GSIS超高维变量选择

摘要

Variable selection plays an important role in high dimensional models .Fan and Lv showed sure independent screening (SIS ) based on simple correlation . But w hen independent variable can be naturally grouped ,SIS does not work .Because SIS is designed for individual variable selection ,but not group selection .In this paper ,we propose group sure independent screening (GSIS ) based on marginal group regression .The method is designed for either variable selection or group selection ,also for both . Monte Carlo simulations indicate that GSIS has superior performance in group and individual variable selection relative to SIS .%变量选择在超高维统计模型中非常重要。Fan 和 Lv 基于简单相关系数提出确保独立筛选法（SIS），但当自变量被分成组时，SIS就会失效。因为SIS只能对单个变量进行选择，不能对组变量进行选择。为此，基于边际组回归提出组确保独立筛选法（GSIS），该方法不仅对组变量有效，对单个变量也有效，或者两者的混合也同样有效。Monte Carlo模拟结果显示，GSIS的表现优于SIS。