Effective Spatial Smoothing of Disease: A Tool for Epidemiologists

Abstract

ISEE-0776 Background and Objective: The availability of routinely collected geocoded data on health, population distribution and environmental pollutants, in conjunction with improved statistical and geographical information science methods, offers an opportunity to assess environmental health risks quickly, at relatively low cost to rapidly address public concerns about perceived environmental health risks. Visualising mortality or morbidity rates and spatial patterns of health outcomes can highlight areas of good or poor health event ascertainment and possibly identify important public health problems. Methods: An application for disease mapping and risk analysis, the Rapid Inquiry Facility (RIF) has been developed by the Small Area Health Statistics Unit (SAHSU) at Imperial College London, in collaboration with the US CDC Environmental Public Health Tracking Program and the EUROHEIS project. Disease mapping is a valuable method for exploring spatial patterns of health outcomes, however, when dealing with low population numbers and/or rare diseases such methods can be problematic. Both rates and SMRs become numerically unstable and typically the less populated areas will show the highest risks. One way of addressing these problems is using Bayesian smoothing methods. Results: The RIF allows users to carry out Empirical Bayes smoothing and to automatically link to WinBUGs to run a number of other smoothing models. We will outline four different smoothing models that can be used with the RIF for: Empirical Bayes, heterogeneity, conditional autoregressive modelling (CAR) and the Besag, York and Mollie (BYM) model and discuss, with examples, advantages of each approach. Conclusion: We will demonstrate that a number of different models should always be run for more accurate interpretation. Furthermore, when adopting the Bayesian approach we will outline the key issues that should be included for the choice of prior distributions of the parameters, in particular for variance parameters and the checking of the convergence.