Classification with High-Dimensional Predictors and Qualitative Response: Applications to Genetics

• Activity Number: 273
• Type: Invited
• Date/Time: Tuesday, August 5, 2003 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Jing Huang*+ and Richard A. Olshen and Chao A. Hsiung
• Companies: Affymetrix Inc. and Stanford University and National Health Research Institutes Taiwan
• Keywords: classification tree ; polygenic ; complex interaction

Abstract

In many real-world situations, multiple factors combine to determine outcome; those factors have complicated and influential interactions, though insignificant individual contributions. One good example is the association between polygenic disease and many genetic and environmental risk factors. Traditional approaches focusing on individual effect have proven to be inadequate in this case. My approach, on the contrary, treats all risk factors together. It considers both interactions and complex combined effects simultaneously. The technique stems from the binary classification tree method. It uses the tree as framework and employs penalized linear regression to define the partitioning rule. Joining regression and classification tree not only allows us to consider combined effects and interactions simultaneously, but also provides us a simple, easy-to-interpret model. In addition, its predictive power and robustness are improved by the variable selection technique. This algorithm has been applied to simulated data and also a large scale genetic study. It demonstrates substantial improvement in performance over prior competitive tree-structured methodologies.