A Shared Latent Process Model to Correct for Preferential Sampling in Disease Surveillance Systems

Disease surveillance systems are crucial to monitor and predict outbreaks, epidemics and pandemics, as well as to understand the dynamics and trends of diseases over space and time. For zoonotic diseases, i.e., diseases that spread from animals to humans, surveillance systems often rely on complex data collection mechanisms which present particular challenges to the statistician, including sampling processes that commonly violate key assumptions of standard statistical methods. One such mechanism is preferential sampling, referring to a stochastic dependency between a spatial process of interest and the locations at which it is observed, commonly arising out of practical considerations related to a limited sampling budget and a rare outcome. While this sampling strategy can lead to considerably biased spatial predictions, few solutions to address preferential sampling have been proposed in the context of disease surveillance. We propose a novel approach to correct for preferential sampling in disease surveillance applications and show by simulation the practical benefits of reduced bias in parameter estimates and greater accuracy of the estimated risk surface. We conclude with an application of the model to a disease surveillance dataset targeting plague (Yersinia pestis) in the sylvatic rodent populations in California.