Predicting Well-Being with Mobile Phone Data: Evidence from Four Countries
M. Merritt Smith, Emily Aiken, Joshua E. Blumenstock, Sveta Milusheva
TL;DR
This study investigates predicting household well-being from mobile phone data across four LMICs (Afghanistan, Côte d'Ivoire, Malawi, and Togo) by linking survey-based measures to rich phone metadata. It uses standardized ML pipelines with three models—LASSO, random forest, and gradient boosting—and five-fold cross-validation, reporting held-out test performance to predict asset index ($\rho$ 0.20-0.59) and multidimensional poverty ($\rho$ 0.29-0.57) as well as consumption ($\rho$ 0.04-0.54), food security ($\rho$ 0.04-0.17), and mental health ($\rho$ 0.01-0.23). Results show asset index and multidimensional poverty are easier to predict than consumption and vulnerability, with call/text and mobility data providing the strongest signals. Training data size matters, with rapid gains up to ~1,000-2,000 observations and continued improvements beyond ~4,500, and nationally representative samples outperform urban/rural-only samples by roughly 20-70%.
Abstract
We provide systematic evidence on the potential for estimating household well-being from mobile phone data. Using data from four countries - Afghanistan, Cote d'Ivoire, Malawi, and Togo - we conduct parallel, standardized machine learning experiments to assess which measures of welfare can be most accurately predicted, which types of phone data are most useful, and how much training data is required. We find that long-term poverty measures such as wealth indices (Pearson's rho = 0.20-0.59) and multidimensional poverty (rho = 0.29-0.57) can be predicted more accurately than consumption (rho = 0.04 - 0.54); transient vulnerability measures like food security and mental health are very difficult to predict. Models using calls and text message behavior are more predictive than those using metadata on mobile internet usage, mobile money transactions, and airtime top-ups. Predictive accuracy improves rapidly through the first 1,000-2,000 training observations, with continued gains beyond 4,500 observations. Model performance depends strongly on sample heterogeneity: nationally-representative samples yield 20-70 percent higher accuracy than urban-only or rural-only samples.