Bayesian Additive Regression Trees for Variable Selection in Biological Data

• Activity Number: 230
• Type: Topic Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Shane T. Jensen*+ and Justin Bleich and Adam Kapelner and Edward George
• Companies: The Wharton School, University of Pennsylvania and Wharton School, UPenn and Statistics Department, The Wharton School at University of Pennsylvania and Wharton University of Pennsylvania
• Keywords: variable selection ; biological data ; trees ; regression

Abstract

There is a crucial need for effective variable selection procedures in high dimensional biological data, where it is difficult to subtle individual effects and interactions between factors. Bayesian Additive Regression Trees are a promising alternative to more parametric regression approaches, such as the lasso or Bayesian latent indicator models. BART constructs an ensemble of decision trees from the set of possible genomic factors underlying a biological outcome. We develop principled methodology that adapts BART to variable selection as well as incorporating additional data as prior information. We evaluate the performance of our BART-based approach in simulation settings as well as an application to the gene regulatory network in yeast (Saccharomyces cerevisiae).