Print Page
West Nile virus (WNV) was introduced to the United States in 1999 and has subsequently become endemic throughout the Western hemisphere. In the United States, the predominant hosts of WNV are passerine birds; vectors are primarily mosquitoes in the genus Culex . To aid the efficient deployment of interventions, we sought to develop a dynamical early warning system informed by the ecology of these host and vector species. We first constructed series of weekly maximum temperature and rainfall at time lags of zero to six weeks, then fit boosted regression tree (BRT) models to estimate the relative influence of each covariate on the WNV-seropositive bird or mosquito growth rate throughout the WNV season. Additional models were constructed to investigate the relative importance of seasonality. In principle, we expected a lag between weather conditions producing environments conducive for vector breeding and transient increases in the mosquito population, which would translate into increased vectorial capacity. To estimate infection prevalence, WNV-positive birds and mosquito pools were aggregated by week and divided by the total number of individuals or pools tested to produce subsequent estimates of abundance.
Results/Conclusions
Based on mean total deviance (MTD, 0.672) and mean residual deviance (MRD, 0.431), a model for prevalence in bird hosts excluding seasonality was selected for forecasting. This model explained 36 % of the observed variation in prevalence. The most influential covariates from this model were current temperature (40.08 %), a 2 week temperature lag (12.05 %), and the week of the year (10.96 %). A comparable model with seasonality explained only 28 % of the variation in seroprevalence. These results suggest that little or no temperature lag is required to predict changes in WNV seroprevalence in birds and that seasonality does not have a significant influence on bird seroprevalence. Based on MTD (0.783) and MRD (0.26), the mosquito model including seasonality was best fit, explaining 67 % of the variation in seroprevalence. The most influential covariates from this model were current temperature (31.06 %) and the week of the year (14.57 %). A comparable model excluding seasonality explained only 64 % of the variation in seroprevalence. These results suggest that a temperature lag of zero best predicts these data, and that the addition of seasonality yields slightly better predictive power for modeling mosquito seroprevalence.
