Statistical methodological issues in mapping historical schistosomiasis survey data

Bayesian geostatistical variable selection with a parameter expanded normal mixture of inverse gamma prior identified altitude, rainfall, night land surface temperature and soil acidity as the most important environmental predictors for Schistosoma mansoni infection risk in Côte d’Ivoire. •We discuss Bayesian computational issues for large dataset.•Our modelling approach reveals heterogeneity of historical schistosomiasis survey data.•We explore stationary and isotropy hypotheses.•Bayesian geostatistical variable selection was employed for block of covariates with a parameter expanded normal mixture of inverse-gamma prior.•We present parsimonious models for Schistosoma mansoni risk prediction for Côte d’Ivoire before 2000 and from 2000 onwards. For schistosomiasis and other neglected tropical diseases for which resources for control are still limited, model-based maps are needed for prioritising spatial targeting of control interventions and surveillance of control programmes. Bayesian geostatistical modelling has been widely and effectively used to generate smooth empirical risk maps. In this paper, we review important issues related to the modelling of schistosomiasis risk, including Bayesian computation of large datasets, heterogeneity of historical survey data, stationary and isotropy assumptions and novel approaches for Bayesian geostatistical variable selection. We provide an example of advanced Bayesian geostatistical variable selection based on historical prevalence data of Schistosoma mansoni in Côte d’Ivoire. We include a “parameter expanded normal mixture of inverse-gamma” prior for the regression coefficients, which in turn allows selection of blocks of covariates, particularly categorical variables. The implemented Bayesian geostatistical variable selection provided a rigorous approach for the selection of predictors within a Bayesian geostatistical framework, identified the most important predictors of S. mansoni infection risk and led to a more parsimonious model compared to traditional selection approaches that ignore the spatial structure in the data. In conclusion, statistical advances in Bayesian geostatistical modelling offer unique opportunities to account for important inherent characteristics of the Schistosoma infection, and hence Bayesian geostatistical models can guide the spatial targeting of control interventions.