Small area estimation of sparse disease counts using shared component models-application to birth defect registry data in New South Wales, Australia

In the field of disease mapping, little has been done to address the issue of analysing sparse health datasets. We hypothesised that by modelling two outcomes simultaneously, one would be able to better estimate the outcome with a sparse count. We tested this hypothesis utilising Bayesian models, studying both birth defects and caesarean sections using data from two large, linked birth registries in New South Wales from 1990 to 2004. We compared four spatial models across seven birth defects: spina bifida, ventricular septal defect, OS atrial septal defect, patent ductus arteriosus, cleft lip and or palate, trisomy 21 and hypospadias. For three of the birth defects, the shared component model with a zero-inflated Poisson (ZIP) extension performed better than other simpler models, having a lower deviance information criteria (DIC). With spina bifida, the ratio of relative risk associated with the shared component was 2.82 (95% CI: 1.46–5.67). We found that shared component models are potentially beneficial, but only if there is a reasonably strong spatial correlation in effect for the study and referent outcomes.