Date Published: June 3, 2019
Publisher: Public Library of Science
Author(s): Alexander C. Keyel, Oliver Elison Timm, P. Bryon Backenson, Catharine Prussing, Sarah Quinones, Kathleen A. McDonough, Mathias Vuille, Jan E. Conn, Philip M. Armstrong, Theodore G. Andreadis, Laura D. Kramer, Jeffrey Shaman.
West Nile virus (WNV; Flaviviridae: Flavivirus) is a widely distributed arthropod-borne virus that has negatively affected human health and animal populations. WNV infection rates of mosquitoes and human cases have been shown to be correlated with climate. However, previous studies have been conducted at a variety of spatial and temporal scales, and the scale-dependence of these relationships has been understudied. We tested the hypothesis that climate variables are important to understand these relationships at all spatial scales. We analyzed the influence of climate on WNV infection rate of mosquitoes and number of human cases in New York and Connecticut using Random Forests, a machine learning technique. During model development, 66 climate-related variables based on temperature, precipitation and soil moisture were tested for predictive skill. We also included 20–21 non-climatic variables to account for known environmental effects (e.g., land cover and human population), surveillance related information (e.g., relative mosquito abundance), and to assess the potential explanatory power of other relevant factors (e.g., presence of wastewater treatment plants). Random forest models were used to identify the most important climate variables for explaining spatial-temporal variation in mosquito infection rates (abbreviated as MLE). The results of the cross-validation support our hypothesis that climate variables improve the predictive skill for MLE at county- and trap-scales and for human cases at the county-scale. Of the climate-related variables selected, mean minimum temperature from July–September was selected in all analyses, and soil moisture was selected for the mosquito county-scale analysis. Models demonstrated predictive skill, but still over- and under-estimated WNV MLE and numbers of human cases. Models at fine spatial scales had lower absolute errors but had greater errors relative to the mean infection rates.
West Nile virus (WNV) has caused 46,086 diagnosed cases in the United States, with over 2000 human deaths (1999–2016) . The ecological impacts of WNV have been even more substantial, as WNV has been found in 332 bird species in the United States , caused a 45% decline in American Crows, Corvus brachyrhynchos, in the United States , killed millions of songbirds (e.g., an estimated 29 million Red-eyed Vireos, Vireo olivaceus) , and contributed to the listing of the Yellow-billed Magpie, Pica nuttalli, as ‘Near Threatened’ . In addition to avian hosts, WNV has been reported from reptiles , mammals [7–9], and amphibians . Non-avian impacts have led to substantial economic losses, notably due to infections of horses and farmed alligators [6,11].
We present first a summary of the random forest model fitting results followed by results that address the question of which–if any—climate covariates improve the model skill. In the last part we describe the spatial and temporal variability for selected dependent and independent variables identified by the models as important.
We found that climate variables improved WNV model fit metrics at both coarse and fine scales (with a few minor exceptions, see Tables 2 and 3). Climate variables were especially important for predicting human West Nile cases across all counties (ΔR2 = 0.12), and mosquito MLE at the county- and trap-scales (ΔR2 = 0.19, ΔR2 = 0.17, respectively). We found evidence that some of the climate effects on WNV were an indirect result of climatic effects on mosquito populations (see e.g., ). When climate data were omitted, MLE became more important in predicting human cases and the mosquito abundance index became more important for explaining MLE at the trap scale (Table 5).