Forecasting West Nile Virus with Arbovirus Monitoring and Prediction (ArboMAP) in Multiple US States: A Comparison of Environmental Data Sources

West Nile virus (WNV) remains a persistent public health threat in many parts of the United States. Having advance warning on where and when WNV outbreaks might occur would allow for better targeting of public health responses. The transmission of mosquito-borne diseases, such as WNV, is influenced by environmental conditions that affect many aspects of the disease transmission system. Environmental data therefore have the potential to be indicators of disease risk, and there is a need to determine which types of measurements are most effective for creating accurate forecasting models. We developed the Arbovirus Monitoring and Prediction (ArboMAP) system, which produces weekly, county-level forecasts of human WNV cases using environmental data combined with entomological data. The objective of this research was to compare model accuracy using three different sources of environmental data: 1) GridMET gridded meteorological data derived from the NASA North American Land Data Assimilation System, 2) data from the MODerate resolution Imaging Spectroradiometer (MODIS), including land surface temperature and spectral indices combined with precipitation data from the Global Precipitation Mission (GPM), and 3) Daily Global Land Parameters (DGLP) derived from AMSR-E and AMSR2 passive microwave data. We conducted a retrospective analysis of historical forecasts with ArboMAP to compare model fit and forecast accuracy based on these three datasets and determine which variables were the strongest predictors of WNV. These comparisons were made using data from four U.S. States: South Dakota, Louisiana, Michigan, and Oklahoma. Temperature was the most important environmental driver, and warmer temperatures were associated with more WNV cases in all states. The effects of precipitation and humidity were more geographically variable, and their relationships with WNV varied across states. Each of the three data types had advantages and limitations for use in WNV modeling.