Markov-switching autoregressive latent variable models for longitudinal data

We propose a generalization of the autoregressive latent variable mod- els for longitudinal data based on an AR(1) process to represent the effect of unobservable factors on the response variables. The generalization is based on assuming that the latent process follows a Markov-switching AR(1) process with correlation coefficient depending on the regime of the chain. Some particular cases are discussed in detail and illustrated by an application to a longitudinal dataset about self-evaluation of the health status.