How moral weights and soil life change cause prioritisation

For an exponent x (n in model 1, m in model 2), the cost-effectiveness of one draw, per unit of output of the intervention (a target organism helped, a death averted, a kg of beef bought or a dollar donated), CE(x), and per dollar, CE_$(x), is

• The two weight models are inspired by Vasco Grilo's posts. The neuron-count post sets the welfare range as a fraction of that of humans to “number of neurons as a fraction of that of humans”^“exponent of the number of neurons”, where the welfare range means individual welfare per fully-healthy-animal-year (comment on the terminology). The metabolic-rate post sets individual welfare per fully-healthy-organism-year proportional to ("BMR at 25 ºC of the organisms"/"BMR at 25 ºC of humans")^m. Model 1 gives trees and microorganisms no weight, since they have no neurons.
• Distributions. LN(a, b) is a lognormal distribution with 5th percentile a and 95th percentile b. Normal(μ, σ) is a normal distribution with mean μ and standard deviation σ. Beta(mean μ, concentration c) is a beta distribution with shape parameters μc and (1 − μ)c.
• Draws. Every input is a distribution except the exponent, including neuron counts and metabolic rates (moral weights), pain intensities (pain intensities) and the welfare of soil organisms in each land class (welfare by land class). For each exponent from 0 to 2 in steps of 0.05, 400 k draws are taken. The same draws serve every exponent (common random numbers). So differences across exponents are not simulation noise.
• Costs. The counterfactual is that no money is spent. A funder's real alternative is spending elsewhere. But every option uses the same farmland per dollar of real resources, about 0.04 m²-years before land-market responses and 0.022 after. So this choice shifts every option by the same amount per dollar, which matters only for near-ties. The cost of each intervention is the money spent on it by the donor (costs). That money is the denominator of the cost-effectiveness per dollar and is not counted again as a welfare loss. Costs the intervention imposes on other people, such as higher prices for consumers, are converted to QALYs and counted. Every dollar spent, by anyone, uses resources that would otherwise produce other goods, which reduces farmland (farmland per dollar).
• Time. Effects count when they happen, with no pure time discounting. But every effect t years ahead gets a weight of 1/(1 + t/T) for uncertainty that grows with time. Effects start when they are expected to: about 2 years from now for GiveWell's grant, which funds nets for 2027–2029; 1 year for HSI's stunners; LN(1, 5) years for cage-free campaigns; LN(1, 8) years for broiler campaigns; about 1 year for the All Grants Fund and salt substitution; LN(1, 6) years for GFI; and LN(3, 10) years for fertility-control agents, which need development and registration speculative. Money spent now is unweighted (births). For GiveWell's grant and the All Grants Fund, the lives saved, their descendants, farmland and income gains are all weighted by 1/(1 + t/T) for t years from now. These effects start 2 years and 1 year from now, respectively. The descendants of the people saved by GiveWell's grant, the All Grants Fund and GiveDirectly's programme are followed across all time. For effects spread over several years, the mean weight over those years is applied. This applies to the saved person's own life, GiveWell's income gains from age 15 to 55 and GiveDirectly's 10 years of consumption gains. It also applies to the years by which HSI, cage-free and broiler campaigns and rodent fertility control bring changes forward, and to the 15 years over which GFI's fall in meat production is spread. School Plates' meals and buying beef take effect at once.
• Means use the expected welfare per organism-year in each land class in place of each draw. This is exact because welfare is drawn independently of everything else. This removes noise that would otherwise swamp the means at low exponents in model 2. Percentiles and the chance of a positive effect use the full draws.
• Soil organisms ÷ target beneficiaries. The second set of charts divides the change in the welfare of soil organisms by the expected change in the welfare of the target beneficiaries, at the same exponent. Soil organisms are soil animals in both models, plus bacteria and archaea in model 2. Trees are not counted as soil organisms. Its lower panel is the chance that the soil change exceeds the expected target change in size. The hatching in the summary marks where the expected soil change is larger in size than the expected target change. The target beneficiaries are the people whose deaths are averted or whose health and income improve, the people whose consumption rises with GiveDirectly's transfers, the hens moved out of cages, the broilers raised to the Better Chicken Commitment, the stunned shrimps and the rodents switched to fertility control. They are also the extra beef-herd cattle raised for the beef bought, the broilers, pigs, cattle and sheep not raised because of School Plates and GFI, the people saved by the All Grants Fund and by salt substitution, and, for GWWC, the target beneficiaries of the interventions its donors fund.
• Probability of sentience is not applied separately: the moral-weight models are taken to set each organism's weight.

Both models make an organism's weight relative to a human equal to a ratio raised to an exponent. An exponent of 0 gives every organism with welfare the same weight; between 0 and 1, the weight grows more slowly than the ratio; at 1, in proportion to it; above 1, faster. The range from 0 to 2 covers a wide range of views on how weight grows with neurons or metabolic rate, rather than committing to one exponent.

One implication: by Haller's rule, Smaller individuals have relatively larger brains for their body sizes. So with a weight that grows less than proportionally with brain size, a given biomass of animals will contain more total sentience if it is made of many small animals (Haller's rule and sentience per unit biomass). This is why soil animals matter so much at low exponents.

For a power of neurons

• Neurons do the processing. Experiences are produced by nervous systems, and more neurons can support more, and more varied or intense, experiences. Brian Tomasik models sentience as an increasing function of brain size, for example its square root or brain size itself (Haller's rule and sentience per unit biomass).
• It fits existing estimates. A weight of (neurons/human neurons)^0.188 explains 89.3% of the variance of the log of the mean sentience-adjusted welfare ranges in Bob Fischer's book about comparing animal welfare across species (Table 8.6, P(Sentience)-adjusted welfare ranges: Mixture Model). The fit uses the 6 species which also have neuron counts in Rethink Priorities' neuron count model (figure below).
• Experiences may differ in size, not just intensity. Phil Trammell argues that it seems more straightforward to me to suppose that moral weight tends to scale superlinearly with inputs like neuron count. Otherwise, taking a brain apart and recomposing it into several smaller brains would multiply moral weight, a funny non-monotonicity (Experience size). This supports exponents above 1.

For a power of metabolic rate

• Producing experiences takes energy. In Vasco Grilo's words, Any production requires energy. So I feel like metabolic energy consumption should be relevant for the production of (positive or negative) welfare. (reply to Bob Fischer on "The metabolic rate of the biosphere and its components").
• It extends beyond animals. Metabolic rate is defined for organisms without neurons, such as trees, bacteria and archaea. So it can weigh them if they have welfare at all, which neuron counts cannot.
• It tracks neurons among animals. Brain mass and metabolic rate both scale as roughly a 2/3 to 3/4 power of body mass (Haller's rule and Kleiber's law). So, as Tomasik notes, the ratio of brain size to metabolic activity is roughly constant regardless of body size.

For both

• Biology runs on power laws. Geoffrey West writes that Kleiber's law is valid across almost all taxonomic groups, including mammals, birds, fish, crustacea, bacteria, plants, and cells. He adds that similar scaling laws hold for essentially all physiological quantities and life-history events, including growth rate, heart rate, evolutionary rate, genome length, mitochondrial density, gray matter in the brain, life span, the height of trees and even the number of their leaves. They are all “power laws” (Scale, quoted in a comment by Vasco Grilo). If the capacity for welfare depends on physiology, a power of a physiological quantity is a natural first guess for how it scales.
• Brain metabolism connects them. Brian Tomasik suggests that, to compare how much information processing brains do, we should look at brain metabolism, since faster processing and more complex processing both use more energy. He adds that Brain metabolism is sort of like gross domestic product (GDP), a crude aggregate (Brain metabolism as a key metric). He also notes that an article suggests that the ratio of brain metabolism to total body metabolism may be roughly constant (Haller's rule and sentience per unit biomass). If so, whole-body metabolic rate is a proxy for brain metabolism, and so is brain size, which tracks neurons.

Against

• Metabolism is not cognition. Commenting on “The metabolic rate of the biosphere and its components”, Bob Fischer objects that Endotherms sacrifice energy efficiency for thermal independence; ectotherms sacrifice thermal independence for energy efficiency. But these traits don't obviously have much to do with the cognitive capacities of the animals in question. Most of an animal's metabolism powers muscles, digestion and body temperature rather than its nervous system. This is why model 2 gives warm-blooded animals and trees much more weight relative to insects than model 1 does.
• Neurons are not all alike. Neuron counts ignore neuron size, connectivity and architecture, and capacities such as pain may depend on specific structures rather than totals.
• Power laws describe physiology, not welfare. No theory of consciousness derives the form, and the true relationship between welfare and neurons or metabolic rate could have thresholds or saturate.

Assessment speculative: among animals, neurons are the more plausible proxy for the capacity for welfare. Metabolic rate is mainly useful for asking what happens if organisms without neurons count. Neither is well grounded. Both the exponent and the choice between the models change which intervention looks best. This is why the results are shown for both models across the whole range.

1. GiveWell's largest grant to the Against Malaria Foundation

GiveWell's largest grant to date is $276 million to the Against Malaria Foundation for the procurement and distribution of 92 million nets in the Democratic Republic of the Congo (DRC) from 2027 through 2029, 46 million of which we think would not otherwise be distributed. GiveWell expects it to avert roughly 67,000 deaths, $4.12 k per death averted. Per death averted, the simulation counts the gains to the people saved and their descendants, the losses from the births averted, and the other health gains and long-term income gains that GiveWell also credits to nets. It also counts the farmland used by the extra years of human life and the chickens the extra people eat, the farmland and farmed animals behind the extra income, and the farmland effect of the money spent. Assumptions:

