Thanks for going through the analysis here! I was a bit confused about a few things.
You cite Sarah Morton as thinking the costs will decrease over time and then you you speculate the marginal cost-effectiveness to be 0.5 (seemingly implying costs will increase over time). You cite diminishing returns as your support for this decision, but Sarah’s comment implying this intervention is in a period of increasing returns (yes this is a thing). Additionally, this may be less of an issue of increasing returns and this intervention could just benefit from economies of scale which would also decrease costs and increase cost-effectiveness. If Sarah is right (and thus the marginal cost-effectiveness is >= 1), School Plates is at least twice as effective as you’ve estimated here. However, very happy to hear I’m misunderstanding something on my end.
It seems like the sole unit of impact measurement is time and intensity of suffering. However, it’s not obvious to me this is all that matters. Suppose we care about the killing of an animal differently than the suffering when they’re alive (I don’t think this is entirely unreasonable, though maybe not the most mathematically convenient model). Then School Plates would get a boost in effectiveness. We could also develop models where the number of animals suffering is a component of our overall utility function (there’s a disutility associated with a factory-farmed animal existing at all), in which case School Plates would also get a boost in effectiveness. My point here is I’m not entirely sure such a narrow definition of impact is appropriate here. I know it makes the math easier and this is just meant to be an estimate, but when the definition of impact favors one intervention over the other in basically every component, the difference in the estimates is going to be exaggerated. Then again, I’m not here to insist your ethics are wrong (I think that’s a personal choice), but I think this adds a lot of nuance to your conclusions.
I hope this isn’t coming off as overly critical. I enjoyed reading this post and think it’s a great starting point for further, highly-relevant work. I’m thinking of potentially building out some Monte Carlo Simulations of your model and Saulius’ model (with Open Phil’s comments) to see how accounting for variance impacts these estimates (my hunch is there will be so much uncertainty it will be hard to decide between the interventions). One additional benefit of the Monte Carlo Simulations is how they point to where collecting more evidence and decreasing our uncertainty would improve our estimates most. Thanks again for posting this!
Thanks for the great points, Blake, and welcome to the EA Forum!
On 1, to clarify, my understanding is that Sarah expects “additional meat-free meals in 2023“/”cost in 2023” to be lower than “additional meat-free meals in 2024“/”cost in 2024”, i.e. she expects the ratio between annual benefits and cost to increase. Despite of this, for any year, I expect that “annual additional meat-free meals”/”annual cost” > “additional meat-free meals from spending an extra e.g. 1 k$ in a given year”/”e.g. 1k$” = “marginal cost-effectiveness”. Do you agree? For any given year, I expect School Plates to start spending on the most cost-effective activities 1st, such that the benefits produced by the last dollar spent in any given year are smaller than those produced by the mean dollar that same year.
On 2, at least in principle, I fully endorse expectedtotalhedonisticutilitarianism, which only accounts for wellbeing. However, in practice, I would agree it can still make perfect sense to account for metrics which do not have to do with wellbeing per se. They could inform one about longer term impacts on wellbeing, or be an indicator about near term impacts on wellbeing which are not being captured by the wellbeing metrics for some reason. It is unclear to me whether accounting for other metrics from this perspective would favour School Plates, as corporate campaigns might also generate momentum for further welfare campaigns, and eventually accelerate the phase-out of factory-farming (or the transition to a net positive factory-farming).
I’m thinking of potentially building out some Monte Carlo Simulations of your model and Saulius’ model (with Open Phil’s comments) to see how accounting for variance impacts these estimates (my hunch is there will be so much uncertainty it will be hard to decide between the interventions).
I do not think the Monte Carlo would change the mainline cost-effectiveness estimates much:
The inputs I used are supposed to be means rather than medians or modes. So, if one modelled the distributions such as their means correspond to my point estimates, the cost-effectiveness would be maintained if it was a sum of products, because e.g. E(X1*X2 + X3*X4) = E(X1) E(X2) + E(X3) E(X4) assuming all distributions are independent (as I would if I was to run a Monte Carlo).
In reality, my estimate of the cost-effectiveness involves 2 variables in the denominator, the cost of the program in 2023 and the number of lunches and dinners per person per year, but these are arguably 2 of the most certain variables. So approximating E(1/X) as 1/E(X), as I implicitly did, should be fine. If I wanted to be more rigorous, I could also calculate the mean of the reciprocal of the variables in the denominator from 2 percentiles using the formulas here, thus not having to run the whole Monte Carlo.
A large overlap between the cost-effectiveness distributions of 2 interventions does not necessarily imply uncertainty about which intervention is best if the Monte Carlo assumes all variables are independent. This is because there may be correlations across variables used to calculate the cost-effectiveness of both interventions.
