Covariate Balancing Propensity Score for General Treatment Regimes

Propensity score matching and inverse-probability weighting are popular methods for causal inference in observational studies. Under the assumption of unconfoundedness, these methods enable researchers to estimate causal effects by balancing observed covariates across different treatment values. While their extensions to general treatment regimes exist, a vast majority of applications have been confined to a binary treatment. Moreover, applied researchers often dichotomize a non-binary treatment in order to utilize propensity score methods. Balancing covariates with respect to the dichotomized treatment, however, does not imply that they are balanced regarding the original non-binary treatment variable. In this paper, we extend the covariate balancing propensity score (CBPS) methodology of Imai and Ratkovic (2014) to general treatment regimes. Specifically, we estimate the generalized propensity score such that the resulting association between a treatment and covariates is minimized. Two social science applications are used to demonstrate that the CBPS methodology significantly improves covariate balance and offer substantive insights the original analyses fail to identify. The proposed methodology is implemented through publicly available open-source software.