Which statistical test fits your study?

This is not a neutral lookup table — it encodes methodological positions that reviewers and statisticians hold, so the recommendation is defensible, not just common:

• Welch’s t by default for two independent means — not Student’s t behind a variance pre-test, which inflates type-I error.
• No normality test to pick the test. At moderate n, normality tests over-detect trivial deviations; judge with a Q–Q plot and the estimand, and treat rank tests as a change of estimand, not a free upgrade.
• An odds ratio is not a risk ratio when the outcome is common — cohort or trial data wanting a risk ratio get log-binomial or modified-Poisson regression, not a relabeled OR.
• Matched designs get matched analyses — McNemar’s, conditional logistic, paired t — because breaking the matching is a top reviewer catch.
• Estimation over testing: every recommendation reports an effect size with a confidence interval; the p-value is secondary.

The same decision table runs inside the $10 design plan — there it reads your study description, and couples the test choice to your sample-size calculation, variable list, and events-per-variable budget.

Some designs need more than a table: cluster-randomized trials, non-inferiority and equivalence designs, Bayesian and adaptive designs, and any analysis where confounding strategy is the real question. If you tick “clustered,” the tool tells you a naïve test is wrong and a mixed-effects model or GEE is needed — it does not pretend the simple answer still holds. Where a question is beyond the table, the honest output is “this needs a biostatistician,” not a guess.

The test is one line of an analysis plan. The $10 RigorMD design plan builds the rest from a plain-prose description of your study: hypotheses, the variables and confounders to collect, the sample-size arithmetic for your effect size, and a draft IRB statistical-methods page — with every missing input named as a gap rather than silently guessed.