Spatial and Temporal Modelling of Ross River Virus in Queensland

Ross River virus (RRv), also known as Epidemic Polyarthritis, is a debilitating disease and is the most prevalent vector-borne disease in Australia (Lin et al. 2002). The virus can survive and replicate in humans and other vertebrae hosts, and is transmitted by a variety of mosquito vectors (Russell and Dwyer 2000). The disease in humans is nonfatal and infections can be either asymptomatic or symptomatic, with symptoms including polyarthritis, rash, fever, myalgia, and lethargy (Harley et al. 2001). There has been much recent research into the spatial and temporal nature of Ross River virus in Queensland (Gatton et al. 2004; Kelly-Hope et al. 2004; Tong and Hu 2002). A recent paper by Gatton et al. (2004) focussed on the spatial and temporal nature of outbreak periods, where outbreak periods are defined by comparison against long term incidence rates specific to that area. The spatial and temporal nature of outbreak periods is of public health importance as increased understanding will lead to more targeted public health interventions (Tong 2004). In this paper, we use a Bayesian mixture model to analyse weekly cases of Ross River virus in Queensland from 1984 to 2001. RRv notification data was obtained from the Communicable Diseases Section of Queensland Health. An exploratory analysis revealed an association between climate variables and cases of RRv, so we aggregated the data to fifteen homogenous climate zones representing Queensland. We explore a mixture model to separate the RRv data over time into a number of states or components, and use model choice criteria to choose which number of components is preferable. This is an extension of previous work on RRv which has focussed on two components or states, an outbreak state and non-outbreak state. The method also allows the data to indicate the component (state) in which it belongs, and thereby avoid possibly subjective decision rules. Extensions to more than two components is expected to offer flexibility in cases where, for example, hyperoutbreak periods can be identified. The choice between competing models of a different number of components invariably involves a selection criteria that will take into account both measures of fit and complexity. In this paper we use methodology developed in Celeux et al. (2003) and choose between competing models based on Deviance Information Criterion (DIC) estimates. The parameters for the different models were estimated by Markov Chain Monte Carlo (MCMC) using the software package WinBUGS (Spiegelhalter et al. 2002). We focussed the analysis on two different climate zones which appeared to display different temporal behaviour, and found much variability in the results, with a lower number of components preferred for data from the zone which appeared to show a more distinctive pattern.