• Remaining life: LN(55, 78) years for a young child in the DRC. In 2023, the Democratic Republic of Congo’s life expectancy at birth was 61.9 years (Our World in Data). The UN projects 63.0 years in 2030, 66.0 in 2050 and 71.9 in 2100 (Our World in Data, medium projection). A child born around 2027 who survives its first years has roughly 66–70 years ahead, less an allowance for the higher risk of the children that nets save. The width of the interval is a judgement speculative.
• Deaths at age 5 or older: a share Beta(mean 0.26, concentration 50) of the deaths averted. GiveWell's model for this grant has Total deaths averted among people under age 5 of 64.4 k and Total deaths averted among people age 5 and older of 23.0 k, 26.3% of the total (cost-effectiveness analysis, column "DRC"). These people have LN(20, 55) years ahead, a share Beta(mean 0.6, concentration 10) of a child's descendants, as fewer of their children are still to come, and avert a share Beta(mean 0.5, concentration 10) of a child's births averted, as parents replace the deaths of young children more than older ones speculative.
• Other benefits: GiveWell's model also credits Units of value from long-term income increases (before accounting for leverage and funging) of 2.95 M, against 9.17 M from deaths averted (7.49 M from under-5 deaths and 1.68 M from deaths at 5 or older). It also credits a Total supplemental intervention-level adjustment of 50.5%, including Malaria morbidity 9%, Short-term anemia effects 9%, Prevention of diseases other than malaria (e.g. dengue, yellow fever, zika, encephalitis) 2%, Prevention of stillbirths 9% and Treatment costs averted from prevention 20% (cost-effectiveness analysis). The model adds morbidity, anaemia and other diseases as LN(0.1, 0.5) times the QALYs of the lives saved. These three are 20% of GiveWell's total value, or about a quarter of its value from deaths averted. It adds long-term income gains of LN(10, 80) increases of 1 in ln(consumption) for a person-year per death averted, valued as for GiveDirectly's recipients (cash transfers). GiveWell's Total present value of units of increase in units of annual ln(income) among people under age 15 is 2.05 M. That is 23.4 per death averted at its Discount rate for future benefits of 4%, for gains that start 10 years after the nets are distributed and last 40 years. Undiscounted, that is about 70. The mean here, 34.5 before the time weighting, is lower, since these gains rest on few studies speculative. The gains are then weighted for uncertainty over time, as all effects are (method). The extra income is consumed at LN($300, $800) per person-year. Households and NPISHs Final consumption expenditure per capita (constant 2015 US$) was 304 dollars in the DRC in 2024 (World Bank), about $373 in current dollars. Incomes will be higher when the children protected are adults. It is spent on food and other goods as GiveDirectly's recipients spend their transfers, with the same farmland and farmed animals (cash transfers). Stillbirths prevented and treatment costs averted are left out (not modelled).
• Births averted: parents have fewer later children when child mortality falls. David Roodman's review for GiveWell concludes that the effect varies by context, above all with total fertility, and is rarely greater than 1:1. For countries where fertility stays high, such as Niger and Mali, it concludes that mortality-drop-fertility-drop ratios of 1:0.5 and 1:0.33 appear more plausible. Here, saving a life can be expected to increase population in the short-term. (summary of the review). For nets specifically, the ITN campaigns raised the total fertility rate by 0.71 children in the overall sample, but this large positive effect only lasts for 1-3 years before turning negative (Wilde, Apouey, Coleman and Picone 2020). This suggests that nets shift the timing of births more than their number. Births averted per death averted are Beta(mean 0.42, concentration 10), whose mean lies between those two ratios. The DRC's total fertility rate, 6.05 in 2023 (below), is similar to theirs. The lives not lived count as losses, as the total view used throughout requires.
• Descendants: the children saved, and the children not born, would have children of their own, generation after generation. Each extra person has half the total fertility rate (TFR) in children, since each child has two parents. This is multiplied by survival to reproductive age and by a share Beta(mean 0.5, concentration 5) of children who are additional rather than displacing other people's, for instance through pressure on land. Partner choice is already covered by counting half the TFR, since women's fertility is the binding constraint. Higher density lowers fertility: increases in population density are associated with declines in fertility rates (Rotella et al. 2021). Roodman's paper expects that where fertility is a conscious choice The impact of mortality drops on fertility will be nearly 1:1, so population growth will hardly change. Of the high-fertility countries, it expects that it would be surprising if these few countries do not join the rest of the world. Roodman's near 1:1 replacement concerns parents replacing children who die, which the births averted cover, rather than the saved person's own children. So the share of additional children is set near the short-run share of births not averted (0.58). The width is a judgement speculative. The DRC's total fertility rate is projected to fall from 6.05 in 2023 to 3.66 in 2050, 2.50 in 2080 and 2.16 in 2100 (Our World in Data, UN medium projection). The model uses 3.4 for the children of the cohort saved, born around 2045–2065, and 2.4 for their children, each times LN(0.8, 1.25). The UN's scenarios bracket the medium one: In the low scenario, total fertility is projected to remain 0.5 births below the total fertility in the medium scenario, and the high one is 0.5 above. ±0.5 births is ×0.86–1.14 around 3.66 and ×0.8–1.2 around 2.5. Today, 11% died before the age of 15 (Our World in Data, 2024). Survival to reproductive age is Beta(mean 0.93, concentration 40) for the first generation and Beta(mean 0.95, concentration 40) for the second, as child mortality keeps falling (Our World in Data). After that each generation replaces itself (a net reproduction rate of 1). So an extra lineage shrinks by the share of additional children in every generation, and the effect on population size tends to zero. Generations are LN(27, 33) years apart: DR Congo's mean age at childbearing is 30.08 years as of 2024 (Measured World, from the UN's World Population Prospects 2024). Descendants live LN(60, 80) years, from today's life expectancy of 61.9 years to that of high-income countries today, as it keeps rising (Our World in Data). They live at Beta(mean 0.85, concentration 30) QALYs per year, the ratio of healthy life expectancy to life expectancy in the DRC. That ratio is 53.2 years against 61.6 in 2021 (WHO data on Statistics of the World, healthy life expectancy and life expectancy), 0.863. The saved child's QALYs are its remaining years times the same ratio. Across all time, each extra person has 2.97 descendants on average (median 1.68, 95th percentile 9.33), who live 207 years between them.
• Uncertainty over time: David Bernard models an effect t years ahead with a prior centred on zero and an estimate whose variance grows with t so that the expected effect is the estimate times D(t) ≡ 1/(1 + f(t)/σ²ᵤ). With his preferred estimates, the posterior expected value is 10% of the signal expected value after 561 years and 1% after 6,168 years (Uncertainty over time and Bayesian updating). That is a weight of 1/(1 + t/62.3) on effects t years ahead, and his estimates without fixed effects give about 1/(1 + t/21). The descendants' years are weighted by 1/(1 + t/T), with T from LN(12, 250) years. So are the saved person's own years and farmland, as for every effect t years ahead (method). The interval for T spans Bernard's estimates, including the lower ones implied by his upper bounds without fixed effects. It also allows for longer persistence, since his extrapolation rests on at most 20 years of forecasts. The weighted years of descendants come to 68.2 per extra person, 67.0% of them after 2100. So, net of the births averted, there are 64.0 extra person-years per child death averted.
• Farmland per extra person-year: LN(800, 3.5 k) m²-years at first, a share Beta(mean 0.6, concentration 10) of it cropland and the rest pasture. It falls by LN(0.2%, 1.8%) a year as yields rise and the agricultural frontier closes. So farmland per person tends to zero too. In the DRC it fell by 2.33% a year from 2018 to 2023 (figures below). Worldwide, cropland area per person decreased by about 20 percent, from 0.24 to 0.19 hectares per capita from 2001 to 2023, about 1% a year. But total agricultural land saw only a small (2 percent) contraction (Food and Agriculture Organization (FAO)). So the fall per person has come mostly from population growth, which is projected to end. The lower tail allows for slow progress in sub-Saharan Africa, where Cereal yields did not grow between 2020 and 2024, the latest data point (Marginal Revolution, quoting The Economist). Each child death averted adds 75.8 k m²-years of farmland on average, 49.2% of it during the saved child's own life, and each death at 5 or older adds 64.9 k. These figures come before market responses. The DRC's farmland per extra person is a correlation that also reflects incomes, yields and exports. Part of an extra person's food is also grown elsewhere. So, like every other land change, they are scaled by the share of a land change that survives land-market responses. Agricultural land per person in the DRC was 0.332 ha in 2023, against 0.593 ha for the world (Our World in Data, from FAO). The marginal person matters here. As the population grew from 90.0 M in 2018 to 105.8 M in 2023 (Our World in Data), agricultural land grew from 33.6 M to 35.1 M ha (Our World in Data), 941 m² per extra person. It grew by 1.22 k m² per extra person from 2013 to 2023. All of the increase was cropland, as FAO's figures hold the DRC's pasture constant, while cropland is 48.1% of the total (Our World in Data). The interval runs from below the marginal figures to near the 2023 average of 3.32 k m², and the cropland share lies between the marginal and average shares, as the constant pasture likely reflects missing data speculative.
• Chickens: LN(0.04, 0.3) chicken-years per person-year, with the welfare of village chickens. The DRC had 18.4 M poultry birds in 2018 and 19.1 M in 2024 (Our World in Data), 0.205 per person in 2018 but only 0.0355 more birds per extra person from 2018 to 2023. A standing stock of birds per person equals bird-years per person-year. Marginal demand is partly met by imports, with up to 90% of meat consumed by the urban population being imported (USAID). The interval runs from just above the marginal figure to above the average. Other farmed animals are left out. Supply per person in the DRC in 2023 was only 0.61 kg of pork, 0.51 kg of beef, sheep and goat meat, 0.06 kg of eggs and 3.29 kg of fish and seafood, mostly wild-caught (meat, eggs, Our World in Data, from FAO). In China, it was 42.1 kg of pork and 41.7 kg of fish and seafood (salt substitution). The farmed animals behind extra income are counted under consumption.

In the formula, per death averted: G is the QALYs of the people saved and their descendants, net of the births averted, plus other health gains and long-term income gains, with r_t = 1 for humans. ΔY covers the soil organisms and trees on the farmland used by the extra person-years (extra person-years × farmland per person-year), on the farmland behind the extra income, and on the farmland freed by the money spent (−K × farmland per dollar). It also covers village chickens (extra person-years × chicken-years per person-year) and the farmed animals behind the extra income. C_h = 0, and K is the cost per death averted (costs), with every effect weighted from now, starting 2 years from now, when the nets are in use.

2. The Navigation Fund's cage-free campaigns fund

Per hen helped, for a donation to The Navigation Fund's Bringing an End to Cages fund: All contributions to this expert-advised fund go directly to support high-impact campaign groups holding companies accountable to their cage-free commitments., and The Navigation Fund does not retain any portion. It has room for more: The Navigation Fund anticipates that its General Farm Animal Fund can absorb $40M in additional funding and $31.5M more in restricted funding for its Cage-Free Fund. (Rethink Priorities' methodology for the Cross-Cause Fund). Campaigns funded now secure and enforce pledges by particular companies to buy only cage-free eggs. These pledges move laying hens from conventional cages to barns or aviaries, increasingly in middle-income countries where most eggs still come from caged hens. A hen helped is a hen moved out of a cage without accounting for changes in the consumption of eggs. Per hen helped, the simulation counts the share of these hens that stay out of cages, the hen's gain, the extra feed cropland, and the change in egg demand and the foods that replace eggs. It also counts the extra cage-free hens needed, the cost to consumers, and the farmland effect of that cost and of the campaigns' spending. Evidence and assumptions:

