A Bayesian Linear Mixed Model for Bi-Level Feature Selection

• Activity Number: 424 - Priors and Model Specifications for Variable and Feature Selection
• Type: Contributed
• Date/Time: Wednesday, August 10, 2022 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Daniel Baer* and Andrew Booth Lawson and Yeonhee Park and Sharon Xie and Andreana Benitez and The ADNI
• Companies: University of Pennsylvania and Medical University of South Carolina and University of Wisconsin-Madison and University of Pennsylvania and Medical University of South Carolina and The Alzheimer's Disease Neuroimaging Initiative
• Keywords: Feature selection; Bayesian; Mixture prior distribution; longitudinal data analysis; Linear mixed model; Markov chain Monte Carlo

Abstract

Bayesian feature selection models allow for the identification of feature data that are associated with outcome data. Current Bayesian feature selection models are limited by the complexities of data arising in the observational study of neurological disease. In particular, there are no Bayesian feature selection models that can simultaneously account for irregularly-spaced longitudinal outcome data, account for feature data group structure, and specify time-varying feature parameters. We therefore propose a Bayesian linear mixed model for bi-level feature selection which addresses these limitations. Our model is fit via a Gibbs sampling Markov chain Monte Carlo algorithm, and we disseminate our model via an R software package. We evaluated our model via a simulation study and an analysis of longitudinal outcome data from a study of neurological disease. We found that our model is advantageous in terms of providing improved precision of feature parameter estimates and allowing us to identify subjects with the greatest propensity for neurological disease risk.