Robust Estimation for Average Treatment Effects

We study the probability tail properties of the Inverse Probability Weighting (IPW) estimators of the Average Treatment Effect T when there is limited overlap in the covariate distribution. Our main contribution is a new robust estimator that performs substantially better than existing IPW estimators. In the literature either the propensity score is assumed bounded away from 0 and 1, or a fixed or shrinking sample portion of the random variable Z that identifies the average treatment effect by E[Z] = T is trimmed when covariate values are large. In a general setting we propose an asymptotically normal estimator that negligibly trims Z adaptively by its large values which sidesteps dimensionality, bias and poor correspondence properties associated with trimming by the covariates, and provides a simple solution to the typically ad hoc choice of trimming threshold. The estimator is asymptotically normal and unbiased whether there is limited overlap or not. In the event there is only one covariate, we also propose an improved robust IPW estimator that trims when the covariate is large. We then work within a latent variable model of the treatment assignment and characterize the probability tail decay of Z. We show when Z exhibits power law tail decay due to limited overlap, and when it has an infinite variance in which case existing estimators do not necessarily have a Gaussian distribution limit. We demonstrate the tail decay property of Z, and study the tail-trimmed estimators by Monte Carlo experiments. We show that our estimator has lower bias and mean-squared-error, and is closer to normal than an existing robust IPW estimator in its suggested form, and in the improved form we propose here