Performance of Estimates for Mixtures of Time-Varying Effects Models

• Activity Number: 135
• Type: Contributed
• Date/Time: Monday, August 10, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Nonparametric Statistics
• Author(s): John Dziak* and Runze Li and Xianming Tan and Saul Shiffman
• Companies: Penn State and Penn State and McGill University and University of Pittsburgh
• Keywords: varying-coefficients models ; finite mixture models ; autoregressive models ; software

Abstract

Finite mixture models provide an easily interpretable way to account for sample heterogeneity. Varying-coefficient regression models provide a nonparametric way to depict changes over time both in the overall levels of variables, and in their relationships with other variables. As Lu and Song (2012) suggested, a finite mixture of varying-coefficient regression models combines both advantages and allows rich exploratory analyses of longitudinal datasets. We provide free R and SAS software for a frequentist implementation of such a mixture model, which we call "mixture time-varying effects modeling" or MixTVEM, implemented using B-splines and the EM algorithm. We use an autoregressive correlation structure with nugget to account for within-subject correlation, and do not assume even spacing of observations in time. Intensive longitudinal data from a smoking cessation study is analyzed, revealing different classes of individuals whose rates of recovery from withdrawal symptoms vary. The relationship between negative affect and urge to smoke varies both over time and between classes. We provide simulation results on the accuracy and coverage of estimates.