July 2022
Comparing Methods for Targeting Water Subsidies to the Poorest Households
MOTIVATION
In Ghana, the poorest households tend to have lower access to safe drinking water, in part due to the cost of improved sources such as piped water systems. Subsidizing safe water services for the poorest can help to address these inequities, but water subsidies are commonly ineffective due to the financial constraints of service providers and unsuccessful targeting that benefits high-income groups. It is critical to find appropriate targeting methods that can accurately identify household poverty, are acceptable to community members and other stakeholders, and can be scaled efficiently.
RESEARCH QUESTION
WHAT IS THE BEST METHOD TO IDENTIFY POOR AND VULNERABLE HOUSEHOLDS FOR WATER SUBSIDIES?
In this study, we compared the performance of five methods for targeting the poorest for water subsidies, identified through the literature and existing practice in Ghana.
METHODS
This study took place in three small towns in the Ahafo and Ashanti Regions of southwestern Ghana. We held six community consultation meetings in neighborhoods within these towns and surveyed all 818 households in these communities. Surveys included questions to assess eligibility with respect to all three proxy-means tests (PMT) and enrollment in the Livelihood Empowerment Against Poverty (LEAP) program. To assess the acceptability of each method, we also conducted qualitative interviews with nine households identified as poor through community consultation, eight households not identified as poor, and six local government officials. We also tracked the costs of implementing each approach.
COMPARING PERFORMANCE
We evaluated the accuracy of the PPI and Aquaya’s ML-based PMT by comparing their predictions with true poverty status (relative to the national poverty line) for 2500 households in the GLSS 7 dataset. The ML-based PMT made slightly more accurate predictions that the PPI (87% of the time compared to 81% for the PPI). Importantly, it made fewer exclusion errors, leaving out 27% of truly poor households as opposed to 43% for the PPI (Figure 1a). It also made fewer inclusion errors: 55% of households predicted to be poor were truly poor, compared to only 40% for the PPI (Figure 1a).
