The Statistics Seminar speaker for Wednesday, Oct. 19, 2016, will be Stefan Wager. Dr. Wager is currently a postdoctoral researcher in statistics at Columbia University, and will start as an assistant professor at Stanford Graduate School of Business in the fall of 2017. He completed his PhD in statistics at Stanford University in 2016, under the supervision of Brad Efron and Guenther Walther. Stefan's research focuses on adapting ideas from machine learning to statistical problems that arise in scientific applications.
Title: Solving Heterogeneous Estimating Equations with Gradient Forests
Abstract: There has been a recent surge of interest in using forest-based methods for a wide variety of statistical tasks, including causal inference, survival analysis, and quantile regression. Extending forest-based methods to new statistical settings requires specifying new tree-splitting rules that are targeted to the task at hand; and ad-hoc design of such splitting rules can require considerable effort. In this talk, I will discuss a unified framework for the design of fast splitting rules targeted to arbitrary heterogeneous estimating equations. The resulting "gradient forests" reduce to recursively applying a pre-processing step where we label each observation with gradient-based pseudo-outcomes, followed by a regression step that runs a standard CART regression split on these pseudo-outcomes. We use our framework to develop new algorithms for two important statistical problems, non-parametric quantile regression and heterogeneous treatment effect estimation via instrumental variables, and show that the resulting procedures considerably outperform baseline forests whose splitting rules do not take into account the statistical question at hand. Finally, we prove consistency of gradient forests, and establish a central limit theorem. Our method will be available as an R-package, gradientForest, built on top of the ranger package for random forests.
Refreshments will be served following the seminar in 1181 Comstock Hall.