Reproduce “Program Targeting with Machine Learning and Mobile Phone Data: Evidence from an Anti-Poverty Intervention in Afghanistan” on OpenResearch

About this paper

Emily Aiken, Guadalupe Bedoya, Joshua Blumenstock, Aidan Coville

Researchers at the University of California, Berkeley and the World Bank assessed the efficacy of machine learning models leveraging mobile phone data for identifying ultra-poor households in Afghanistan. The study found that this approach achieved targeting accuracy comparable to or better than traditional methods, particularly when combined with survey data, while offering substantial cost and speed efficiencies for anti-poverty interventions.

Key insights

• Mobile phone data, processed with machine learning, provides a viable and efficient alternative or complement to traditional methods for identifying vulnerable populations, especially where conventional data collection is impractical.
• Combining mobile phone data with traditional survey-based measures (e.g., asset indices) significantly enhances targeting accuracy, achieving superior performance over any single data source.
• The effectiveness of mobile phone data for targeting is critically dependent on mobile phone penetration within the target population and comprehensive access to data from relevant network operators.

Summary by alphaXiv.