Matching using Semiparametric Propensity Scores

Propensity score matching is becoming increasingly common in clinical medicine, demographic and economic research for the evaluation of the magnitude of treatment effects. Existing studies generally use parametric estimators of binary response models such as the probit and logit to estimate the propensity score, which imposes strong distributional assumptions on the error term that are often violated with the underlying data. This paper considers matching using semiparametrically estimated propensity scores. Our approach allows for heterogeneity in response across observed covariates along the conditional willingness to participate in the treatment intervention distribution. Data from the NSW experiment, CPS and PSID are used to evaluate the performance of alternative matching estimators. Preliminary estimates indicate mean absolute bias error reductions between 6.2% and 706% of the experimental treatment impact with stratification matching using semiparametric propensity score estimates relative to matching algorithms that employ parametric propensity scores