Causal inference using generalized empirical likelihood methods

Abstract

In this paper, we propose a one step method for estimating the average treatment effect, when the assignment to treatment is not random. We use a misspecified generalized empirical likelihood setup in which we constrain the sample to be balanced. We show that the implied probabilities that we obtain play a similar role as the weights from the weighting methods based on the propensity score. In Monte Carlo simulations, we show that GEL dominates many existing methods in terms of bias and root mean squared errors. We then aply our method to the training program studied by Lalonde (1986).