📈 economics

Predicting Well-Being with Mobile Phone Data: Evidence from Four Countries

This study demonstrates that mobile phone data can effectively predict long-term household well-being metrics like wealth and multidimensional poverty across four countries, with accuracy improving significantly through larger, nationally representative samples and the use of call and text message behavioral data.

Original authors: M. Merritt Smith, Emily Aiken, Joshua E. Blumenstock, Sveta Milusheva

Published 2026-02-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: M. Merritt Smith, Emily Aiken, Joshua E. Blumenstock, Sveta Milusheva

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to understand how well a family is doing financially. Traditionally, researchers have to knock on doors, sit down with families, and ask them to fill out long, expensive surveys. This is like trying to map a forest by walking every single tree path—it takes forever and costs a fortune. Because of this, many poorer countries often lack up-to-date maps of who is poor and who is not.

This paper suggests a new way to draw that map using something almost everyone carries in their pocket: a mobile phone.

The Core Idea: The Digital Footprint

Think of your mobile phone usage as a digital footprint. Just as your footprints in the sand can tell someone how fast you were walking or how heavy you were carrying, your phone's "call detail records" (metadata) can tell a computer a lot about your life.

The researchers' theory is simple: Wealthy people use their phones differently than poor people. Maybe they make more calls, travel to more places, or use different apps. By feeding this data into a computer "brain" (machine learning) and teaching it to spot these patterns, the computer can guess a person's economic status without ever asking them a single question.

The Experiment: Four Different Forests

To test if this works everywhere, the authors went to four very different countries: Afghanistan, Côte d'Ivoire, Malawi, and Togo. In each place, they linked survey answers from thousands of people with their actual phone records. They then asked three main questions:

What can we predict?
- The "Sturdy" vs. The "Fleeting": The computer was very good at guessing long-term wealth (like what furniture or appliances a family owns) and multi-dimensional poverty (a mix of health, education, and living standards).
- The "Harder" Stuff: It was okay at guessing how much money a family spends (consumption).
- The "Hardest" Stuff: It struggled significantly with predicting food security (will they eat today?) or mental health. These are like trying to guess the weather in five minutes; they change too fast and are too complex for the phone data to catch.
Which phone data matters most?
- The researchers found that calls and text messages are the gold mine. These records contain rich info about where people go and who they talk to.
- Surprisingly, data about mobile internet usage, mobile money, or buying airtime wasn't as helpful on its own. It's like trying to guess someone's job by looking at their coffee receipts (internet usage) versus looking at their commute route (calls/texts). The commute route tells you much more.
- Note: Using all data types together was still the best, but calls and texts alone did almost as well.
How much data do we need to teach the computer?
- You don't need a million people to start. The computer learns very quickly with the first 1,000 to 2,000 people. After that, adding more people helps, but the "magic" of learning slows down.
- However, the type of people matters more than just the number. A model trained on a mixed crowd (rich and poor, city and country) works much better than one trained on just one type of person.
- Analogy: Imagine trying to learn what "fruit" looks like. If you only show a computer pictures of red apples (a rural-only sample), it will fail when you show it a banana (an urban person). But if you show it a basket with apples, bananas, and oranges (a national sample), it learns the concept of "fruit" much better.

The Big Takeaways

The paper concludes with a few practical lessons for anyone wanting to use this technology:

Diversity is Key: The models work best when the training data represents a whole country, not just a village or a city. The more mixed the population, the smarter the model gets.
Keep it Simple: You don't need complex data like internet logs to get good results; standard call and text logs are often enough.
Know the Limits: This method is great for mapping long-term wealth, but it's not a crystal ball for predicting immediate crises like hunger or mental health struggles.
The "Missing" People: The study only looks at people who have phones. If you use this to help the poor, remember that the poorest of the poor might not have phones, so this method might accidentally leave them out of the map.

In short, mobile phone data is a powerful, low-cost tool for drawing economic maps, but it works best when you have a diverse group of people to learn from, and it's best at showing the "big picture" of wealth rather than the daily ups and downs of survival.

Technical Summary: Predicting Well-Being with Mobile Phone Data

Problem Statement

Accurate measurement of economic livelihoods is fundamental to empirical social science and effective policy design. However, traditional survey-based methods are costly and often fail to provide timely, reliable data in low- and middle-income countries (LMICs). While mobile phone call detail records (CDRs) have emerged as a promising alternative for predicting individual wealth and poverty, there is a need to systematically understand the potential and limitations of this approach across diverse contexts. Specifically, researchers and policymakers require clarity on which well-being measures are most predictable, which data types yield the highest accuracy, and the sample sizes required for robust model training.

