Abstract:
|
Covariate balancing methods have recently been studied extensively for the estimation of average treatment effects using observational data. Those methods create weights that balance the pre-treatment covariate distribution between treated and controls, but cannot be directly extended to mediation analysis because mediators are post-treatment variables. We introduce titled covariate and mediator balancing for estimating natural mediation effects, and consider both parametric and nonparametric balancing methods. Robustness and efficiency of the methods are presented.
|