B1335
Title: Data-augmented MCMC for learning spatiotemporal transmission structure in epidemic models
Authors: Jason Xu - Duke University (United States) [presenting]
Abstract: A new class of flexible spatiotemporal stochastic epidemic models amenable to scalable full Bayesian inference is introduced. Drawing on methods developed for Poisson data, the transmission rate of the epidemic model is developed with a dynamic multiscale structure. The method tracks changes in the transmission rate of a stochastic epidemic model over time while smoothing across regions. Borrowing information in a hierarchical manner stabilizes the transmission rate estimates for regions where the observed data are sparse and improves generalization error. Further, through the use of time-specific discount factors developed in the time series literature, both gradual and abrupt changes are captured over time and between regions. It is shown how inference under the exact model posterior is possible using a block Gibbs sampler relying on an efficient forward-filtering backwards-sampling algorithm, enabling Bayesian analysis for large outbreaks for challenging missing data settings such as incidence data.