I would discourage you from doing guesswork this rough and legitimising it by calling it a “Fermi estimate”.
Interesting. I was thinking that calling it a “Fermi estimate” would highlight it is a rough calculation instead of legitimising it. I also did not want to say something like “please note these numbers are super speculative, and you should not trust them at all”, because I have only guessed parameters for which there is basically no data, so it seems hard to easily improve on my estimate. However, I definitely encourage people to make a copy of my sheet, and use their own numbers.
Most cells in your spreadsheet are commented “my guess”.
Nitpick. Not “most”, as only 30 % (= 6⁄20) of the inputs of the tab you were referring to say “My guess.” in the notes. The fraction is lower considering all the cells in the tab, and even lower considering all the cells of all tabs.
The 6 cells in the tab you printed saying “My guess.” in the notes refer to the following assumptions (which I mentioned in the post; I did not hide in the sheet any of my assumptions) for both air asphyxiation and electrical stunning slaughter:
Time in excruciating pain as a fraction of that in disabling pain equal to that of ice slurry.
Time in hurtful pain as a fraction of that in disabling pain equal to that of ice slurry.
Time in annoying pain as a fraction of that in hurtful pain equal to that of ice slurry.
I have now clarified these in the notes of the relevant cells in the tab. I made them because I only had data about the time in disabling pain for air asphyxiation and electrical stunning slaughter. Assuming these have the same proportion of time in pain across categories as ice slurry slaughter felt the most agnostic assumption to me (i.e. the one which would minimise over or underestimating cost-effectiveness). Do you have a better assumption?
Even if I assume air asphyxiation slaughter does not involve annoying, hurtful or excruciating pain to minimise the badness of the initial conditions, thus underestimating cost-effectiveness, I still conclude the past cost-effectiveness of HSI is 1.36 k times the marginal cost-effectiveness of GiveWell’s top charities. @Henry Howard🔸[1], I encourage you to make a copy of the sheet, and try to get to a ratio lower than 1. I believe this is very hard, which makes me think HSI is robustly more cost-effective than GiveWell’s top charities.
The definition of Fermi estimate linked in this post defines a Fermi estimate as aiming to be within 1 magnitude of true. Given just the Rethink Priorities welfare range estimates span several magnitudes (infinite really, given lower bound is 0), this at least is incorrect.
This sort of chaining of EV calculations is common on this forum. I think it’s counterproductive. Show the confidence intervals and it becomes clear that the result is as good as “I have no idea”, which is a fine thing to say. Just say that.
Could you clarify the argument you are making? I agree the 5th percentile past cost-effectiveness of HSI is 0 given this is RP’s 5th percentile welfare range of shrimps. However, I think what matters is the expected cost-effectiveness. Are you suggesting one should disregard interventions whose 5th percentile cost-effectiveness is 0? Imagine one could pay 1 k$ to save 0 lives with 10 % probability, and 1 life with 90 % probability. The 5th percentile cost-effectiveness is 0 (the 5th percentile cost-effectiveness of deworming programs could also be super low?), but the expected cost-effectiveness is 0.9 life/k$, i.e. around 4.5 times the cost-effectiveness of GiveWell’s top charities of 0.2 life/k$ (= 1/(5*10^3)).
Thanks for looking into my calculations, Henry.
Interesting. I was thinking that calling it a “Fermi estimate” would highlight it is a rough calculation instead of legitimising it. I also did not want to say something like “please note these numbers are super speculative, and you should not trust them at all”, because I have only guessed parameters for which there is basically no data, so it seems hard to easily improve on my estimate. However, I definitely encourage people to make a copy of my sheet, and use their own numbers.
Nitpick. Not “most”, as only 30 % (= 6⁄20) of the inputs of the tab you were referring to say “My guess.” in the notes. The fraction is lower considering all the cells in the tab, and even lower considering all the cells of all tabs.
The 6 cells in the tab you printed saying “My guess.” in the notes refer to the following assumptions (which I mentioned in the post; I did not hide in the sheet any of my assumptions) for both air asphyxiation and electrical stunning slaughter:
Time in excruciating pain as a fraction of that in disabling pain equal to that of ice slurry.
Time in hurtful pain as a fraction of that in disabling pain equal to that of ice slurry.
Time in annoying pain as a fraction of that in hurtful pain equal to that of ice slurry.
I have now clarified these in the notes of the relevant cells in the tab. I made them because I only had data about the time in disabling pain for air asphyxiation and electrical stunning slaughter. Assuming these have the same proportion of time in pain across categories as ice slurry slaughter felt the most agnostic assumption to me (i.e. the one which would minimise over or underestimating cost-effectiveness). Do you have a better assumption?
Even if I assume air asphyxiation slaughter does not involve annoying, hurtful or excruciating pain to minimise the badness of the initial conditions, thus underestimating cost-effectiveness, I still conclude the past cost-effectiveness of HSI is 1.36 k times the marginal cost-effectiveness of GiveWell’s top charities. @Henry Howard🔸[1], I encourage you to make a copy of the sheet, and try to get to a ratio lower than 1. I believe this is very hard, which makes me think HSI is robustly more cost-effective than GiveWell’s top charities.
I am tagging you here because I have added this sentence to the comment after posting it for the 1st time.
The definition of Fermi estimate linked in this post defines a Fermi estimate as aiming to be within 1 magnitude of true. Given just the Rethink Priorities welfare range estimates span several magnitudes (infinite really, given lower bound is 0), this at least is incorrect.
This sort of chaining of EV calculations is common on this forum. I think it’s counterproductive. Show the confidence intervals and it becomes clear that the result is as good as “I have no idea”, which is a fine thing to say. Just say that.
Could you clarify the argument you are making? I agree the 5th percentile past cost-effectiveness of HSI is 0 given this is RP’s 5th percentile welfare range of shrimps. However, I think what matters is the expected cost-effectiveness. Are you suggesting one should disregard interventions whose 5th percentile cost-effectiveness is 0? Imagine one could pay 1 k$ to save 0 lives with 10 % probability, and 1 life with 90 % probability. The 5th percentile cost-effectiveness is 0 (the 5th percentile cost-effectiveness of deworming programs could also be super low?), but the expected cost-effectiveness is 0.9 life/k$, i.e. around 4.5 times the cost-effectiveness of GiveWell’s top charities of 0.2 life/k$ (= 1/(5*10^3)).