• Cost: the EU's average cost of producing eggs is 98.5 euro cents per kg in enriched cages and 120 in barns, 21.8% more, in the Wageningen report Competitiveness of the EU egg sector (van Horne and Bondt 2023, Tables 3.2 and 3.4). In the US, Matthews and Sumner (2015) find the cage-free eggs costing 36 percent more to produce than conventional eggs (UC Davis review). The model uses LN($0.20, $0.55) more per kg of eggs. At 0.841 euros per dollar in 2021 (Wageningen report, Table A1.1, row USA 0.880 0.841), the EU gap is $0.256 per kg. That is $0.304 in 2025 dollars at US consumer price indices of 255.657 in 2019, 270.970 in 2021 and 321.943 in 2025 (Bureau of Labor Statistics (BLS) data). In the US study, total costs per dozen are $0.670, $0.913 and $0.756 for conventional, aviary and enriched housing, about 36% higher for the aviary than the conventional house (Matthews and Sumner 2015). A later estimate gives $0.644 for conventional production and $0.908 for cage-free, taken here as 2019 dollars, since its capital costs were re-expressed as of the end of 2019 according to the Consumer Price Index (Trejo-Pech and Thompson 2021). The gap is $0.489 per kg in 2025 dollars, at 680 g per dozen.
• Feed and hens: the Wageningen report notes that In barn/aviary systems egg production is slightly lower and feed intake is higher than in the cage system. The UC Davis review notes that Twice as many hens died in the cage-free system compared to conventional. In the US study of the Coalition for Sustainable Egg Supply, Total accumulated mortality was highest in the aviary system (11.5 percent) against 4.7 percent in conventional cages. In the same study, production from the enriched colony system was approximately 3 percent higher than that of the conventional system, with the enriched colony 6–10% above the aviary (WATTPoultry, preliminary results). That is 2.83–6.36% fewer eggs per aviary hen, or 0.0291–0.0680 extra hens per hen moved. But a meta-analysis of 6.04 k commercial flocks from 16 countries finds that each year of experience with cage-free aviaries was associated with a 0.35–0.65% average drop in cumulative mortality, with no differences in mortality between caged and cage-free systems in more recent years (Schuck-Paim et al. 2021). Feed cost per kg of eggs is 9.69–15.1% higher in barns than in enriched cages across eight EU countries, 11.8% on average (Wageningen report, Tables 3.2 and 3.4). In the US study, Feed costs are higher for the aviary by 2.6% compared with the conventional house (Matthews and Sumner 2015). The model uses LN(4%, 15%) more feed per kg of eggs and LN(0.02, 0.12) extra cage-free hens per hen moved, each living LN(1.1, 1.6) years. The laying phase was assumed to last 40 to 60 weeks after about 20 weeks of rearing (Welfare Footprint Institute), 1.15–1.53 years in total. The eggs no longer eaten remove cage-free hens. Feed cropland per kg of eggs is the stunning model's (Poore & Nemecek, 6.27 m²-years per kg, almost all cropland).
• Eggs per hen: LN(16, 27) kg. Enriched cages in the EU produce 345 to 438 eggs per hen (Wageningen report, Table 3.1), 21.4–26.7 kg at the table's egg weights. In the US, The annual average production per layer on hand in 2024 was 290 eggs (USDA). Brian Tomasik's figures give 16 kg per hen lifetime (How Much Direct Suffering Is Caused by Various Animal Foods?). Lower-yield systems in middle-income countries produce less.
• Demand: |own-price elasticity of total egg demand| LN(0.25, 1.2), mean 0.614. A review of 14 estimates gives eggs a mean of 0.27 (0.08, 0.45) with a range of 0.06–1.28 (Andreyeva, Long and Brownell 2010, Table 1). A US study estimates −0.2348 for conventional eggs and −1.0018 for cage-free eggs (Bakhtavoryan et al. 2021). A campaign funded now raises the price of the targeted companies' eggs only. Their customers can switch to cheaper caged eggs sold by others, which is what the high elasticity for cage-free eggs mostly reflects. That switching leaves total egg consumption unchanged. So the fall in egg consumption per kg of eggs affected is set by the elasticity of total egg demand. But that switching puts some hens back in cages. A global review puts the elasticity for eggs at –0.54 (–0.67 to –0.42), –0.48 (–0.61 to –0.35) and –0.36 (–0.49 to –0.23) in low-, middle- and high-income countries (Green et al. 2013, Table 2), above the US figure. Campaigns now work in high- and middle-income countries. Vasco Grilo's spreadsheet takes 0.970 for eggs in the EU from Table 6 of Bouyssou et al. (2024), in the upper part of the interval. For switching, the elasticity is LN(0.3, 3). In the US study, the compensated cross-price elasticity of cage-free eggs with respect to the price of conventional eggs is 0.8626. The compensated own-price elasticity of conventional eggs is −0.1849. So the Morishima elasticity of substitution is 1.05, close to the median here, 0.948. The width reflects that switching away from the companies targeted, rather than between kinds of eggs, may be easier or harder. So the share of these hens that is lost to caged eggs is that elasticity times the price rise not matched by willingness to pay (below): 9.70% on average. The retail price is LN($2, $6) per kg. In the US it was 2.272 dollars per dozen large eggs in August 2026 (Federal Reserve Economic Data (FRED)), $3.34 per kg at 680 g per dozen. EU eggs sold at packing stations for €272/100 kg in October 2025 (Zootecnica), below retail. So the extra cost raises it by 10.6% on average. Consumers who keep buying from the companies targeted value cage-free eggs at LN(0.02, 1) times the extra cost. In a US survey, half of consumers are willing to pay no more than a $0.30/dozen premium for cage free eggs and the mean premium is $1.16/dozen (Lusk 2018), 0.968 and 3.74 times a cost gap of $0.31 per dozen. Purchases reveal less: most shoppers are not willing to pay such high prices for cage-free and organic attributes (Chang, Lusk and Norwood 2010). Stated values also exceed actual ones, with a median value of the ratio of hypothetical to actual value of 1.35 (Murphy et al. 2003). This brings the survey's median to 0.717 times the cost gap. The median here, 0.141, lies between that and the revealed preferences, as the companies targeted switch because of campaigns rather than their customers' demand speculative. The eggs not eaten are partly replaced by the same substitution basket as shrimp, scaled to egg protein, since people regulate protein intake more strongly than any other dietary component (Scimex, on the protein leverage hypothesis). The basket's composition for eggs is a judgement speculative.
• Net land change. Egg consumption falls by 3.98% of the hens' original eggs on average. Per hen helped, extra feed adds 5.24 m²-years of cropland, lower egg demand removes 1.46 net of the cropland used by substitute foods, and the substitutes add 1.86 of pasture. In total, these add 5.63 m²-years of farmland, after land-market responses. Vasco Grilo's spreadsheet for his post on chicken welfare reforms finds the opposite sign. It takes a Price elasticity of the demand for eggs of -0.97 and a ratio of production costs of 1.21 (1.15/0.95) as the rise in price. So egg demand falls to 0.831 of its level. Then, neglecting increases in the consumption of other foods, cropland falls by 0.529 m²-years per kg of eggs, −0.828 times the increase without the fall in demand. All three steps are questionable. The elasticity of total egg demand is likely lower, with a mean of 0.614 here. Retail prices rise by less than production costs in percentage terms, because retail margins are a large share of the price. Foods also replace part of the eggs not eaten. With his elasticity and price rise but with substitution, the cropland change per kg of eggs is close to zero and total farmland still rises through pasture for beef.

In the formula, per hen helped: G is the gain of a hen moved to an aviary times the share of hens that stay out of cages. ΔY covers the soil organisms and trees on the extra feed cropland, on the land of the eggs not eaten (a loss) and of the foods replacing them, and on the farmland freed by the extra production cost and the campaigns' spending. It also covers the farmed animals behind the replacement foods, and the extra cage-free hens, less those no longer needed for the eggs not eaten. C_h is the extra production cost × (1 − willingness to pay), borne by consumers. K is 1 ÷ hens helped per dollar, divided by the mean weight for uncertainty over the years in which campaigns help hens.

3. The Shrimp Welfare Project's Humane Slaughter Initiative

Per shrimp stunned, for a donation to SWP restricted to HSI. SWP places electrical stunners with shrimp producers through HSI. The share of the cost producers pay raises the price of shrimp, and consumers buy less of it. Part of the shrimp they forgo is replaced by other foods that use more cropland and pasture. That land holds different numbers of soil organisms. The chain is the one in the stunning model, whose script is adapted below as model.py. Its main inputs:

The producer's mean cost of stunning is $1.31×10⁻⁴ per shrimp. SWP pays for the stunners, a share σ = 1 ÷ (1 + running-cost multiple) of the producer's cost (62.7% on average), and producers pay the running costs, the rest. SWP pays for each stunner once. A stunner is counterfactual only for the LN(3, 20) years by which HSI brings stunning forward, capped at its working life: Vasco Grilo's analysis uses 10 years of acceleration of the adoption of electrical stunning, as used by Open Philanthropy (costs) speculative. In this adaptation of the stunning model the fall in shrimp demand is proportional to pass-through × (1 − σ) − ω. The cost of the intervention is the stunner's price × (1 + advocacy and programme costs) ÷ (shrimps stunned a year × years brought forward), divided by the mean weight for uncertainty over those years (costs). That is 5.48 k shrimps stunned per dollar on average. This cost is not a welfare loss (method). Producers and consumers bear cost × (1 − σ − ω), net of consumers' willingness to pay, a welfare loss. As the small price rise exceeds consumers' willingness to pay in expectation, demand falls a little: the mean land change is −3.65×10⁻⁶ m²-years of cropland, −2.93×10⁻⁶ of pasture and −2.76×10⁻⁷ of pond per shrimp stunned, including the farmland freed by the real resources used, which leave the economy with fewer other goods (farmland per dollar). The same change in food demand also changes the numbers of farmed broilers, hens, pigs, fish, cattle and shrimps (other inputs).

In the formula, per shrimp stunned: G is the gain per shrimp stunned, with r_t for a slaughter-weight shrimp. ΔY covers the soil organisms and trees on the land of the extra shrimp eaten and of the foods it replaces, and on the farmland freed by the real resources used. It also covers the farmed animals behind those foods and the extra farmed shrimps. C_h is cost × (1 − σ − ω). K is the stunner's price × (1 + advocacy and programme costs) per shrimp stunned in the years brought forward, divided by the mean weight for uncertainty over those years.

4. Conservation X Labs' rodent fertility-control work

Per rodent switched from rodenticide to fertility control, in a pest population brought under control either way. The fundable route is research, through a donation to Conservation X Labs restricted to its work on rodent fertility control. The Wildlife Technology Fund at Renaissance Philanthropy plans a programme to Develop agents that achieve high fertility reduction, substantial reductions in suffering versus rodenticides, comparable ecological effects, and high palatability., This program is being carried forward from Wild Animal Initiative’s existing work with Conservation X Labs on the topic. We have six CRP concepts ready to develop, and we aim to launch 2–3 of them in Year 1. They are run by program managers who commit to fixed-term programs (typically 3–5 years) (Wildlife Technology Fund). Anticoagulant rodenticides kill over days: Vasco Grilo's post on rodent fertility control cites times to death of 1-3 days for rats and up to 11 days for mice. A rodenticide guideline gives Minimum time to death is two to three days, with average times to death being four to eight days (Rodenticide Resistance Action Group 2021). Fertility control lets the rodents live on and die of other causes, while the population declines through fewer births. Assumptions:

