FLEXIBLE MODELLING IN ENVIRONMENTAL EPIDEMIOLOGY WHEN COVARIATES ARE NONLINEAR, WITH APPLICATIONS TO MALARIA.
Lawrence Ndekeleni Kazembe, Immo Kleinschmidt · Epidemiology · 2005
ISEE-370 Introduction: Malaria is a disease that is influenced by environmental factors such as topography, temperature and rainfall. Often the relationship between the response and covariates is continuous and nonlinear. This paper describes a flexible method for dealing with such scenario. Methods: Using two malaria prevalence datasets from Malawi and the Gambia, we predict malaria risk by using climatic and environmental covariates that are continuous as well as nonlinear. We propose the use of generalized additive models (GAM). GAMs are flexible methods that fit “non-rigid” regression curves adapted to the behaviour of the covariate to the response. A similar model is fitted for comparison, where continuous variables are assumed categorical and fixed. Results: In both models, the covariates were significant, however the goodness of fit of the GAM approach was better than the linear regression and generalized linear models. Conclusion: We have demonstrated use of GAMs in malaria Epidemiology. These models have become popular in environmental and ecological modelling, but their use in malaria is not widespread. GAMs are data-driven rather than model-driven, thereby allow for nonlinearity and non-constant variance structures and are well suited for prediction