Stable and practical semi-Markov modelling of intermittently-observed data

Multi-state models are commonly used for intermittent observations of a state over time, but these are generally based on the Markov assumption, that transition rates are independent of the time spent in current and previous states. In a semi-Markov model, the rates can depend on the time spent in the current state, though available methods for this are either restricted to specific state structures or lack general software. This paper develops the approach of using a "phase-type" distribution for the sojourn time in a state, which expresses a semi-Markov model as a hidden Markov model, allowing the likelihood to be calculated easily for any state structure. While this approach involves a proliferation of latent parameters, identifiability can be improved by restricting the phase-type family to one which approximates a simpler distribution such as the Gamma or Weibull. This paper proposes a moment-matching method to obtain this approximation, making general semi-Markov models for intermittent data accessible in software for the first time. The method is implemented in a new R package, "msmbayes", which implements Bayesian or maximum likelihood estimation for multi-state models with general state structures and covariates. The software is tested using simulation-based calibration, and an application to cognitive function decline illustrates the use of the method in a typical modelling workflow.