Improving national level spatial mapping of malaria through alternative spatial and spatio-temporal models

The most common approach to create spatial prediction of malaria in the literature is to approximate a Gaussian process model using stochastic partial differential equation (SPDE). We compared SPDE to computationally faster alternatives, generalized additive model (GAM) and state-of-the-art machine learning method gradient boosted trees (GBM), with respect to their predictive skill for country-level malaria prevalence mapping. We also evaluated the intuition that incorporation of past data and the use of spatio-temporal models may improve predictive accuracy of present spatial distribution of malaria. Model performances varied among the countries and setting with SPDE and GAM performed well generally. The inclusion of past data is beneficial for GAM and GBM, but not for SPDE. We further investigated the weaknesses of SPDE at spatio-temporal setting and GAM at the edges of the countries. Taken together, we believe that spatial/spatio-temporal SPDE models should be evaluated alongside with the alternatives or at least GAM.