Value of Infomation, an example with GiveDirectly
Disclaimer: I do not represent GiveWell or the Quantified Uncertainty Research Institute. All opinions are my own.
In information economics, you can calculate the Value of Information. The Value of Information is how much one should pay to gain information before deciding. I’ve created a simple annotated example of the value of knowing GiveDirectly’s cost-effectiveness in funding decisions between GiveWell charities.
You can read more in this Squiggle Notebook.
https://observablehq.com/@hazelfire/value-of-information-givedirectly.
I found very little value in knowing GiveDirectly’s actual cost-effectiveness. More precisely, I find it’s only valuable to know GiveDirectly’s cost-effectiveness if the amount of funding in the decision is 7 trillion times the cost of obtaining the information. The information value is low because GiveDirectly is almost certainly not more cost-effective than the current expected value of AMF or other GiveWell charities (that need to meet a 10x GiveDirectly bar).
I think that this is a promising technique for quantifying the value of EA research.
For this calculation, I use my quantification of uncertainty of GiveDirectly. This calculation makes all the assumptions in that quantification.
Acknowledgements
I completed this work with help and funding from the Quantified Uncertainty Research Institute. Thanks to Nuño Sempere, Ozzie Gooen, David Reinstein, Quinn Dougherty, Misha Yagudin, and Edo Arad for their feedback.
- Quantifying Uncertainty in GiveWell Cost-Effectiveness Analyses by (31 Oct 2022 14:31 UTC; 118 points)
- 's comment on Methods for improving uncertainty analysis in EA cost-effectiveness models by (30 Aug 2022 20:51 UTC; 17 points)
- 's comment on Estimating the cost-effectiveness of scientific research by (4 Nov 2022 10:17 UTC; 4 points)
While I agree that GiveDirectly is almost certainly not more cost-effective than AMF or other GiveWell top charities, the quote
rang alarm bells in my head in the sense of conflating confidence levels inside and outside an argument. But this is an extremely minor nitpick; let me also balance it out by saying I’m a fan of your uncertainty quantification of GW’s GiveDirectly CEA, which is a lot better than what I would’ve come up with (as someone who’s also spent long hours poring over their CEAs and, like Froolow, found their model architecture frustratingly confusing).