Mobile phone data targeting offers a way to make decisions when conventional records are incomplete or out of date. A study of Togo’s Novissi emergency cash-transfer program examined how machine learning applied to mobile network data could help identify people most likely to be poor. The findings are relevant beyond humanitarian programs: they illustrate both the value and the operational limits of building decisions from indirect, high-volume data.
The data gap behind emergency assistance
When COVID-19 restrictions disrupted work and incomes in Togo, the government needed to direct assistance quickly. It lacked a current, comprehensive social registry, and the 2011 census did not include household poverty information. Existing surveys covered only a fraction of households, making it difficult to assess need across the country.
Novissi delivered cash digitally through mobile money. In its rural expansion, the government aimed to prioritize people in the country’s 100 poorest cantons and then identify the poorest eligible individuals within those areas. That created two related decisions: where to focus the program and whom to prioritize within each selected location.
How mobile phone data targeting worked
The approach combined geographic estimates with individual-level predictions. First, satellite-based estimates of relative wealth for small geographic grid cells were aggregated to estimate average wealth in each canton. These estimates helped identify poorer areas.
Next, researchers linked survey responses to mobile phone metadata for consenting participants. Survey data provided measures of wealth or consumption that could be used to train supervised machine learning models. The models learned associations between those measures and patterns in phone activity, including calls and texts, mobile money use, mobile data use, and location-related features.
After training, the model generated a poverty score for mobile subscribers. Program administrators could then rank eligible subscribers and prioritize those estimated to be poorest. This resembles a proxy means test, but uses a larger set of phone-use features and a machine learning model trained to predict outcomes beyond the training observations.
What the evaluation found
In the rural Novissi scenario, the phone-based approach performed better than the feasible alternatives evaluated, including geographic blanket targeting and occupation-based targeting. When the program aimed to reach the poorest 29% of eligible people, the phone-based method had an area under the curve (AUC) of 0.70. Geographic approaches had AUC values of 0.59 to 0.64. The study reported exclusion errors of 53% for phone-based targeting, compared with 59% to 78% for feasible alternatives.
A similar advantage appeared in a hypothetical nationwide program: phone-based targeting reached the poorest more effectively than most geographic alternatives. However, an “optimal” occupation-based method—one that prioritized the poorest occupational category using consumption information—slightly outperformed it in that simulation. Results therefore depend on the population, available information and policy objective; the model was not uniformly best in every comparison.
A simpler method that ranked people by estimated phone expenditures performed substantially worse than the machine learning model. It required less data, but phone spending alone was a weaker proxy for poverty. The researchers also noted that repeatedly using a simple expenditure rule could create opportunities for people to change behavior strategically.
Data freshness and inclusion matter
Model performance depends on how well the training data represent the people and conditions at the time of deployment. When the model or phone records were about 18 months out of date, predictive accuracy declined by 4% to 6%, while precision fell by 10% to 14%. These losses were close to the gains the model had achieved over geographic targeting. For practical deployments, current, representative training data are therefore not a minor detail: they can determine whether a data-driven approach retains its advantage.
Prediction quality is also only one part of program access. Novissi required people to have access to a phone and SIM, appear in the voter registry, attempt digital registration, complete the process, and be selected as eligible. Each requirement could exclude people before the model made a decision. The study estimated that 65% of adults and 85% of households had a phone, underscoring that a digital channel does not automatically reach everyone in need.
The fairness analysis found no systematic pattern of women being more likely than men to be incorrectly excluded by the phone-based method, and no significant exclusion errors for particular ethnic groups, religions, ages or household types. But small differences remained, and none of the targeting methods achieved perfect parity across demographic groups. Fairness needs to be evaluated explicitly rather than assumed from the use of neutral-looking data.
Operational lessons for data-driven systems
For technical and operational decision-makers, the case highlights several considerations that also apply when introducing predictive models into complex workflows:
- Define the decision and its ground truth. A model can only be evaluated against a clearly specified outcome. In this study, the rural evaluation used a proxy means test, while the national simulation used consumption data.
- Compare against realistic alternatives. The useful question is not whether a model predicts well in isolation, but whether it improves on methods that can actually be implemented under the same constraints.
- Track data coverage. Phone access, registration and data availability shaped who could enter the process at all. Model accuracy cannot correct for people missing from the underlying system.
- Plan for privacy and governance. Phone records can reveal sensitive patterns, including location and social connections. The study used pseudonymized data, secure storage and consent for survey matching, while recognizing that re-identification remains a concern.
- Monitor changes over time. Economic conditions, user behavior and data coverage can shift. A model that performs well at launch may need new representative data and renewed evaluation.
Conclusion
Togo’s Novissi experience shows that machine learning and mobile network data can complement traditional targeting when registries and surveys are unavailable or outdated. In the rural program, the approach reduced exclusion errors relative to feasible geographic and occupation-based alternatives, but it remained imperfect and depended on timely data, digital access and careful privacy safeguards. The practical lesson is to treat predictive targeting as one component of an inclusive decision system—not a replacement for reliable records, human oversight or ongoing evaluation.
Access Official Technical PDF