The MNet Estimator

• Activity Number: 182
• Type: Contributed
• Date/Time: Monday, August 2, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Patrick Breheny*+ and Jian Huang and Shuangge Ma and Cun-Hui Zhang
• Companies: University of Kentucky and The University of Iowa and Yale University and Rutgers University
• Keywords: Elastic net ; Minimax concave penalty ; Lasso ; Variable selection ; Oracle property ; Shrinkage

Abstract

We propose the MNet estimator, a new penalized method for variable selection using a combination of minimax concave and ridge penalties. The MNet method is designed to deal with p > n problems with highly correlated predictors. Similar to the elastic net, the proposed method also tends to select or drop highly correlated predictors together. In addition, the MNet is sign consistent and equals the oracle ridge estimator with high probability under some reasonable conditions. We apply the coordinate descent algorithm to compute the MNet estimates. Simulation studies suggest that the MNet has superior performance in variable selection in the presence of highly correlated predictors relative to existing methods.