Introduction
I created a simple web-based tool which ranks animal species according to the harm caused by consuming them. The user can specify the relative priority of two subscales of harm: animal suffering and greenhouse gas emissions.
Numerous analyses have been published on how much suffering is caused by eating various animals. For example by Peter Hurford, Brian Tomasik, Charity Entrepreneurship and Dominik Peters. Results of these analyses hint at the small animal replacement problem which is the concern that advocating for reduced meat consumption for environmental reasons leads people to replace beef with smaller animals such as chicken and fish. This increases total suffering because more farmed animals are consumed for the same amount of calories.
I was inspired by Dominik Peters' tool and was wondering if a similar ranking could be developed which accounts for animal suffering, greenhouse gas emissions and human health. My main motivation was to better understand the climate change/animal welfare trade-off when deciding which animals to leave off our plates. Due to difficulties with modeling health effects I eventually narrowed down the harms to just animal suffering and greenhouse gas emissions and developed a tool based on Dominik's model.
Methods
A simple model is used to calculate the animal suffering and greenhouse gas emissions subscale scores of each species in the data set. The subscale scores are then combined into a single score which is used to rank the species.
The animal suffering subscale estimates the number of hours spent on a farm to produce 2000 kcal of food energy. The climate change subscale estimates CO2-equivalent greenhouse gas emissions produced per 2000 kcal of food. The suffering subscale can be adjusted according to the relative suffering intensity of the species and brain size/neuron count. Both subscales can be adjusted by supply and demand elasticity.
The subscale scores are exponentiated using the subscale priorities that the user has provided and then multiplied to get a single score (weighted product model). The combined scores are normalised to the 0-100 range and used to display a ranking of the species based on the estimated harm.
The user interface allows the user to set the subscale priorities, toggle the adjustments, change the relative suffering intensities and choose the brain weighting function. The goal is to enable the user to specify their beliefs if they don't agree with the default parameters.
When playing around with the sliders it seems that the model is generally consistent with the welfare/climate trade-off. If climate is prioritised, ruminants rank higher on the combined scale. If welfare is prioritised, smaller animals rank higher on the combined scale.
Limitations
The model does not consider indirect effects on wild animal welfare. The suffering of wild animals could significantly exceed that of farmed animals. Indirect effects of farming contribute to wild animal suffering. It would be interesting to also analyse how changes in animal consumption affect wild animals through indirect effects on feed crop production.
It is difficult to come up with meaningful subscale priorities. It would make sense to measure the disvalue of emissions and suffering based on the underlying values which cause us to be concerned about these issues in the first place. If, for example I am motivated by improving welfare, it would be helpful to estimate the welfare impacts of climate change and factory farming on a common scale which seems difficult.
Heather Browning's doctoral thesis outlines several issues with common methods of measuring animal welfare. This includes the hours lived on a farm and relative suffering intensity methods that were used in this work. Jason Schukraft has also written about difficulties in measuring the intensity of valenced experiences.
Notes
More detailed information about the website is provided on the methods page.
I want to thank Dominik Peters for providing the data and model that he used for ethical.diet.
Cool idea!
I'm not sure I understand how it works, but isn't one of the most important parameters how someone would want to trade 1 tonne of CO2 for 1 h of suffering on a factory farm? I.e. I could imagine that ratio could vary by orders of magnitude, and could make either the suffering or the carbon effects dominate.
It seems like your current approach is to normalize both scales and then add them. This will be implicitly making some tradeoff between the two units, but that tradeoff is hidden from the user, which seems like a problem if it's going to be one of the main things driving the results.
Moreover, (apologies if I've misunderstood) but as far as I can see, the way the tradeoff is made is effectively that whichever animal is worst is set to 100 on each dimension. This doesn't seem likely to give the right results to me.
For instance: Perhaps I think beef = 10 CO2, and chicken = 1 CO2 Beef = 1 unit suffering, chicken = 100 units of suffering
In your process, I would normalize both scales so the worst is '100 points', so I'd need to increase beef to 100 and chicken to 10 on the CO2 scale.
If I weight each at 50%, I end up with overall harm scores of: Beef = 100 + 1 = 101 Chicken = 10 + 100 = 110
However, suppose my view is that 1 tonne of CO2 doesn't result in much animal suffering, so I think 1 unit of suffering = 100 CO2.
