- Published
- Updated
Labels such as full analysis set, safety set or per-protocol set do not themselves identify included participants. Inclusion, exclusion and treatment-group assignment rules need to be explicit, justified and reproducible for each analytical purpose.
Define the question and assumptions
Define the objective, required information, entry condition and group classification for every set. Analysis by randomised assignment and analysis by treatment received may contribute to different questions. Rules for protocol deviations and missing measurements should be clear enough to apply without selecting decisions according to observed treatment outcomes.
Hypothetical study example
In a hypothetical trial, one randomised participant receives no administration while another receives the wrong product. The team does not impose one classification across every table. Each decision follows the scientific purpose and prespecified treatment-assignment rule for that particular analysis population and is recorded transparently.
Design review checks
- Derive participant-level inclusion flags for every specified analysis population.
- Check exclusion reasons for duplicate, inconsistent or contradictory classification decisions.
- Reconcile participant-flow counts with datasets and the corresponding analysis tables.
Record the decision and its limits
Acceptance means an independent programmer can reproduce the participant list using the same rules. Record blinded-review decisions, subsequent changes and their timing. A familiar population label cannot justify unexplained exclusions or selective reporting. The narrative should make clear which population addresses which question so readers can assess the estimate appropriately.
Frequently asked questions
- Must every analysis use the same population?
- Different objectives may require different populations, with differences explained explicitly.
- Is the population label a sufficient definition?
- No. Operational inclusion, exclusion and group-assignment rules are required.