An evaluation of non-parametric relative risk estimators for disease maps

Abstract In geographical epidemiology it is often required to produce a map of the risk of disease over a study region, a disease map. This paper reviews a variety of approaches to produce disease maps when individual address locations are observed. These methods vary from kernel based smoothing approaches, e.g. Nadaraya–Watson, local linear and GAMs, to Bayesian partition models. The kernel-based methods have the advantage of speed, but the partition model has the advantage of being able to adapt to local features (i.e. clustering) of the surface. Another advantage of the kernel-based methodology is that the local linear model has a built in edge correction. A simulation study designed to assess the benefits of relative merits of each approach for relative risk estimation is described.