A Practical Review of Variable Selection Methods

• Activity Number: 289
• Type: Topic Contributed
• Date/Time: Tuesday, August 3, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Zhen Zhang*+
• Companies: Abbott Vascular
• Keywords: Variable selection ; Correlation ; Stepwise ; Bootstrap ; LASSO

Abstract

Variable selection is of great interest in medical research, be it identifying important factors or predicting future events. PROSPECT is a prospective natural history study, where 700 patients treated for Acute Coronary Syndrome (ACS) were followed for over 3 years for Major Adverse Cardiac Events (MACE). Huge amount of factors are available for each patient and the need for variable selection is obvious when one is examining the relationship between these measurements and MACE. We review statistical variable selection in general and several variable selection methods, including stepwise and subset selection, bootstrap model selection, penalized regression (e.g., LASSO), etc, for different model settings. We evaluate the performances of various methods by both simulations and application to the PROSPECT data, followed by suggestions for practical variable selection.