Nonparametric Bayes Approaches for High-Dimensional Data in Biomedical Applications

• Activity Number: 522
• Type: Topic Contributed
• Date/Time: Wednesday, August 7, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): David Kessler*+ and David B. Dunson
• Companies: The University of North Carolina, Chapel Hill and Duke University
• Keywords: nonparametric bayes ; high-dimensional ; feature selection

Abstract

Biomedical investigators must frequently describe the relationship between a response and a large set of features. Those features important in the description of the response are not usually known in advance. In addition, many studies of disease etiology rely on data gathered from related individuals, further obscuring the relationship between certain endogenous features and the response. Finally, the conditional probability of extreme events or the description of quantiles may be of greater interest than the conditional mean. In this paper, we develop and demonstrate nonparametric Bayes approaches to these complex settings and questions. Our approach includes separate stages for the identification of promising features, the selection of a feature set, and the development of a predictive model. The underlying model is based upon a tensor factorization of feature-dependent weights; this flexible specification accommodates complex interactions between features and facilitates the investigation of behavior beyond the conditional mean. We illustrate the method through simulated data sets and real data sets from toxicology and quantitative trait loci studies.