Then, my overall harm scores would be:
Beef = 10/100 + 1 = 1.1 Chicken = 1/100 + 100 = 100.1
So the picture is totally different.
(If instead I had a human-centric view that didn't put much weight on reducing animal suffering, the picture would be reversed.)
I could try to fix the results for myself by changing the relative weighting, but given that I'm not given any units, it's hard for me to know I'm doing this correctly.
Thank you for the thoughtful feedback, Benjamin! I will try to explain the model a bit more thouroughly than the methods section of the post.
Let's forget normalising and weights for a moment. If we measure suffering in hours/kcal and emissions in CO2eq/kcal then the subscales have different units and can't be added (unless we have a conversion formula from one unit to the other somehow). A common solution in this case is to multiply the subscale values. If we do this a 1% change in suffering changes the combined score by the same amount that a 1% change in emissions would.
We still might want to prioritise some subscales more than others. If we would have added subscale scores we could have multiplied the subscale scores by some constant weights beforehand. If instead we multiply subscale scores we would exponentiate the subscale scores by weights beforehand. This simple idea is called a weighted product model (WPM) in the multiple-criteria decision analysis discipline which studies how to make decisions when we have multiple conflicting criteria.
This tool uses a weighted product model. The unnormalised suffering and emissions scores are:
1. exponentiated by their corresponding weight,
2. multiplied together to get a combined score,
3. the combined score is normalised to the 0-100 range for cleaner display.
WPM is a dimensionless method used for ranking options when making decisions. That is, to answer questions like "is it more important to avoid chicken or beef" not "what is the cardinal utility of avoiding chicken". This model is only useful for prioritising if I have decided to reduce meat consumption but am only able to leave one species off my plate. I understand now that I should have made it more clear.
Somehow measuring the utility of leaving a species off my plate would be much more interesting but seemed difficult considering the time and skills I had. I did consider using something like DALYs. There is research on converting emissions to DALYs which would allow us to use a parameter for converting non-human animal DALYs to human DALYs but I decided for the simpler ranking-only model.
That makes sense. The point I'm trying to make, though, is that the choice of how to do the conversion from CO2/kcal to hours/kcal is probably the most important bit that drives the results. I'd prefer to make that clearer to users, and get them to make their own assessment.
Instead, the WPM ends up coming up with an implicit conversion rate, which could be way different from what the person would say if asked. Given this, it seems like the results can't be trusted.
(I expect a WPM would be fine in domains where there are multiple difficult-to-compare criteria and we're not sure which criteria are most important – as in many daily decisions – but in this case, it could easily be that either CO2 or suffering should totally dominate your ranking, and it just depends on your worldview.)
You are right. I spent time thinking about your comments and I agree that making the tradeoff clearer is one of the most important improvements I can make. Thank you for bringing it out.
Also, have you considered adding other farmed fish species to this tool, like tuna, milkfish, and tilapia? I'm from the Philippines where milkfish and tilapia are by far the two most consumed farmed fish species, based on this report by Fish Welfare Initiative. If anyone can point me to sources on relative welfare or climate impact between different farmed fish species, I'd be interested to read :)
I will try to look for data on more species but it can be difficult to find.
This is awesome! I like the model, and the UI is intuitive and clean. Two requests/suggestions:
I was about to say the same thing - I skimmed for "eggs" and it took me a bit to figure it out.
Thank you for the feedback Ben!
I completely agree about "caged hen" and "broiler". The issue was that there's not a lot of room on the plot (especially on phones). Initially I even had S-G broiler for slow-growth broiler which was even more misleading. I will try to come up with a better layout for the plot so I can use longer labels.
Currently the domain is a sub-domain on my personal website. Do you have suggestions for a new domain name?
more-ethical.diet :)
I am not good at coming up with names for things, but do agree a specific URL would be nice. Maybe "meat impacts estimator" (although it's not just about meat, but "animal products impacts estimator" or "farmed animals impacts estimator" are pretty long), or some other combination of words like:
Love the endeavor. But the calculation method really should be changed before anyone interested in the quantification of the combined CO2+animal suffering harm should use it, in my opinion: a weighted product model is inappropriate to express the total harm level of two independent harms, I really think you want to not multiply CO2 and animal suffering harm, but instead separately sum them, with whichever weight the user chooses. In that sense, I fully agree with what MichaelStJules also mentioned. But I want to give an example that makes this very clear - and please let me know if instead, it seems like I misread your calculation details in https://foodimpacts.org/methods :
Imagine a product A with 0 CO2 but a huge animal suffering impact, B with huge CO2 but 0 suffering, and C with non-zero but tiny impact on both dimensions. Your weighting would favor either A or B (or both), while for any rational person C would necessarily be preferable. Your WPM may sound nicer in theory but it cannot be applied here, I'd really want to see it changed before considering the model usable for quantitative indications of the harm on a general level!
