Analysis of Longitudinal Counts from a Population with Two Latent Classes

• Activity Number: 243
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Huijing Wang*+ and X. Joan Hu
• Companies: Simon Fraser University and Simon Fraser University
• Keywords: EM-Algorithm ; Extended GEE Approach ; Likelihood-based Approach ; Mixture of Poisson Processes

Abstract

Motivated by young cancer survivor's physician visits data, we consider a cohort with two latent classes: subjects live as general population and subjects, named as non-cured group, with potentially higher rates of physician visits. This paper attempts to evaluate the portion of the non-cured and their frequency of physician visits over time, and to identify the associated risk factors. It introduces a latent binary variable to indicate if a subject belongs to non-cured. We specify that with a regression model, and the conditional expectation of the physician visits over time of non-cured into a parametric form. Taking advantages of Canadian medical insurance databases, we assume that the distribution of physician visits in the general population is known. We first consider a mixture model of two Poisson processes, and derive the MLE using EM algorithm. The procedure motivates an extended GEE approach, which does not require to fully specify the underlying probability model for non-cured. We establish the consistency and asymptotic normality of the extended GEE estimator, and examine its small sample properties via simulation. The methodology is illustrated with the motivated data.