Subgroup analysis (forest plots)
Splitting a trial's patients into groups (by age, sex, biomarker, region) to see whether the treatment worked the same in each. Shown as a forest plot. Genuine differences are rare and most striking subgroup results are noise, so they need confirmation.
Trials are powered for the whole population, so subgroups have wide confidence intervals and, with a dozen subgroups, one will look different by chance; the proper test is for interaction, not for significance within each subgroup. Regulators have nonetheless used subgroups to restrict labels (PD-L1 cut-offs in gastro-oesophageal cancer, IMpassion130's PD-L1-positive population) and to question generalisability (regional differences in STARGLO, ex-US benefit in IMbrave050). Biologically plausible, pre-specified subgroups with a stratified randomisation carry more weight; a post-hoc TP53 wild-type subgroup with no interaction test (selinexor in endometrial cancer) needed its own trial.