Background Good maps of malaria risk have long been recognized as an import
ant tool for malaria control. The production of such maps relies on modelli
ng to predict the risk for most of the map, with actual observations of mal
aria prevalence usually only known at a limited number of specific location
s. Estimation is complicated by the fact that there is often local variatio
n of risk that cannot be accounted for by the known covariates and because
data points of measured malaria prevalence are not evenly or randomly sprea
d across the area to be mapped.
Methods We describe, by way of an example, a simple two-stage procedure for
producing maps of predicted risk: we use logistic regression modelling to
determine approximate risk on a larger scale and we employ gee-statistical
('kriging') approaches to improve prediction at a local level. Malaria prev
alence in children under 10 was modelled using climatic, population and top
ographic variables as potential predictors. After the regression analysis,
spatial dependence of the model residuals was investigated. Kriging on the
residuals was used to model local variation in malaria risk over and above
that which is predicted by the regression model.
Results The method is illustrated by a map showing the improvement of risk
prediction brought about by the second stage. The advantages and shortcomin
gs of this approach are discussed in the context of the need for further de
velopment of methodology and software.