Variable Selection via a Multi-Stage Strategy

• Activity Number: 22
• Type: Topic Contributed
• Date/Time: Sunday, July 31, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Jing Chang*+ and Herbert Lee
• Companies: University of California at Santa Cruz and University of California at Santa Cruz
• Keywords: variable selection ; computer experiment ; BART the sum of trees model ; adaptive sampling ; sensitivity analysis

Abstract

Due to the rapid increases in computer power, scientists are able to do discovery and engineering analysis with the help of large scale computer simulations. Variable selection is a crucial step in the early stages of a computer experiment. However, there is no existing method which can successfully reduce the total number of variables from several hundreds to a few (< 10) truly significant ones, when we are faced with potential non-linear relationships and limited sample sizes. This challenging setting frequently occurs in simulation designs for complicated scientific phenomena. In this paper, a multi-stage strategy is developed which incorporates a state-of-the-art technique in each stage. By making the best use of the property of each technique, in combination they can achieve a sophisticated goal that can not be achieved by any single method. We combine an extension of the BART sum of trees model with adaptive sampling techniques and sensitivity analysis to select variables in a highly precise manner.