Confounding effects via spatially-varying demographic factors upon West Nile virus surveillance in Illinois identified using a Bayesian functional data analysis model

In previous work, we applied a spatially-aware graph neural network model to West Nile virus data collected in Illinois for mosquito surveillance. The purpose of this was to aid mosquito abatement efforts within the state while accounting for environmental variability. Other studies have also taken this data to examine links between West Nile virus levels and spatially-varying demographic factors, but did not identify any evidence of such links. This raises the question of whether the inability to identify such links could be due to confounding via surveillance levels related to the demographic factors. That is to say, such links may be disguised by varying West Nile virus surveillance data quality across Illinois, and this variance in data quality may be related to the demographic factors themselves. In our work, we examine the spatial trends of surveillance frequency in this data within the Chicago area using a functional data analysis model within a Bayesian hierarchical model that is implemented in the Stan probabilistic programming language. We then relate these trends to zipcode-level demographic factors and present the likelihood for the existence of the statistical relationships in question. The use of a functional data analysis method allows for increased flexibility in our choice of model, such that it more closely reflects the reality supported by our sources from various fields in the literature. Furthermore, our use of a Bayesian hierarchical framework allows for greater interpretability of our findings, at the cost of greater required computational resources for model fitting.