Identification of Significant Features in High-Dimensional Data Using Supervised Principal Components

• Activity Number: 467
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Flora Mulkey*+ and Eric Bair
• Companies: The University of North Carolina at Chapel Hill and The University of North Carolina at Chapel Hill
• Keywords: Feature Selection ; Microarray ; High-Dimensional Data ; Semi-Supervised Methods ; Genetics

Abstract

We consider the problem of identifying features that are associated with an outcome of interest in high-dimensional data. This problem is particularly relevant in the analysis of DNA microarray data and genome-wide association data, where measurements from thousands of genes can be evaluated to further elucidate the genetic causes of complex diseases. We show that a method called supervised principal components can successfully identify features associated with a specific outcome more accurately than conventional methods. It is particularly effective in situations where the outcome variable is measured with a high degree of error or when the outcome variable is merely a surrogate for an unobserved variable. For example, the survival time of cancer patients is frequently used as a surrogate variable for unobserved tumor subtypes. We apply our procedure to a series of simulated and real genetic data sets and show that it outperforms conventional methods and several competing approaches.