A Semiparametric Approach to Variable Importance with Multiple Testing Corrections

• Activity Number: 292 - Inferential Thinking in a Machine Learning World
• Type: Invited
• Date/Time: Wednesday, August 11, 2021 : 3:30 PM to 5:20 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Aaditya K Ramdas*
• Companies: Carnegie Mellon University
• Keywords: variable importance; conditional independence; FDR; confidence intervals; Model-X; semiparametrics

Abstract

In modern machine learning, it often makes no sense to assume that a model for Y|X is well-specified. But there are some settings in genetics and neuroscience where modeling the covariates X might be reasonable instead. We discuss the MX(2) model, where we know the first two moments of X, and discuss how a prediction-based perspective can neatly interface with classical semiparametric theory to deliver an interpretable notion of variable importance for all variables, along with valid confidence intervals and controlled false discovery rate.