Mobile phone data for poverty targeting: what we learned from four countries

Timely and accurate data is essential to effective poverty reduction and social protection. Whether designing a cash transfer program or directing infrastructure investment, policymakers need reliable information on where the poor are and what resources they lack.

Yet in many low- and middle-income countries (LMICs), the traditional gold standard for measuring living standards — household surveys — is expensive and infrequent. As a result, many policy decisions are made based on data that is low quality, nonrepresentative, and several years old.

One emerging and promising alternative is mobile phone data (MPD): anonymized metadata from calls, text messages, and data usage. These data do not contain information about the content of calls or texts, but they do record the frequency with which mobile subscribers place and receive calls and texts and use mobile data. The intuition for using these data is straightforward: people with different levels of economic well-being often use their phones in different ways. These differences may reflect variation in social networks, mobility, access to services, and economic activity. Machine learning models can use these patterns to estimate the wealth or poverty of individual subscribers, or of small regions of a country.

A growing body of research has shown that MPD can help estimate poverty. But much of the evidence has come from one-off studies in individual countries.  In our recent paper, we use rich data from four countries — Afghanistan, Côte d'Ivoire, Malawi, and Togo — to systematically explore the potential and limitations of using MPD to estimate economic livelihoods. We link survey-based measures of well-being for a sample of individuals in each country to comprehensive metadata capturing their phone usage patterns, allowing for a parallel, standardized analysis.

For governments and development practitioners considering using these new methods for poverty measurement, the findings offer five pragmatic insights into what to measure and which data to use.

1. Mobile data is better suited for predicting long-term wealth than transient income

A critical question for policymakers is what type of economic well-being these algorithms can actually measure. We find that MPD is significantly more successful at predicting stable, long-term measures of well-being, such as an asset index or multidimensional poverty status. The models are less successful at capturing more transient, dynamic indicators like daily consumption, short-term income, or temporary food insecurity.

The Policy Takeaway: Mobile data is a robust tool for programs that rely on identifying the chronically poor, such as targeting long-term social safety nets. It can also inform emergency response efforts in scenarios where the chronically poor are expected to be the most vulnerable and require the most immediate support.

Figure 1. Cross-country model performance by target and data available

2. Basic call and text records offer high predictive value

In the era of big data, there is a common assumption that more complex data inputs like mobile internet usage or mobile money transactions are required to make accurate predictions. Our findings suggest otherwise.

Machine learning models leveraging metadata derived solely from phone calls and text messages produce predictions that are nearly as accurate as models utilizing a full suite of phone metadata types (including mobile money transactions, mobile data usage, and airtime top-ups). Call and text records inherently contain rich, predictive information about an individual's social networks, mobility patterns, and broader economic activity.

The Policy Takeaway: Even in regions where smartphone and mobile money adoption are still low, effective targeting models can be built using basic 'feature phone' data, simplifying data processing and reducing computational costs for national statistical offices.
 

3. A heterogeneous target population is critical for accuracy

Perhaps the most crucial operational finding is that performance depends largely on the heterogeneity of the target population. The predictive models performed notably better in Togo and Côte d'Ivoire, where the target populations were nationally representative, compared to more homogeneous contexts in Afghanistan (rural-only) and Malawi (urban-only).

The Policy Takeaway: Policymakers must carefully consider the demographic makeup of their target population. MPD is most effective for diverse, national populations but requires more caution when targeting highly homogeneous groups.

Figure 2. Varying sample size and composition

4. Flexible, nonparametric machine learning models perform best

We tested three different machine learning models: LASSO regression, random forest, and gradient boosting. We find that the random forest and gradient boosting approaches perform comparably; the LASSO model performs somewhat worse. The difference between the nonparametric models and the linear approach is largest in countries where our data cover a homogeneous target population.

The Policy Takeaway: Governments should test different machine learning models to determine what works best in their target population. However, a range of off-the-shelf tools now make this relatively straightforward to implement. For instance, all of the models used in our study were based on the open-source cider toolkit.

5. Small initial samples can work, but scale improves precision

Finally, the study provides guidance on the initial investment in training data required. We found that it is possible to train a reasonably accurate predictive model with a relatively small sample of ~1,000 surveyed individuals. However, we also find a clear positive correlation between sample size and accuracy, with predictive gains continuing as the sample size scales past 4,500 individuals.

The Policy Takeaway: Governments do not need to wait for massive, resource-intensive data collection efforts to begin exploring mobile data targeting. Pilot programs can be launched cost-effectively with modest initial surveys to prove the concept. As the system demonstrates value, it can be scaled and refined over time, leveraging larger datasets to iteratively improve precision.

Moving forward: a complement, not a replacement

Mobile phone data is not a replacement for traditional household surveys. Survey data remain essential for measuring welfare directly, validating models, and anchoring poverty statistics in representative evidence. But mobile phone data can be a valuable complement, especially in settings where household surveys are infrequent, expensive, or difficult to field.

The four-country evidence points to a practical message: mobile phone data can help measure poverty and vulnerability, but the accuracy of these predictions depends on the setting and the policy objective. The approach performs best when there is meaningful variation in the target population, as in nationally representative samples that include both rural and urban areas. It also works better for relatively stable measures of welfare, such as wealth or multidimensional poverty, than for short-term outcomes like income, food security, or mental health. Encouragingly, much of the predictive signal comes from basic call and text metadata, and useful models can be trained with modest survey samples — though larger samples continue to improve performance.

At the same time, the evidence points to important limits. The four-country analysis focuses on mobile phone subscribers, excluding people without phones. In some contexts, this may not be of great consequence; in others, it affects who is represented and how well programs reach the poorest households. The analysis is also cross-sectional, based on one wave of survey data in each country, so more work is needed to understand whether mobile phone data can reliably detect changes in livelihoods over time.

These insights, produced under the Global Data Facility – Mobile Phone Data for Policy program, provide a foundation for more systematic use of mobile phone data in low- and middle-income countries. By leveraging this guidance, government agencies across the world, including in the 24 LMICs currently supported by the program, can thoughtfully update their poverty maps and ensure critical resources reach those who need them most.

The GDF-MPD window is a multi-partner initiative funded by the Spanish Government, Ministry of Economy, Commerce and Business. The GDF-MPD program is jointly managed by DEC (DECDG, DECDI), Poverty and Digital Development Global Departments in technical partnership with the International Telecommunications Union (ITU).