A0334
Title: Policy evaluation with many outcomes
Authors: Maria Nareklishvili - Stanford University (United States) [presenting]
Abstract: Probabilistic causal inference is proposed for observational data. The test statistics are designed to discover treatment effects in the presence of heterogeneous effects, noisy outcomes, heavy tails, or complex nonlinear relationships among outcomes. The key idea is to predict treatment status using all outcome variables jointly or individually under an appropriate balancing mechanism or residualization scheme, and test whether the distribution of predicted treatment scores differs significantly between treated and control groups. Simulations and empirical results demonstrate that the method is highly powerful and outperforms the Wald test in settings with small samples, heterogeneous treatment effects, or extreme values.
