Generalized Forward Selection: Subset Selection in High Dimensions

• Activity Number: 81
• Type: Contributed
• Date/Time: Sunday, August 1, 2010 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Derick R. Peterson*+ and Alexander T. Pearson
• Companies: University of Rochester and University of Rochester
• Keywords: model selection ; subset selection ; proportional hazards ; high-dimensional data ; Cox model ; survival analysis

Abstract

We consider subset selection in high dimensions. We propose a generalized forward selection procedure that conceptually lies between the greedy traditional forward selection method and the computationally infeasible all-subsets search. In contrast to LASSO and other shrinkage-based approaches to this problem, we allow standard unconstrained parameter estimation in the selected models, and we further allow prespecified predictors to be forced into all candidate models. Since it offers more complexities than most other model frameworks, we focus our simulations and data analyses on the Cox model for censored survival outcomes. Our simulation and data analysis results demonstrate that our generalized forward selection method has a number of advantageous properties for selecting sets of predictive variables, compared with forward selection, univariate screening, and the LASSO.