|
Activity Number:
|
141
|
|
Type:
|
Invited
|
|
Date/Time:
|
Monday, August 3, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #302986 |
|
Title:
|
Orthant-Normal Shrinkage Priors in Regression
|
|
Author(s):
|
Christopher Hans*+
|
|
Companies:
|
The Ohio State University
|
|
Address:
|
Department of Statistics, 1958 Neil Ave, Columbus, OH, 43210-1247,
|
|
Keywords:
|
Elastic net ; lasso ; MCMC ; model uncertainty ; regularization ; sparsity
|
|
Abstract:
|
This talk introduces a class of regression shrinkage priors that is based on collections of restricted normal distributions. These orthant-normal prior distributions are shown to give rise to Bayesian models whose posterior modes correspond to several popular regularization procedures including the lasso and the elastic net. Special attention is given to the Bayesian elastic net model: a complete characterization of the posterior distribution is introduced and computational approaches for posterior sampling are provided, allowing for inference that moves beyond simple use of the posterior mode. Bayesian predictive inference based on the posterior predictive distribution is compared with predictions based on the usual elastic net procedure. Variable selection, one of the motivations behind the elastic net, is considered from a Bayesian perspective.
|
