|
Activity Number:
|
163
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 3, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistical Learning and Data Mining
|
| Abstract - #303696 |
|
Title:
|
Automatic Variable Shrinkage and Selection via Confidence Regions
|
|
Author(s):
|
Funda Gunes*+ and Howard D. Bondell
|
|
Companies:
|
North Carolina State University and North Carolina State University
|
|
Address:
|
2501 Founders Drive, Statistics Department, Raleigh, NC, 27695-8203,
|
|
Keywords:
|
Variable Shrinkage and Selection ; Confidence Regions ; Tuning ; Sparse Solution ; Lasso
|
|
Abstract:
|
We propose a new approach to variable shrinkage and selection in regression models based on confidence regions. The approach finds a point within a given confidence region closest to the origin with respect to some measure. Tuning regression coefficients using confidence regions is intuitive and easy to interpret to non-statisticians. Unlike typical methods, the tuning parameter has a simple interpretation as the confidence level, and is thus chosen a priori, as usual by specifying the value such as 95%. The proposed method can be used with a variety of statistical methods where confidence regions can be created for model coefficients. This method can be easily implemented using available computer packages specific for penalized regression techniques. Simulation studies show the new method generally outperforms competing tuning methods in terms of identifying the correct sparse models.
|
