Statistical Modelling of Mosquito Abundance and West Nile Virus Risk with Weather Conditions

Weather affects the abundance of mosquito vectors of mosquito-borne infectious diseases such as West Nile virus (WNv). Study and prediction of these effects could be used to develop disease forecasting methods. In this dissertation, we analyzed the frequency distribution of mosquito surveillance data and built the statistical forecasting models to predict the West Nile virus risk. In the first part, using mosquito data from the surveillance program in Peel Region, Ontario, we studied the distribution properties of Culex mosquito abundance data for the period from 2004 to 2012. We first employed statistical clustering method to identify two clusters of mosquito traps. The validation against landuse data supported the hypothesis that the clustering result successfully captured the influence of geographic variation in habitat effects on mosquito abundance. Accounting for the occurrence of these clusters, distribution analysis showed that Culex mosquito abundance in Peel Region followed a gamma distribution. Further analysis showed that summer mean temperature, but not precipitation has a significant effect on mosquito distribution properties. We defined a normal weather threshold under which the mosquito abundance followed a gamma distribution and abnormal weather conditions under which the mosquito abundance deviated from a gamma distribution. A predictive statistical model by clusters to forecast mosquito abundance in Peel Region using weather conditions was developed. In the second part, we developed forecasting models to predict the Culex mosquito abundance, the WNv risk and human incidence in Great Toronto Area (GTA) under weather changes by model selection. The predictions were in a good agreement with the observations for the period from 2002 to 2012. The model selection was demonstrated to be an effective way to compare different models. In the final part, finite mixture model and Markov regression models were combined to develop model-based clustering with generalized linear regression to cluster time series. Quasi-likelihood approach was adopted to deal with the Markov chain in the data generating process and Estimation-Expectation algorithm was used to estimate the parameters. The proposed algorithm was tested on simulated data and applied to mosquito surveillance data in Peel Region.