Two-Layer Propensity Score Weighting: Multiple Treatments, Heterogeneous Treatment Effects, and a Causality Fallback Position

• Activity Number: 40
• Type: Contributed
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Amelia Haviland*+ and David Choi
• Companies: Carnegie Mellon University and Carnegie Mellon University
• Keywords: propensity score ; heterogeneous treatment effects ; causal inference ; multiple treatments

Abstract

In a study of the impacts of specific health plans on costs, we had multiple treatment groups of interest and concern about selection bias/heterogeneous treatment effects. The goal was estimation of treatment on the treated effects for each treatment but on a constant 'treated' population. We performed 2-layer propensity score (PS) weighting to address this goal. First we used PS to weight the control group to balance with the covariate distribution of the pooled treatment group. Second we used PS from multiple models to weight each treatment type to balance with the covariate distribution of the pooled treatment group. In this setting, the assumption of no unobserved confounders is less likely to hold for the first layer of weighting than the second. The double-layer then provides a built in causality fallback position: while the overall treatment size effect may be biased it is considerably less likely that the differences in treatment effects between treatment types are biased. In addition, in the context of potentially heterogeneous treatment effects, the differences in effect by treatment type are more likely to be due to treatment type than to selection by treatment type.