Site-specific prediction of mosquito abundance using spatio-temporal geostatistics

Adult mosquito surveillance programs provide a primary means to understand mosquito vector population dynamics, but such data are typically sparse in space and irregular in time, due to the limit to the available resources for trapping and/or extreme physical conditions. Particularly, missing observations often encountered in long-term surveillance data may hinder a comprehensive analysis of the landscape epidemiology of vector-borne disease and limit our efforts to establish significant and stationary relationship with weather conditions. We developed a West Nile virus (WNV) mosquito abundance model for a sitespecific prediction and associated uncertainty measure, where any point prediction is obtained by a linear combination of a spatio-temporal drift estimate and a stochastic residual. The effects of meteorological and environmental conditions on mosquito population are incorporated in the drift model, while the variations around the drift is modeled by a spatio-temporal random field. The proposed model accounts for discrete counts in nature of the mosquito surveillance data within a generalized linear mixed model and tackles the non-stationarity in WNV mosquito abundance data by limiting the decision of stationarity only to local neighborhoods around any prediction point where the target prediction is sought after.