bioRxiv2019
Geospatial statistical models play an important role in malaria control and prevention; they are widely used to produce malaria risk maps, which are essential to guide efficient resource allocation for intervention. Although many models are available for spatial mapping, the most commonly used model in the literature is the Bayesian geostatistical model (BGM), which is based on an underlying Gaussian process. To our knowledge, methods such as splines and decision trees ensemble methods have not been compared relative to their predictive skill for country level malaria prevalence mapping. Moreover, as more countries now have multiple datasets collected throughout the past decade, it is critical to evaluate if the inclusion of past datasets and the use of spatio-temporal models improve the prediction accuracy of present spatial distribution of malaria. Here we compare the prediction accuracy of five models under spatial and spatio-temporal settings in five African countries. The five models are stepwise logistic regression, generalized additive model (GAM), gradient boosted trees (GBM), Bayesian additive regression trees (BART) and the BGM. There is not a single best model to predict malaria prevalence on a national scale. The model performances varied from country to country, and from spatial to spatio-temporal setting. In general, BGM, GAM and BART models performed well, with BGM being the most consistent. The inclusion of past data is not always beneficial: the predictive performance of GAM and GBM increased under spatio-temporal setting, but BGMs performance decreased in most of the countries. An accurate depiction of malaria risk is important and statistical assumptions that are suitable for a country does not always fit other countries and a wide range of models and settings should be used. It ensures that we find the best modeling approach possible and can provide additional insight to the spatial distribution of malaria risk.\n\nAuthor summaryMalaria is still affecting hundreds of millions of people every year, and killing hundreds of thousands. As the majority of malaria intervention and control policies are developed at the national level, accurate spatial prediction of malaria risk is important. Choosing the best modeling approach for prediction is not straightforward. Here we compare the predictive performance of five models in five countries, with and without dataset from multiple surveys, to provide empirical evidence on whether there is a single best model for national level malaria prediction, and whether inclusion of past dataset may be beneficial in predicting current distribution of malaria risk. We find that models performances vary from country to country and there is no single best model. Although Bayesian geostatistical model is widely and commonly used in the literatures, its performance is not necessarily superior to other simpler-to-fit methods such as general additive model (splines) and Bayesian additive regression tree. Importantly, we also show that the incorporation of past data does not always improve spatial predictions of current disease risk. Together, this demonstrate the importance of fitting wide range of models as part of the prediction mapping process, instead of relying on a one-size-fit-all model.