The graph below shows the costs and benefits of each GiveWell top charity on a log scale (except New Incentives, which I ran out of time to add). The dotted black line shows the cost-effectiveness of the most cost-effective charity; anything above this line is less effective than this charity and anything below this line is more effective than this charity. On the left is effectively GiveWell’s published conclusions – the base case analysis of my refactoring of their model. You can see it recommends Sightsavers, HKI and AMF as the most cost-effective charities, with very little to pick between them. On the right is the same graph with full uncertainty analysis added (see Section 4 for details on the ‘PSA’ technique). This shows that any charity except GiveDirectly could potentially be the most cost-effective, with varying degrees of plausibility (and GiveDirectly could easily be most cost-effective if we add in some risk adjustment, as we do in Section 4).
This is important because it shows that without uncertainty analysis, we are at a material risk of picking the wrong charities to fund.
… By means of a broad conclusion, uncertainty analysis is supportive of GiveWell’s current approach of not simply donating to the charity with the highest cost-effectiveness ratio in the base case model. There is a high degree of uncertainty surrounding this model and any charity in GiveWell’s top recommendations has the potential to be cost-effective within the bounds of some very reasonable sensitivity analysis. The table below suggests some plausible scenarios by which each charity might be the most – or at least significantly more – cost-effective
You might be interested in Quantifying Uncertainty in GiveWell Cost-Effectiveness Analyses (2022)
as well as Methods for improving uncertainty analysis in EA cost-effectiveness models (also 2022, it was a good year for this stuff)