NB: I actually have an interest in using your model in the medium-term future! We're trying to set up an animal food welfare compensation scheme, and happen to have CO2 on our list in addition to animal suffering itself, www.foodoffset.org (very much work in progress).
I think it's the other way around. Under a weighted product model (WPM), the overall impact of both A and B is zero because either component is zero, so the WPM favors A and B over C. Whereas summing the climate and welfare components (with "reasonable" weights) would result in C being the most favorable.
You were of course right; I now fixed the A B & C round to make them consistent. Thanks!
Thank you for the feedback Florian! I will move this issue upwards in my to do list since so many of you have explained the issues with WPM.
This is very cool! Good job!
Some thoughts:
Also, it might be worth taking a look at this model, too, for more ideas.
Thank you for the feedback, MichaelStJules! I added all of your ideas to my todo list.
I definitely should have added probability of sentience to the model. I looked at Brian Tomasik's model which included sentience multipliers and I have read the OPP report you linked so I don't know why I didn't consider it.
Jason Schukraft's "Differences in the Intensity of Valenced Experience across Species" was great and I will be happy to study his other research. Thank you for linking to it.
I wish I was aware of kbog's post before I started. I managed to find multiple analyses of suffering but I didn't know someone had already devised a combined model!
This is really neat, thank you for making it!
Question: Did you take into account the following?:
And a couple requests:
Thank you for the feedback, Gina!
You are the first person to mention both culling of male chicks and fish who are fed to farmed salmon. You are correct that these are shortcomings. We must also be mindful of breeder chickens used in broiler farming. Unfortunately, accounting for all of these effects would be difficult and would require data that could be difficult to find. About your requests:
1. I will consider it.
2. I assume that understanding the impacts of eating wild animals would be even more difficult so for the sake of keeping it doable I would rather not venture there even though wild animal welfare is an important area to me personally.
3. Brian Tomasik has written an interesting in-depth article on this question and he remains undecided. It seems that this is unfortunately a fundamentally difficult question. It troubles me as well that I don't know much about such a sensitive parameter.
Great work! I think it might be a good idea for you to state on the page that the numbers are per kcal of energy. I clicked the link before reading your post and initially assumed it was the impact of eliminating the category from a standard diet. For what it's worth, I think it could be useful to have "impact of category in a standard diet" as an option on the page.
Thank you for the feedback! I took note about the per kcal issue and will try to fix it soon.
It would definitely be useful to also account for current levels of consumption. The impact of avoiding broilers would probably increase.
Multiple people suggested getting a simpler domain name. The website is now also accessible at foodimpacts.org.
This tool is great! I thought I had read enough about animal welfare and this topic but apparently not, since I had these reactions:
The comments above were when I was playing with the relative priorities, and I didn't make any changes yet to other parameters. I now see that the brain size and neuron count checkboxes can also drastically affect the tool. But I assume that the default you set is our best guess on how to determine sentience?
Also, something I'd be curious on is how much climate impact certain crops or plant-based products have, and if they could be put on the same tool. I'd be interested to know how much better both for climate and animal welfare consuming various plant-based products are compared to the various animal products.
Thank you for the feedback Brian!
1. Shrimp have a high suffering score because the suffering subscale measures hours lived on a farm to produce 2000 kcal of energy. Since shrimp are small you would have to eat a large number of small beings to gain 2000 kcal of energy. A beef cow's carcass yield is over 200 kg so 2000 kcal worth of beef is just a fraction of the yield.
If you turn off the "consider brain size" parameter you will see that the differences are even larger. This illustrates the concern of the small animal replacement problem: we might not be very certain if shrimp are sentient but if they are then eating shrimp would cause suffering to a huge number of animals. By expected disvalue it would make sense to avoid shrimp even if the probability of sentience is low.
