Spatio-temporal landscape models for West Nile virus vector population abundance and distribution

In response to the rapid spread of West Nile virus throughout North America there has been an increase in the number of mosquito vector surveillance and control programs. However, mosquito surveillance programs are costly and time-consuming for the districts that maintain them and only provide data for the specific locations sampled. Extracting maximum information about mosquito population abundance and distribution from such sparse data is essential. Mosquito vector abundance and distribution is dependent upon landscape factors that provide habitats specific to each vector species. Variation in vector mosquito abundance over time and space confounds landscape modeling of vector population distributions; however, utilizing a combination of current statistical techniques and a unique dataset, I was able to successfully identify important meteorological and landscape factors for prediction of abundance and population distributions of two important vector species, Aedes vexans and Culex pipiens-restuans. Both vector species were adequately forecasted for a fixed trapping site with multivariate SARIMA models but the time series analysis revealed that meteorological conditions are more important for predicting Ae. vexans abundance, explaining 55% in the variation of this important bridge vector. Weather factors impacted population abundance of Cx. pipiens-restuans to a lesser extent, explaining only 15% of the variation in abundance. Three true meteorological predictors were identified for Ae. vexans: a cooling degree day (CDD)-precipitation index, evapotranspiration x evapotranspiration, and CDD base 65°F (CDD65) x CDD65. The only true predictor identified for Cx. pipiens-restuans was CDD base 63°F. The results of the time series analysis were applied to the more spatially intensive Amherst data set using Classification and Regression Trees (CART) to categorize the spatial data by appropriate temporal and meteorological parameters. Mosquito trap sites are often located too far apart to detect spatial dependence but the results showed that integration of spatial data over time for Cx. pipiens-restuans and by meteorological conditions for Ae. vexans enables spatial analysis of sparse sample data. Spatial autocorrelation was quantified for both Cx. pipiens-restuans and Ae. vexans with more spatial dependence evident in Cx. pipiens-restuans populations than in Ae. vexans populations. The range of spatial dependence for Ae. vexans was relatively constant, only varying between 3250 to 3750 meters, but was only detectable for three groups of Ae. vexans at the scale measured. Spatial dependence was stronger in Cx. pipiens-restuans populations and was shown to vary over time, with ranges of spatial dependence between 2000 and 6500 meters. Identification of important landscape factors was conducted by development of aspatial models for each group of vector mosquito with stepwise linear regression analysis. Generally, urban areas, vacant and agricultural land, and wetlands were important covariates of Ae. vexans. Urban land, hydrography, forested land, and areas outside the 500-year floodplain were important covariates for Cx. pipiens-restuans. Age of housing units and housing unit density were important subclasses of urban areas for predicting Cx. pipiens-restuans abundance distribution. However, accurate determination of predictive landscape elements for mosquito abundance requires that spatial structure be accounted for in the model. I developed spatial regression models for Cx. pipiens-restuans and Ae. vexans population abundance with a mixed model regression framework that incorporates parameters of spatial correlation and landscape variables that best explain the spatial structure in mosquito abundance. More accurate estimates of landscape effects were derived from the spatial regression model and spatially explicit predictions were produced for the study area.