Performance of a novel machine learning-based proxy means test in comparison to other methods for targeting pro-poor water subsidies in Ghana

Abstract

Access to safe drinking water is still very low among the poorest households in sub-Saharan Africa, and economic shocks can make water access even more difficult for poor consumers. Water subsidies can be a solution to enhance access to safe water services, but they are often ineffective as they regularly fail to reach the very poor. In this study, we developed a new Machine Learning-based proxy means test (ML-based PMT) to identify the poorest households and field-tested it in comparison to four other methods (the Demographic and Health Survey (DHS) wealth index, the Poverty Probability Index (PPI), Community Based Targeting (CBT) and the Ghana Government's Livelihood Empowerment Against Poverty (LEAP) program). We first developed our new ML-based PMT by applying machine learning techniques to the nationally-representative 2016–2017 Ghana Living Standards Survey and compared its performance with an existing PMT (the PPI). We then compared the strengths and weaknesses of this new method in three rural towns of southwestern Ghana against the four other methods, with respect to the characteristics of households they identified, their ease of implementation, their cost, and their acceptability among local stakeholders. In our field assessment we found that our new ML-based PMT performed better than most other approaches at screening out households having assets associated with wealth, but it had higher implementation costs than CBT and LEAP. Local government officials considered CBT to be more transparent than the PMTs, while community members perceived the PMTs to be fairer.

By highlighting the strengths and weaknesses of five different targeting methods, this study provides guidance to practitioners in choosing the most appropriate methods to target poor households eligible for water subsidies in rural Ghana.