• Fertility control achieves control for a share Beta(mean 0.5, concentration 2) of the rodents switched, uniform from 0 to 1. Its field evidence is thin and mixed. A review concludes that the field efficacy of ContraPest© cannot be assessed from the published literature and that Systematic field studies using orally delivered contraceptives are rare (Massei, Jacob and Hinds 2024). A study by the manufacturer's staff found that Rat populations on protein production farms decreased by an average of 46% following 100 days of treatment (Pyzyna et al. 2018). But in that trial both groups also got rodenticide: The control group was baited with rodenticide only. Meanwhile, in New York City's test no clear and consistent decrease in rat activity was detected over time (NBC Washington). A modelling study finds rodent fertility control effective in reducing population size only when sustained over multiple generations and was ineffective when application was episodic (Shuster et al. 2023). With one field result for fertility control added to rodenticide and one null result, a uniform share expresses that anything from failure to full control is plausible speculative. For the rest, rodenticide is used anyway. So its bait and deaths remain.
• Baits. Rodenticide bait is mostly oats (61.0%), palm oil (14.0%), sugar (10.7%) and wheat (9.42%). Fertility-control bait is palm oil (36.8%), sugar (10%) and water. These are the compositions in the post above. With land use per kg from Poore & Nemecek of 7.6 m²-years for oatmeal, 3.85 for wheat and rye and 2.04 for cane sugar (Our World in Data), and 2.66 for palm oil, the baits occupy 5.59 and 1.18 m²-years per kg. For palm oil, Poore & Nemecek's Data S2 gives a mean of 2.42 m²-years per litre, 2.66 per kg at 0.91 kg per litre. For comparison, From each hectare of land, you can produce about 2.9 tonnes of palm oil (Our World in Data), 3.45 m² per kg. Each is times a lognormal with mean 1 and 5th and 95th percentiles a factor of 1.5 either side of its median, for uncertainty about the average producer, since Impact can vary 50-fold among producers of the same product (Poore & Nemecek). As for other foods, only the share of the change that survives land-market responses counts. Each rodent switched needs LN(0.03, 0.3) kg of rodenticide bait or LN(0.1, 2) kg of fertility-control bait. Rats eat about 10 percent of their body weight per day (University of Nebraska–Lincoln), 25–30 g, for the 2–4 days before anticoagulant poisoning stops them feeding. Bait is also eaten by rats that would die anyway and by other animals. Fertility-control bait must be eaten for months while the population declines: 1 lb of Evolve will bait 25 rats per month, and rats will regain their fertility within approximately 6 weeks (SenesTech). That is 18.1 g per rat-month, or 0.109–0.327 kg over 6–18 rat-months per rodent switched. The liquid version, ContraPest, comes in 8 fl. oz. (237 ml) refills, a 1-Month supply based on average consumption per station (SenesTech). The widths are a judgement speculative. That is more than the post's estimate of about 56 g of each. So the expected land change has the opposite sign to the post's.
• Extra cost of LN($2, $40) per rodent switched for bait stations, refills and labour, borne by pest-control clients. A 12 lb pail of Evolve costs $209.32 (DoMyOwn), $0.70 per rat-month, or $4.20–12.6 of bait over 6–18 rat-months, before stations and the monthly servicing. The city of Newton reports spending about five times as much on ContraPest as they had on rodenticide (Rethink Priorities). This cost and the programme's spending use real resources (farmland per dollar).
• Cost per rodent switched: the programme's philanthropic cost, LN($1 M, $10 M), over the rodents it switches in expectation. These are the share of its potential use it reaches, times a potential use of LN(3 M, 100 M) rodents a year across countries and farms, times LN(2, 15) years by which it brings the switch forward. The share is LN(0.1%, 20%) capped at 100%, mean 4.79%. It includes the chance that development and registration succeed, and spans negligible to wide uptake rather than all or nothing. Effects over those years are weighted for uncertainty (method). Conservation X Labs' related open innovation process expects around $400k annually for the consortium partners (CXL, WAI, BIWFC) to design and run the process and ideally $1M for the innovation prize (Nitin Sekar). That is $2.20 M–3.00 M over the 3–5 years of a programme, before development and registration. On scale, Rodenticides cause immense suffering to (likely) hundreds of millions of rats and mice annually (Nitin Sekar). Wild Animal Initiative's Cameron Meyer Shorb, in a comment, expects effects at fairly large scales relative to other wild animal animal welfare interventions (i.e., hundreds of thousands to tens of millions a year, possibly much more once non-target species are accounted for). In addition, about 31 million rodents are exposed to rodenticides each year in the U.S. in urban areas alone (Rethink Priorities). Existing products have spread little, after about a decade on the market. SenesTech's Revenue increased by 20% to $2.2 million in 2025 (SenesTech results). Nitin Sekar, replying to a question about Loretta Mayer's work behind SenesTech's products, notes that we are yet to see rigorous peer-reviewed evidence that the interventions associated with her work are effective. Philanthropy is likely to bring the switch forward rather than cause it outright. The share, scale and years are judgements speculative. That is a median of $2.89 per rodent switched and 3.18 rodents switched per dollar on average.

In the formula, per rodent switched: G is the suffering of a poisoning death avoided, less that of the natural death replacing it, plus the rodents' extra years of life at the welfare of wild rats, when fertility control works. ΔY covers the soil organisms and trees on the cropland of the fertility-control bait, less that of the rodenticide bait not used, and on the farmland freed by the extra cost and the programme's spending. The wildlife spared secondary poisoning enters as a welfare gain weighted as wild small mammals. C_h is the extra pest-control cost per rodent. K is the programme's cost per rodent switched, divided by the mean weight for uncertainty over the years it brings the switch forward.

5. GiveDirectly's Cash for Poverty Relief programme

Per dollar donated. GiveDirectly's Cash for Poverty Relief programme gives households in Kenya, Malawi, Mozambique, Rwanda and Uganda $865–1.10 k each. GiveWell's 2024 cost-effectiveness analysis puts it at 3.3 times its cash benchmark (its Aggregated cost-effectiveness estimate (x cash)), with the five countries weighted by its Estimate of where marginal $ goes: 16%, 33%, 2%, 30% and 18%. The human inputs below start from that analysis, with the same country weights. The farmland and farmed animals behind the extra consumption are added. The consumption model below is also used for the income gains of GiveWell's grant. Assumptions:

• Transfers: a share Beta(mean 0.82, concentration 200) of the cost. GiveWell's Transfers as a percentage of total cost are 84%, 84%, 75%, 85% and 71%. The programme costs use real resources (farmland per dollar).
• Recipients' consumption: LN(0.0025, 0.0075) increases of 1 in the natural log of consumption for a person-year per dollar, mean 0.00458. That is a little above GiveWell's 0.0042, since the published evidence suggests that consumption gains gradually dissipate across time, though at a slower rate than we've previously assumed, and the unpublished findings we've been sent, which suggest almost no fade-out in consumption gains across five to seven years (GiveWell blog). GiveWell's 0.0042 is the sum of its yearly increases in ln(consumption) over its 10 years, undiscounted, computed from its consumption path per person in the spreadsheet. In Kenya, that path is $278 in the first year on a baseline of $652, falling to $14 in the tenth on $779. Per person reached, the sum is 0.855 in Kenya to 1.54 in Mozambique. GiveWell's figure multiplies it by the people reached per dollar and by its Distributional adjustment (accounting for consumption gains accruing to richer recipients) of 90%.
• Non-recipients: gains of LN(0.07, 1.5) times the recipients', mean 0.500, a little below GiveWell's 59–68%. GiveWell gets these from the Egger et al. economic multiplier of 2.5 after adjustments for double counting, internal and external validity and richer non-recipients. Egger et al. estimate a local transfer multiplier of 2.5 and document large positive spillovers on non‐recipient households and firms, and minimal price inflation (Econometrica 2022). But The average of the two multipliers is 2.40 (SE = 1.38, p = 0.15) (working paper). GiveWell concludes that these positive consumption spillovers are about 60% to 70% as large as the direct consumption benefits to recipients (GiveWell blog). It also notes that they are much larger than those found in four other studies of the GiveDirectly program, which found no spillovers or mildly negative spillovers. The interval reflects that standard error and those studies.
• QALYs per unit of ln(consumption) for a person-year: LN(0.12, 1.0), mean 0.426. GiveWell values that unit at 1.44 against 116 for averting an under-5 death. At the 35.8 weighted QALYs per child saved used here, that is 0.445 QALYs. GiveWell also gives a moral weight of 2.3 to averting one year of healthy life lost to disability (GiveWell). Against that, the unit is 0.626 QALYs. The Happier Lives Institute puts the ratio of an under-5 death to doubling consumption at 154:1 on deprivationism, and 33:1 on TRIA (the time-relative interest account), against GiveWell's ratio of 100:1 (Estimating moral weights). Measured by life satisfaction, the effect is smaller: the life satisfaction-log permanent income gradient falls between 0.3 and 0.5 standard deviations (Stevenson and Wolfers 2013), about 0.6–1 point out of 10. One QALY is then associated with a 7- point change in life satisfaction (HM Treasury). That gives 0.09–0.14 QALYs per unit. The Happier Lives Institute's estimate from trials of cash transfers, for every $1,000 donated, the organisation creates 7.55 WELLBYs (Happier Lives Institute; a WELLBY is a point of life satisfaction for a year), is about 1.80 WELLBYs per unit at the recipients' 0.0042 units per dollar. That is 0.257 QALYs at 7 WELLBYs per QALY. But a QALY here is a year of healthy life measured from the neutral point. The Happier Lives Institute writes that In the countries that AMF operates in, the average life satisfaction is 4.5/10 (The elephant in the bednet). With a neutral point of 0–2 and 0.85 QALYs per year of life there, a QALY is 2.94–5.29 WELLBYs, which gives 0.340–0.612 QALYs per unit. The interval spans these estimates, centred above the life-satisfaction figures. GiveWell's supplementary adjustments for morbidity, development and other effects, a Total adjustment factor of 108%, are kept as Normal(1.08, 0.04), applied to the consumption and mortality benefits.
• Deaths averted: LN(2, 13) under-5 deaths per $1 M, mean 5.99, against GiveWell's Under-5 deaths averted among recipient households of 6.0, 6.4, 9.8, 4.8 and 6.7 per $1 M by country, 5.96 with its country weights (cost-effectiveness analysis). GiveWell's figures apply a 50% Internal validity adjustment to a % reduction in under 5 mortality due to transfer, unadjusted of 45% from Walker et al.'s results. Their paper finds 45% fewer child deaths before age five, at a cost for targeted transfers of about USD PPP 92,000 (or $39,000 in nominal dollars) per child death averted (PPP is purchasing power parity). Untargeted, the cost is USD PPP 299,418 per death averted (Walker et al. 2025, August 2025 working paper). That is 7.88 deaths per $1 M transferred untargeted at the paper's own ratio of nominal to PPP dollars, or 6.46 per $1 M donated at 82% transferred. A meta-analysis finds a smaller effect, an 8% reduction in the risk of death among children aged <5 years (Richterman et al. 2023), which supports the lower tail. Each child saved gets the births model of GiveWell's grant (births). Adults' gains are LN(0.15, 1) times the QALYs of the child saved, mean 0.457, against GiveWell's Adult mortality benefits (as % of child mortality benefits) of 50%. They add person-years but no descendants. The same meta-analysis finds a 20% reduction in the overall risk of death among adult women. Each adult death averted is worth fewer years of life than a child's.
• Extra consumption in dollars: recipients consume LN(0.9, 1.8) times the transfer over 10 years, mean 1.30. GiveWell writes We expect consumption gains to last for 10 years (GiveDirectly review). In its path for Kenya, $488 per person brings $278 more consumption in the first year and $363 over the next nine, 1.31 times the transfer. Non-recipients consume the share above divided by 0.8 times that, as GiveWell's spillover share is net of a Distributional adjustment (accounting for consumption spillovers accruing to richer non-recipients) of 80%.
• Food: a share Beta(mean 0.45, concentration 10) of the extra consumption, of which a share Beta(mean 0.33, concentration 8) goes on animal-source foods. In Kenya, transfers raised spending on food (19.46 USD PPP), of which meat and fish (5.05 USD PPP), within higher non-durable expenditure (35.66 USD PPP) (Haushofer and Shapiro, replication by 3ie). So food is 0.546 of the extra non-durable spending, less once durables and livestock are counted. Meat and fish alone are 0.260 of the extra food spending, before dairy and eggs. Among low-income households in Nairobi, these animal-source goods make up nearly 40% of the food budget and half of this is spent on dairy products (International Livestock Research Institute (ILRI)). Plant foods occupy LN(3, 20) m²-years of cropland per dollar. The Average smallholder farmer yield of maize in Kenya is 1.7t/ha (International Maize and Wheat Improvement Center (CIMMYT)). The cereals board pays Sh4,000 for a 90 kg bag (The Star), about $0.34 per kg, against a 2025 average retail price of Sh67.58, presumably per kg, (The Star), about $0.52. That gives 11–17 m²-years per dollar of maize, and higher-value plant foods use less. Animal-source foods occupy LN(5, 60) m²-years per dollar, a share Beta(mean 0.7, concentration 10) of it pasture. Poore & Nemecek's land per kg (Our World in Data) is 326 m²-years for beef, 8.95 for milk and 12.2 for poultry meat. Kenyan retail prices of beef and milk are about $5.35 per kg and $0.87 per litre (The Star, above), and locally produced (commercial) poultry meat costs $4.30 per kg at retail in the DRC (USAID). With these, the mean split below gives about 20 m²-years per dollar, nearly all of it for beef and milk, which mostly use grazing land. The intervals around these are judgements speculative. As for other foods, only the share of the change that survives land-market responses counts. Other spending has the land intensity of any dollar of real resources.
• Farmed animals: spending on animal-source foods is split by Dirichlet weights 2, 4, 2.5 and 1.5 among poultry and eggs, dairy, red meat and fish, giving dairy 40% on average after the Nairobi figure above. A share Beta(mean 0.7, concentration 6) of the extra demand is met by extra production. The transfers caused little inflation: Average price inflation is 0.1% (Egger et al., working paper above). In the US, a unit given up cuts total beef consumption by 0.68 lbs, chicken by 0.76 lbs and milk by 0.56 lbs (Norwood and Lusk, quoted by Brian Tomasik). Village chickens live LN(0.3, 2) years per kg at LN($2.5, $6) per kg. In Kenya, indigenous chickens and their crosses number about 22 million (77%) and produce about 11,400 tons (61%) of poultry meat. Hens lay about 45 eggs per year (Kenya Agricultural and Livestock Research Organization (KALRO)). That is 1.93 chicken-years per kg of meat, with eggs as well. Dairy cows live LN(1/2000, 1/400) years per kg of milk at LN($0.35, $0.9) per kg. Smallholder yields over 305 days range from 823 to 2453L (CGIAR, a global agricultural research partnership), before counting heifers and dry periods. The farm-gate price was Sh49.9 per litre (Capital FM), about $0.39. Cattle, standing in for goats and sheep too, live LN(0.04, 0.3) years per kg of red meat at LN($3, $8) per kg. A Cameroonian herd with an annual offtake rate of 11.3 percent, a mature live weight of 320 kg and a carcass yield of 52 percent (FAO) gives 0.0532 cattle-years per kg of carcass. Herds with lower offtake, and goats and sheep, give more animal-years per kg. Kenya has 17M cattle, 53M sheep and goats (Strathmore Agri Atlas). Fish are mostly wild-caught and not modelled.

