Very interesting argument. As someone fond of Bayesian modeling, I am almost always in favor of replacing point estimates with distributions.
GiveWell links to this paper in their CEA spreadsheet. Their recommendation is a discount rate of 4% for upper-middle-income countries and 5% for low-income countries. This recommendation seems to be based on three factors: Past growth of GDP, implied social discount rate, and projected GDP growth. All three, are measured with uncertainty and will vary by country. I think it would be very interesting to take that variability into account!
Thanks! I think most of the value would be captured by just switching to a single distribution for all countries, but yes having slightly different distributions for each country could be slightly better. My (fairly uninformed) guess would be that the socio-economic status of each of GiveWell’s countries is similar enough that we wouldn’t have much reason to use different distributions.
Very interesting argument. As someone fond of Bayesian modeling, I am almost always in favor of replacing point estimates with distributions.
GiveWell links to this paper in their CEA spreadsheet. Their recommendation is a discount rate of 4% for upper-middle-income countries and 5% for low-income countries. This recommendation seems to be based on three factors: Past growth of GDP, implied social discount rate, and projected GDP growth. All three, are measured with uncertainty and will vary by country. I think it would be very interesting to take that variability into account!
Thanks! I think most of the value would be captured by just switching to a single distribution for all countries, but yes having slightly different distributions for each country could be slightly better. My (fairly uninformed) guess would be that the socio-economic status of each of GiveWell’s countries is similar enough that we wouldn’t have much reason to use different distributions.