|
Abstract:
|
The estimation of heterogeneous treatment effects in pre-specified subgroups is uniquely challenging in observational data, where methods such as propensity score weighting must achieve covariate balance both overall and within subgroups. Propensity score estimation via machine learning methods can improve subgroup-specific mean balance. The use of overlap weights further reduces bias and variability of subgroup treatment effect estimation in linear models. However, these findings have not been evaluated for survival outcomes. Here, we use simulation studies to evaluate various propensity score estimation methods (logistic regression, random forests, LASSO, or GBM) with different weighting methods (IPW or OW) for estimating the subgroup marginal hazard ratio and subgroup restricted average causal effect. We vary the effect of covariates on treatment-selection process, the prevalence of treatment and underlying proportional hazards assumption. Performance metrics include relative bias, root mean square error, and empirical coverage rate. Our results suggest that LASSO paired with overlap weighting may reduce bias and improve variance in subgroup survival analysis.
|
