Bayesian forecasting of disease spread with little or no local data

Abstract

Rapid and targeted management actions are a prerequisite to efficiently mitigate disease outbreaks. Targeted actions, however, require accurate spatial information on disease occurrence and spread. Frequently, targeted management actions are guided by non-statistical approaches that define the affected area by a pre-determined distance surrounding a small number of disease detections. As an alternative, we present a long-recognized but underutilized Bayesian technique that uses limited local data and informative priors to make statistically valid predictions and forecasts about disease occurrence and spread. As a case study, we use limited local data that were available after the detection of chronic wasting disease in Michigan, U.S. along with information rich priors obtained from a previous study in a neighboring state. Using these limited local data and informative priors, we generate statistically valid predictions of disease occurrence and spread for the Michigan study area. This Bayesian technique is conceptually and computationally simple, relies on little to no local data, and is competitive with non-statistical distance-based metrics in all performance evaluations. Bayesian modeling has added benefits because it allows practitioners to generate immediate forecasts of future disease conditions and provides a principled framework to incorporate new data as they accumulate. We contend that the Bayesian technique offers broad-scale benefits and opportunities to make statistical inference across a diversity of data-deficient systems, not limited to disease.