MobileAid: Transforming Aid Delivery for Global Crises with Big Data

Our solution identifies millions of the most economically disadvantaged individuals using machine learning and sends them cash remotely at unprecedented speed. This represents a paradigm shift for the social protection sector, with data science fundamentally transforming how aid actors can deliver relief to society’s most vulnerable groups when crises hits.

Han Sheng Chia, GiveDirectly’s Innovation VP, has grown cash operations from $1M in seed funding to $120M+ globally and advises governments, philanthropists, and agencies on the transformation of aid delivery. 

COVID-19 has pushed an estimated 124M people into extreme poverty globally, the first increase in 20 years. Cash transfers are one of the most proven and effective approaches to combating poverty. While the use of cash transfers in social protection programs was growing steadily prior to COVID-19, the pandemic galvanized aid actors to move faster and expand program coverage.

Traditional social protection programs delivering cash rely on extremely expensive and time consuming models often subject to exclusion errors for identifying/enrolling recipients and delivering aid. Low-income countries often lack the resources to maintain national social registries - and, as reported by the World Bank, such registries often exclude the most vulnerable population groups. At a time when people are rapidly falling into poverty, using outdated registries to dispense relief, relying on door-to-door visits to screen recipients, and moving slowly to deliver cash can leave behind those most impacted and compound the secondary negative impacts of a crisis. 

These flaws in traditional cash-based social protection models, while exacerbated by the COVID-19 pandemic, are not new. The growing number of governmental and aid actors relying on cash to provide social protection need new tools to move faster and reach the most vulnerable, now.

The target audience for this solution are global aid actors, including governments, INGOs, and NGOs, that operate in the social protection sector and can apply this model and/or our findings to improve the accuracy, inclusion, speed, and scale of their cash programs. Our solution provides an alternative to expensive, time consuming, and exclusionary traditional models (like occupation-based targeting or door-to-door enrollment) that aid actors have relied on in the past to identify people in economic distress. 

The target beneficiaries of our solution are the individuals most economically affected by the COVID-19 pandemic, many of whom have lost their livelihoods entirely. We will deliver contactless, unconditional cash transfers to them, giving them the resources and dignity to decide how to best recover from this crisis. Recipients will be engaged throughout the implementation period, via SMSes, radio, community leaders, and surveys.

We are currently working with the government of Togo, other countries, and various UN/multilateral agencies to further develop this model. Policymakers from several other countries have also expressed interest. Furthermore, we will coordinate an international Community of Practice (CoP) to effectively disseminate our findings and support collaborative, machine-learning-based models to global crises.

All code used in our machine-learning-based poverty prediction models will be made available under an open source license through a public Git repository. 

We will also document all of our work in academic journal articles and public-facing blog posts, as well as disseminate this information through our international Community of Practice (CoP).. 

All academic journal articles produced will also be added to GiveDirectly’s free-to-use cash research explorer platform, which makes it easy for people to access peer-reviewed papers, working papers, and program reports about the impacts of cash transfers.

First and foremost, our project will enable aid actors to identify individuals in need and deliver aid to them at a much greater speed and scale than ever before, based on our success in Togo. Our solution will demonstrate an improved model for identifying, enrolling, and paying the extreme poor and inform future social protection programming - empowering governments, INGOs, and NGOs to recover better from COVID-19 and respond faster to future crises. 

We will also make all code publicly available under an open source license and document our findings in public-facing articles and blogs. Connecting aid actors to these materials will be facilitated through our international CoP, governments partnerships, and ongoing dissemination activities. 

Secondly, our solution will disperse millions into the hands of recipients reeling from the economic impacts of COVID-19, some of whom were excluded from traditional programs, while giving them the autonomy to decide how best to meet their needs. We expect cash to have a broad range of positive impacts, including: (1) coverage of basic needs (based on our COVID-19 programs in Kenya, Malawi, Rwanda, and Liberia), (2) investments in livelihoods, (3) investments in human capital, and (4) debt repayment.

After our successful pilot in Uganda, we chose to work in Togo because of the high need (50%+ of the population lives in poverty) and strong on-the-ground partners, including a forward-thinking Ministry of Digital Economy. We have paid ~115K recipients in Togo and in the next year plan to double the geographic reach of the program while refining our targeting model. 

