Annals of epidemiology
-
Annals of epidemiology · Apr 2013
Confounding control in a nonexperimental study of STAR*D data: logistic regression balanced covariates better than boosted CART.
Propensity scores (PSs), a powerful bias-reduction tool, can balance treatment groups on measured covariates in nonexperimental studies. We demonstrate the use of multiple PS estimation methods to optimize covariate balance. ⋯ Comparing multiple PS estimates is a pragmatic way to optimize balance. Logistic regression remains valuable for this purpose. Simulation studies are needed to compare PS models under varying conditions. Such studies should consider more flexible estimation methods, such as logistic models with automated selection of interactions or hybrid models using main effects logistic regression instead of a constant log-odds as the initial model for BCART.