Rainfall, mosquito density and the transmission of Ross River virus: A time-series forecasting model

Abstract

This paper attempted to develop an epidemic forecasting model using local data on rainfall and mosquito density to predict outbreaks of Ross River virus (RRV) disease in Brisbane, Australia. We obtained monthly data on the counts of RRV cases, monthly total rainfall, human population size and mosquito density (i.e., average number of mosquitoes trapped in all mosquito monitoring stations per month) between 1 November 1998 and 31 December 2001 from the Queensland Department of Health, Australian Bureau of Meteorology, Australia Bureau of Statistics and Brisbane City Council, respectively. Both polynomial distributed lag (PDL) time-series regression and seasonal auto-regressive integrated moving average (SARIMA) models were used to examine associations of RRV transmission with rainfall and mosquito density after adjustment for seasonality and auto-correlation. The results show that 85% and 95% of the variance in the RRV transmission was accounted for by rainfall and mosquito density, respectively. Both rainfall and mosquito density were strong predictors of the RRV transmission in simple models. However, multivariate PDL models show that only mosquito density at lags of 0 and 1 month was significantly associated with the transmission of RRV disease. The SARIMA models show similar results. The findings of this study may facilitate the development of early warning systems for the control and prevention of this disease and other similar vector-borne diseases using local rainfall and/or vector data.

Introduction

Ross River virus (RRV) infection is the most prevalent vector-borne disease in Australia (Harley et al., 2001, Russell, 2002, Gatton et al., 2004). It caused a large epidemic in 1979 and 1980 involving some Pacific Island nations (Aaskov et al., 1981). A recent study suggests that RRV is recurring in Fiji (Klapsing et al., 2005). RRV causes about 5000 cases in Australia per annum. The disease is characterized by headache, fever, rash, lethargy and muscle and joint point, with approximately 50% of patients presenting with rash (MacKenzie and Smith, 1996). The arthritic symptoms may persist for months and can be severe and debilitating. There is no effective treatment for the disease and, in the absence of a vaccine, prevention remains the sole vital public health strategy. Conservatively estimated, the yearly cost is A$ 2.7–5.6 million (Harley et al., 2001). In general terms, RRV activity appears to have increased in Australia in the past decade (Curran et al., 1997, Tong et al., 2001). There is an increasing concern that this disease may also pose a significant risk to other countries due to economic globalisation and increased international travel (Kelly-Hope et al., 2002).

Time-series methodology has a long history of application in econometrics, particularly in the domain of forecasting. Recently it has been used extensively in the assessment of the health effects of environmental exposures (e.g., air pollution and mortality/morbidity) (Bowie and Prothero, 1981, Catalano and Serxner, 1987, Helfenstein, 1991). In environmental health research, there is often an obvious time lag between response and explanatory variables (Schwartz et al., 1996). Some studies approach this by examining models with simultaneous multiple lags of the explanatory variables (Schwartz, 2000). However, serial correlation between these variables may produce unstable estimates (Schwartz et al., 1996). The polynomial distributed lag (PDL) time-series models can reduce the effect of temporal multicollinearity. These models have been used for decades in econometrics (Judge et al., 1980) and recently have been applied in epidemiologic research (Pope and Schwartz, 1996, Schwartz, 2000, Teklehaimanot et al., 2004a, Teklehaimanot et al., 2004b).An advantage of the PDL model is that it does not require a priori the specification of the temporal relationship between the response and explanatory variables, since the degree of the polynomial term can be identified as part of the analysis. This, combined with the flexibility of PDL model in describing a very large range of temporal patterns, makes it an ideal ‘semi-parametric’ choice for epidemiological modelling (Chatfield, 1975).

We have endeavoured to develop an epidemic forecasting system for RRV disease using local weather conditions (Tong et al., 1998, Tong et al., 2004, Tong and Hu, 2001, Tong and Hu, 2002, Hu et al., 2004). In previous studies, we found that disease response to climate variability varied with geographic area (Tong and Hu, 2002) and different climate variables appeared to play different roles in the disease transmission cycles (Tong et al., 1998, Tong et al., 2004, Hu et al., 2004). For example, rainfall seemed to be an important determinant of RRV transmission in Brisbane (Hu et al., 2004). However, the usefulness of vector data in the development of the predictive model still remains to be determined. This paper aims to develop an epidemic forecasting model using local data on rainfall and mosquito density to predict outbreaks of RRV disease in Brisbane, Australia, and to test the applicability of this model in the control and prevention of RRV transmission.