Comparing Community-Based Targeting to Big Data and Machine Learning in Bangladesh

Emily Aiken, Anik Ashraf, Joshua E. Blumenstock, Raymond P. Guiteras and A. Mushfiq Mobarak

Manuscript, December 2023. 

Write for early draft with preliminary results.

Abstract: In recent years, the rise of ``big data'' and machine learning has enabled new paradigms for the targeting of social protections and humanitarian aid. However, these centralized and top-down approaches typically do not involve community participation or feedback, and thus may not capture nuanced conceptions of poverty as well as local, community-based approaches. This paper uses a wealth of data from Bangladesh --- including mobile phone records from the four major mobile network operators in Bangladesh, community-based wealth ranking data from 180 communities, and survey data from 5,000 households --- to directly compare "big data" targeting (based on mobile phone records) with community-based targeting. We find that phone-based targeting is more accurate than community-based targeting at identifying the consumption-poorest households, but both methods perform substantially worse than traditional proxy-means testing. We also explore the extent to which different data sources can be combined to more effectively identify poor and vulnerable households, and the extent to which different targeting methods work better for different types of communities and households.