Howdy, Sam (Joel here). I assume what's going on here is you're saying for $1 (???) AMF increases economic benefits by 0.04 log units, while for GiveDirectly, the distribution doesn't really go past that. E.g., the 95% percentile is 0.0095 which is quite a bit less than 0.04 so there's not really any chance that GiveDirectly is more cost-effective, so the VOI is smaller. I'd rescale the units such that it's log Econ units per $1000 or something so GiveDirectly isn't in scientific notation. This comparison also misses out on comparing BOTH distributions uncertainties. It would be relatively more valuable to research if AMF has uncertainty, right?

It would be more valuable to research AMF if we had AMF's uncertainty. That being said, we don't currently have any distributions for any other charities that are not GiveDirectly, so I've stuck with this implementation for now.

Is this for the value of reducing the uncertainty at all or entirely? I'm not sure how to think about this. I think GiveWell spends ~1 million on its research staff. Maybe more. It spends about a billion so for most GiveWell cause prioritization between existing causes they need this ratio to be above 100x?

Is this 1/10th heuristic sensible? I haven't tried to reason about this myself.

Is it straightforward to go from this to the value of reducing uncertainty by 10%, 25%, 50%? (Which is more doable...)

No, but VPI gives a good upper bound. You can do those calculations but they are quite a bit more complicated.

Given the assumptions of this exercise it may have a low VOI. However, in a more general model that takes other effects into account (not just the things in the GW model) I think that the VOI for GD's effectiveness could actually be very high. It is still a large recipient of funds, it's used as a benchmark, and it represents a close to pure form of a more general philosophy, strategy and policy of direct cash transfers and 'let the recipient decide what's best for them.' Note: (I'm not personally advocating this philosophy.)