Predictive assessment of a non-linear random effects model for space-time surveillance data

Notification data collected by national surveillance systems are typically available as weekly time series of counts of confirmed new cases, stratified e.g. by geographic areas. This work outlines the statistical modeling framework in Paul and Held (2011) for the analysis of such data. Inherent (spatio-)temporal dependencies are incorporated via an observation-driven formulation. Using region-specific and possibly spatially correlated random effects, we are able to address heterogeneous incidence levels. Inference is based on penalized likelihood methodology for mixed models. The predictive performance of models is assessed using probabilistic one-step-ahead predictions and proper scoring rules.