In particular, the welfare range of chickens, and welfare of chickens per time as a fraction of their welfare range are 2 of the most uncertain variables, and will be same for corporate campaigns and School Plates.
Here is an extreme example illustrating the issue I am pointing to:
Intervention A has a cost-effectiveness given by X1*Y, where X1 is a super uncertain distribution spanning many orders of magnitude, and Y is practically a point estimate of 1.
Intervention B has a cost-effectiveness given by X2*Z, where X2 is distributed as X1, and Z is practically a point estimate of 1.1.
Running a Monte Carlo with all the distributions as independent would result in almost perfect overlap between the cost-effectiveness distributions of A and B. So one may be tempted to conclude it is unclear which interventions is best.
However, if X1 and X2 are perfectly correlated, the correct conclusion is that B is almost certainly better than A (1.1 > 1).
In any case, the issue could be solved by setting all uncertain distributions shared across interventions to their means. I guess modelling the correlations within the Monte Carlo would be significantly more work for little benefit.
Before running a Monte Carlo with some guesses for the distributions of each variable, I think it would be good to get more information about the variables with the most reducible uncertainty. Yet, the conclusion that corporate campaigns are more cost-effective is pretty robust:
Corporate campaigns are more cost-effective even for most favourable compositions of baseline meals. Playing around with the Sheet:
Assuming all the meat in the baseline meals is poultry, and the meat-free meals are all plant-based[1], corporate campaigns are still 70.3 times as cost-effective as School Plates.
It makes sense the ratio is smaller than that of 186 I estimated.
Poultry is the meat with the highest suffering-adjusted animal living time (animal living live times welfare range) per mass, equal to 9.53 year/kg (= 28.7*0.332).
Assuming all the meat in the baseline meals is farmed fish and seafood, and the meat-free meals are all plant-based[2], corporate campaigns are still 89.4 times as cost-effective as School Plates.
It makes sense the ratio is smaller than that of 186 I estimated, but higher than that of 70.3 I got just above.
Farmed fish and seafood is the meat with 2nd highest suffering-adjusted animal living time per mass, equal to 7.31 year/kg (= 82.1*0.089).
Reality check. The meat in the baseline meals being poultry is 1.27 (= 89.4/70.3) times as cost-effective as it being farmed fish and seafood, which is quite close to the ratio of 1.30 (= 9.53/7.31) suggested by the suffering-adjusted animal living time per mass. The values do not have to be exactly the same due to the presence of eggs in the baseline meals.
Supposing lunches and dinners account for 100 % of the animal living time instead of 75 % only makes the cost-effectiveness of School Plates 1.33 (= 1⁄0.75) times as high, so corporate campaigns would still be 140 (= 186⁄1.33) times as cost-effective.
Assuming all the meat in the baseline meals is poultry, that lunches and dinners account for 100 % of the animal living time instead of 75 %, and very optimistically that the marginal cost-effectiveness of School Plates is equal to the ratio in 2023 between benefits and cost, corporate campaigns would still be 26.4 (= 70.3/1.33*0.5) times as cost-effective as School Plates.
In general, this cost-effectiveness analysis is quite linear, and I am not envisionaing any plausible combination of factors which would make the cost-effectiveness of School Plates increasing by a factor of 186.
I hope this isn’t coming off as overly critical.
I strongly upvoted your comment, as it made me think a little bit more about relevant questions. Sorry if I sound overly critical above too!
This can be defined in the Sheet by assuming here a consumption of poultry of 82.3 kg/person/year (meat consumption per capita in the UK in 2021), and no consumption of other meats, hard coding eggs’ suffering-adjusted animal living time per annual food consumption to my original value of 0.198 year/person (in the original sheet, it is defined from the value for poultry), and making the cost per meal swapped to meat-free 2⁄3 as high as my original value (since I assumed only 2⁄3 of meat-free meals are plant-based, and meat-free non-plant-based meals to be as bad as the conventional ones).
This can be defined in the Sheet by assuming here a consumption of farmed fish and seafood of 123 kg/person/year, and no consumption of other meats, hard coding eggs’ suffering-adjusted animal living time per annual food consumption to my original value of 0.198 year/person, and making the cost per meal swapped to meat-free 2⁄3 as high as my original value (see previous footnote for details).
Consequently, the cost per additional plant-based meal in 2023 was 0.0600 $ (= 248*10^3/(4.13*10^6)). Sarah expects this to decrease in the future.
Therefore, I just assumed the “this” referred to the “cost per additional plant-based meal” and not the effectiveness per dollar. This is a factor which can change the effectiveness of School Plates pretty radically, so I’d be careful. Obviously, from a comparison point of view, this alone probably won’t make up the difference, but claiming 186 times as cost-effective is very different from saying 93 times as cost-effective or even 47 times as cost-effective.
