THE IMPACT OF DIFFERENT DROUGHT METRICS ON WEST NILE VIRUS PREDICTION IN OKLAHOMA

Abstract

West Nile virus (WNV) is an arbovirus originating in birds and transmitted to humans via infected mosquitoes. Since its introduction into the U.S. in 1999, research has concluded that weather variables, including temperature, precipitation, humidity, and wind speed, influence transmission. Previous research has also identified drought as a potential factor in arbovirus transmission, however, the metrics selected to represent drought in previous works are inconsistent and often chosen without sufficient exploration of alternatives. This study aims to explore which drought metrics have the most significant impact on WNV prediction in Oklahoma, while also expanding knowledge of how weather and land cover influence WNV transmission within the state. In order to accomplish this broad goal, this study presents three main objectives: (1) Identify the optimal meteorological predictors of WNV transmission in Oklahoma from the standard base weather variables (2) Identify the optimal drought metric predictor for WNV transmission in Oklahoma and determine whether it improves the base weather variable model, and (3) Determine which land cover classifications exhibit the most influence over the spatial patterns of WNV transmission in Oklahoma. Data was obtained on an epi-week timescale for each Oklahoma county between 2004-2023, and included Oklahoma State Department of Health de-identified human WNV case data, GridMet average minimum, maximum, and mean temperature, mean relative humidity, vapor pressure deficit, average and total precipitation, and the percentage of reclassified National Land Cover Database (NLCD) land cover classes. Drought metrics were similarly obtained through GridMet and included the Palmer Drought Severity Index (PDSI), Standardized Precipitation Index (SPI), Standardized Precipitation Evapotranspiration Index (SPEI), and Evaporative Demand Drought Index (EDDI). The SPI, SPEI, and EDDI were all obtained for 14 day, 30 day, 90 day, 180 day, one year, two year, and five year aggregations. Generalized additive models (GAMs) with embedded distributed lag nonlinear models (DLNMs) were used to test the relationships of the variables. Best performing models were selected using the Bayesian Information Criterion (BIC) and validated using the area under the curve (AUC) and Spearman’s rank correlation coefficient (SRCC). The best overall model included minimum temperature, wind speed, developed land cover, and the EDDI at 30 day calculations. Furthermore, the model lags indicated that there were multiple pathways through which dry conditions could influence WNV risk. These findings solidify drought, especially in terms of evapotranspiration, as an important factor in WNV prediction, imply that models containing anomalies of weather variables and variables combining multiple weather phenomena may perform well for WNV prediction, and present a possible modelling scheme for Oklahoma that should be evaluated in other locations.