Examining medical products' benefits and risks in different population subsets is often necessary for informing public health decisions. In observational cohort studies, safety analyses by pre-specified subgroup can be powered, and are informative about different population subsets' risks if the study designs or analyses adequately control for confounding. However, few guidelines exist on how to simultaneously control for confounding and conduct subgroup analyses. In this simulation study, we evaluated the performance, in terms of bias, efficiency and coverage, of six propensity score methods in 24 scenarios by estimating subgroup-specific hazard ratios of average treatment effect in the treated with Cox regression models. The subgroup analysis methods control for confounding either by propensity score matching or by inverse probability treatment weighting. These methods vary as to whether they subset information or borrow it across subgroups to estimate the propensity score. Simulation scenarios varied by size of subgroup, strength of association of subgroup with exposure, strength of association of subgroup with outcome (simulated survival), and outcome incidence. Results indicated that subsetting the data by the subgrouping variable, to estimate the propensity score and hazard ratio, has the smallest bias, far exceeding any penalty in precision. Moreover, weighting methods pay a heavier price in bias than do matching methods when the propensity score model is misspecified and the subgrouping variable is a strong confounder.
Keywords: Subgroup analyses; cox regression; incident exposure cohort studies; inverse probability of treatment weighting; propensity score matching.