If GiveDirectly does end up using the methods described in this paper for village selection, it would be very interesting to learn how accurate the method is (of using thatched vs. non-thatched roofs to estimate depth of poverty).
The paper indicates that GiveDirectly did pilot this algorithm in 2014 but does not seem to have numbers on an accuracy increase/decrease and assigns a probable increase to consistency and elimination of human error.
We use a rigorous audit process. 100% of the households we enroll are audited by a second field team that operates completely independently from the team responsible for enrollment. Senior staff audit a smaller proportion of recipients. We also call our recipients after sending them money to learn more about their experiences. To date, 1% of recipients called report paying a bribe to a village official, and less than 1% report paying a bribe to a mobile money agent.
I am surprised they didn't address the perverse incentives this may create. Surely, if trying to help people out of poverty, you shouldn't make their help _contingent on them continuing to use thatched roofs_. Worth noting, because if the program is successful, it will change which communities are most in need of help, so you can't just run the algorithm once and forget about it.
I demand giving be analyzed in an open data sort of sense empirically over the long term for viable effectiveness. I question all methods that seemingly signal 'good person' and fail to produce long term results.