Using AI to Target Emergency COVID-19 Assistance in Togo
This post, written by Program Associate Anya Marchenko, describes an AI-based approach to targeting cash transfers that is currently being implemented in partnership with the Government of Togo. The project is part of CEGA’s Targeting Aid Better initiative, which supports innovative approaches to targeting social protection programs in low- and middle-income countries.
“People should not have to choose between death by Covid-19 or by hunger”
— Faure E. Gnassingbé, President of Togo
Poverty and the pandemic
COVID-19 and related lockdowns have been especially hard for the hundreds of millions of people living in poverty around the world. In low-income countries, lack of access to savings and robust social protection programs leave families uniquely vulnerable to economic disruptions. For example, in a separate post, I wrote about how after the COVID-19 lockdown started, spending on food by families in rural Western Kenya dropped by almost half compared to before the pandemic started.
To soften the blow, most countries have launched some kind of emergency assistance program. Ugo Gentilini, an economist at the World Bank, has been diligently keeping track of every social protection policy that countries have launched in response to COVID-19. Gentilini estimates that cash transfers represent fully one-third of all COVID-related social protection programs around the world. But delivering cash is expensive, and there’s only so long governments can keep funding these programs. When funding is tight, it’s important that it gets into the hands of the people who need it the most.
Unfortunately, low-income governments often lack reliable information about who and where their poorest citizens are, making the task of delivering emergency aid extremely challenging. Many of those needing assistance are unbanked and self-employed (or employed in the informal sector), with no tax records or formal identification for officials to reference. Government censuses and other administrative records are often incomplete or out of date, and conducting surveys on short notice — particularly in the middle of a pandemic — is simply not a viable option.
Targeting aid “intelligently”
As it turns out, wealth — and by extension, poverty — can be predicted with a fairly high degree of accuracy using data that already exist. Research by CEGA faculty has shown how high-frequency data from mobile phone operators and satellite networks, combined with machine learning approaches, can be used to rapidly identify households in direst need. These and related methods carry enormous potential to help policymakers quickly and effectively respond to the current crisis.
A research team led by Joshua Blumenstock, CEGA Faculty Co-Director and Associate Professor at the UC Berkeley School of Information, and involving researchers from Northwestern University and Innovations for Poverty Action, has been working with the Government of Togo and the NGO GiveDirectly to develop and test an approach to targeting COVID-19 aid that leverages big data from satellites, mobile phone data, and other sources. Their approach, described in greater detail here, has two key components.
First, they use deep-learning algorithms (machine learning algorithms that use multiple layers to progressively extract higher-level features from the raw input) to process satellite images. Satellite images show the quality of roofing on a house or the density of roads, which correlate well with actual economic conditions. So long as a sufficient amount of high-quality “ground-truth” data is available, algorithms can be trained to generate high-resolution poverty maps from the satellite images. These maps can help governments understand patterns of wealth and poverty in their country, and aid them in making important resource allocation decisions.
Blumenstock and his colleagues then use mobile phone metadata to target emergency aid to individuals within the areas already flagged with the satellite approach. These data (referred to as call detail records, or “CDR”) are useful because people living in poverty tend to use their phones differently from wealthier individuals: low-income users make shorter calls, have fewer contacts, and carry around a lower balance in their mobile money accounts. Blumenstock’s team has worked with CDR data from Afghanistan and other countries to develop algorithms that — when trained with high-quality survey data — can effectively identify which phone users are most likely to be below a certain income level.
Since April, Blumenstock’s team has been leveraging these two approaches in concert in an effort to help the Government of Togo identify which individuals they should prioritize for cash transfers. Of course, the use of such data— especially private consumer data from mobile phone operators — requires very careful consideration. The team is taking extensive precautions to protect the privacy of beneficiaries and to ensure that algorithms used to decide consequential aid allocations are developed in a way that is fair and equitable. While technology-based approaches to targeting have the potential to be far more accurate than conventional methods, it’s imperative that these tools are developed with careful human oversight.
A new approach to doing research
Blumenstock’s work represents a new way of doing business for CEGA. Findings from rigorous impact evaluations (CEGA’s core “product”) often take years to materialize — much longer than most policymakers tend to find useful. In this case, the research team is providing direct support to a partner government, drawing on insights from past research. In this way, the team is able to provide immediate value to Togo, while carefully generating insights that could inform similar targeting approaches in other countries and contexts.
CEGA’s Targeting Aid Better initiative
With generous support from an anonymous donor, CEGA launched our Targeting Aid Better (“Targeting”) initiative in spring 2020 to facilitate the development and dissemination of novel methods that governments and development partners can use to dynamically target emergency aid to the poorest and most vulnerable. Stay tuned for the blog post announcing our first set of competitively funded projects in a few weeks. We have also launched a Targeting Aid Better newsletter that will provide regular updates to subscribers regarding efforts to develop and test novel targeting approaches. We encourage you to subscribe here by selecting our “Data Science for Development (Targeting)” group.