Methodology

The authors conduct a parallel, standardized analysis across four countries: Afghanistan, Côte d'Ivoire, Malawi, and Togo. The study links household survey data (hundreds to thousands of individuals per country) with comprehensive mobile phone metadata obtained from network operators.

Data Characteristics:

Metadata Types: The study utilizes four transaction types: calls/text messages, airtime top-ups, mobile data usage, and mobile money transactions.
Survey Linkage: Phone numbers of household members are linked to survey responses via informed consent.
Variation: The datasets vary significantly in sample size (528 in Afghanistan to 5,469 in Malawi), sample composition (rural-only, urban-only, or nationally representative), survey modality (in-person vs. phone), and metadata duration (one to three months).

Experimental Design:
The authors employ machine learning to predict survey-based outcomes using features derived from phone metadata.

Models: Three algorithms are trained and tuned using five-fold cross-validation: LASSO regression, Random Forest, and Gradient Boosting.
Evaluation: The best-performing model (lowest RMSE on a validation set) is evaluated on a separate, held-out test set.
Metrics: Predictive accuracy is measured using Pearson's correlation coefficient ( $\rho$ ) between predicted and actual values.
Variables: Experiments vary the outcome variable (asset index, consumption, income, food security, mental health, etc.), the size of the training dataset, and the specific mobile phone data features used.

Key Results

1. Predictability of Different Well-Being Measures

The accuracy of predictions varies significantly depending on the target variable.

High Accuracy: Long-term poverty measures, specifically asset indices ( $\rho = 0.20–0.59$ ) and multi-dimensional poverty ( $\rho = 0.29–0.57$ ), are the most accurately predicted.
Moderate Accuracy: Consumption estimates show lower accuracy ( $\rho = 0.09–0.54$ ).
Low Accuracy: Transient or vulnerability measures, such as food security ( $\rho = 0.04–0.17$ ) and mental health ( $\rho = 0.01–0.23$ ), are difficult to predict. Income prediction performs poorly ( $\rho = -0.01–0.03$ ).

2. Utility of Data Types

Call and Text Metadata: Data containing location and mobility information (calls and texts) are the most predictive. Models using only this data perform nearly as well as models using all available data types.
Other Data Types: Mobile money, mobile data usage, and airtime top-ups are less predictive in isolation.
Combined vs. Naive Baselines: While models using all data types consistently outperform those using single types, even the most comprehensive phone-data models substantially outperform a naive baseline that assigns average local poverty based on location fixed effects.

3. Training Sample Size

Diminishing Returns: Predictive accuracy improves rapidly with the first 1,000–2,000 surveyed individuals.
Continued Improvement: Accuracy continues to increase, albeit more slowly, up to at least 4,500 individuals.
Model Sensitivity: Random Forest and Gradient Boosting models show comparable accuracy, while LASSO generally performs worse, except in scenarios with very limited data.

4. Impact of Population Heterogeneity

Predictive accuracy varies sharply across countries, driven largely by the heterogeneity of the underlying population.

National vs. Sub-national: Models trained on nationally representative samples (Côte d'Ivoire and Togo) perform significantly better than those trained on homogeneous subsets (rural-only in Afghanistan or urban-only in Malawi).
Urban/Rural Split: When restricting nationally representative samples to only urban or only rural populations, predictive accuracy drops substantially. This suggests that a significant portion of the predictive power comes from the model's ability to distinguish between rural and urban contexts.

Significance and Claims

The paper positions itself as a systematic exploration of the "potential and limitations" of using mobile phone data for economic estimation. The authors claim the following contributions:

Contextual Nuance: The study demonstrates that predictive performance is not uniform; it is heavily dependent on population heterogeneity. Models perform best when the training data reflects a diverse, nationally representative population rather than a narrow demographic slice.
Data Efficiency: The findings suggest that in settings with computational or storage constraints, utilizing only standard call and text metadata may be sufficient to generate accurate poverty estimates, as these data types capture the most predictive signals.
Feasibility of Small Samples: The authors establish that reasonably accurate models can be trained with relatively small datasets (approx. 1,000 observations), though larger samples continue to yield marginal gains.
Scope of Application: The paper clarifies that mobile phone data is best suited for predicting long-term structural poverty (wealth/assets) rather than transient vulnerability or specific income/consumption metrics.

Limitations Acknowledged:
The authors explicitly note that their conclusions are based on four countries with idiosyncratic datasets, limiting generalizability. They also highlight that their analysis focuses on the well-being of phone subscribers and does not address the implications for non-phone owners, which could lead to systematic exclusions in policy targeting. Finally, the study is cross-sectional; it does not yet demonstrate the ability of phone data to track dynamic changes in livelihoods over time.