Title:The Role of the Propensity Score in Fixed Effect Models

Abstract:We develop a new approach for estimating average treatment effects in observational studies with unobserved group-level heterogeneity. We consider a general model with group-level unconfoundedness and provide conditions under which aggregate statistics -- group-level averages of functions of treatments and covariates -- are sufficient to eliminate differences between groups. Building on these results, we generalize commonly used linear fixed-effect regression estimators in three ways. First, we allow researchers to explicitly select sufficient statistics themselves, whereas the standard specifications make this choice implicit. Second, we suggest using flexible adjustments for sufficient statistics. Finally, we propose robustifying the regression estimators using inverse propensity score weighting. In practice, we recommend researchers use all three modifications.