Title: A bivariate treed linear model for causal inference from observational studies
Authors: Richard Hahn - University of Chicago (United States) [presenting]
Carlos Carvalho - The University of Texas at Austin (United States)
Abstract: The aim is to explain how regularized regressions (linear and nonlinear) can yield poor estimators of treatment effects due to a phenomenon we call ``regularization induced confounding''. How to overcome this problem will be explained, starting first with the linear model and then extending the approach to the case of nonlinear, heterogeneous treatment effects using treed linear models. We show that our approach dramatically outperforms common alternatives across a range of plausible data generating processes.