Novissi: How machine learning and phone data helped Togo target emergency COVID-19 cash payments

The problem

When the COVID-19 pandemic reached Togo in early 2020, lockdown measures cut off many people's sources of income. Work in the informal economy dried up, and the risk of hunger rose as households lost income. The government wanted to quickly move emergency cash to the people in greatest need, without in-person contact that could spread the virus.

That aim ran into a practical obstacle. To send money to the poorest communities, a government first has to know who they are and where they are. Togo had no current, comprehensive register of households to draw on, the kind of database that is normally used to decide who qualifies for support. The most recent national census was conducted in 2011, and building a new register during a pandemic was not feasible. Without one, identifying the communities most in need, especially in rural areas, was the central difficulty.

The solution

The Government of Togo built and launched a contactless system called Novissi: people enrolled from a basic mobile phone, no smartphone required, and payments arrived through mobile money. Recipients received roughly US$20 a month for three months. In the first phase, around the capital, the government set eligibility based on a recently updated voter database in which people had stated their location and occupation.

Reaching rural areas posed a harder question, since there was no registry that revealed who was most in need. The government set two goals: direct money to the poorest districts, and within them reach the most poorest individuals. A research team gave technical support, and the government used the method they developed to set eligibility for the rural phase. It worked in two steps, each using a form of artificial intelligence that learns patterns from examples.

Step one: identifying the poorest districts from satellite imagery

The first step used deep learning, a type of artificial intelligence that uses layered networks well suited to reading images. The team paired high-resolution satellite photographs with a pre-pandemic survey of 6,171 households from 2018 and 2019 that recorded the wealth of the surveyed locations. By learning which features visible from above tended to accompany wealth or poverty, the system could estimate the relative wealth of places for which it had no survey. It produced an estimate for each square of land about 2.4 kilometres across, and combined these with population data to rank areas and select the 100 poorest cantons, with a canton being the smallest administrative unit in Togo.

Step two: identifying the poorest individuals

The second step used machine learning trained on mobile-phone records, meaning the pattern of each individual's mobile usage over time. The team first surveyed about 8,915 subscribers in September 2020 to measure each one's level of consumption, then matched those answers to phone-use records from Togo's two mobile networks. The system learned the relationship between how someone used their phone and their consumption, which let it estimate the likely wealth of other individuals from their records alone. The government prioritised those it estimated to be living on less than US$1.25 a day.

The rural expansion was delivered in partnership with the charity GiveDirectly, and the surveys and data work were financed by the World Bank under a six-country West African identification programme, within which about US$72 million was directed to Togo toward social protection delivery systems, including the payments platform.

The impact

1. Cash reached people quickly during the emergency
The first phase of the programme paid 572,852 informal workers between April and September 2020, of whom 373,858 were women. This phase set eligibility from the voter database before the machine-learning method was introduced.

2. The machine-learning method directed the rural phase
To extend the scheme into rural areas, models trained on phone records estimated living standards based on consumption for about 5.7 million people, around 70% of the population. The government used the method to prioritise 57,000 recipients between November 2020 and March 2021.

3. It missed fewer of the eligible recipients than the simpler alternative
For the rural phase, the government's other realistic option was to pay everyone living in the chosen cantons. The drawback of that approach is the risk that payments run out before the poorest individuals are covered. A peer-reviewed evaluation found that the phone-based method reduced the share of eligible people wrongly left out by 4 to 21% compared with the pay-everyone approach. The range is wide because the gain depends partly on how much money is available to distribute.

4. It did not appear to disadvantage particular groups
A risk of using an algorithm to decide who receives aid is that it could be worse at recognising poverty in some groups than in others, meaning eligible people in some groups may be missed more often. The same evaluation checked how often the method wrongly excluded women compared with men, and likewise across ethnic, religious, and age groups, measured against each group's actual poverty rate, and found no such gap.

5. A side effect was wider financial inclusion
The programme opened 170,278 new mobile money accounts, raising mobile money use in Togo by about 7%.

Lessons learned

• New data sources can serve as substitutes when records are missing. Where a government has no current register of households, the study showed that satellite and phone data, read by machine learning, can identify those most in need quickly enough to act in a crisis. The findings suggest this works best as a complement to conventional methods.
• It suits a data emergency more than normal times. The researchers estimated that a conventional approach, a proxy-means test that gauges poverty from visible household characteristics such as assets, would have done better had a detailed social register existed. The phone-and-satellite method earns its place when current data is unavailable and time is short.
• Test an aid algorithm for what it leaves out. Because the model decided who was prioritised for money, the team examined whether it disadvantaged specific groups before relying on it. Building that fairness check into the design, rather than after the fact, is applicable to any government that uses data to allocate support.
• A phone-based system cannot reach people without a phone. Enrolment and targeting both relied on mobile phone use, so people without phones, who are often among the most in need, could be missed. A government adopting this approach needs a parallel route to reach them.

Launch year: 2020

This case study was written with assistance from artificial intelligence.