Sylvatic plague (Yersinia pestis) causes frequent catastrophic declines in prairie dog colonies and regional complexes. Previous studies have shown that the probability of plague outbreaks in black-tailed prairie dog (Cynomys ludovicianus) colonies can be affected by temperature, precipitation, topography and colony attributes (e.g., total colony area). However, a framework for predicting plague outbreaks on the basis of these intrinsic and extrinsic covariates remains elusive, as previous results are somewhat inconsistent and difficult to reconcile due to discrepancies in methodology and spatial scale. To enable more accurate prediction of when and where plague die-offs will occur, we are undertaking a broad re-analysis of raw datasets from across the black-tailed prairie dog’s range. Thus far we have gathered time-series’ of prairie dog colony extents for 9 colony complexes in 6 states, originally collected by a variety of agency biologists and academic researchers. To maximize predictive inference, we are using an analytical framework (Random Forest) that allows for non-linear relationships and complex interactive effects. Finally, we are assessing predictive performance using rigorous cross-validation.
Results/Conclusions
Preliminary results from a well-studied site in Colorado agree with previously reported results, with plague outbreak probability positively influenced by colony size and summer temperatures. However, effects of other covariates are modulated by previously undescribed interactions. For example, an indirect effect of precipitation (higher precipitation in the summer of the previous year leading to die-offs) is only apparent in large colonies, potentially providing an explanation for why published models constructed using the same dataset (but not including interactions) failed to detect this effect. Predictive performance under rigorous cross-validation (withholding entire colonies for validation) was fair (AUC=0.78), and we believe model performance will improve with additional covariates, data sets, and refinements to our modeling framework. The results of this validated model could help managers to more precisely target plague mitigation efforts such as dusting (pulicide applications) and oral vaccination.