Sensitivity of Various Estimators of Average Treatment Effects to Double Misspecification

This paper examines various estimators of average treatment effects (ATE) and their sensitivity to double model misspecification under the assumption of unconfoundedness. The consistency of these estimators of ATE is generally associated with the specification of two models: one model for the potential outcome and the other one for the probability of being assigned to the treatment group (i.e., the propensity score). It is unknown which estimator of ATE performs systematically better when the model for the outcome and the model for the propensity score are both misspecified. In this study, we compare the six estimators of ATE under double misspecification in a simulation study and find which estimators yield the smallest errors. We find that nearest neighbor matching performs systematically better, while augmented inverse probability weighting yields the largest errors, and the two doubly robust estimators fail to improve the single regression adjustment approach. We also examine how different levels of overlap affect the biases in estimates and how sensitive the estimators are to additional functional forms in the model.Bachelor of Art