Social protection programs are essential to assisting the poor, but governments and humanitarian agencies are rarely resourced to provide aid to all those in need, so accurate targeting of benefits is critical. In developed economies, targeting decisions typically rely on administrative income data or broad survey-based social registries. In low-income countries, however, poverty information is rarely reliable, comprehensive, or up-to-date. Novel sources of digital data — from mobile phones and satellites, in particular — are well suited to fill this gap: they are predictive of wealth in low-income contexts and ubiquitously collected. The research studies in this dissertation design and evaluate new methods for targeting aid in low-resource contexts using machine learning, satellite imagery, and mobile phone data, and evaluate these methods in large, real-world interventions. Across social protection programs in Togo, Afghanistan, and Bangladesh, the studies in this dissertation show that targeting methods based on machine learning and digital data sources identify poor households more accurately than methods based on categorical eligibility criteria like geography or occupation, but typically less accurately than traditional survey-based poverty measurement approaches. These results highlight the potential for digital data and machine learning to improve the targeting of humanitarian aid, particularly when traditional poverty data are unavailable or out-of-date and in settings where conflict, environmental conditions, or health concerns render primary data collection infeasible. These studies also provide empirical evidence on the limitations and risks of digital and algorithmic targeting approaches, including privacy, transparency, fairness, and digital exclusion. 

Personal mobility data from mobile phones and other sensors are increasingly used to inform policymaking during pandemics,natural disasters, and other humanitarian crises. However, even aggregated mobility traces can reveal private informationabout individual movements to potentially malicious actors. This paper develops and tests an approach for releasing privatemobility data, which provides formal guarantees over the privacy of the underlying subjects. Specifically, we (1) introduce an algorithm for constructing differentially private mobility matrices, and derive privacy and accuracy bounds on this algorithm; (2) use real-world data from mobile phone operators in Afghanistan and Rwanda to show how this algorithm can enable the use of private mobility data in two high-stakes policy decisions: pandemic response and the distribution of humanitarian aid; and (3) discuss practical decisions that need to be made when implementing this approach, such as how to optimally balance privacy and accuracy. Taken together, these results can help enable the responsible use of private mobility data in humanitarian response.

The COVID-19 pandemic has devastated many low- and middle-income countries, causing widespread food insecurity and a sharp decline in living standards¹. In response to this crisis, governments and humanitarian organizations worldwide have distributed social assistance to more than 1.5 billion people². Targeting is a central challenge in administering these programmes: it remains a difficult task to rapidly identify those with the greatest need given available data^3,4. Here we show that data from mobile phone networks can improve the targeting of humanitarian assistance. Our approach uses traditional survey data to train machine-learning algorithms to recognize patterns of poverty in mobile phone data; the trained algorithms can then prioritize aid to the poorest mobile subscribers. We evaluate this approach by studying a flagship emergency cash transfer program in Togo, which used these algorithms to disburse millions of US dollars worth of COVID-19 relief aid. Our analysis compares outcomes-including exclusion errors, total social welfare and measures of fairness-under different targeting regimes. Relative to the geographic targeting options considered by the Government of Togo, the machine-learning approach reduces errors of exclusion by 4-21%. Relative to methods requiring a comprehensive social registry (a hypothetical exercise; no such registry exists in Togo), the machine-learning approach increases exclusion errors by 9-35%. These results highlight the potential for new data sources to complement traditional methods for targeting humanitarian assistance, particularly in crisis settings in which traditional data are missing or out of date.

The COVID-19 pandemic has devastated many low- and middle-income countries (LMICs), causing widespread food insecurity and a sharp decline in living standards. In response to this crisis, governments and humanitarian organizations worldwide have mobilized targeted social assistance programs. Targeting is a central challenge in the administration of these programs: given available data, how does one rapidly identify the individuals and families with the greatest need? This challenge is particularly acute in the large number of LMICs that lack recent and comprehensive data on household income and wealth. Here we show that non-traditional “big” data from satellites and mobile phone networks can improve the targeting of anti-poverty programs. Our approach uses traditional survey-based measures of consumption and wealth to train machine learning algorithms that recognize patterns of poverty in non-traditional data; the trained algorithms are then used to prioritize aid to the poorest regions and mobile subscribers. We evaluate this approach by studying Novissi, Togo’s flagship emergency cash transfer program, which used these algorithms to determine eligibility for a rural assistance program that disbursed millions of dollars in COVID-19 relief aid. Our analysis compares outcomes – including exclusion errors, total social welfare, and measures of fairness – under different targeting regimes. Relative to the geographic targeting options considered by the Government of Togo at the time, the machine learning approach reduces errors of exclusion by 4-21%. Relative to methods that require a comprehensive social registry (a hypothetical exercise; no such registry exists in Togo), the machine learning approach increases exclusion errors by 9-35%. These results highlight the potential for new data sources to contribute to humanitarian response efforts, particularly in crisis settings when traditional data are missing or out of date.