Learning Overlapping Representations for the Estimation of\n Individualized Treatment Effects

Abstract

The choice of making an intervention depends on its potential benefit or harm\nin comparison to alternatives. Estimating the likely outcome of alternatives\nfrom observational data is a challenging problem as all outcomes are never\nobserved, and selection bias precludes the direct comparison of differently\nintervened groups. Despite their empirical success, we show that algorithms\nthat learn domain-invariant representations of inputs (on which to make\npredictions) are often inappropriate, and develop generalization bounds that\ndemonstrate the dependence on domain overlap and highlight the need for\ninvertible latent maps. Based on these results, we develop a deep kernel\nregression algorithm and posterior regularization framework that substantially\noutperforms the state-of-the-art on a variety of benchmarks data sets.\n