Per dollar donated, people gain 0.00329 QALYs on average, 86.1% of it from consumption, and farmland rises by 5.67 m²-years, 93.6% of it for food.

In the formula, per dollar donated: G is the QALYs from the consumption of recipients and non-recipients, plus those of the children and adults whose deaths are averted, with the children's descendants, net of the births averted. ΔY covers the soil organisms and trees on the farmland for the extra food and other consumption, for the extra people, and freed by the programme costs. It also covers the village chickens and cattle behind the extra animal-source foods, and the village chickens of the extra people. C_h = 0, and K = $1.

6. Buying beef

Per kg of beef bought by a donor and eaten in place of other foods. Vasco Grilo argues that this increases welfare if soil animals have negative lives, because [C]onsuming unprocessed red meat in the US increases agricultural land by 178 m²-year/(100 meat-g), 1.72 k times the reduction in it from the reduction in life expectancy (Reduction in life expectancy and agricultural land due to eating red meat; see also Cost-effectiveness accounting for soil nematodes, mites and springtails). Assumptions:

• Extra production: a share Beta(mean 0.7, concentration 6) of the beef bought is newly produced. In Norwood and Lusk's equilibrium model, a unit of beef given up cuts total consumption by 0.68 lbs (quoted by Brian Tomasik). The same holds for a unit bought. It is produced as beef-herd beef with 326 m²-years of land per kg (Poore & Nemecek), whose production also uses resources that would otherwise produce other goods. The rest comes out of other people's consumption, and they replace the same share of its protein with other foods as the eater displaces (below).
• Displaced food: the eater cuts other foods by a share Beta(mean 0.7, concentration 5) of the beef's protein. This is split among foods by Dirichlet weights, by protein, of 3 for poultry, 2.5 for pig meat, 1.5 for milk, 1 each for beef and staples, 0.5 each for farmed fish and eggs, and 0.25 each for wild fish and pulses, as extra beef mostly displaces other meat and dairy speculative. Net of the people whose beef was bought instead of newly produced, other foods fall by the share newly produced times this share. People regulate protein intake more strongly than any other dietary component (Scimex), which favours high displacement. But the regulation is asymmetric. On a 10% protein diet people consumed 12% more energy over four days than they did on a 15% protein diet, while on 25% protein researchers observed no change in behavior relative to the 15% protein diet (Gosby et al. 2011, ScienceDaily). So extra beef may displace energy rather than other protein foods. The wide interval is a judgement speculative.
• Health: the eater loses LN(5, 150) minutes of healthy life per kg of red meat, at 0.85 QALYs per year, the ratio of healthy life expectancy to life expectancy used for the DRC. Vasco Grilo's estimate, assuming 15% of the epidemiological association is causal, is about 91 minutes per kg. For processed meat, which carries more risk, 0.45 minutes are lost per gram of any processed meat that a person eats in the US (Stylianou et al. 2021, University of Michigan), 450 minutes per kg. But a systematic review of the evidence found that All evidence on red meat and its links to disease were weak (the Burden of Proof study of unprocessed red meat, Lescinsky et al. 2022, via STAT). The interval runs from a weak effect to above Vasco Grilo's estimate, and below the figure for processed meat.
• Target beneficiaries: the extra beef-herd cattle, at LN(0.01, 0.05) cattle-years per kg of extra beef, mean 0.0252. In the US, All cattle and calves in the United States as of Jan. 1, 2025, totaled 86.7 million head, with Milk cows, at 9.35 million head (DTN, an agricultural news service) and only 3.914 million replacement heifers (DairyReporter), against total beef production to 27.0 billion pounds (Farm Progress). That is 0.00600 cattle-years per kg of carcass, or 0.0109–0.0133 per kg of boneless retail beef at 45–55% of the carcass, as for the 756 lb carcass in Table 2 of a Penn State Extension guide. The production figure is a forecast for 2024 and includes beef from culled dairy cows, which slightly understates cattle-years per kg. Extensive systems elsewhere need more cattle per kg. In Brazil, O efetivo bovino atingiu 238,6 milhões de cabeças em 2023 (cattle reached 238.6 M head), with 15,7 milhões de vacas ordenhadas (15.7 M cows milked) (Compre Rural, on data from Brazil's statistics institute, IBGE). Also, o país abateu 33,9 milhões de cabeças, o que resultou em uma produção 8,91 milhões de toneladas de carne (33.9 M head slaughtered for 8.91 M tonnes of meat) (Canal Rural). That is about 0.0250 cattle-years per kg of carcass, or 0.0455–0.0556 per kg of boneless beef. This is somewhat overstated, as the herd includes dairy heifers and part of the slaughter is uninspected. The land per kg of beef is Poore & Nemecek's global mean, which extensive systems raise. The price of beef is a rich-country retail price, as paid by a donor buying it. Under the welfare guess for cattle here, Normal(+0.15, 0.6), they have positive lives in expectation.
• Other animals: changes in the farmed animals behind the displaced foods.

In the formula, per kg of beef bought: G is the welfare of the extra beef-herd cattle-years. ΔY covers the soil organisms and trees on the land of the extra beef, less that of the foods displaced, and on the farmland freed by the resources used to produce the extra beef. It also covers the farmed animals behind the displaced foods, and the eater's lost years of life, at 0.85 QALYs per year with r = 1. C_h = 0, and K is the price per kg.

Buying beef does not have the highest mean cost-effectiveness at any exponent, nor in any of the seven combinations of flipping the sign of soil animals' welfare per year, of their welfare shift on farmland, and of the welfare of soil bacteria and archaea. At exponent 0, it ranks fifth at best across the base case and those combinations, and eighth under model 2 in the base case, behind GWWC, cage-free campaigns, GiveWell's grant, broiler campaigns, GiveWell's All Grants Fund, GiveDirectly's programme and salt substitution. Per dollar, it adds 7.01 m²-years of farmland after market responses, against 18.3 for GiveWell's grant, through the people it saves, their descendants and their income, and 20.3 for cage-free campaigns. This contrasts with Brian Tomasik's cost-effectiveness comparison for different ways to reduce insect suffering, which also turns on land. In it, Buy beef directly from Brazil is joint highest by Rough expected insect-years of suffering prevented per US$, about 1 M, while Donate to Against Malaria Foundation gets 1.4 * 10⁴, sign is very unclear. The script for the sign flips is sens_flip.py, below.

7. ProVeg UK's School Plates programme in 2023

Per dollar donated. School Plates helps schools and local authorities in the UK serve more plant-based meals. ProVeg UK reports helping an estimated 12.4 million school meals become meat-free or plant-based across the UK in 2023 (ProVeg UK impact report). According to Colette Fox, head of School Plates, this swapped 12.4 M meals to meat-free in 2023 at a cost of 155 k£ (Cost-effectiveness of School Plates). From these, Vasco Grilo writes that I calculated School Plates swapped 64.5 lunches/dinners per $ in 2023 and that this decreased cropland by 610 m²-years per dollar (Cost-effectiveness accounting for soil nematodes, mites and springtails; spreadsheet). The arithmetic reproduces, but four steps overstate the change:

• Pasture counted as cropland. The spreadsheet takes the Value for beef (beef herd) from Poore and Nemecek (2018), 326 m²-years per kg, as cropland. But Poore & Nemecek's land includes grazing land. Their Data S2 splits beef from beef herds into 20.507 m² of arable land, 5.067 of fallow and 168.092 of permanent pasture per kg (stunning inputs). With Data S2's shares, 83.5% of the 610 is pasture.
• Meat per meal. It takes 0.104 kg per meal, the UK's consumption of meat, fish and seafood at lunches and dinners, including fish, for adults and children. School portions are smaller. The guidance for primary pupils is Raw, 50g to 75g. of red meat or poultry in dishes such as stews, and for roast poultry Raw, 60g to 85g. For secondary pupils, it is Raw, 75g to 90g. in such dishes (primary and secondary guidance, Department for Education).
• Claimed meals. The 12.4 M is ProVeg's own estimate, and its public figures disagree. ProVeg International's annual review attributes the same number to three countries, as school programs in the UK, Germany, and Poland resulted in 12.4 million meals being changed from meat-based to plant-based (vegconomist). ProVeg UK reported over 7 million meals since launch by the end of 2022 (Annual Review 2022), against a cumulative 12.3 M on the Food Planet Prize page. Many caterers would also have cut meat anyway (share caused, below).
• Market responses. Lower demand lowers prices, and others eat part of the meat no longer bought. Land markets also absorb part of any change in farmland.

