East Coast Ballroom
Generalizability of subgroup effects (307923)
Hwanhee Hong, Duke University Medical Center*Marissa J Seamans, UCLA Fielding School of Public Health
Elizabeth Stuart, Johns Hopkins Bloomberg School of Public Health
Keywords: generalizability, causal inference, randomized trials, treatment effect heterogeneity
Generalizability methods are increasingly used to make population inferences of treatment effects using results from randomized trials. These methods rely on the implicit assumption that effects within subgroups (e.g., men) transport from the trial to the target population. However, there may be concerns that effects within subgroups are not constant; for example, men in the trial differ from the men in the target population by an unmeasured factor (e.g., smoking status) that modifies the treatment effect. When there is unmeasured treatment effect heterogeneity within subgroups, trial results may not generalize well to the target population. In this work, we present the bias formula and use simulations to show the potential for bias due to unmeasured differences between subgroups in the trial and target population. Finally, we present a data example using trial data from a study of lifestyle interventions for blood pressure and target population data from a cross-sectional epidemiologic study. Our goal is to provide guidance to policymakers when transporting results from trials with participant subgroups that differ from the populations in which the interventions may be applied.