A dynamic latent space time series model to assess the spread of mumps in England

Abstract

This work is motivated by an original dataset of reported mumps cases across nine regions of England, and focuses on the modeling of temporal dynamics and time-varying dependency patterns between the observed time series. The goal is to discover the possible presence of latent routes of contagion that go beyond the geographical locations of the regions, and instead may be explained through other non directly observable socio-economic factors. We build upon the recent statistics literature and extend the existing count time series network models by adopting a time-varying latent distance network model. This approach can efficiently capture across-series and across-time dependencies, which are both not directly observed from the data. We adopt a Bayesian hierarchical framework and perform parameter estimation using L-BFGS optimization and Hamiltonian Monte Carlo. We demonstrate with several simulation experiments that the model parameters can be accurately estimated under a variety of realistic dependency settings. Our real data application on mumps cases leads to a detailed view of some possible contagion routes. A critical advantage of our methodology is that it permits clear and interpretable visualizations of the complex relations between the time series and how these relations may evolve over time. The geometric nature of the latent embedding provides useful model based summaries. In particular, we show how to extract a measure of contraction of the inferred latent space, which can be interpreted as an overall risk for the escalation of contagion, at each point in time. Ultimately, the results highlight some possible critical transmission pathways and the role of key regions in driving infection dynamics, offering valuable perspectives that may be considered when designing public health strategies.