Estimating Average Treatment Effects with Continuous and Discrete Covariates: The Case of Swan-Ganz Catheterization
Qi Li,J. Racine,J. Wooldridge
DOI: https://doi.org/10.1257/AER.98.2.357
2008-04-01
Abstract:In recent years, propensity score methods have become popular for estimating average treatment effects across a variety of disciplines. The propensity score is simply the probability, conditional on a set of covariates, of being in the treatment group. The two most common estimation methods using the propensity score are matching based on the propensity score—as described in Paul R. Rosenbaum and Donald B. Rubin (1983) and James J. Heckman, Hidehiko Ichimura, and Petra Todd (1997) (HIT (1997) for short)—and propensity score weighting (PSW)—as described, for example, in Wooldridge (2002) and Guido W. Imbens (2004). Matching based on estimated propensity scores, while often used, has both practical and theoretical shortcomings. Inference is difficult, at least partly because bootstrapping methods cannot be justified. Further, propensity score matching is not known to be asymptotically efficient. Recently, estimators based on propensity score weighting, where the propensity scores are estimated nonparametrically, have been shown to achieve the lowest possible asymptotic variance. Further, bootstrapping can be justified for inference. Propensity score weighting dates back to D. G. Horvitz and D. J. Thomson (1952), who showed that, under the assumption of ignorable sample selection, inverse probability weighting can be used to obtain unbiased estimators of population moments. Wooldridge (2007) contains a unified framework for inverse probability weighting and describes many applications. Estimating Average Treatment Effects with Continuous and Discrete Covariates: The Case of Swan-Ganz Catheterization
Economics