2017 ESA Annual Meeting (August 6 -- 11)

COS 108-3 - Statistical prediction of human West Nile virus incidence based on changes in land cover pattern and composition

Wednesday, August 9, 2017: 2:10 PM
D137, Oregon Convention Center
Sarah E. Bowden, Vijay Ramesh and Barbara Han, Cary Institute of Ecosystem Studies, Millbrook, NY
Background/Question/Methods

Mosquito species are often associated with specific land cover types that contain their preferred habitat for oviposition, such as artificial containers in urban areas or irrigation ditches in agricultural areas. Changes in land cover composition can result in changes in distribution and abundance of mosquito vectors. Further, these changes in distribution and abundance of vectors can then alter disease transmission to hosts, including humans. Land cover change can be quantified in different ways (e.g., changes in composition vs. changes in pattern), each of which has different ecological implications. Therefore, we aimed to answer the following questions: Which land cover change metrics best predict human vector-borne disease incidence? Which type of change is a bigger driver of disease incidence: land cover composition or land cover pattern? We used machine learning algorithms in an attempt to predict yearly human West Nile virus (WNV) disease incidence at the county level in the U.S. (2001-2012) using a suite of land cover change covariates as predictors. Our model was trained on an 80% subset of the data to obtain the top covariates of importance. We then ran the remaining 20% of data through the model to obtain a measure of predictive accuracy.

Results/Conclusions

Our initial model covariates included yearly lagged changes in land cover composition (e.g., proportion of a county covered in a specific land cover type) for 17 land cover types, as well as county size and geographic location (XY centroid). We identified lags of two to six years as predictors with the highest relative contribution to county-level WNV incidence. However, the specific land cover types of importance varied geographically, with increases in urban land cover as an important predictor primarily in the eastern U.S. and increases in agricultural land cover an important predictor primarily in the western U.S. When we assessed predictive accuracy on the 20% test data, we obtained a pseudo-R2 of 0.41. Our next step is to add additional covariates to our dataset that describe changes in land cover patterns, such as patch size and perimeter, the number of like adjacencies (a measure of landscape connectivity), and land cover type richness, evenness, and diversity. This will allow us to determine the relative contribution of changes in land cover composition versus land cover pattern in determining vector-borne disease incidence.