Bias and bias amplification in treatment effect estimates based on misspecified propensity score models

Nonparametric propensity score methods are being used more frequently in social research to address misspecification bias in parametric methods such as linear regression. However, these methods can still be vulnerable to misspecification if the propensity score is estimated using parametric methods. Various methodological innovations have been developed to reduce or eliminate misspecification bias in propensity score methods; however, they are underused in sociological research. A comprehensive Monte Carlo simulation study was conducted to evaluate the performance of these innovations compared with that of standard methods. The results show that while some of the more recently developed extensions or alternatives to propensity score methods can substantially reduce misspecification bias, some are biased even in the absence of misspecification. In addition, most methods are subject to bias amplification due to "hidden dual misspecification", a problem previously overlooked in methodological research. Among the estimators evaluated, entropy balancing was the most successful in both eliminating regular misspecification bias and reducing bias amplification. The covariate balancing propensity score and augmented inverse probability weighting performed well. This study concludes that these estimators deserve more attention in applied social research.