I would almost insist this is not a situation where the law of diminishing marginal returns is applicable. From my work at The Mission Motor, I’ve found most organizations are not able to identify where their most-effective activities may be. This isn’t to say anything bad about the organizations themselves, just that these activities are not at all obvious to identify. I spend a fair bit of my time trying to brainstorm reasonable target groups with organizations, but there’s so little information to work from and contexts vary so widely that it just becomes a semi-educated guess for most organizations.
To clarify this point a bit, it’s not too tricky to identify the biggest targets, but it’s really challenging to estimate the resources needed to obtain a commitment and enforce it before you start a campaign. However, once you start a campaign, you can’t really stop because that hurts the success of all future campaigns. Disclaimer: Some organizations are more equipped with MEL or data support than others and ProVeg does have an MEL team who can assist with some of this work. However, ProVeg has two people working on MEL and there are many dozens of interventions being carried out in many different countries; as great as they are, I doubt they can identify the most cost-effective targets very reliably for all their interventions given doing this for one intervention is challenging enough.
Additionally, it appears to me you calculated the average cost per meal rather than the marginal cost per meal in this calculation “0.0600 $ (= 248*10^3/(4.13*10^6))”. I imagine the $248,000 is the total budget for the program serving 4,130,000 meals. At early phases of interventions, it is very common for the marginal cost to be below the average cost and it is also not uncommon for the derivative of the marginal cost function to be negative (it’s kind of expected at the very beginning of a venture). So, I would argue the marginal cost per meal is possibly far less than $0.06 instead of the $0.12 you estimated.
Overall for point 1, over time organizations generally get better at identifying the best targets, staff are upskill and improve their tactics, and many of the supporting materials and tools for an intervention can be reused once created. All of these components would suggest the marginal cost for an intervention would decrease over time rather than increase. Additionally, as an organization grows, specialization and other aspects of economies of scale could continue to decrease costs. We’re also dealing with a social movement, so there may just be less pushback over time as well. I would probably use a marginal cost of $0.03 instead of $0.12 if I had to pick a point estimate here. This changes the comparison to corporate campaigns being 47 times as cost-effective as School Plates—still a large margin, but feels a bit different.
On point 2, many of your points are well taken—namely the linearity of your model. I’m not a huge fan of sheets and would have written then model in Python where it would be relatively easy to turn the model into a MCS, so I would have just done that first instead of thinking through everything you wrote in your comment (different work styles and it seems yours is much more efficient here). Additionally, I may want a lot of this modeling sitting in Python anyway for comparing other interventions or tactics so building the MCS has other benefits for me (not to mention the fact that having a computer program spit out some examples is a nice communication tool for people without a background in probability).
I started playing around in your sheet to get a better sense of why this result seems so counterintuitive to me (nothing wrong with counterintuitive, but if I can understand why, I can learn how to update my ideas in this space). While I am a bit skeptical of the 8.2 chicken lives affected per dollar, I’m not going to jump into all these calculations at the moment, so I’ll just have to accept it for the point of conversation.
However, it does appear to me there is an additional point of major uncertainty for corporate campaigns not present in the School Plates model—the improvement of conditions from conventional systems to cage-free systems. You get to an estimate that cage-free systems generate 22.3% as much suffering as conventional systems. But this is a point estimate on many very uncertain variables. While your point estimate is probably reasonable as a point estimate, I know people who would try to argue this number should be more like 95%. I’m not saying they’re correct or endorsing these estimates in any way, but I feel the need to keep that uncertainty. With your particular ethical slant (particularly the EXPECTED component of your utilitarianism), this probably isn’t very relevant to you personally. Additionally, even using the 95% estimate AND the $0.03 marginal cost estimate would not be enough to make School Plates more effective than corporate campaigns, but the estimate changes to 3 times as effective, which is considerably different.
I think there are other factors such as how much these interventions can shape society in the long run and whatnot which could make the School Plates intervention more effective than corporate campaigns. However, a lot of things would need to go right.
On a higher level, while I am a Bayesian, I still believe there is a “true value” as I think most Bayesians do, even if they don’t talk about it much because Frequentists are so obsessed with this theoretical “true value”. Because there is so much uncertainty in many of these calculations, and corporate campaigns will inevitably never lead to a world without animal exploitation (I know this may not be perfectly utilitarian but I’m not certain of this either) without other complementary interventions, I think abolitionist interventions have their place in the movement—even if just to lay the groundwork for the future. Additionally, I have heard numerous accounts of corporate campaigners sharing how much easier the more extreme abolitionists make their job. After corporations work with an extreme abolitionist, working with THL is so much more attractive.
Overall, I think these interventions do and must work together towards the world we want to create for animals, even if there may be some disagreement about what that ultimate world looks like exactly. This leads me to prefer a pluralistic movement and err on the side of endorsing less effective interventions (at least in the short run) if they are of a different “flavor”. By different “flavors”, I basically mean the tactics and the theories of change are not very related and may even be complementary.
