Spatial interpolation of health and demographic variables: Predicting malaria indicators with and without covariates

Accurate mapping and disaggregation of key health and demographic risk factors have become increasingly important for disease surveillance, which can reveal geographical social inequalities for improved health interventions and for monitoring progress on relevant Sustainable Development Goals (SDGs). Household surveys like the Demographic and Health Surveys have been widely used as a proxy for mapping SDG-related household characteristics. However, there is no consensus on the workflow to be used, and different methods have been implemented with varying complexities. This study aims to compare multiple modelling frameworks to model indicators of human vulnerability to malaria (SDG Target 3.3) in Senegal. These indicators were categorised into socioeconomic (e.g., stunting prevalence, wealth index) and malaria prevention indicators (e.g., indoor residual spraying, insecticide-treated net ownership). We compared three categories of the commonly used methods: (1) spatial interpolation methods (i.e., inverse distance weighting, thin plate splines, kriging), (2) ensemble methods (i.e., random forest), and (3) Bayesian geostatistical models. Most indicators could be modelled with medium to high predictive accuracy, with R² values ranging from 0.40 to 0.86. No method or method category emerged as the best, but performance varied widely. Overall, socioeconomic indicators were generally better predicted by covariate-based models (e.g., random forest and Bayesian models), while methods using spatial autocorrelation alone (e.g., thin plate splines) performed better for variables with heterogeneous spatial structure, such as ethnicity and malaria prevention indicators. Increasing the complexity of the models did not always improve predictive performance, e.g., thin plate splines sometimes outperformed random forest or Bayesian geostatistical models. Beyond performance, we compared the different methods using other criteria (e.g., the ability to constrain the prediction range or to quantify prediction uncertainty) and discussed their implications for selecting a modelling approach tailored to the needs of the end user.