Expanding disease modeling beyond the spatial footprint of the data: regional and national models of human West Nile virus cases

Abstract:

Human West Nile virus (WNV) cases have been modeled and predicted using weather, entomological, and other data. It is often assumed that predictive covariates, such as the rate of infection of the mosquito with the virus, are only predictive over small spatial scales; a person’s risk in a park is assumed to be best estimated by collecting nearby mosquitoes. Mosquito control decisions are often based on similar scales; spraying will occur in response to positives in mosquitoes or humans, for example.

In previous work, however, we have repeatedly found that the infection status of mosquitoes sampled several counties away are strongly predictive of human risk in counties across an entire state. There are strong state-wide relationships among environmental conditions, mosquito infection rates, and human risk that transcend the assumption that you need data from here to predict risk here.

Using mosquito infection data obtained from South Dakota and Louisiana, human data from ArboNET, and GridMET meteorological data derived from NASA’s National Land Data Assimilation System, we developed a national model of per-county human risk in which mosquito infections were found to predict human cases outside the boundaries of these two states. We found that sampling mosquitoes in a small number of counties has improved predictions of WNV risk in neighboring states, contrary to the usual but unspoken assumption that the data’s utility stops at a county’s boundary.

To generalize, we considered models in which states other than SD and LA were allowed act as sentinels in a hidden-variable model, as if mosquito infection data were available as predictors for a national audience. Human WNV cases reported to ArboNET by county and year were modeled as a function of publicly available environmental data and mosquito data (with estimated derived from their observed human cases), and we found combinations of sentinel states that predict regional and national human WNV risk parsimoniously. We described the relationship between participation and the ability to predict human risk, and considered various scenarios in which data-sharing would permit an accurate national-level model of human WNV risk.