Alternative response: If someone told me that there was somewhere between a 0.00001 and 0.5 chance that I was to be struck by lightning tomorrow, it would not be reasonable for me to say “well almost everywhere within that confidence interval I have a >1% chance of being hit by lightning tomorrow”
Most of these CIs start at zero and they can’t go below zero so shouldn’t we consider these on a log scale? In which case the scale goes back to negative infinity and “almost everywhere within” is meaningless.
I don’t know of any reasonable justification for caring about expected log-welfare rather than expected welfare. For a welfare range estimate, the thing that matters is the expected value.
For a welfare range estimate, the thing that matters is the expected value.
Right. I have been using Rethink Priorities’ (RP’s) median welfare ranges, but I care about expected welfare. RP thinks their median welfare ranges are a better proxy for the actual means than the means of the distributions they got, and I tend to agree.
I think it’s very relevant that animal welfare interventions look better than global health interventions almost everywhere within the RP intervals.
I think this point is stronger than inferred from the graph because the 90 % confidence interval of the median is narrower than the range from the 5th to the 95th percentile.
Even if one takes the midpoint of the RP intervals as established fact, there are a lot of other assumptions Vasco’s arguments depend on, like the magnitude and duration of suffering a particular creature experiences with pain scales with thousands of points to cancel out the RP weights, and the cost-effectiveness of brand new charities in a field (campaigning) where marginal cost-effectiveness is relatively difficult to measure.
Unlike for RP we don’t have published estimates of distributions or confidence intervals for these, but if we did they’d also be extremely wide and I’m not sure that animal welfare interventions would look better across most of the distribution for them.
That argument is weak to me because you could take any intervention we are clueless about and it would look better than global health interventions within most of the interval. If our interval spans zero to close to infinity then global health interventions are going to be a speck near the bottom of that interval.
That argument is weak to me because you could take any intervention we are clueless about and it would look better than global health interventions within most of the interval.
The overall effect of global health and development (GHD) interventions depends on effects of animals due to the meat-eating problem. So I think clueless about the benefits of helping animals implies cluelessness about whether GHD interventions are beneficial or harmful.
Besides not touching on all of these considerations, many of my modelled inputs are highly uncertain too. However, this means extending human lives globally, and in China, India and Nigeria may be, in the nearterm, not only beneficial, but also hugely harmful. Using RP’s [Rethink Priorities’] 5th and 95th percentile welfare range of shrimp of 0 and 1.15, and maintaining all the other inputs, the harms caused to poultry birds and farmed aquatic animals as a fraction of the direct benefits of human life in 2022 would be:
Globally, 5.61 to 372.
In China, 12.3 and 841.
In India, 1.70 and 131.
In Nigeria, 1.12 and 45.3.
If our interval spans zero to close to infinity then global health interventions are going to be a speck near the bottom of that interval.
The cluelessness for GHD interventions would in that case be way more severe. It would go from minus to plus infinity instead of 0 to infinity. Less abstractly, one can be much more confident that humane slaughter interventions are beneficial than that saving human lives is beneficial. Humane slaughter interventions may have negligible benefits if the animals helped turn out to have a negligible welfare range, but it is very hard for them to be harmful in expectation, because they have minor effects on human- and animal-years. In contrast, saving human lives may well increase the animal-years with negative lives, thus potentially being harmful.
That might be true by your lights Vasco, but we are discussing a specific issue here (GiveWell vs. Animal Welfare confidence intervals) and I think its a bit disingenuous to bring adjacent arguments like the meat eating problem into this here.
If you’re clueless about an intervention and you use a fat-tailed prior, then the expected value might be very large but the median value will be very small, and most of the probability mass will be close to 0. For the RP welfare estimates, the median values make animal welfare interventions look highly effective.
Hi Henry. That graph represents the welfare range distributions (minimum, 25th percentile, median, 75th percentile, and maximum), not the confidence intervals of the medians. I think these are what matters, and that they would be much narrower.
The Rethink Priorities Welfare Ranges have absurdly wide confidence intervals. So wide that I would argue they’re almost worthless.
I think it’s very relevant that animal welfare interventions look better than global health interventions almost everywhere within the RP intervals.
Alternative response: If someone told me that there was somewhere between a 0.00001 and 0.5 chance that I was to be struck by lightning tomorrow, it would not be reasonable for me to say “well almost everywhere within that confidence interval I have a >1% chance of being hit by lightning tomorrow”
Most of these CIs start at zero and they can’t go below zero so shouldn’t we consider these on a log scale? In which case the scale goes back to negative infinity and “almost everywhere within” is meaningless.
I don’t know of any reasonable justification for caring about expected log-welfare rather than expected welfare. For a welfare range estimate, the thing that matters is the expected value.
Agreed, Michael!
Right. I have been using Rethink Priorities’ (RP’s) median welfare ranges, but I care about expected welfare. RP thinks their median welfare ranges are a better proxy for the actual means than the means of the distributions they got, and I tend to agree.
I think this point is stronger than inferred from the graph because the 90 % confidence interval of the median is narrower than the range from the 5th to the 95th percentile.
Even if one takes the midpoint of the RP intervals as established fact, there are a lot of other assumptions Vasco’s arguments depend on, like the magnitude and duration of suffering a particular creature experiences with pain scales with thousands of points to cancel out the RP weights, and the cost-effectiveness of brand new charities in a field (campaigning) where marginal cost-effectiveness is relatively difficult to measure.
Unlike for RP we don’t have published estimates of distributions or confidence intervals for these, but if we did they’d also be extremely wide and I’m not sure that animal welfare interventions would look better across most of the distribution for them.
That argument is weak to me because you could take any intervention we are clueless about and it would look better than global health interventions within most of the interval. If our interval spans zero to close to infinity then global health interventions are going to be a speck near the bottom of that interval.
Hi Nick.
The overall effect of global health and development (GHD) interventions depends on effects of animals due to the meat-eating problem. So I think clueless about the benefits of helping animals implies cluelessness about whether GHD interventions are beneficial or harmful.
The cluelessness for GHD interventions would in that case be way more severe. It would go from minus to plus infinity instead of 0 to infinity. Less abstractly, one can be much more confident that humane slaughter interventions are beneficial than that saving human lives is beneficial. Humane slaughter interventions may have negligible benefits if the animals helped turn out to have a negligible welfare range, but it is very hard for them to be harmful in expectation, because they have minor effects on human- and animal-years. In contrast, saving human lives may well increase the animal-years with negative lives, thus potentially being harmful.
That might be true by your lights Vasco, but we are discussing a specific issue here (GiveWell vs. Animal Welfare confidence intervals) and I think its a bit disingenuous to bring adjacent arguments like the meat eating problem into this here.
If you’re clueless about an intervention and you use a fat-tailed prior, then the expected value might be very large but the median value will be very small, and most of the probability mass will be close to 0. For the RP welfare estimates, the median values make animal welfare interventions look highly effective.
Hi Henry! The reason why the intervals are so wide is because they’re mixing together several models. I’ve explained more about this modeling choice and result here: https://forum.effectivealtruism.org/posts/rLLRo9C4efeJMYWFM/welfare-ranges-per-calorie-consumption?commentId=Wc2xksAF3Ctmi4cXY
Hi Henry. That graph represents the welfare range distributions (minimum, 25th percentile, median, 75th percentile, and maximum), not the confidence intervals of the medians. I think these are what matters, and that they would be much narrower.