A Comparative Study of Variable Screening Methods: Univariate vs. Multivariate Screening

• Activity Number: 189
• Type: Invited
• Date/Time: Monday, August 2, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Cong Liu*+ and Tao Shi and Yoonkyung Lee
• Companies: The Ohio State University and The Ohio State University and The Ohio State University
• Keywords: LASSO ; variable screening ; correlation method

Abstract

We consider the problem of screening variables for regression where a large number of variables are given as potential predictors of a response of interest. To examine the relative merits and drawbacks of screening methods, we compare univariate screening on the basis of correlation between each variable and the response with multivariate screening via a penalized least squares method.

A comprehensive simulation study is carried out for comparison of the two approaches under various settings. We vary several factors which may affect their performance in the study. Results are summarized by the ROC (Receiver Operating Characteristic) analysis. We highlight the situations where the performance of the two approaches differs, offering a guideline for proper choice of a method in practice. We also draw some connection to theoretical results regarding the penalized least square method.