B1643
Title: A Bayesian multilevel hidden Markov model with time-varying covariates for multivariate count time series
Authors: Sebastian Mildiner Moraga - Utrecht University (Netherlands) [presenting]
Emmeke Aarts - Utrecht University (Netherlands)
Abstract: Technological advances such as accelerometers, tracking devices, automatic coding of video recordings, and in vivo experimental set-ups made it easier and more affordable to collect data on multiple subjects or animals with a high resolution over time. However, the high dimensionality of these data combined with their nested structure makes them challenging to analyse. Moreover, extracting insights about the dynamics of processes underlying these data also requires statistical models capable of tracking changes on multiple individuals across time. The hidden Markov models (HMMs) are a promising approach to this end, as they can summarize complex processes with a set of hidden states that switch over time. We present a novel parametric multilevel HMM (MHMM) with continuously distributed random effects and a Poisson-LogNormal emission distribution. Our model deals with the nested structure of the data and allows for including time-varying covariates in the transition distribution. We illustrate the use of our model with a small simulation based on an empirical dataset with multi-electrode electrophysiological measurements in monkeys. In addition, we show how the MHMM can be used to obtain the forward probabilities of the states to investigate changes in a latent process over time.
