New Feature Screening Method for Time-Varying Coefficient Model with Ultrahigh-Dimensional Longitudinal Data

• Activity Number: 90
• Type: Invited
• Date/Time: Sunday, August 9, 2015 : 8:30 PM to 9:15 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Wanghuan Chu* and Runze Li and Matthew Reimherr
• Companies: Penn State and Penn State and Penn State
• Keywords: feature screening ; time-varying coefficient model ; longitudinal data ; mixed-effects model

Abstract

This project is concerned with feature screening method for time-varying coefficient models with ultrahigh dimensional longitudinal data. While some covariates are truly associated with the response mean function, there are other covariates that are potentially responsible for the variation in the response. We propose a two-step screening method that identifies important fixed effects in the first step and selects random effects in the second step. We study its sure screening property, and examine the finite sample performance via Monte Carlo simulations under generalized linear model framework for both continuous and categorical response. In an empirical analysis of a genetic dataset, we advocate a two-stage approach by first reducing the ultrahigh dimensionality for both fixed and random effects to moderate sizes using the proposed two-step procedure, and then applying model selection techniques to make statistical inference on the coefficient functions, variance components and covariance structure.