Ticks in the genus Ixodes are the principal vector of Borrelia burgdorferi, the causative bacterium of Lyme disease. Ticks acquire and spread the pathogen between vertebrate hosts when taking a blood meal, which is required for the tick to molt and reproduce. This host-use trait is responsible for Lyme infection in ticks. Previous work has shown that tick abundance is correlated with Lyme disease incidence, and thus human risk. Given that reported cases of the disease have doubled in the last decade, accurate near-term forecasts of tick populations are needed to infer human disease risk. Here, we constructed several bayesian matrix models to forecast Ixodes scapularis populations at the Cary Institute of Ecosystem Studies. Priors for daily survival rate of larvae and nymphs were built by fitting a binomial process model to survival experiment data. We evaluated forecasts that use daily weather and host population abundance to drive tick survival and transition probabilities, respectively. The abundance of the host population, white-footed mice (Peromyscus leucopus), was also forecasted. Models were evaluated by their forecast uncertainty and predictive ability.
Results/Conclusions
All tick forecasts evaluated used informative priors constructed from the survival experiments, as they increased identifiability and significantly reduced uncertainty of demographic parameters. Additionally, the introduction of weather covariates and mouse abundance reduce forecast uncertainty compared to the random intercept model. Relative humidity is the best single driver of daily survival for all life stages of Ixodes scapularis, while the current number of mice (no time lag) best explains larva-to-nymph and nymph-to-adult transitions. However, the uncertainty introduced in forecasting the mouse populations contributed to the increasing uncertainty in the tick forecast, becoming no more certain than random by the end of the forecast. In general, process error is the most dominant form of variance in all forecasts regardless of driver(s) used. By partitioning uncertainty we show that constraining drivers and using highly informed priors can improve forecasts. Future work should focus on constraining process uncertainty with additional experiments and through assimilation of new data sources, which could further improve forecasts to inform human risk.