Scalable Targeting of Social Protection: When Do Algorithms Out-Perform Surveys and Community Knowledge?

Summary

Primary academic source. Cowles Foundation discussion paper presenting a head-to-head comparison of phone-based targeting (PBT), proxy means testing (PMT), and community-based targeting (CBT) for a cash transfer programme in Cox's Bazar, Bangladesh. Documents the gradient boosting ML model trained on 1,578 CDR features from all four MNOs. Reports PMT as most accurate (AUC 0.82), PBT second (AUC 0.61), CBT third (AUC 0.58). Introduces cost-effectiveness framework showing PBT is welfare-maximising for large-scale programmes with tight budgets. Details IRB approvals, data handling protocols (pseudonymisation, isolated a2i server), and census of 106,000 households.