Improving the prediction of arbovirus outbreaks: A comparison of climate-driven models for West Nile virus in an endemic region of the United States.

Models that forecast the timing and location of human arboviral disease have the potential to make mosquito control and disease prevention more effective. A common approach is to use statistical time-series models that predict disease cases as lagged functions of environmental variables. However, the simplifying assumptions required for standard modeling approaches may not capture important aspects of complex, non-linear transmission cycles. Here, we compared a set of alternative models of human West Nile virus (WNV) in 2004-2017 in South Dakota, USA. We used county-level logistic regressions to model historical human case data as functions of distributed lag summaries of air temperature and several moisture indices. We tested two variations of the standard model in which 1) the distributed lag functions were allowed to change over the transmission season, so that dependence on past meteorological conditions was time varying rather than static, and 2) an additional predictor was included that quantified the mosquito infection growth rate estimated from mosquito surveillance data. The best-fitting model included temperature and vapor pressure deficit as meteorological predictors, and also incorporated time-varying lags and the mosquito infection growth rate. The time-varying lags helped to predict the seasonal pattern of WNV cases, whereas the mosquito infection growth rate improved the prediction of year-to-year variability in WNV risk. These relatively simple and practical enhancements may be particularly helpful for developing data-driven time series models for use in arbovirus forecasting applications.