Dispersal in a hurry: Bayesian learning from surveillance to establish area freedom from plant pests with early dispersal

Abstract

Declaration of area freedom from plant pests is crucial for the agricultural sector, since it promotes continuing domestic and international trade of crops at risk from exotic pests. Freedom from plant pests may also enhance environmental health, with indirect effects on agricultural productivity. Every year, several new exotic plant pest species are reported for the first time. In the face of this continual pressure and growing globalization, resources to undertake surveillance are limited. Design of surveillance is critical for determining how to allocate these limited resources. Designs for surveillance that help assess area freedom have focused on the colonization process, captured by a prevalence model that does not accommodate dispersal. In this paper, we extend these designs to accommodate early dispersal from a few colonization points. This provides a basis for evaluating the effectiveness of surveillance over multiple sampling occasions. To achieve this we harness a Bayesian statistical framework. Although there are some computational overheads, this provides several benefits: (i) an intuitive hierarchical structure that helps separate then link modelling components; (ii) the facility to incorporate expert knowledge; and (iii) inference that directly addresses the questions of farmers and biosecurity managers, in a way that the range of plausible outcomes is provided together with point estimates. Finally the Bayesian framework facilitates a natural cycle of learrning, that readily incorporates new information - from surveillance snapshots - as it becomes available. Firstly, we harness the natural hierarchical structure of the Bayesian statistical framework to separate the model—for the spatio-temporal dynamics of dispersal underlying prevalence—from the model for detection, which depends on prevalence. Secondly, expert knowledge on both point estimates and variability can be explicitly incorporated as Bayesian prior distributions, and in each phase, these priors are updated into new posteriors as more surveillance data becomes available. This is important since much of the data informing design of surveillance for exotic plant pests relies heavily on expert judgment, especially during the early phases of plant biosecurity—when establishing area freedom. Thirdly, the Bayesian posterior approach used here automatically answers the question of If we detect nothing, how many infested plants could we have missed? This approach provides a ready mechanism for including information about dispersal on the infested plants (both missed and detected). Finally, the Bayesian framework facilitates an adaptive cycle of learning. We can apply Bayesian inference to analyze the first surveillance snapshot and learn about prevalence and detectability parameters. Then Bayesian predictions can be used to progress the pest status before analysis of the next snapshot. This flexibly provides a basis for incorporating new knowledge as it is obtained.