Covariance Estimation and Variable Selection for High-Dimensional Psychiatric Data

• Activity Number: 311
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 8:30 AM to 10:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Vivian H. Shih*+ and Catherine Ann Sugar
• Companies: University of California at Los Angeles and University of California at Los Angeles
• Keywords: covariance estimation ; variable selection ; high dimensionality ; neurocognition ; pediatric psychiatry

Abstract

Clinical investigators now routinely collect data on a large number of measures relative to their sample size. This is particularly true in psychiatry where test batteries include a multiplicity of data types such as clinical symptoms, behaviors, and neurocognitive performance, as well as measures derived from EEG, MRI and fMRI. It is of particular interest to identify phenotypes or patterns of deficits characterizing specific disorders. CIDAR: Translational Research to Enhance Cognitive Control is a study examining cognitive control deficits in children with ADHD and Tourettes. We explore patterns of baseline characteristics based on a restrictive subset of ~500 variables for 367 subjects. In this preliminary analysis, we apply variants of cutting-edge tools (e.g. sparse covariance estimation, graphical/network models, variable selection methods based on adaptive shrinkage), tailored to incorporate the known superstructure of our data. We use these modified techniques to look for differences in the relational structure of our measures across groups and to identify predictors of behavioral symptoms and academic performance for use in subsequent longitudinal analyses.