Abstract:
|
Finite mixture models are increasingly used to model the distribution of mixtures of subpopulations in multivariate regression settings. Since observations collected from the same subject over time may be dependent, mixture model-based methods that assume independence of observations may not be valid for modeling and clustering of longitudinal data, such as growth curve trajectories. We consider finite mixture regression models with random effects to capture the variation between subjects and dependencies within units. We extend nonlinear mixed-effects mixture regression models to allow the mixing proportion to depend on covariates. We propose a version of stochastic approximation EM (NR-SAEM) algorithm for maximum likelihood estimation. The implementation of the algorithm may also include drawing unobservable random effects from conditional distributions. The proposed method is demonstrated using simulated and real data of repeated hormone levels over a part of the menstrual cycle. We also compare the standard EM approach and the NR-SAEM algorithm for parameter estimation in a setting of finite mixture regression model.
|