Therefore, I just assumed the “this” referred to the “cost per additional plant-based meal” and not the effectiveness per dollar.
“Sarah expects this to decrease in the future” is indeed supposed to mean that Sarah expects the cost per additional plant-based meal to decrease in the future. In addition, the cost per additional plant-based meal is supposed to refer to the ratio between the annual cost and benefits.
I would almost insist this is not a situation where the law of diminishing marginal returns is applicable.
Thanks for the context around this! I have just asked Sarah about whether she has any guesses for how many x % more additional meat-free meals they would have had in 2023 if they had spent 10 % more in 2023.
Additionally, it appears to me you calculated the average cost per meal rather than the marginal cost per meal in this calculation “0.0600 $ (= 248*10^3/(4.13*10^6))”.
Yes, 0.0600 $ is the average cost per additional plant-based meal in 2023 (in ratio between total cost and benefits in 2023), which I then adjust to get the marginal one.
All of these components would suggest the marginal cost for an intervention would decrease over time rather than increase.
This would suggest the marginal cost-effectiveness would increase over time. In this case, it would make sense for School Plates to save more (and funders to support it less) since spending later (and funders supporting it later) would be more impactful at the margin. Ideally, spending should be moved from the worst to the best years until the marginal cost-effectiveness is the same across years (in the same way that it makes sense at any point in time to move money from the worst to best charities until their marginal cost-effectiveness is the same).
I would probably use a marginal cost of $0.03 instead of $0.12 if I had to pick a point estimate here.
In this case, I believe you are suggesting that the marginal cos-effectiveness of Shool Plates is 2 (= 0.0600/0.03) times their ratio between benefits and cost in 2023, i.e. that them spending 10 % more then would have resulted in 20 % (= 0.1*2) more additional plant-based meals in 2023. This would be surprising to me.
On point 2, many of your points are well taken—namely the linearity of your model. I’m not a huge fan of sheets and would have written then model in Python where it would be relatively easy to turn the model into a MCS, so I would have just done that first instead of thinking through everything you wrote in your comment (different work styles and it seems yours is much more efficient here).
I used to run Monte Carlos more often, but now I tend to prefer Sheets, at least for linear problems where I can foresee the effects of changing inputs, because then more people can review/update the model.
Additionally, I may want a lot of this modeling sitting in Python anyway for comparing other interventions or tactics so building the MCS has other benefits for me (not to mention the fact that having a computer program spit out some examples is a nice communication tool for people without a background in probability).
Fair!
While I am a bit skeptical of the 8.2 chicken lives affected per dollar, I’m not going to jump into all these calculations at the moment, so I’ll just have to accept it for the point of conversation.
Quick clarification. I assume corporate campaigns affect 8.20 years of chicken life per $, not 8.20 chicken lives per $.
While your point estimate is probably reasonable as a point estimate, I know people who would try to argue this number should be more like 95%. I’m not saying they’re correct or endorsing these estimates in any way, but I feel the need to keep that uncertainty. With your particular ethical slant (particularly the EXPECTED component of your utilitarianism), this probably isn’t very relevant to you personally.
For reference, this cell calculates the welfare per unit time of a broiler in a reformed scenario relative to that of one in a conventional scenario, as a function of conversion factors between 4 types of pain, being asleep, and being alive (I assume being alive with hurtful pain is neutral). It is very hard to arrive to that 95 % because, based on data from the Welfare Footprint Project, when a broiler goes from a conventional to a reformed scenario, the time experiencing (the types of pain are defined here):
Excruciating pain decreases 81.5 %. This is close to my assumption that overall suffering decreases 77.7 % (= 1 − 0.223) because I weight extreme suffering very heavily.
Disabling pain decreases 65.7 %.
Hurtful pain decreases 23.4 %.
Annoying pain increases 4.51 %.
For the pain per unit time of a broiler in a reformed scenario being 95 % of that of one in a conventional scenario, one would have to give negligible weight to the above reductions in excruciating, disabling and hurtful pain, and then for some reason weight very heavily a reduction in pain with intensity somewhere between those of annoying and hurtful pain. I think this would be unjustifiable under expected total hedonistic utilitarianism. I would not want to assume the suffering per time in a reformed scenario is more than 50 % as bad as in a conventional one[1], in which case corporate campaigns would be more than 120 times as cost-effective as School Plates.
Additionally, I have heard numerous accounts of corporate campaigners sharing how much easier the more extreme abolitionists make their job. After corporations work with an extreme abolitionist, working with THL is so much more attractive.
That makes intuitive sense to be. On the other hand, I think it would mostly push one towards supporting activities optimising for radical flank effects, not ones aiming to increase the number of plant-based meals at schools and universities.
