|
Activity Number:
|
461
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Wednesday, August 1, 2007 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #308795 |
|
Title:
|
An Extended BIC for Model Selection
|
|
Author(s):
|
Surajit Ray*+ and James Berger and Susie Bayarri and Woncheol Jang and Luis R. Pericchi
|
|
Companies:
|
Boston University and Duke University and University of Valencia and University of Georgia and University of Puerto Rico, San Juan
|
|
Address:
|
111 Cummington St, Boston, MA, 02215,
|
|
Keywords:
|
Bayes factor ; Cauchy Priors ; Consistency ; Model Selection ; Effective sample size ; Fisher Information
|
|
Abstract:
|
We present a new approach to Bayes factors based on Laplace expansions (as BIC) which we call EBIC (Extended BIC). In our approach, we do not include the prior in the Laplace expansion, but choose it appropriately so that it produces close-form expressions for the resulting EBIC. We explore both joint priors and independent priors for the component parameters. To help choose the scale of the prior, we use a novel definition of effective sample size which allows for different effective sample sizes for the parameters. The new EBIC avoids many of the difficulties commonly associated with BIC, and can often be shown to be consistent. We also produce a modified EBIC which is more favorable to complex models while still retaining consistency. A comparative study of a range of model selection tools including the EBIC will be presented during the talk.
|
