We use unique data from over 600 Indonesian communities on what individuals know about the poverty status of others to study how network structure influences information aggregation. We develop a model of semi-Bayesian learning on networks, which we structurally estimate using within-village data. The model generates qualitative predictions about how cross-village patterns of learning relate to network structure, which we show are borne out in the data. We apply our findings to a community-based targeting program, where citizens chose households to receive aid, and show that the networks that the model predicts to be more diffusive differentially benefit from community targeting. (JEL D14, D83, D85, I32, O12, Z13)

This paper shows that adding a small application cost to a transfer program can substantially improve targeting through self-selection. Our village-level experiment in Indonesia finds that requiring beneficiaries to apply for benefits results in substantially poorer beneficiaries than automatic enrollment using the same asset test. Marginally increasing application costs on an experimental basis does not further improve targeting. Estimating a model of the application decision implies that the results are largely driven by the nonpoor, who make up the bulk of the population, forecasting that they are unlikely to pass the asset test and therefore not bothering to apply.