We provide systematic evidence on estimating household well-being from mobile phone data across four countries (Afghanistan, Côte d’Ivoire, Malawi, Togo). Using parallel, standardized machine learning experiments, we assess which welfare measures are most predictable and which data types most useful. Long-term poverty measures—wealth indices (Pearson’s ρ = 0.20–0.59) and multidimensional poverty (ρ = 0.29–0.57)—are predicted more accurately than consumption (ρ = 0.04–0.54); transient measures like food security are difficult to predict. Call and text behavior outperforms internet, mobile money, and airtime metadata. Nationally representative samples yield 20–70 percent higher accuracy than urban- or rural-only samples.
Aiken et al. (Fri,) studied this question.