In response to the COVID-19, we design Togo’s flagship social protection program, Novissi, using insights from machine learning. Research team then open-sourced the approach, creating Cider, a set of software tools for estimating the poverty status of individuals and households using mobile phone data, for use in the targeting and delivery of humanitarian aid.
The team first began by applying machine learning algorithms to high-resolution satellite imagery to develop micro-estimates of the relative wealth in Togo. This initial step relied on nationally representative survey data from 6,171 Togolese citizens collected before the start of the Covid-19 pandemic in 2018 and 2019. Using these data, the team built an algorithm that could predict relative wealth in Togo at a more granular level than the original survey: the grid predictions were an improvement over the prefecture-level representative sample collected. The team leveraged the models to create initial poverty maps of rural Togo which helped inform the subsequent 2020 phone-based survey sampling strategy and provided the information needed for the Togolese government to geographically target Novissi to the 100 poorest cantons.
The research team buttressed this initial geographic targeting with an additional phone-based survey of a representative sample of 8,915 individual cell phone subscribers in September 2020. The sampling strategy inferred these individuals lived in rural cantons eligible for Novissi from their mobile phone data. As planned, this second survey provided researchers with a more accurate picture of canton-level individual variation in wealth and consumption.
Blumenstock et al found that these algorithms generated estimates that correlated strongly with survey and satellite-based estimates of wealth at the canton and prefecture level. Indeed, researchers calculated that this approach improved the precision of the social assistance program targeting by 42% relative to a naive geographic targeting of the 100 poorest cantons in Togo.

https://app.digitalpublicgoods.net/a/10169

• Africa
• Asia and Pacific

• The poor
• Remote and rural communities

There is no central bottleneck to more entities using Cider; each runs it locally. The approach has already been tested in several new contexts since its first application in Togo. The Global Policy Lab at UC Berkeley is actively looking for partners that want to integrate the tool for their own social programs. Cider is also registered as digital public good and is completely open sourced and available to anyone that wants to use the code. See here: https://app.digitalpublicgoods.net/a/10169

To implement Cider, social benefit/cash transfer program administrators need to have access to comprehensive call detail record data for the preceding 3-6 months prior to payments. Telephone surveys provide "ground-truth" measures of individual wealth and consumption for the places the program will run. The target beneficiaries are the poorest individuals in LMICs. Once the model is trained on the joined survey data and call detail record data, Cider assesses predictive performance by calculating inclusion/exclusion rates both globally as well as across important demographic groups (e.g., gender or ethnicity). If performance improves upon the existing targeting approach, the model is used to develop wealth/consumption predictions for the entirety of the cellphone numbers included in the call detail record data and individual subscribers that fall below the wealth level set by the program receive their payments via digital payment if they enroll and consent to the program via a digital subscription.