Exploring space-time models for West Nile virus mosquito abundance data

This paper explores variant space-time models for log-transformed West Nile virus (WNv) mosquito data, which explicitly account for both local environmental conditions and complex dependent structures. Four space-time models take various forms to accommodate correlated structure in space and time, nested data, and nonstationarity. The average WNv mosquito abundance is captured by a global trend across all four models, but different model assumptions are imposed on the stochastic component of the proposed models: a simple multivariate linear regression model with independent and identical errors, a site-specific linear mixed model with temporally correlated errors, a week-specific linear mixed model with spatially correlated errors, and a local space-time kriging model. In a case study, the predictive performance of the four models was assessed using data collected in 2007 and 2008 for the Greater Toronto Area by the mosquito surveillance program of Ontario Ministry of Health and Long-term Care: the local space-time kriging model outperforms others, but closely followed by a site-specific linear mixed model with temporal correlation. Our findings suggest that the predictive accuracy of space-time WNv mosquito abundance models can be enhanced by explicitly taking into account spatiotemporal correlation, nonstationarity, and the data collection procedure, such as surveillance design, based on sound understanding of mosquito behavior and population dynamics.