Overall, I think these interventions do and must work together towards the world we want to create for animals, even if there may be some disagreement about what that ultimate world looks like exactly. This leads me to prefer a pluralistic movement and err on the side of endorsing less effective interventions (at least in the short run) if they are of a different “flavor”. By different “flavors”, I basically mean the tactics and the theories of change are not very related and may even be complementary.
I agree it is better to support interventions whose cost-effectiveness is less correlated everything else equal[2]. At the same time, conditional on corporate campaigns being something like 186 times as cost-effective as School Plates, I do not see how one could justify supporting School Plates over corporate campaigns. If, for the same cost, one could save 186 human lives via intervention A or just 1 via B, it would feel very wrong to me to support B over A on grounds that A and B are quite different.
In any case, thanks for outlining some reasons the ratio of 186 may be too high!
I get 53.2 % here assuming excruciating pain is 10 times as bad as disabling pain (instead of my original 1 k times), that disabling pain is 10 times as bad as hurtful pain (instead of 100 times), and that there are no positive experiences (instead of being alive with hurtful pain being neutral).
Thanks for going through the analysis here! I was a bit confused about a few things.
You cite Sarah Morton as thinking the costs will decrease over time and then you you speculate the marginal cost-effectiveness to be 0.5 (seemingly implying costs will increase over time). You cite diminishing returns as your support for this decision, but Sarah’s comment implying this intervention is in a period of increasing returns (yes this is a thing). Additionally, this may be less of an issue of increasing returns and this intervention could just benefit from economies of scale which would also decrease costs and increase cost-effectiveness. If Sarah is right (and thus the marginal cost-effectiveness is >= 1), School Plates is at least twice as effective as you’ve estimated here. However, very happy to hear I’m misunderstanding something on my end.
It seems like the sole unit of impact measurement is time and intensity of suffering. However, it’s not obvious to me this is all that matters. Suppose we care about the killing of an animal differently than the suffering when they’re alive (I don’t think this is entirely unreasonable, though maybe not the most mathematically convenient model). Then School Plates would get a boost in effectiveness. We could also develop models where the number of animals suffering is a component of our overall utility function (there’s a disutility associated with a factory-farmed animal existing at all), in which case School Plates would also get a boost in effectiveness. My point here is I’m not entirely sure such a narrow definition of impact is appropriate here. I know it makes the math easier and this is just meant to be an estimate, but when the definition of impact favors one intervention over the other in basically every component, the difference in the estimates is going to be exaggerated. Then again, I’m not here to insist your ethics are wrong (I think that’s a personal choice), but I think this adds a lot of nuance to your conclusions.
I hope this isn’t coming off as overly critical. I enjoyed reading this post and think it’s a great starting point for further, highly-relevant work. I’m thinking of potentially building out some Monte Carlo Simulations of your model and Saulius’ model (with Open Phil’s comments) to see how accounting for variance impacts these estimates (my hunch is there will be so much uncertainty it will be hard to decide between the interventions). One additional benefit of the Monte Carlo Simulations is how they point to where collecting more evidence and decreasing our uncertainty would improve our estimates most. Thanks again for posting this!
Thanks for the great points, Blake, and welcome to the EA Forum!
On 1, to clarify, my understanding is that Sarah expects “additional meat-free meals in 2023“/”cost in 2023” to be lower than “additional meat-free meals in 2024“/”cost in 2024”, i.e. she expects the ratio between annual benefits and cost to increase. Despite of this, for any year, I expect that “annual additional meat-free meals”/”annual cost” > “additional meat-free meals from spending an extra e.g. 1 k$ in a given year”/”e.g. 1k$” = “marginal cost-effectiveness”. Do you agree? For any given year, I expect School Plates to start spending on the most cost-effective activities 1st, such that the benefits produced by the last dollar spent in any given year are smaller than those produced by the mean dollar that same year.
On 2, at least in principle, I fully endorse expected total hedonistic utilitarianism, which only accounts for wellbeing. However, in practice, I would agree it can still make perfect sense to account for metrics which do not have to do with wellbeing per se. They could inform one about longer term impacts on wellbeing, or be an indicator about near term impacts on wellbeing which are not being captured by the wellbeing metrics for some reason. It is unclear to me whether accounting for other metrics from this perspective would favour School Plates, as corporate campaigns might also generate momentum for further welfare campaigns, and eventually accelerate the phase-out of factory-farming (or the transition to a net positive factory-farming).
I do not think the Monte Carlo would change the mainline cost-effectiveness estimates much:
The inputs I used are supposed to be means rather than medians or modes. So, if one modelled the distributions such as their means correspond to my point estimates, the cost-effectiveness would be maintained if it was a sum of products, because e.g. E(X1*X2 + X3*X4) = E(X1) E(X2) + E(X3) E(X4) assuming all distributions are independent (as I would if I was to run a Monte Carlo).
