Discrete-Time Multistate Models: Estimation, Inference, Properties

Christian Dudel , Max Planck Institute for Demographic Research
Daniel C. Schneider, Max Planck Institute for Demographic Research
Angelo Lorenti, Max Planck Institute for Demographic Research
Peng Li, Max Planck Institute for Demographic Research
Mikko Myrskyla, Max Planck Institute for Demographic Research

Multi-state models generalize survival analysis to transitions back-and-forth between several states. The majority of the literature conceptualizes multi-state models in continuous time. Discrete-time approaches are rare, despite being well-suited for many practical applications. In this paper, we collect existing and provide new results on discrete-time multi-state models. These models have desirable properties and are easy to apply. Specifically, we set up the model as an absorbing Markov chain, and we discuss estimation and inference. Given certain assumptions on functional form, estimation is straightforward, and can use standard methods widely implemented in statistical software. Moreover, we show that Markov chain multi-state models provide consistent estimates of several estimands even when the underlying data generating process is non-Markovian. We conduct simulations which show that small-sample bias is negligible, and that bias is moderate even when the estimate is not consistent under a non-Markovian data generating process.

See paper

 Presented in Session 73. Innovations in Demographic Modelling and Projections