

Suppose a study compares mean pain scores between two different treatment groups at six weeks. If scores behave sufficiently like a continuous measure and assumptions are reasonable, consider an independent-samples t-test; where variance equality is doubtful, Welch's t-test may be preferable. If the same patients have baseline and six-week scores, compare within-patient differences instead. The observation structure changes the analysis even though the measured variable is unchanged.

What to report

State the exact test, sample sizes, summary measures, magnitude and direction of the effect, a confidence interval where appropriate, and the p-value. Explain missing data, prespecified comparisons and any changes to the analysis plan. Correlation is not causation; statistical significance is not the same as clinical importance.

Common mistakes

Choosing a test only because another thesis used it.
Treating repeated observations as independent.
Reporting p = 0.000 rather than p < 0.001.
Running many unplanned tests without discussing multiplicity.
Reporting only p-values and leaving out effect sizes and uncertainty.

Next step: Prepare one row for every objective with its variables, measurement scales, and planned analysis before uploading data to ResearchTry.

Frequently asked questions

Does a normality test decide the analysis on its own?

No. Consider the distribution of residuals or paired differences, sample size, outliers, measurement scale, independence and the estimand alongside graphical checks.

What if my study has repeated measures?

Use a method that represents within-person correlation, such as an appropriate repeated-measures model; do not treat linked observations as independent.