Post Hoc Analysis
A post hoc analysis is one conceived after the data were collected or unblinded, which makes it hypothesis-generating rather than confirmatory however plausible its result appears.
Post hoc means after the fact, covering any comparison, subgroup, endpoint or analysis population not written into the protocol and statistical analysis plan before unblinding. The distinction is procedural rather than statistical: the same calculation on the same numbers is confirmatory if specified in advance and exploratory if not, because only then does the reported error rate describe the procedure that produced the result. Prespecified secondary endpoints sit between the two, planned but usually outside the alpha-protecting hierarchy.
Regulators act on the distinction: labelling claims rest on prespecified primary endpoints, and post hoc findings are treated in review documents as supportive at best. The pattern that vindicates the discipline is a post hoc or meta-analytic signal that survives a trial built to test it: cardiovascular signals in pooled early diabetes data were examined prospectively in dedicated outcome trials such as LEADER for liraglutide. Plenty of comparable signals did not survive that step.
Post hoc analysis is legitimate and often necessary. It is where safety signals surface and where the next trial's population is chosen. What it cannot do is convert a negative trial into a positive one. A result read off the data can generate a hypothesis or test one, never both, and an honest report says which analyses were planned and how many were run.
The telling phrase is a trial that missed its primary endpoint but found a significant benefit in a defined group; that sentence describes a failed trial plus an unplanned search. Two variants are worth spotting: retrospective relabelling, where an analysis appears in the paper as prespecified with no registry record behind it, and a per-protocol population invented once dropouts were known, discarding randomisation where it mattered.