There is currently no parameter for sentience probabilities. Brain size/neuron count weighting is the only proxy for this. Maybe a "sentience probability" adjustment would be useful as well?
2. The difference between caged hen egg and cage-free egg is sensitive to the "relative suffering intensity" parameter which is an informed opinion but not empirical data. It originates from an old version of Brian Tomasik's similar analysis. You should change the scores for hens if you have reason to believe that the difference is greater.
Measuring the welfare of farm animals in different farming systems on a cardinal scale is very difficult and is discussed in Heather Browning's PhD thesis which I linked in the post. I hope that this area develops further and in the future we have data and would not have to resort to beliefs anymore when adjusting this parameter.
3. Higher welfare egg production systems have lower production which means you need more birds and feed to produce the same amount of food. This partially negates both the welfare and climate benefits. Higher welfare systems also have more space per bird which increases heating and lighting costs. If I googled "cage free egg LCA" then the first result seemed like a useful discussion of the environmental impacts of different egg production systems.
Thanks for the tool, this post and these explanations. Perhaps, given how counter-intuitive the problem of shrimp suffering is for non-EAs, it’d be better to have a paragraph with some remarks about it on the front page, linking to the methodology section and to CE’s report on this area. Also, you could explicit that it does not account for by-catch in wild shrimp fishing (which is also a major source of harm).
I notice that shrimp is usually consumed as a delicacy; unlike beef, it’s not going to be the main source of nutrients in the corresponding dish, and people will not consider the weight of the product necessary for 2000kcal. Thus, I wonder if it wouldn’t be interesting to have the option of disregarding the refweight parameter for computing the suffering score.
Thank you for the feedback Ramiro! Since people have asked so many questions about the results of this kind of analysis I was thinking about adding a FAQ section to the website. I could add the remarks about shrimp welfare there.
User RandomEA suggested here that there should be an option to use the number of animals consumed as part of a common diet instead of the current animals-per-2000kcal logic. Would you consider that a good solution to the issue that shrimp is realistically a delicacy?
This is really nicely done and it is exactly what many are looking for. Thank you so much!
If it is to be shared more widely it might help to add a remark about how sensitive the results are to which country the animal products are from and whether they're organic or not. The reason for this being that many in the public sphere (and not infrequently wrongly) assume that this makes a crucial difference.
Thank you Dominic! There is already a paragraph about transportation on the methods page of the website. I will add a paragraph about organic food as well.
Great!
And just to add a small comment: The country of origin does not only affect transport distance but also the legal standards for animal welfare (and to a lesser extent how much GHGs are involved in production). My impression is that many people overrate this. They think "Oh yes, there is horrible animal farming elsewhere - but I only eat meat from my own country and surely everything is much better here." It would be nice to have something to counter this objection.
Just wanted to say that this looks extremely cool, and is something I will definitely be sharing with people in my life! (Especially those who insist on still eating shrimp and almost no other animal products...)
Thank you for the kind words Bella!
Thanks for making this!
I'd like to understand the suffering involved in egg production of free-range hens where the male chicks aren't killed.
What assumptions does "cage free" make here?
Thank you for commenting, Devon. Culling of male chicks is not considered. Unfortunately the tool will not include every effect because that would be rather difficult. As someone pointed out the breeder chickens used in broiler farms are not considered nor are the fish who are killed to feed farmed salmon.
I'm not sure the maceration of male chicks induces any suffering. IIRC, it's approved as a humane killing method by the SPCA or someone like that.
Surprised to see that suffering levels for non-cage eggs are so similar to those from cage eggs. Where do you get the data on that? I skimmed a Brian Tomasik piece on suffering per kg of food that you cited, but it only examines caged hens.
This is an amazing tool by the way. Thanks for making it.
I'm glad to see the inclusion of anthropic units as a function of neuron count/brain mass. Turns out that makes a huge difference. Ideally I'd use brain mass*square(neuron count), but that would be overkill...
In building this, did you come across literature about this question of how anthropic measure relates to mass and neuron configuration? I'd love to see any if you have that. I've got quite an interest in the anthropic measure binding question, my somewhat unconventional stance influences my decisions regarding animal welfare, so I really ought to read whatever's out there.