Extended Data Table 4.
Table compares three scenarios: (1) “Best case”: when the model is calibrated using survey data and phone data gathered just before deployment; these results are comparable to the paper’s main analysis (slight differences are due to the sample restrictions described below); (2) “Old model”: when the model is trained using a survey conducted two years before deployment, but the phone data are collected just before deployment; and (3) “Old model and data”: when the phone-based wealth estimates are generated using survey and phone data from two years prior. Rows (4) and (5) show geographic targeting results using the same sample as in rows (1) – (3). In the simulations, the “old” data are from the 2018-19 national household survey and corresponding 2019 phone dataset; the 2020 phone survey PMT is used as the ground truth measure of poverty (restricted to respondents for whom CDR are available in 2019 and 2020, N = 7,064).