Calibrated Power Model Selection

• Activity Number: 81
• Type: Contributed
• Date/Time: Sunday, August 1, 2010 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Hugh Crews and Clayton Barker*+
• Companies: University of North Carolina Wilmington and SAS Institute
• Keywords: False selection rate ; Model selection ; Practical significance ; Variable screening ; Power

Abstract

Recent developments in variable selection have focused on controlling false selection of uninformative variables and maximizing predictive ability. In some situations, such as screening studies, the goal is to select the most important variables for further research. We develop a general method for model selection that achieves a specified power to detect these important variables. By repeatedly adding phony informative variables to the data, we gain insight into the tuning parameter needed to select practically significant variables with a desired power. The new model selection technique is general enough to work with virtually any regression problem.