Abstract
Vector-borne diseases (VBDs) account for more than 17% of all infectious diseases, with an annual global burden of more than 700,000 deaths. In this context, disease risk maps are a key decision-making tool to target control interventions. Yet, these maps are often not sufficiently accurate, as they rarely capture all disease determinants. They usually rely on environmental factors that influence vector presence and vectorial capacity, e.g. climatic and land cover variables. However, socioeconomic factors, such as housing quality or education, and human behaviors, such as the use of preventive measures, are also known risk factors. Environmental and socioeconomic determinants are still rarely combined in prediction models. While environmental factors are commonly derived from satellite-based earth observation, socioeconomic factors are measured through household surveys, which are more time-consuming and challenging to collect. Demographic and Health Surveys (DHS) provide estimates of key demographic and health variables based on nationally representative samples. These include indicators of respondents’ socioeconomic level (e.g. wealth index and education) and use of preventive measures for malaria (e.g. bed nets and indoor spraying). Geolocations for survey clusters are available, and these can be interpolated to create grids that can be further used in spatial models of VBDs. The production of interpolated surfaces of DHS indicators has been widely studied. However, studies do not reach a consensus on the modelling workflow to be used and different types of models have been implemented with different levels of complexity and information required as input. Overall, three types of approaches have been used: (1) spatial interpolation methods, (2) ensemble methods, and (3) Bayesian models. In this research, we focus on malaria as an example of VBDs, and we aim at comparing several methods for predicting DHS indicators influencing malaria risk.
We selected DHS indicators falling into two categories: socioeconomic variables and malaria preventive measures. Using kriging, random forest modelling and spatial Bayesian models (INLA-SPDE implementation), we modelled and predicted these indicators at a high spatial resolution across several sub-Saharan African countries. A set of predictors representing climate, land use and land cover was compiled for use in some of the models.
Different categories of DHS indicators required different modelling approaches; indicators of malaria prevention were best modelled by capturing the spatial autocorrelation pattern, while socioeconomic variables were best predicted with spatial predictors.
These findings highlight the need to test different modelling approaches when mapping human determinants of VBDs risk. Ultimately, these factors can be integrated into a single modelling framework with environmental factors to map malaria risk and identify key drivers. As DHS are conducted consistently in many countries, the methods used in this research are applicable beyond sub-Saharan Africa and could be replicated for other infectious diseases.