In reality, my estimate of the cost-effectiveness involves 2 variables in the denominator, the cost of the program in 2023 and the number of lunches and dinners per person per year, but these are arguably 2 of the most certain variables. So approximating E(1/X) as 1/E(X), as I implicitly did, should be fine. If I wanted to be more rigorous, I could also calculate the mean of the reciprocal of the variables in the denominator from 2 percentiles using the formulas here, thus not having to run the whole Monte Carlo.
A large overlap between the cost-effectiveness distributions of 2 interventions does not necessarily imply uncertainty about which intervention is best if the Monte Carlo assumes all variables are independent. This is because there may be correlations across variables used to calculate the cost-effectiveness of both interventions.
In particular, the welfare range of chickens, and welfare of chickens per time as a fraction of their welfare range are 2 of the most uncertain variables, and will be same for corporate campaigns and School Plates.
Here is an extreme example illustrating the issue I am pointing to:
Intervention A has a cost-effectiveness given by X1*Y, where X1 is a super uncertain distribution spanning many orders of magnitude, and Y is practically a point estimate of 1.
Intervention B has a cost-effectiveness given by X2*Z, where X2 is distributed as X1, and Z is practically a point estimate of 1.1.
Running a Monte Carlo with all the distributions as independent would result in almost perfect overlap between the cost-effectiveness distributions of A and B. So one may be tempted to conclude it is unclear which interventions is best.
However, if X1 and X2 are perfectly correlated, the correct conclusion is that B is almost certainly better than A (1.1 > 1).
In any case, the issue could be solved by setting all uncertain distributions shared across interventions to their means. I guess modelling the correlations within the Monte Carlo would be significantly more work for little benefit.
Before running a Monte Carlo with some guesses for the distributions of each variable, I think it would be good to get more information about the variables with the most reducible uncertainty. Yet, the conclusion that corporate campaigns are more cost-effective is pretty robust:
Corporate campaigns are more cost-effective even for most favourable compositions of baseline meals. Playing around with the Sheet:
Assuming all the meat in the baseline meals is poultry, and the meat-free meals are all plant-based[1], corporate campaigns are still 70.3 times as cost-effective as School Plates.
It makes sense the ratio is smaller than that of 186 I estimated.
Poultry is the meat with the highest suffering-adjusted animal living time (animal living live times welfare range) per mass, equal to 9.53 year/kg (= 28.7*0.332).
Assuming all the meat in the baseline meals is farmed fish and seafood, and the meat-free meals are all plant-based[2], corporate campaigns are still 89.4 times as cost-effective as School Plates.
It makes sense the ratio is smaller than that of 186 I estimated, but higher than that of 70.3 I got just above.
Farmed fish and seafood is the meat with 2nd highest suffering-adjusted animal living time per mass, equal to 7.31 year/kg (= 82.1*0.089).
Reality check. The meat in the baseline meals being poultry is 1.27 (= 89.4/70.3) times as cost-effective as it being farmed fish and seafood, which is quite close to the ratio of 1.30 (= 9.53/7.31) suggested by the suffering-adjusted animal living time per mass. The values do not have to be exactly the same due to the presence of eggs in the baseline meals.
Supposing lunches and dinners account for 100 % of the animal living time instead of 75 % only makes the cost-effectiveness of School Plates 1.33 (= 1⁄0.75) times as high, so corporate campaigns would still be 140 (= 186⁄1.33) times as cost-effective.
Assuming all the meat in the baseline meals is poultry, that lunches and dinners account for 100 % of the animal living time instead of 75 %, and very optimistically that the marginal cost-effectiveness of School Plates is equal to the ratio in 2023 between benefits and cost, corporate campaigns would still be 26.4 (= 70.3/1.33*0.5) times as cost-effective as School Plates.
In general, this cost-effectiveness analysis is quite linear, and I am not envisionaing any plausible combination of factors which would make the cost-effectiveness of School Plates increasing by a factor of 186.
I strongly upvoted your comment, as it made me think a little bit more about relevant questions. Sorry if I sound overly critical above too!
This can be defined in the Sheet by assuming here a consumption of poultry of 82.3 kg/person/year (meat consumption per capita in the UK in 2021), and no consumption of other meats, hard coding eggs’ suffering-adjusted animal living time per annual food consumption to my original value of 0.198 year/person (in the original sheet, it is defined from the value for poultry), and making the cost per meal swapped to meat-free 2⁄3 as high as my original value (since I assumed only 2⁄3 of meat-free meals are plant-based, and meat-free non-plant-based meals to be as bad as the conventional ones).
