Water, Sanitation, and Cholera in Sub-Saharan Africa

Abstract

Improvements in water and sanitation should reduce cholera risk though the associations between cholera and specific water and sanitation access measures remain unclear. We estimated the association between eight water and sanitation measures and annual cholera incidence access across sub-Saharan Africa (2010-2016) for data aggregated at the country and district levels. We fit random forest regression and classification models to understand how well these measures combined might be able to predict cholera incidence rates and identify high cholera incidence areas. Across spatial scales, piped or "other improved" water access was inversely associated with cholera incidence. Access to piped water, septic or sewer sanitation, and septic, sewer, or "other improved" sanitation were associated with decreased district-level cholera incidence. The classification model had moderate performance in identifying high cholera incidence areas (cross-validated-AUC 0.81, 95% CI 0.78-0.83) with high negative predictive values (93-100%) indicating the utility of water and sanitation measures for screening out areas that are unlikely to be at high cholera risk. While comprehensive cholera risk assessments must incorporate other data sources (e.g., historical incidence), our results suggest that water and sanitation measures could alone be useful in narrowing the geographic focus for detailed risk assessments.

Keywords: geographic classification; infrastructure access; population-level analysis; random forest; risk analysis.

Figure 1

Water and sanitation measures and incidence rate of suspected cholera (2010–2016): (A) mean water and sanitation measures and log of mean annual incidence of suspected cholera per 1000 population by country; scatter plots with point color indicating mean annual incidence of suspected cholera cases per 1000 people as a function of (B) reliance on surface water and open defecation (extremes) by district; and (C) piped or other improved water and septic, sewer, or other improved sanitation by district. Univariate histograms of district-level measures shown along the axes of panels (B) and (C) are in blue and orange.

Figure 2

Risk ratio of water and sanitation measures from univariate models at country (Quasi-Poisson) and district scales (Poisson GEE). Reference groups (denoted “ref”) for each model are included to compare the risk ratios.

Figure 3

Summary of random forest models’ performance. Panel A illustrates the observed mean annual incidence versus the predicted values in cross-validation in the heatmap with each hexagon showing count of points. The root-mean-square error (RMSE) between the observed and model-predicted estimates and the correlation (\rho) between the two are noted in the panel. Panel B illustrates the cross-validated receiver operator characteristic curve from the random forest classification model used to identify high incidence areas. Youden cutoff point (jointly maximized sensitivity and specificity) shown as a red dot in panel B.