Keywords: heterogeneity of treatment effect, subgroup analysis, personalized medicine, benefit-harm trade-off, clinical trials
Subgroup analysis is a frequently used tool for evaluating heterogeneity of treatment effect and heterogeneity in treatment harm across observed baseline patient characteristics. While treatment efficacy and adverse event measures are often reported separately for each subgroup, analyzing their within-subgroup joint distribution is critical for better informed patient decision-making. In this talk, we describe Bayesian models for performing a subgroup analysis to compare the joint occurrence of a primary endpoint and an adverse event between two treatment arms. Our approaches emphasize estimation of heterogeneity in this joint distribution across subgroups, and our approaches directly accommodate subgroups with small numbers of observed primary and adverse event combinations. In addition, we describe several ways in which our models may be used to generate interpretable summary measures of benefit-risk trade-offs for each subgroup. We demonstrate how our models may be used to assess heterogeneity with respect to either a composite outcome which weights adverse event-free and adverse event-occurring survival differently, and how our model may be used to evaluate heterogeneity with respect to a treatment effect based on the probability of achieving an improved outcome. We illustrate the methods using a large cardiovascular trial ($N = 9,361$) investigating the efficacy of an intervention for reducing systolic blood pressure to a lower-than-usual target.