Bayesian Methods for Variable Selection in Generalized/Nonlinear Models

点击翻译按钮获取中文摘要

The current proposal builds upon the P.I.'s experitize in variable selection and summarizes her current and future directions in the development of Bayesian methodologies. In particular, the P.I. plans to consider extensions to generalized models and to models that allow arbitrary nonlinear associations of a set of variables to a response. Inferential strategies for the proposed models are more challenging than the typical linear settings addressed by the P.I. in previous work. In addition, nonparametric priors will be investigated with the purpose of sharpening the selection and relaxing distributional assumptions of the models, such as those on random effects and on the error terms, often questionable in real-data applications.Methodologies developed through this project carry potential for significant impact in statistics and in applied fields in which high-dimension/low-sample-size data arise. Applications to data arising from interdisciplinary collaborations will demonstrate the practical usefulness of the proposed methods. In particular, the P.I. plans to build upon her recent interest in brain imaging data by investigating applications and extensions of Bayesian methods for variable selection to generalized linear and mixed models currently used for the analysis of such data. The broader impacts of this proposal are in its educational and training objectives, in its efforts to disseminate results and in the collaborative nature of the proposed research. The P.I. maintains an updated webpage on her research activities where papers and accompanying software are posted in a timely manner.