Estimating average treatment effect by model averaging

•We propose to use a model average method to improve the estimation of average treatment effects.•The proposed model average estimator selects weight optimally to minimize estimation mean squared errors.•Simulation results show that the model average estimator exhibits smaller estimation mean squared errors in post-treatment prediction than AIC or AICC methods.

In this paper, we propose to use a model average method to improve the estimation performance of Hsiao et al. (2012) panel data approach for program evaluation. Instead of using the two-step model selection strategy which chooses one best model according to a criterion such as AIC or AICC, we average over a set of candidate models. Simulation results show that the model average estimator exhibits smaller estimation errors in post-treatment prediction than AIC or AICC method.