Variable Selection Methods for Population Mixtures

• Activity Number: 461
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Mental Health Statistics Section
• Author(s): Tian Chen*+ and Xin M. Tu
• Companies: and University of Rochester
• Keywords: SCAD ; mixture model ; zero-inflated count outcomes ; longitudinal data ; missing data

Abstract

Variable selection methods are widely used to help determine optimal sets of predictors in regression models.However, many outcomes in studies, particularly in the behavioral and social sciences, follow mixture distributions and as such most existing variable selection methods developed for single-mode distributions (in the absence of covariates) do not apply to such population mixtures. Although some methods address population mixtures, they depend on parametric distributions such as normal and Poisson, with estimates sensitive to departures from assumed models.We propose a variable selection approach for population mixtures without imposing such strong parametric distributions to provide more robust inference in a longitudinal setting with missing data.Our approach utilizes the SCAD technique and BIC criteria to determine the optimal tuning parameter for variable selection.We investigate the Oracle property of the SCAD as well as consistency of model estimates for the proposed approach.We evaluate the performance of the new approach through extensive simulation studies and apply it to a real study on HIV prevention intervention to help determine different treatment moderators.