Continuous time hidden Markov model for longitudinal data

Abstract

Hidden Markov models (HMMs) describe the relationship between two stochastic processes, namely, an observed outcome process and an unobservable finite-state transition process. Given their ability to model dynamic heterogeneity, HMMs are extensively used to analyze heterogeneous longitudinal data. A majority of early developments in HMMs assume that observation times are discrete and regular. This assumption is often unrealistic in substantive research settings where subjects are intermittently seen and the observation times are continuous or not predetermined. However, available works in this direction restricted only to certain special cases with a homogeneous generating matrix for the Markov process. Moreover, early developments have mainly assumed that the number of hidden states of an HMM is fixed and predetermined based on the knowledge of the subjects or a certain criterion. In this article, we consider a general continuous-time HMM with a covariate specific generating matrix and an unknown number of hidden states. The proposed model is highly flexible, thereby enabling it to accommodate different types of longitudinal data that are regularly, irregularly, or continuously collected. We develop a maximum likelihood approach along with an efficient computer algorithm for parameter estimation. We propose a new penalized procedure to select the number of hidden states. The asymptotic properties of the estimators of the parameters and number of hidden states are established. Various satisfactory features, including the finite sample performance of the proposed methodology, are demonstrated through simulation studies. The application of the proposed model to a dataset of bladder tumors is presented.