Title: A latent Markov model approach for flexible clustering of longitudinal data
Authors: Zhivko Taushanov - University of Geneva (Switzerland) [presenting]
Abstract: Latent Markov models have been successfully used to model and, in some cases, also cluster continuous longitudinal data. We will present a type of approach to longitudinal data, that combines modelling and clustering together. Such a ``flexible'' clustering would allow some groups to change behaviour over time while keeping them separated from others. The proposed framework consists of a two-level model with a Mixture Transition Distribution in the visible part and a Markovian model driving the latent part, the transition matrix of the latter having a central role. This combined approach would be possible by constraining the transition matrix in a way that it considers data features such as a one- or two-way transition between some pairs of latent states, an impossible transition between others, etc. We will discuss the estimation procedure with its challenges, and possible applications to various domains such as social sciences, demography, psychology etc. Simple examples may include identifying distinct groups while they change temporarily or permanently their behaviour (transition from work to retirement, life-changing events etc.). In such examples, more than one latent state may be associated with the same cluster, aiming to capture changing behaviour of the same group.