Evaluating statistical models for predicting animal-vehicle collisions
There is growing concern over the impact of urban expansion and increasing road development on the conservation of biodiversity and the condition of global ecosystems. Loss of habitat and fragmentation restrict biodiversity to isolated refugia that are increasingly bisected by roads, especially in peri-urban landscapes. The increasing frequency of collisions between animals and vehicles are a significant threat to fauna and hence has received considerable attention.
Modelling approaches to predicting locations of animal-vehicle collisions are conducted to a) infer those factors contributing to collisions and b) to identify hotspots where mitigation efforts would be most effective (given limited resources). Published models exist for a wide range of species at a range of different scales. Models have typically relied upon variables that characterise the road environment for prediction; such as road sinuosity, road-verge attributes and spatial and temporal traffic variation. Landscape variables are also included but often their incorporation is a generic attempt to describe environmental variation, although a few examples exist where animal population parameters and spatial patterns of landscape use are incorporated into collision prediction. The majority of models have been constructed using standard linear regression techniques despite the fact that the intrinsic spatial structure of collision locations invalidates the assumption of independence. Consequently, many models are poor predictors of crash locations.
In this paper I review past attempts at modelling animal-vehicle collisions and place them in the context of current statistical theory. I provide a number of examples of how improvements in collision prediction can be made by addressing attributes of the animal populations and by explicitly incorporating spatial dependence. Reliance on models that only describe the road environment and broad landscape variables are limited in their capacity to identify hotspots and to inform managers of locations to engage in mitigation activities.