This can be defined in the Sheet by assuming here a consumption of farmed fish and seafood of 123 kg/person/year, and no consumption of other meats, hard coding eggs’ suffering-adjusted animal living time per annual food consumption to my original value of 0.198 year/person, and making the cost per meal swapped to meat-free 2⁄3 as high as my original value (see previous footnote for details).
Thanks for the welcome!
For 1, you wrote
Therefore, I just assumed the “this” referred to the “cost per additional plant-based meal” and not the effectiveness per dollar. This is a factor which can change the effectiveness of School Plates pretty radically, so I’d be careful. Obviously, from a comparison point of view, this alone probably won’t make up the difference, but claiming 186 times as cost-effective is very different from saying 93 times as cost-effective or even 47 times as cost-effective.
I would almost insist this is not a situation where the law of diminishing marginal returns is applicable. From my work at The Mission Motor, I’ve found most organizations are not able to identify where their most-effective activities may be. This isn’t to say anything bad about the organizations themselves, just that these activities are not at all obvious to identify. I spend a fair bit of my time trying to brainstorm reasonable target groups with organizations, but there’s so little information to work from and contexts vary so widely that it just becomes a semi-educated guess for most organizations.
To clarify this point a bit, it’s not too tricky to identify the biggest targets, but it’s really challenging to estimate the resources needed to obtain a commitment and enforce it before you start a campaign. However, once you start a campaign, you can’t really stop because that hurts the success of all future campaigns. Disclaimer: Some organizations are more equipped with MEL or data support than others and ProVeg does have an MEL team who can assist with some of this work. However, ProVeg has two people working on MEL and there are many dozens of interventions being carried out in many different countries; as great as they are, I doubt they can identify the most cost-effective targets very reliably for all their interventions given doing this for one intervention is challenging enough.
Additionally, it appears to me you calculated the average cost per meal rather than the marginal cost per meal in this calculation “0.0600 $ (= 248*10^3/(4.13*10^6))”. I imagine the $248,000 is the total budget for the program serving 4,130,000 meals. At early phases of interventions, it is very common for the marginal cost to be below the average cost and it is also not uncommon for the derivative of the marginal cost function to be negative (it’s kind of expected at the very beginning of a venture). So, I would argue the marginal cost per meal is possibly far less than $0.06 instead of the $0.12 you estimated.
Overall for point 1, over time organizations generally get better at identifying the best targets, staff are upskill and improve their tactics, and many of the supporting materials and tools for an intervention can be reused once created. All of these components would suggest the marginal cost for an intervention would decrease over time rather than increase. Additionally, as an organization grows, specialization and other aspects of economies of scale could continue to decrease costs. We’re also dealing with a social movement, so there may just be less pushback over time as well. I would probably use a marginal cost of $0.03 instead of $0.12 if I had to pick a point estimate here. This changes the comparison to corporate campaigns being 47 times as cost-effective as School Plates—still a large margin, but feels a bit different.
On point 2, many of your points are well taken—namely the linearity of your model. I’m not a huge fan of sheets and would have written then model in Python where it would be relatively easy to turn the model into a MCS, so I would have just done that first instead of thinking through everything you wrote in your comment (different work styles and it seems yours is much more efficient here). Additionally, I may want a lot of this modeling sitting in Python anyway for comparing other interventions or tactics so building the MCS has other benefits for me (not to mention the fact that having a computer program spit out some examples is a nice communication tool for people without a background in probability).
I started playing around in your sheet to get a better sense of why this result seems so counterintuitive to me (nothing wrong with counterintuitive, but if I can understand why, I can learn how to update my ideas in this space). While I am a bit skeptical of the 8.2 chicken lives affected per dollar, I’m not going to jump into all these calculations at the moment, so I’ll just have to accept it for the point of conversation.
However, it does appear to me there is an additional point of major uncertainty for corporate campaigns not present in the School Plates model—the improvement of conditions from conventional systems to cage-free systems. You get to an estimate that cage-free systems generate 22.3% as much suffering as conventional systems. But this is a point estimate on many very uncertain variables. While your point estimate is probably reasonable as a point estimate, I know people who would try to argue this number should be more like 95%. I’m not saying they’re correct or endorsing these estimates in any way, but I feel the need to keep that uncertainty. With your particular ethical slant (particularly the EXPECTED component of your utilitarianism), this probably isn’t very relevant to you personally. Additionally, even using the 95% estimate AND the $0.03 marginal cost estimate would not be enough to make School Plates more effective than corporate campaigns, but the estimate changes to 3 times as effective, which is considerably different.
I think there are other factors such as how much these interventions can shape society in the long run and whatnot which could make the School Plates intervention more effective than corporate campaigns. However, a lot of things would need to go right.