To scale our impact and preposition aid actors to respond rapidly to future crises, we have to expand further. In the next three years, we aim to work in new LMICs, focusing on those that are most populous (Bangladesh, Nigeria, DRC) and/or prone to repeat crises (Malawi, Uganda, Kenya). We are already in discussion with the several governments and multilateral agencies to replicate our solution. We are also actively seeking additional funding streams to support this expansion.

While many high quality studies have shown that cash has positive impacts on diverse recipient outcomes, it has been unclear how aid actors can effectively scale identification, enrollment, and payment processes. Our solution provides an answer. Once the upfront costs of launching our model are covered, the underlying technology can be scaled rapidly and inexpensively, as demonstrated in Uganda and Togo.

We have two sets of indicators for monitoring impact and evaluating success: 

(1) Targeting performance indicators include measures of inclusion and exclusion error like precision and recall. We also measure targeting performance based on the magnitude of algorithmic bias and underrepresentation of specific vulnerable populations. In Togo, our model had a precision of 0.9 (of the 5M individuals identified as poor, 90% were indeed poor) and a recall of 0.35 (of all the projected poor individuals in Togo, 35% were classified as poor by our algorithm). We showed that our model was 10% more precise and 2.5x more inclusive than the Togolese government’s next best alternative model.

(2) Cash transfer performance indicators include measures like time taken to enroll the target number of recipients, transfer success rates, number of adverse events reported, and more. These metrics are tracked and stored in our data management system. In Togo, our data shows that >95% of the target number of recipients were enrolled on time, 96% of transfers were successful on the first try, and <1% of recipients reported adverse events. We are currently rolling out our long-form follow-up surveys in Togo to collect additional cash transfer performance data directly from recipients.

Three primary barriers exist:

1. Policy adoption takes time: Achieving our goal of helping countries preposition for the next global crisis requires more than technology development. It also requires planning with policymakers and getting buy-in from key stakeholders from the start; garnering international attention, thought leadership, and press; and driving momentum across multiple levels of government. GiveDirectly and CEGA have accomplished this to date in Togo, and are working towards global expansion.
2. Context-specific regulations can lead to higher costs: We have operated across various contexts; some countries require telecommunications data to be managed in-country, while others allow data to be stored in a secure, international facility or in the cloud. The latter is more cost effective for our team, as we do not have to place our experts in-country for prolonged periods. However, higher costs can be mitigated through strong upfront project management, negotiation, and cost management.
3. There exists political instability in target countries: Many of the countries in which we work or hope to work face political instability. This is not new for GiveDirectly or CEGA. We mitigate negative potentialities by maintaining strong government relations and establishing playbooks on managing key stakeholder relationships and emergency events.

We are applying to the Trinity Challenge to ensure that the economic devastation caused by this global pandemic never happens again. We aim to fundamentally transform how countries respond to and recover from the economic impact of humanitarian crises, and agree with Professor Dame Sally Davies that we must apply the best ideas to prevent a similar disaster in the future. Never again should someone in a poor country have to choose between infection by a deadly virus or feeding their children. Never again should long lines of hungry families wait outside of aid offices for in-kind aid because lockdowns have destroyed their incomes. Economic resilience during a major health crisis means responsive, prepositioned systems that deliver cash immediately.

To achieve this vision at scale, the specific barrier that the Trinity Challenge can help us with is that policy adoption takes time and achieving our goal requires early adoption from key stakeholders. The Trinity Challenge can help us convene stakeholders across multiple sectors to further the policy conversation and drive political will, thought leadership, and press. We believe the Challenge’s emphasis on building global coalitions and deep ties across academic, public, and private sectors will help us achieve our goal. 

We are currently partnering with Google and Facebook, two of the Trinity Challenge’s founding member organizations, who have provided datasets, funding, and technical expertise. Google has provided engineers to help strengthen some of our technology and funding over the last eight years, while Facebook has provided datasets on population density. Some of our non-member partner organizations include the Brookings Institute, U.C. Berkeley, and telecommunications and satellite companies.

We would like to partner with global academic institutions like the University of Cambridge, the London School of Economics, HKU Med, Nanyang Technological University, Tsinghua University, and the University of Melbourne. While we have a strong academic partner in U.C. Berkeley, a global solution requires global academic validation and support. We also seek to demonstrate to an international community that data science can deliver great social impact at a fraction of the cost compared to more traditional models, proving the case that data science can be the driver for social impact at scale. Global academic partners will be integral to furthering this goal.