Reducing Bias in Treatment Effect Estimation in Observational Studies Suffering from Missing Data

Matching based on estimated propensity scores (that is, the estimated conditional probability of being treated) has become an increasingly popular technique for causal inference over the past decade. By balancing observed covariates, propensity score methods reduce the risk of confounding causal processes. Estimation of propensity scores in the complete data case is generally straightforward since it uses standard methods (e.g. logistic regression or discriminant analysis) and relies on diagnostics that are relatively easy to calculate and interpret. Most studies, however, have missing data. This paper illustrates a principled approach to handling missing data when estimating propensity scores makes use of multiple imputation (MI). Placing the problem within the framework of the Rubin Causal Model makes the assumptions explicit by illustrating the interaction between the treatment assignment mechanism and the missing data mechanism. Several approaches for estimating propensity scores with incomplete data using MI are presented. Results demonstrating improved efficacy compared with existing methodology are discussed. These advantages include greater bias reduction and increased facility in model choice