Rank-based Model Average Estimator for Average Treatment Effects

Abstract

Model averaging emerges as an attractive ensemble technique that strategically synthesizes all available information from potential candidate models to enhance estimation accuracy. Despite of having been widely practiced in various areas in biostatistics and economics, its application to causal inference is rather limited. We propose to estimate the average treatment effect (ATE) by forming the weighted combination of the robust rank-based estimators obtained from a series of candidate models, where the model weights, selected via the growing window cross-validation, dynamically adjust the importance of candidate models. The proposed robust rank-based model averaging ({\sf RRMA}) is not only robust to handle potential outliers and heavy-tailed error distributions but also capable to reduce the risk of model mis-specification and enhance the flexibility and the accuracy of estimation. Theoretically, we establish the desirable asymptotic properties of {\sf RRMA}, including the asymptotic optimality, consistency and asymptotic distribution. Moreover, motivated by the asymptotic distribution of {\sf RRMA}, an easy-to-implement confidence interval is also developed for the ATE. Simulation experiments and an empirical application demonstrate the superiority of {\sf RRMA} over existing competitors in estimation accuracy, coverage probabilities and widths of confidence intervals.