Nonstationary geostatistical modelling of malaria prevalence using ecological covariates: A five-country analysis of national survey data in sub-saharan Africa

Abstract

Geostatistical models are widely used to analyze malaria data and generate spatial predictions based on the principle that nearby locations are more similar than those farther apart. However, when environmental factors influence both the prevalence and variability of malaria, the stationarity assumption may be violated. This study addresses this limitation by integrating ecological covariates, specifically the Enhanced Vegetation Index (EVI), directly into the covariance structure of a geostatistical model to capture spatio-environmental dependency and nonstationarity. Using data from national malaria intervention surveys conducted in Angola, Burkina Faso, Malawi, Mozambique, and Tanzania, we compared the proposed nonstationary model with a conventional stationary alternative. The average malaria prevalence among children under five was 20.77%, ranging from 7.28% in Tanzania to 38.9% in Mozambique. The nonstationary model demonstrated improved model fit, as indicated by lower Akaike Information Criterion (AIC) values, and enhanced predictive performance, with reductions in mean absolute error (MAE) and root mean square error (RMSE) in most countries. The resulting risk maps identified areas of elevated malaria prevalence, including the northwest and southeast of Tanzania and the central regions of Burkina Faso, Malawi, and Mozambique. By explicitly modeling the influence of environmental factors on spatial dependence, this approach provides a more flexible and informative framework for malaria risk mapping in environmentally heterogeneous settings across sub-Saharan Africa. Future research should explore incorporating multiple covariates into the covariance structure to improve predictive accuracy further.