Machine learning and phone data can improve targeting of humanitarian aid

The COVID-19 pandemic has devastated many low- and middle-income countries, causing widespread food insecurity and a sharp decline in living standards. In response, governments and humanitarian organizations distributed social assistance to more than 1.5 billion people. Targeting—the challenge of rapidly identifying those with greatest need—remains difficult given limited data. This study shows that mobile phone network data can improve the targeting of humanitarian assistance. Using traditional survey data to train machine-learning algorithms to recognize patterns of poverty in mobile phone data, the trained models can prioritize aid to the poorest mobile subscribers. The approach is evaluated via Togo’s Novissi emergency cash transfer program, which used these algorithms to disburse millions of USD in COVID-19 relief. Outcomes—including exclusion errors, social welfare, and fairness—are compared under different targeting regimes. Relative to feasible geographic targeting options, the phone-based machine-learning approach reduces exclusion errors by 4–21%; relative to methods requiring a comprehensive social registry (hypothetical), it increases exclusion errors by 9–35%. These results highlight the potential for new data sources to complement traditional methods for targeting assistance, particularly when traditional data are missing or outdated.

Emily Aiken, Suzanne Bellue, Dean Karlan, Chris Udry, Joshua E. Blumenstock