The model uses, per dollar:

• Claimed meals: 12.4 M for 155 k£, at 1.24 dollars per pound, from the US Internal Revenue Service's 2023 average of 0.804 pounds per dollar (To convert from foreign currency to U.S. dollars, divide the foreign currency amount by the applicable yearly average exchange rate in the table below., IRS). The total cost is LN(1.1, 1.7) times the direct cost, mean 1.38. ProVeg UK is a non-profit organisation. We are registered as a community-interest company and are entirely donor-funded. (ProVeg UK). Its accounts are filed at Companies House, where they could not be read for this analysis. For comparison, ProVeg's US entity, ProVeg International Inc., spent $544 k on administration and $51.5 k on fundraising against $3.73 M on programmes in 2023, 16.0% of programme spending (Impala, from its tax return). Veganuary spent £94.7 k on raising funds and £19.3 k on governance against £1.95 M on charitable activities in the year to February 2023, 5.84%. This understates its overheads, as UK charities report most support costs under charitable activities. The 155 k£ probably leaves out shared management, communications and services from ProVeg International. So the interval sits above these ratios speculative. That is 47.6 claimed meals per dollar on average.
• Share caused by School Plates: Beta(mean 0.35, concentration 5), 5th to 95th percentiles 0.0695–0.703, the product of three judgements speculative:
  • Counting, 0.8. ProVeg's figures are estimates from partners' reports, Based on estimated overall impact data (School Plates guide, footnote 1), and no method has been published. The yearly totals implied by This is up from 24.9 million meals at the end of 2023, 12.3 million at the end of 2022, and 5.6 million at the end of 2021. (Food Planet Prize) seem to include menus changed in earlier years and still served, set against spending in 2023 alone. But swaps made in 2023 may last beyond it. Partners report measured changes, such as a 9% shift from meat-based to meat-free or plant-based meals served and a larger shift of around 29% towards meat-free and plant-based (School Plates guide). This suggests counts of meals eaten rather than of menu slots.
  • Change that would not have happened anyway, 0.5. Public-sector caterers were already cutting meat: Nearly 80% of those polled have decided to reduce meat content across menus, and 48% of those polled are introducing a meat-free day (Eating Better, 2021). ProVeg itself cites a study finding over 80% of public-sector caterers have committed to meat reduction (School Plates). As food prices rose, 27% of respondents confirmed that they are serving less meat to cut costs (LACA, the school caterers' association). Caterers choose to work with ProVeg. The law caps meat-free days, as A portion of meat or poultry must be provided on three or more days every week. (School Food Regulations 2014).
  • Pupils not making up for it, 0.8. In a university canteen, Opting for a vegetarian lunch did not make a compensatory meat-heavy dinner any more likely (Garnett et al. 2019, University of Cambridge). In 67 randomised meat-free days at a Swiss university, MFD also led to a 16.8% drop in on-campus meal sales on treated days., with a 3.5% rebound in animal-based meal consumption on days immediately following treated days (Russo et al. 2025). A weekly vegetarian day in Helsinki schools produced a mixture of intended and unintended effects (Lombardini & Lankoski 2013), pupils taking less food. No study of pupils eating more meat at home was found.
  So 16.7 net meals per dollar on average.
• Meat per swapped meal: LN(0.04, 0.09) kg, mean 0.0619, around the portions above and ProVeg's flagship bolognese, Made with 50g beef mince per portion and served with 55g wholewheat spaghetti (School Plates guide). The meat is split by Dirichlet weights proportional to the UK's consumption of poultry, pig meat, beef and lamb, 35.2, 25.9, 17.1 and 4.06 kg per person in 2022 (Vasco Grilo's spreadsheet, from Our World in Data). The total concentration is 10, since no data on the meat mix of school menus was found. That is 1.03 kg of meat per dollar on average.
• Replacement foods: LN(0.4, 1.2) kg of plant food per kg of meat, dry as bought, mean 0.733, a share Beta(mean 0.3, concentration 6) of it staples and the rest pulses, soy and other protein foods, at the land of pulses. The flagship bolognese becomes one Made with 30g lentils and 20g dried soya mince per portion and served with 55g wholewheat spaghetti (School Plates guide). That is 0.64–1 kg of dry pulses and soy per kg of beef, depending on whether the lentils are weighed cooked or dry, and no extra pasta. For primary pupils, a portion is Raw, 20g to 25g. of pulses or Cooked, 50g to 70g. of meat alternatives (Department for Education), against 50–75 g of raw meat. That is 0.27–0.5 kg of dry pulses per kg of meat, and less dry soya mince, which absorbs about twice its weight in water. Meat-free meals still carry a protein food, since A portion of food from group C must be provided every day. (School Food Regulations 2014), the group of eggs, pulses and other non-dairy protein. Caterers also report that 83% have increased beans and pulses in dishes by up to a fifth (Eating Better). Some meat-free dishes, such as jacket potatoes, add starch instead, hence the staples share. Mycoprotein uses less land than pulses, 1.8 m²-years per kg (Carbon Trust, table 9) speculative.
• Cheese: LN(0.0015, 0.012) kg per swapped meal, mean 0.00518, at 9.81 kg of milk per kg of cheese (Poore & Nemecek's 87.79 m²-years per kg of cheese over 8.95 per kg of milk). Jimmy Pierson, of ProVeg UK, says that meals are meat-free rather than plant-based because of a small amount of dairy cheese. It's very rarely, if ever, a result of eggs in the meal (quoted by Vasco Grilo). A primary portion of cheese is 20g to 30g. (Department for Education). If 10–40% of swapped meals add 10–25 g of cheese, net of the cheese in meat dishes such as lasagne, that is 1–10 g per meal speculative.
• Production: a share Beta(mean 0.7, concentration 6) of the change in demand is met by a change in production, as for buying beef (above).
• Lamb: 369.81 m²-years per kg (Poore & Nemecek), a cropland share Beta(mean 0.05, concentration 25) after Data S2's 0.0491, and LN(0.10, 0.20) sheep-years per kg of edible lamb, mean 0.145, at the welfare of cattle, with sheep's own neuron counts and metabolic rates speculative. In the UK, the Department for Environment, Food and Rural Affairs (Defra) reports that Defra census figures recorded the UK sheep flock at 31.8 million head as of 1 June 2023. Of these, lambs under a year were down 998,000 head (6.1%) year-on-year to stand at 15.5 million (Agriculture and Horticulture Development Board (AHDB)). Production reduced by 5,300t, or 1.8%, to 286,000t, with the lamb kill recorded at 12.28m head (Farmers Journal). Lambs are born in spring and are ready for the food chain around six months later (National Sheep Association). Adults (16.3 M) plus slaughter lambs at about 7.5 months on average (7.68 M sheep-years), replacement ewe lambs (about 1.9 M) and lambs that die are about 26.0 M sheep-years a year. That is 0.0909 per kg of carcass, or 0.152 per kg of edible meat at 60% of the carcass. That share lies between boneless cuts at 43–50% and bone-in cuts at 65–75% of the carcass (University of Wisconsin). The interval covers 0.113–0.196 from other yields and flock assumptions.

Farmland falls by 5.43 m²-years of cropland and 31.5 of pasture per dollar on average, after market responses, 6.05% of 610.

In the formula, per dollar donated: G is the welfare of the broiler-, pig-, cattle- and sheep-years not lived. ΔY covers the soil organisms and trees on the land of the meat not eaten (a gain) and of the foods replacing it, and on the farmland freed by the dollar spent. It also covers the dairy cows behind the cheese, less the beef-herd cattle its beef displaces. C_h = 0, and K = $1.

8. The Good Food Institute

Per dollar donated. Impact-focused evaluators have analysed land use mostly through climate. Giving Green recommends policy work over forest protection, as we generally recommend donating to policy organizations over purchasing carbon offsets (forestry report). For biodiversity, it notes that we are less likely to recommend narrowly targeted conservation projects, since establishing protected areas can displace hunting, fishing, and deforestation to unprotected areas (biodiversity report). Founders Pledge recommended the Coalition for Rainforest Nations and later stopped: The fact that FP stopped recommending CfRN is additional evidence of this kind (Johannes Ackva, comment). None of these publishes land per dollar. Derived from their figures, avoided deforestation keeps about 100 to 350 m²-years of natural land per dollar. In a randomised trial in Uganda, tree cover declined by 4.2% during the study period in treatment villages, compared to 9.1% in control villages, at $0.46 per averted MT of CO2. The average carbon per hectare of forest cover in the study villages was 153.5 metric tons (MT) (Jayachandran et al. 2017). That gives 116 m²-years per dollar for the 3 years by which the trial delays clearing on average, or 341 if payments continued, at a net present cost to permanently avert a MT of CO2 of $2.60. Rethink Priorities estimates that REDD+ (Reducing Emissions from Deforestation and forest Degradation) would cost $6-$62 per tonne of CO2 (REDD+ report). That is 107 m²-years per dollar on average at 367 tonnes of CO2 per hectare over 50 years.

The Good Food Institute (GFI) is a top pick of Giving Green for both climate and biodiversity, whose work can help reduce future habitat loss (Giving Green). It decreases farmland by a similar amount: Our best guess for GFI's cost-effectiveness is $2.98 per metric ton of CO2e in expectation (range: $0.68 to $192) (GFI evaluation). Giving Green's analysis estimates the effect that GFI's existence from 2016 to 2022 is expected to have on emissions from 2022 onward. It adds We use this outcome as a proxy for the cost-effectiveness of a 2024 donation to GFI, and warns This CEA includes highly subjective guess parameters and should not be taken literally. Its spreadsheet multiplies a fall in meat production by GFI's change in the probability of that fall and by a fixed 14.0 tonnes of CO2e per tonne. It then divides by GFI's budget of $101 M in 2016–2022. The range is not a 90% interval: its ends set every guess to its lowest or its highest value at once. The model takes the spreadsheet's physical inputs rather than its emissions, since land and animals depend on the meat displaced. Per dollar:

• Fall in meat production: LN(70 M, 600 M) tonnes, mean 253 M, if alternative proteins move from low to high innovation. The spreadsheet's The difference in livestock production between adjusted high- and adjusted low-innovation scenarios over years of interest (million tons per year) is 430 in its best guess, 50.0 in its low-impact case and 710 in its high-impact case. These are for 15, 5 and 20 # of years before the high-innovation scenario is roughly similar to BAU in the absence of GFI, a Guess. Each figure is a sum of yearly gaps between Vivid Economics' scenarios (livestock tab), 430 for 2022–2035, with no discounting. The series is meat. The spreadsheet splits Conventional livestock and dairy into Conventional livestock, 327 M tonnes a year in 2020, and Conventional dairy, 108 M. It uses the first, close to the 358 M tonnes of poultry, pig, bovine and ovine meat in FAO's carcass-weight figures for 2023 (meat mix, below). The 5th percentile sits above the low-impact case, as Giving Green's best guess is 15 years. But the high-innovation path holds livestock flat from 2020. Meanwhile, World meat production is forecast to reach 384 million tonnes (carcass weight equivalent) in 2025, up 1.4 percent from 2024 (FAO, via 3tres3). So the gaps for 2022–2025, about 50 M tonnes, did not happen, and the distribution is centred below Giving Green's case speculative.
• GFI's change in the probability of high innovation: LN(0.04%, 1%), mean 0.323%. In the spreadsheet, it is the change in probability from R&D and talent, 12.5% (5–20%), times GFI's share of it, 3% (1–5%), plus the change from public funding, 12.5% (5–20%), times GFI's share, 1.5% (0.5–2.5%). That is 0.563% in the best guess, and 0.075% and 1.5% in the low- and high-impact cases, all guesses. Taking the corner cases as the 5th and 95th percentiles is wider than combining independent guesses would give, which allows for the guesses themselves being uncertain. The market has fallen behind. In the US, Retail sales of plant-based meat and seafood fell 10% to $1 billion in 2025 with units down 11% (AgFunderNews). Private investment in alternative proteins as a whole fell by 20% in 2025, attracting just $881M, though plant-based funding rose (Green Queen, on GFI's report). So the range is shifted down from the spreadsheet's speculative.
• Marginal dollar: LN(0.2, 1) times as effective as the average dollar of 2016–2022, mean 0.504. GFI's spending rose to $30 M in 2022 (spreadsheet), and Giving Green reports a 2024 budget of $40.7 million (deep dive). Returns diminish as an organisation grows, and the early years of a field are likely to have been the most influential; the spreadsheet's Effectiveness adjustment for marginal donation going to international efforts instead of US-based R&D is 1 in every case speculative.
• Edible meat per kg of carcass: Beta(mean 0.7, concentration 20). Boneless retail beef is 45–55% of the carcass (buying beef), and lamb 43–50% boneless or 65–75% bone-in (University of Wisconsin). Poultry and pig meat, most of the mix, lose less to bone speculative. That is 4.10 kg of carcass, and 2.87 kg of edible meat, per dollar on average before the weight for uncertainty over time (5th to 95th percentiles of carcass weight 0.224–14.8 kg). That weight covers the 15 years starting LN(1, 6) years from now, with mean 0.815. Giving Green's best guesses give 23.9 kg of carcass per dollar (430 M tonnes × 0.563% ÷ $101 M).
• Meat mix: Dirichlet weights proportional to world production in 2023, with total concentration 20. In the FAO's figures, Global poultry meat output is forecast to reach 142 million tonnes in 2023, Global pig meat production is pegged at 123.1 million tonnes in 2023, World bovine meat production in 2023 is pegged at 76 million tonnes and Global ovine meat output is forecast at 17.0 million tonnes in 2023 (FAO Meat Market Review). Studies of which meats plant-based meat replaces disagree in sign. Rethink Priorities concludes that any effects of changes in plant-based meat prices seem to have only very small effects on animal-based meat sales (Peacock 2023).
• Land of the alternative protein: LN(1.8, 6.0) m²-years of cropland per kg, mean 3.51. Tofu uses 3.52 (Our World in Data, from Poore & Nemecek). The Beyond Burger uses 0.45 m²-years per 0.113 kg patty, 3.98 per kg, and 93% less land use than US beef by ecosystem damage (Heller & Keoleian 2018). Quorn's mycoprotein, mince and pieces use 1.8×10⁻⁴, 3.8×10⁻⁴ and 2.5×10⁻⁴ ha per kg, 1.8–3.8 m²-years (Carbon Trust, table 9). On average, plant-based meat uses 79% less land than animal meat per kg (GFI life-cycle assessment). Giving Green's 41-98% less land comes from a comparison made Per 100 grams of protein (Center for Strategic and International Studies (CSIS)). Protein isolates, added fats and cultivated meat give the upper tail.
• Alternative protein per kg of meat displaced: LN(0.75, 1.75) kg, mean 1.18. By Poore & Nemecek's land per kg and per 100 g of protein (per kg, per protein), tofu has 160 g of protein per kg, poultry 173, pig meat 162, beef from beef herds 199 and lamb 200. So equal protein takes 1.01–1.25 kg of tofu per kg of meat, while life-cycle assessments compare products kilogram for kilogram. Sales of alternatives displace less than their own weight of meat. At a University of California, Los Angeles dining hall, one-for-one replacement predicted an 11.4 percentage point reduction in beef entrees, but the study found a much smaller decline of 3.3 percentage points (Malan et al. 2022, via Peacock 2023). Consumers only partially substitute plant-based meat for conventional meat (Lusk et al. 2022, via Breakthrough Institute). Much of the rest replaces other plant foods, with little net land. So the upper tail is modest speculative.
• Production: Giving Green's figures are already changes in production. So all of the displacement changes production. Lamb is as for School Plates.

Farmland falls by 21.2 m²-years of cropland and 99.3 of pasture per dollar on average after land-market responses, about 219 m²-years before them. Both chains, from advocacy and research to meat displaced, are highly uncertain speculative.

In the formula, per dollar donated: G is the welfare of the broiler-, pig- and cattle-years not lived. ΔY covers the soil organisms and trees on the land of the meat displaced (a gain) and of the products replacing it, and on the farmland freed by the dollar spent. C_h = 0, and K = $1.

9. The Humane League UK's Better Chicken Commitment work

Per broiler helped, for a donation to The Humane League (THL) UK restricted to its work holding companies to the Better Chicken Commitment, which Animal Charity Evaluators has analysed. The EA Animal Welfare Fund is not used here: its 3-year grantmaking strategy for 2026–2028 starts with Ending cage confinement globally and lists no broiler work (EA Animal Welfare Fund). So a donation to it would not fund broiler campaigns at the margin. Each broiler helped is moved to the standards of the Better Chicken Commitment (slower-growing breeds, lower stocking density, enrichment and better slaughter) before any change in chicken consumption. The model follows the one for cage-free campaigns, and organisations running cage-free campaigns run these too. Slower-growing breeds eat more feed per kg of meat, which needs more cropland, and chicken becomes more expensive. So people eat a little less of it. Per broiler:

• Welfare gain: the Welfare Footprint Institute estimates that the commitment is expected to prevent “at least” 33 [13 to 53] hours of Disabling pain, 79 [-99 to 260] hours of Hurtful and 25 [5 to 45] seconds of Excruciating pain for every bird affected by this intervention (Welfare Footprint Institute). Each is drawn from a normal distribution with those 5th and 95th percentiles, as the page does not define its intervals. The total is multiplied by LN(1, 2.5) for welfare issues the estimates leave out, as for hens. So the gain is 0.254 AQALYs on average (median 0.123) speculative.
• Broilers helped per dollar: LN(0.4, 25), mean 6.99, over the LN(4, 36) years by which campaigns bring the commitment forward, starting LN(1, 8) years from now. Animal Charity Evaluators' analysis of THL UK puts its Program 4 (Better Chicken Commitment) cost-benefit (animals helped/$) at 46 in its best guess, 4 in its conservative case and 140 in its optimistic one, for spending in 2024. It puts the Expected probability of companies meeting their commitments at all at 25–30% in its best guess (Animal Charity Evaluators). That is an average over THL UK's past spending, not the effect of a marginal dollar. The marginal factors used for cage-free campaigns, 1/5 to 0.3, give 9.20–13.8 broilers per dollar. Fulfilment so far is low: just 6% of global BCCs with a deadline of 2025 or earlier have been fulfilled, against 64% for cage-free commitments (Faunalytics). In Europe, Of the reporting companies, 36 report less than 10% progress on adopting slower‑growing breeds, with 17 at 0%. (Compassion in World Farming's ChickenTrack, via Food & Agribusiness). So eventual fulfilment is likely below the 25–30% Animal Charity Evaluators assumes, and the distribution is centred lower than 9.20–13.8. Animal Charity Evaluators' figure already spreads the broilers over the years brought forward. The weight for uncertainty over time stands in for its discount rate speculative.
• Extra feed per kg of chicken: LN(8%, 32%). Torrey et al. (2021) find a 22 to 43-point difference in feed conversion ratios, 14.3–27.9% more feed per kg for slower-growing strains to the same weight. Wageningen University found that the feed conversion rate for Redbro birds was 14% higher than for Ross 308 birds (Faunalytics). These are per kg of live weight. Carcass yield is also lower for slower-growing breeds. In Table 1 of Weimer et al. (2022), carcass weight is 1.75 and 2.07 kg for slow-growing birds of 2.43 and 2.88 kg of live weight (72.0% and 71.7%). It is 2.15 and 2.06 kg for fast-growing birds of 2.86 and 2.70 kg (75.4% and 76.0%). So the ratio of carcass yields is LN(0.92, 0.98), and each kg of chicken needs 1 ÷ that ratio more live weight, feed and birds. With it, feed per kg of chicken rises by 23.8% on average. The cropland of chicken is Poore & Nemecek's 12.22 m²-years per kg, almost all feed.
• Rise in the cost of chicken: LN(5%, 25%) of the retail price of LN($3, $9) per kg, mean 12.6%. Wageningen found that the European commitment caused an 18.5% increase in costs at the farm level, a 16% increase after slaughter, and a 19% increase after further processing into consumer meat products (Faunalytics). A study for the industry gives higher figures. These are cost rises along the chain, and retail prices rise by less in percentage terms, as retail margins are a large share of the price. Vasco Grilo's spreadsheet takes a cost ratio of 1.19 between chicken from Redbro and Ross 308, the mean across six EU countries from Figure S.1 of van Horne et al. (2025).
• Willingness to pay: LN(0.02, 0.8) times the extra cost, mean 0.237. In a US survey, the mean WTP premium is $0.58/lb ($0.85/lb with brands) (Lusk 2018), about 15% of the prices shown, before correcting for the gap between stated and actual purchases, which is large, as for eggs speculative.
• Own-price elasticity of chicken demand: LN(0.4, 1.4) in absolute value, mean 0.805. A review finds 0.68 (0.44, 0.92) for poultry (Andreyeva et al. 2010). Vasco Grilo's spreadsheet takes 0.840 for poultry meat in the EU from Table 6 of Bouyssou et al. (2024). Customers can also switch to conventional chicken sold by others, with an elasticity LN(0.3, 3) as for eggs. So 11.9% of the broilers helped are lost on average, and chicken consumption falls by 7.73% of the chicken produced to the commitment.
• Longer lives: LN(20%, 45%) more days per broiler. The Welfare Footprint Institute assumes a slaughter weight of 2.5 Kg at 42 days for conventional birds, and that the same slaughter weight would be reached in approximately 56 days for slower-growing ones, 33.3% more (Welfare Footprint Institute). Torrey et al. report that At TW2, CONV was 48 days of age and the other categories were 62 d of age (Torrey et al. 2021), 29.2% more. The Welfare Footprint Institute's estimates per bird already cover the longer life. So the longer lives only enter through the broilers no longer raised as chicken consumption falls and through the extra birds needed per kg (below). These birds have the welfare of broilers raised to the commitment, Normal(−0.7, 1.0), as for cage-free hens speculative.
• Birds per kg of chicken: the relevant comparison is per kg of chicken eaten, as for eggs rather than layers. At the same slaughter weight, the gain per bird is also the gain per 2.5 kg of live weight, and the Welfare Footprint Institute's comparison holds the weight fixed. Slower-growing breeds are often slaughtered lighter, which needs more birds per kg: Beta(mean 0.05, concentration 20) more. The lower carcass yield above adds more again, each living as a broiler raised to the commitment. Per kg of chicken, Vasco Grilo's spreadsheet has 33.9 chicken-days for Redbro against 27.8 for Ross 308, 21.9% more. Here, it is 37.9% more on average, from the longer lives and the lower carcass yield speculative.

