Bayesian hierarchical modelling to enhance the epidemiological value of abattoir surveys for bovine fasciolosis

Four classes of Bayesian hierarchical models were evaluated using an historical dataset from an abattoir survey for fasciolosis conducted in Victoria, Australia. The purpose of this analysis was to identify areas of high prevalence and to explain these in terms of environmental covariates. The simplest of the Bayesian models, with a single random effect, validated the use of smoothed maps for cartographic display when the sample sizes vary. The model was then extended to partition the random effect into spatially structured and unstructured components, thus allowing for spatial autocorrelation. Rainfall, irrigation, temperature-adjusted rainfall and a remotely sensed surrogate for rainfall, the normalised difference vegetation index (NDVI), were then introduced into the models as explanatory variables. The variable that best explained the observed distribution was irrigation. Associations between prevalence and both rainfall and NDVI that were significant in fixed effects models were shown to be due to spatial confounding. Nevertheless, provided they are used cautiously, confounded variables may be valid predictors for the prevalence of disease.