A non-homogeneous hidden Markov model for partially observed longitudinal responses

TL;DR

This paper introduces a semi-parametric hidden Markov model for longitudinal data with non-ignorable dropout, capturing time-varying heterogeneity and dependence between responses and missing data through non-homogeneous Markov chains.

Contribution

It proposes a novel dynamic, semi-parametric model with non-homogeneous hidden Markov chains for random coefficients, allowing flexible dependence modeling in longitudinal studies with dropout.

Findings

Model effectively captures dependence between responses and missing data.

Applied to elderly cognitive data, demonstrating practical utility.

Flexible nonparametric estimation of joint random coefficient distribution.

Abstract

Dropout represents a typical issue to be addressed when dealing with longitudinal studies. If the mechanism leading to missing information is non-ignorable, inference based on the observed data only may be severely biased. A frequent strategy to obtain reliable parameter estimates is based on the use of individual-specific random coefficients that help capture sources of unobserved heterogeneity and, at the same time, define a reasonable structure of dependence between the longitudinal and the missing data process. We refer to elements in this class as random coefficient based dropout models (RCBDMs). We propose a dynamic, semi-parametric, version of the standard RCBDM to deal with discrete time to event. Time-varying random coefficients that evolve over time according to a non-homogeneous hidden Markov chain are considered to model dependence between longitudinal responses recorded from the same subject. A separate set of random coefficients is considered to model dependence between missing data indicators. Last, the joint distribution of the random coefficients in the two equations helps describe the dependence between the two processes. To ensure model flexibility and avoid unverifiable assumptions, we leave the joint distribution of the random coefficients unspecified and estimate it via nonparametric maximum likelihood. The proposal is applied to data from the Leiden 85+ study on the evolution of cognitive functioning in the elderly.