Define Analysis Populations Before Building Tables

Begin with the question each analysis answers

A label such as “evaluable” does not specify who belongs in an analysis. Write the intended comparison, outcome, time point, and population before creating a dataset. State whether participants are grouped by randomized assignment or another prespecified rule, and distinguish membership from the method used to address missing outcomes.

The ICH E9 statistical principles discuss full analysis and per-protocol sets, the importance of preserving randomization, and documenting exclusions. A dataset containing only complete observed records should not be called intention-to-treat merely because its remaining participants retain their assigned groups.

Turn labels into explicit rules

Create a population specification with a stable name, inclusion conditions, exclusion conditions, grouping rule, and source fields. For each condition, explain how unknown or conflicting information is handled. A rule requiring treatment exposure needs a defined exposure field; a rule requiring an outcome needs an explicit outcome and time window.

Avoid one general “exclude” flag applied to every table. A participant may meet the prespecified definition for one analysis and not another. Keep the reason, applicable analysis, and decision date beside each membership flag so reviewers can reconstruct the logic.

Test a fictional participant

Imagine participant P014 was randomized to arm A, received no study treatment, and supplied the scheduled primary outcome. In an assignment-based efficacy analysis intended to include all randomized participants, lack of treatment receipt does not by itself justify dropping that person. A separately defined exposed safety population could handle the same participant differently under its specified rules.

This fictional example is a test of definitions, not a universal instruction for every trial. The point is to resolve membership through the planned question and rules rather than through a convenient database filter.

Reconcile decisions before interpretation

Compare population counts with the enrollment and flow records, then inspect reasons for differences by group. Check that each participant has one consistent assignment and that duplicated visits do not create duplicated people. Review unresolved membership before generating final outcome comparisons, with appropriate blinding where feasible.

Record protocol deviations separately from automatic exclusions. A per-protocol analysis can be affected by selection after randomization, so agreement with another analysis does not remove every possible bias. Prespecify and justify relevant sensitivity analyses rather than choosing the population that gives the strongest result.

Finish with a versioned population table, executable rules, reason counts, and a list of unresolved cases. When rules change, retain the earlier version and explain the change. Clear population definitions make a result reproducible and interpretable; they do not by themselves validate the causal assumptions or missing-data method.

Source checked on 31 October 2026; participant scenario is fictional.