Note that GiveDirectly in general is a bit of a weird outlier in terms of GiveWell top recommendations, because it’s a lot less cost-effective than the other charities, but is very useful as a “standard candle” for evaluating whether an intervention is potentially a good target for donations. I think being better than GiveDirectly is not sufficient to be a top recommendation for a cause area.
Methodologically, I do think there are a variety of reasons for why you should estimate a regression to the mean in these impact estimates, more so than for GiveDirectly, in large parts because the number of studies in the space is lot lower, and the method of impact is a lot more complicated in a way that allows for selective reporting.
Note that GiveDirectly in general is a bit of a weird outlier in terms of GiveWell top recommendations, because it’s a lot less cost-effective than the other charities, but is very useful as a “standard candle” for evaluating whether an intervention is potentially a good target for donations. I think being better than GiveDirectly is not sufficient to be a top recommendation for a cause area.
Methodologically, I do think there are a variety of reasons for why you should estimate a regression to the mean in these impact estimates, more so than for GiveDirectly, in large parts because the number of studies in the space is lot lower, and the method of impact is a lot more complicated in a way that allows for selective reporting.