Abstract

Difference-in-differences, one of the most commonly used research designs in empirical analyses of program evaluation, infers the ex post counterfactual of what would have happened, but does not inform ex ante policy choice of what we should do. This paper develops a framework of learning ex ante policy allocation in settings where available data are panel observational data with treatments allocated at group level. Assuming that the common trend of the potential outcomes conditional on individual's characteristics holds in the data, we study how to estimate an optimal treatment allocation policy for a group of individuals who are not yet exposed to the treatment. To allow for an environment in which the distribution of potential outcomes can differ between the periods that the data were obtained and the period that the policy will be implemented, we introduce an intermediate welfare criterion that can inform an optimal policy for the target group. We construct a sample analogue of the intermediate welfare using the available data and estimate an optimal policy for the target population by maximizing it.