Clinical heterogeneity and average treatment effects: consequences for superiority and non-inferiority trials

Clinical heterogeneity is common in acute infections, but differences in prognosis, phenotype, or outcomes do not imply different comparative treatment responses. Randomized trials estimate population-average contrasts, whereas bedside decisions concern individual patients characterized before treatment.

To distinguish routes to decision-relevant heterogeneity and their consequences in superiority and active-control non-inferiority trials.

A narrative search of PubMed and regulatory guidance addressing treatment-effect heterogeneity, risk and effect modelling, estimands, transportability, non-inferiority margins, constancy, and assay sensitivity.

Three related routes make heterogeneity relevant to decisions. Baseline prognosis can change absolute benefit despite a constant relative effect; biological or exposure-related modifiers can change comparative response; and population composition reweights the distributions underlying these two routes across settings. Each requires an explicit effect scale. In superiority trials, heterogeneity affects power, interpretation, and transportability without invalidating randomization. Active-control non-inferiority trials also depend on a historical active-control effect not observed in the trial; population differences may therefore threaten constancy, assay sensitivity, and the margin. Risk-based analyses should report outcomes under both strategies and absolute and relative effects across prespecified multivariable risk strata. Proposed modifiers require treatment-by-covariate contrasts, support, and validation. Conditional-average and clustering methods do not identify individual effects or, by themselves, establish treatment-responsive phenotypes.

Heterogeneity should be addressed prospectively by aligning the clinical decision with a clearly specified treatment question, population, effect scale, and decision threshold. Reporting pre-trial practice clarifies what the control strategy represents. A heterogeneous effect should guide treatment only when available before the decision, reproducible, and large enough to matter.