A0472
Title: Controlling false discovery rate in high-dimensional linear regression: The Gaussian mirror approach
Authors: Xin Xing - Virginia Tech University (United States) [presenting]
Abstract: Identifying key variables that influence outcomes in linear regression models while also maintaining control over the false discovery rate (FDR) poses a significant challenge in statistical analysis. The aim is to introduce the Gaussian Mirror (GM) method, a novel approach for identifying crucial variables in linear regression models while controlling the false discovery rate (FDR). By creating a pair of mirror variables for each predictor using Gaussian perturbations, the GM method improves variable selection. It's compatible with standard regression techniques like ordinary least squares and Lasso and offers the flexibility of applying mirror variables pre or post-selection. The key advancement of the GM method lies in its capacity to generate test statistics that effectively maintain the FDR at a predefined level under realistic covariate dependence assumptions. The analysis showcases the GM method's superiority in managing FDR constraints, especially in situations with high covariate correlation and a dense array of influential variables. The GM method's innovative approach is covered, theoretical foundation, and empirical efficacy, offering attendees valuable insights into tackling complex statistical challenges.