On a higher level, while I am a Bayesian, I still believe there is a “true value” as I think most Bayesians do, even if they don’t talk about it much because Frequentists are so obsessed with this theoretical “true value”. Because there is so much uncertainty in many of these calculations, and corporate campaigns will inevitably never lead to a world without animal exploitation (I know this may not be perfectly utilitarian but I’m not certain of this either) without other complementary interventions, I think abolitionist interventions have their place in the movement—even if just to lay the groundwork for the future. Additionally, I have heard numerous accounts of corporate campaigners sharing how much easier the more extreme abolitionists make their job. After corporations work with an extreme abolitionist, working with THL is so much more attractive.
Overall, I think these interventions do and must work together towards the world we want to create for animals, even if there may be some disagreement about what that ultimate world looks like exactly. This leads me to prefer a pluralistic movement and err on the side of endorsing less effective interventions (at least in the short run) if they are of a different “flavor”. By different “flavors”, I basically mean the tactics and the theories of change are not very related and may even be complementary.
Thanks for the follow up! Strongly upvoted.
“Sarah expects this to decrease in the future” is indeed supposed to mean that Sarah expects the cost per additional plant-based meal to decrease in the future. In addition, the cost per additional plant-based meal is supposed to refer to the ratio between the annual cost and benefits.
Thanks for the context around this! I have just asked Sarah about whether she has any guesses for how many x % more additional meat-free meals they would have had in 2023 if they had spent 10 % more in 2023.
Yes, 0.0600 $ is the average cost per additional plant-based meal in 2023 (in ratio between total cost and benefits in 2023), which I then adjust to get the marginal one.
This would suggest the marginal cost-effectiveness would increase over time. In this case, it would make sense for School Plates to save more (and funders to support it less) since spending later (and funders supporting it later) would be more impactful at the margin. Ideally, spending should be moved from the worst to the best years until the marginal cost-effectiveness is the same across years (in the same way that it makes sense at any point in time to move money from the worst to best charities until their marginal cost-effectiveness is the same).
In this case, I believe you are suggesting that the marginal cos-effectiveness of Shool Plates is 2 (= 0.0600/0.03) times their ratio between benefits and cost in 2023, i.e. that them spending 10 % more then would have resulted in 20 % (= 0.1*2) more additional plant-based meals in 2023. This would be surprising to me.
I used to run Monte Carlos more often, but now I tend to prefer Sheets, at least for linear problems where I can foresee the effects of changing inputs, because then more people can review/update the model.
Fair!
Quick clarification. I assume corporate campaigns affect 8.20 years of chicken life per $, not 8.20 chicken lives per $.
For reference, this cell calculates the welfare per unit time of a broiler in a reformed scenario relative to that of one in a conventional scenario, as a function of conversion factors between 4 types of pain, being asleep, and being alive (I assume being alive with hurtful pain is neutral). It is very hard to arrive to that 95 % because, based on data from the Welfare Footprint Project, when a broiler goes from a conventional to a reformed scenario, the time experiencing (the types of pain are defined here):
Excruciating pain decreases 81.5 %. This is close to my assumption that overall suffering decreases 77.7 % (= 1 − 0.223) because I weight extreme suffering very heavily.
Disabling pain decreases 65.7 %.
Hurtful pain decreases 23.4 %.
Annoying pain increases 4.51 %.
For the pain per unit time of a broiler in a reformed scenario being 95 % of that of one in a conventional scenario, one would have to give negligible weight to the above reductions in excruciating, disabling and hurtful pain, and then for some reason weight very heavily a reduction in pain with intensity somewhere between those of annoying and hurtful pain. I think this would be unjustifiable under expected total hedonistic utilitarianism. I would not want to assume the suffering per time in a reformed scenario is more than 50 % as bad as in a conventional one[1], in which case corporate campaigns would be more than 120 times as cost-effective as School Plates.
That makes intuitive sense to be. On the other hand, I think it would mostly push one towards supporting activities optimising for radical flank effects, not ones aiming to increase the number of plant-based meals at schools and universities.
I agree it is better to support interventions whose cost-effectiveness is less correlated everything else equal[2]. At the same time, conditional on corporate campaigns being something like 186 times as cost-effective as School Plates, I do not see how one could justify supporting School Plates over corporate campaigns. If, for the same cost, one could save 186 human lives via intervention A or just 1 via B, it would feel very wrong to me to support B over A on grounds that A and B are quite different.
In any case, thanks for outlining some reasons the ratio of 186 may be too high!
I get 53.2 % here assuming excruciating pain is 10 times as bad as disabling pain (instead of my original 1 k times), that disabling pain is 10 times as bad as hurtful pain (instead of 100 times), and that there are no positive experiences (instead of being alive with hurtful pain being neutral).
Basically, for similar reasons that lead me to invest in global stocks rather than the single company I expect to grow the most.