|
Activity Number:
|
318
|
|
Type:
|
Invited
|
|
Date/Time:
|
Tuesday, August 8, 2006 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Technometrics
|
| Abstract - #304889 |
|
Title:
|
A New Strategy for Variable Selection
|
|
Author(s):
|
Xiaohui Luo and Leonard A. Stefanski and Dennis A. Boos*+
|
|
Companies:
|
Merck & Co., Inc. and North Carolina State University and North Carolina State University
|
|
Address:
|
Department of Statistics, Raleigh, NC, 27695-8203,
|
|
Keywords:
|
false selection rate ; forward selection ; mean squared error ; noise addition ; phony variables ; SIMEX
|
|
Abstract:
|
We describe an approach to variable selection wherein we calibrate a tunable selection method to achieve desirable properties of models selected. We focus on forward selection where the tuning parameter is alpha-to-enter, aka SLENTRY to SAS users. Calibration is achieved by adding "noise" to the data and tracking its effect on the models selected. In one version of the strategy, parametric bootstrap-like data sets are generated by adding Gaussian noise to the response variable. Then SLENTRY is tuned by tracking the effect of the added noise on the MSEs of models selected for different SLENTRY values. An alternative means of "adding noise" to the data gives rise to a second method wherein random phony predictor variables are appended to the data, and SLENTRY is tuned by tracking the proportion of falsely-included phony variables in the models selected for different SLENTRY values.
|
