There is no study-wide minimum sample size independent of the planned analysis. Calculate or simulate for the primary decision and document every assumption that materially changes the result.
Minimum inputs
- Primary estimand and model. Define the comparison, coefficient, contrast or precision target.
- Effect or precision target. Justify the smallest effect of interest or acceptable confidence-interval width.
- Error criteria. State alpha, desired power and any multiplicity adjustment.
- Design structure. Include allocation ratio, repeated measures, clustering, stratification and covariate adjustment.
- Operational losses. Inflate transparently for attrition, non-response, exclusions or unusable records.
Use ranges when inputs are uncertain
Effect sizes borrowed from small or selective studies can be unstable. Scenario analysis across plausible values is often more informative than one apparently precise number. For complex models, simulation can reflect non-normal outcomes, unequal clusters and model-specific decision rules.
Worked use case
A clustered programme evaluation uses an individually randomised formula. Recalculate using expected cluster size, intraclass correlation, number of clusters, attrition and the planned cluster-adjusted analysis. Report the sensitivity of required sample size to the intraclass correlation.
Evidence boundary
Achieving nominal power does not rescue biased measurement, poor implementation or an unidentified causal contrast. Power addresses a defined probability under assumptions; it is not a general quality score.
Primary sources
- NIH Research Methods Resources. Sample size and power resources. Retrieved 5 August 2026.
- CONSORT. CONSORT reporting guidance. Retrieved 5 August 2026.