|
Activity Number:
|
241
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, August 8, 2006 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #306840 |
|
Title:
|
Novel Bayesian Variable-Selection Priors for "Large p Small n" Data Analysis
|
|
Author(s):
|
Joseph Lucas*+
|
|
Companies:
|
Duke University
|
|
Address:
|
, Durham, NC, 27708-0251,
|
|
Keywords:
|
variable selection ; sparcity ; microarray ; low sample ; model selection ; high dimensional
|
|
Abstract:
|
Standard Bayesian variable-selection priors for regression coefficients involve mixing a "point mass" at zero with a normal (or other) distribution assuming an unknown mixing proportion, q. We show that the use of the traditional prior for q can lead to over-estimation and significant false-positive bias. The problem is particularly apparent in highly multivariate regression and ANOVA modeling such as arises in the analysis of gene expression experiments. We describe a novel hierarchical extension of the traditional approach involving observation-variable specific indicators of inclusion and which alleviates these issues. Resulting posterior distributions, computed with custom MCMC methods, induce conservative inferences of "significant" effects consistent with the expectation of variable-selection priors. Examples in analysis of micro-array data demonstrate these issues and advances.
|
- The address information is for the authors that have a + after their name.
- Authors who are presenting talks have a * after their name.
Back to the full JSM 2006 program |
