Events and Seminars : Upcoming Seminars

VARIABLE SELECTION AND MACHINE LEARNING METHODS FOR CAUSAL INFERENCE PROBLEMS

DEBASHIS GHOSH, PH.D.
Professor and Chair
Department of Biostatistics and Informatics
Colorado School of Public Health
University of Colorado
TUESDAY, OCTOBER 20, 2015
11:00 a.m.–12:00 p.m., CRB 692

In many observational data settings, there has been tremendous recent interest in trying to estimate causal effects. Two natural questions that arise involve (1) how to conduct model selection in this setting and (2) how to adjust for confounders. What makes the first question nonstandard relative to regression modellingis that typically two models, a mean outcome model and a propensity score model, are fit to the data, and these models play different roles in the causal modeling process. In this talk, we will describe two methodologies for addressing question (1). The first will be based on what we term a “predictive LASSO”, while the second approach will entail using L1-penalized regression in order to identify necessary conditional independencies for valid causal inference to hold. For question (2), we will develop a kernel machine-based approach to confounder adjustment that reveals the utility of probability metrics as a tool for evaluating covariate balance in causal inference problems. This is joint work with YeyingZhu at the University of Waterloo.