|
Activity Number:
|
445
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 6, 2008 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #301930 |
|
Title:
|
Early Thresholding for High-Dimensional Linear Regression and Variable Selection
|
|
Author(s):
|
Xinge J. Jeng*+ and Michael Y. Zhu
|
|
Companies:
|
Purdue University and Purdue University
|
|
Address:
|
250 N. University Street, West Lafayette, IN, 47907-2066,
|
|
Keywords:
|
Early Thresholding ; Covariance Regularization ; Estimation Efficiency ; Variable Selection
|
|
Abstract:
|
We introduce a general procedure called Early Thresholding to improve the performance of shrinkage methods such as Lasso and Dantzig selector in high dimensional regression. It is known that in high dimensional regression, singularity of sample covariance matrices can compromise the performance of most estimators. On the other hand, in many applications, a large amount of predictors are weekly correlated. Early Thresholding applies an additional regularization step in estimating the sample covariance matrix, in order to induce a sparse structure. Both theoretical and simulation results show that Early Thresholding can improve estimation and variable selection when combined with Lasso and Dantzig selector.
|