The extra feed adds 1.92 m²-years of cropland per broiler after land-market responses, and the fall in consumption removes a little, net of the foods replacing the chicken not eaten. Vasco Grilo's spreadsheet finds that cropland falls, by 0.102 m²-years per kg of chicken from Ross 308, neglecting increases in the consumption of other foods. In it, demand falls to 86.1% from a cost ratio of 1.19 and an elasticity of 0.840, while cropland per kg rises from 6.63 to 7.58 m²-years. Here, demand falls less, by 7.73% on average, as the retail price rises by less than the cost ratio, the elasticity is a little lower, and some consumers value the change. The foods replacing chicken also add land back. Even without that replacement, cropland rises here, as feed per kg of chicken rises by 23.8% once the lower carcass yield is counted, against a fall in consumption of 7.73%. In total, farmland rises by 1.83 m²-years of farmland per broiler and 10.5 per dollar on average (5th to 95th percentiles 0.131 to 41.1). This compares with 20.3 for cage-free campaigns, which help more chicken-years per dollar.

In the formula, per broiler helped: G is the pain averted for the broilers that stay moved. ΔY covers the soil organisms and trees on the land of the extra feed and of the chicken not eaten and the foods replacing it, and on the farmland freed by the money spent. The extra broiler-years and the farmed animals behind the replacement foods are also counted. C_h is the extra cost not matched by consumers' willingness to pay. K is 1 ÷ broilers helped per dollar, divided by the mean weight for uncertainty over the years in which campaigns help broilers.

10. GiveWell's All Grants Fund

Per death-averted equivalent: a bundle of deaths averted, other health gains and income gains like that of GiveWell's grant, bought at GiveWell's bar. The fund provides grants to incubate newer programs, provide organizational support, promote policy change, fund relevant scoping and research, or support other potentially high-impact and cost-effective initiatives. GiveWell writes that we believe that—in aggregate—these grants represent the highest-impact use of marginal funding. (All Grants Fund). Its marginal grants are at the bar: As of May 2026, our bar for funding programs is 6x our benchmark. (GiveWell), against about 16 times for the grant to the Against Malaria Foundation (costs). Marginal cost-effectiveness should be similar across countries and programmes, as GiveWell funds the best opportunities first. So a marginal dollar donated through GiveWell is better represented by this fund than by the grant, which was well above the bar. For comparison, GiveWell writes that We directed $427 million to support outstanding opportunities to do good in 32 countries. and We estimate that the programs supported by the funds we directed this year will save an additional 86,000 lives in 2025, $4.97 k per life. But The calculations for the estimate of lives saved cover only the 2025 grants that had reduced mortality as a primary outcome. (GiveWell's impact). The model is the grant's, with these changes:

• Cost per death-averted equivalent: LN($3.5 k, $50 k), mean of 1 ÷ cost 1 ÷ $9.55 k, against $4.12 k × 16 ÷ 6 = $11.0 k for the grant's bundle at the bar. So it averts 2.35 times fewer deaths per dollar than the grant.
• Who is saved: a share Beta(mean 0.15, concentration 30) at age 5 or older, as the fund's large grants go mostly to children's health, such as malaria vaccines, seasonal malaria chemoprevention and vitamin A; other health gains LN(0.08, 0.55) times the QALYs of the lives saved; and income gains LN(10, 80) × LN(0.5, 2) ln(consumption)-years per death averted, at LN($300, $1 k) of consumption per person-year speculative.
• Farmland and descendants: LN(0.4, 1.4) times the DRC's farmland per extra person, mean 0.804, and LN(0.5, 1) times its descendants per person saved, mean 0.723, as the fund's grants also go to countries with less land per person and lower fertility, such as Nigeria, Pakistan and India speculative.
• Timing: effects start about a year from now.

The fund adds 7.61 m²-years of farmland per dollar on average (5th to 95th percentiles 0.421 to 26.7), 41.7% of the grant's, mostly because it averts fewer deaths per dollar. So it does not increase farmland more cost-effectively than the grant.

In the formula, as for GiveWell's grant, with K the cost per death-averted equivalent with every effect weighted from now, starting a year from now.

11. Resolve to Save Lives' salt substitution work in China and India

Per death averted, for a donation to Resolve to Save Lives restricted to dietary salt modification in China and India. The deaths averted are mostly cardiovascular deaths of older adults, from replacing salt with potassium-enriched salt substitutes. GiveWell writes that Overall, we estimate that it costs about $855 to avert a death via dietary salt modification in China and India. But We have not yet considered the impact of leverage and funging due to limited details about the intervention, and we think it is likely to reduce cost-effectiveness when taken into account. It also writes that We have identified one charity that is seeking funding to implement this type of intervention, Resolve to Save Lives. (GiveWell). In GiveWell's table, its Moral weight of averting one cardiovascular death, units of value is 30 (salt page), against about 116 for a death under 5 in its models. Per death averted:

• Years of life gained: LN(6, 15), mean 9.86, starting about a year from now, 8.61 after the weight for uncertainty over time, at Beta(mean 0.75, concentration 30) QALYs per year, plus LN(1, 3.5) years lived with disability averted, from non-fatal strokes and heart attacks speculative. The people saved are mostly past having children. So no births are added or averted.
• Farmland per extra person-year: LN(250, 3 k) m²-years, a share Beta(mean 0.55, concentration 10) of it cropland. Agricultural land per person was 0.124 ha in India and 0.366 ha in China in 2023, and 24.5% of agricultural land was cropland in China and 94.3% in India (per person, cropland and agricultural land, Our World in Data, from FAO). Farmland fell in both countries from 2013 to 2023 while their populations grew. So the change in farmland per extra person is negative there, reflecting rising yields rather than people's food needs. The interval centres on the average, lowered for the smaller diets of older people. In the DRC, by contrast, farmland keeps expanding at the frontier, and its change per extra person is used. Both are then scaled by the share of a land change that survives land-market responses speculative.
• Cost per death averted: LN($500, $5 k), mean of 1 ÷ cost 1 ÷ $1.24 k, above GiveWell's $855, for leverage, funging and a marginal grant speculative.
• Farmed animals: the people saved eat animal foods for the years they gain. Supply per person in 2023 was 19.2 kg of poultry, 22.9 kg of eggs, 42.1 kg of pork, 11.9 kg of beef, sheep and goat meat, and 41.7 kg of fish and seafood in China. In India, it was 3.52, 4.72, 0.26, 4.21 and 8.73 kg (meat, eggs, Our World in Data, from FAO). Of the fish and seafood, 3.35 and 0.515 kg are shrimp, and 2.41 and 0.371 kg farmed shrimp (Vasco Grilo's spreadsheet, from FAO). For his post "GiveWell may have made 1 billion dollars of harmful grants", the same spreadsheet puts the Harms caused to poultry birds and farmed aquatic animals as a fraction of the direct benefits of human life in 2022 at 34.6 in China, 5.17 in India and 2.31 in Nigeria, with his welfare levels and weights. Here, all farmed animals behind the diet count (broilers, hens, pigs, cattle, farmed fish and farmed shrimps), wild-caught seafood does not, and the sign depends on the exponent. Fish and seafood other than shrimp also include molluscs and other invertebrates: a share Beta(mean 0.7, concentration 10) of them is taken as finfish speculative. The model weights the two countries by China's share of the deaths averted, Beta(mean 0.5, concentration 4), multiplies by LN(0.6, 1) for older adults' smaller diets, takes a share Beta(mean 0.6, concentration 10) of fish as farmed, and counts the share of extra demand met by extra production, as for consumption elsewhere, at the animal-years per kg used throughout speculative.

It adds 4.38 m²-years of farmland per dollar on average (5th to 95th percentiles 0.283 to 15.1): the people saved are older, with fewer years left and no descendants.

In the formula, per death averted: G is the QALYs of the years gained and the years lived with disability averted. ΔY covers the soil organisms and trees on the farmland of the extra person-years and on the farmland freed by the money spent. It also covers the farmed animals behind the food the people saved eat (person-years × share met by extra production × kg per person-year × animal-years per kg). C_h = 0, and K is the cost per death averted.

12. Giving What We Can

Per dollar donated to Giving What We Can (GWWC), which promotes pledges to give to effective charities. Its latest evaluation finds that GWWC achieved a giving multiplier of 7.1x in 2025 — producing $16.1M in value for highly effective charities at a cost of $2.3M., and that Even combining all conservative assumptions, our giving multiplier is 1.5x (up from 0.9x in 2023–2024). Global health and wellbeing remains the largest cause area, accounting for 48% of pledge donations by volume, while Animal welfare accounts for 12% of recorded pledge donations and Reducing global catastrophic risks accounts for 14% of recorded pledge donations. The multiplier counts only highly effective charities. The shares of donations to them are high for the first two and lower for the third: A high proportion of global health donations go to charities we classify as highly effective (94% best guess, 70% conservative)., A high proportion of these donations go to charities we classify as highly effective (94% best guess, 55% conservative). for animal welfare, and a lower fraction of these are classified as high impact (59% best guess, 30% conservative) for catastrophic risks. Besides, Under our best-guess assumptions, we also classify selected programs in cause areas where we do not yet have active recommendations — including climate change and movement building — as high impact. (GWWC's 2025 impact evaluation). Per dollar:

• Dollars moved to effective charities: LN(1, 12), mean 4.61, median 3.47, between the conservative and best-guess multipliers. GWWC notes that its multiplier does not tell us whether the next dollar we spend would be equally productive (our marginal cost effectiveness), which is arguably the more decision-relevant quantity for funders (GWWC's 2025 impact evaluation). A marginal dollar likely moves less than the average one. Most of the value comes from future donations by new pledgers, which GWWC discounts: For our best-guess estimate, we apply a discount rate of 3.5% per year. That is close enough to the weight for uncertainty over time used here for the dollars moved to be treated as given now speculative.
• Where the money goes: the value for highly effective charities is 45.1% (48% × 94%) global health, 11.3% (12% × 94%) animal welfare and 8.26% (14% × 59%) catastrophic risks of recorded pledge donations. Climate change and movement building add about 10%, read from the evaluation's chart of best-guess high-impact donations by cause area. Normalised, the Dirichlet shares have total concentration 50. Their means are 54.0% for life-saving global health, modelled as the All Grants Fund, and 6.44% for cash transfers, as GiveDirectly's programme. For animal welfare, they are 9.54% for corporate campaigns, as cage-free campaigns; 1.92% for alternative proteins, as GFI; and 3.63% for shrimp and fish welfare, as HSI. The rest, 24.5%, goes to catastrophic risks, climate change and movement building, which only use resources here. The splits within global health and within animal welfare are guesses speculative. So 2.49 dollars per dollar go to life-saving global health on average.

GWWC adds 18.9 m²-years of farmland per dollar on average, with a wide spread (5th to 95th percentiles −18.6 to 94.6). The farmland added by the All Grants Fund and cage-free campaigns is partly offset by that decreased by GFI. By source, in m²-years per dollar: the All Grants Fund +18.9, cage-free campaigns +8.94, GiveDirectly +1.69, GFI −10.7 and HSI −0.00375. So GWWC adds 2.48 times as much farmland per dollar as the All Grants Fund, not 4.61 times. Only 53.9% of the money moved goes to the fund, and GFI's share cancels the gain from cage-free campaigns. The 24.5% moved to charities not modelled, on catastrophic risks, climate change and movement building, is counted only as resources used, as its effects on farmland are unclear speculative.

In the formula, per dollar donated: G and ΔY are those of the interventions funded, each times the dollars moved to it per dollar ÷ its cost per unit. ΔY also includes the farmland freed by the dollar spent by GWWC and by the dollars moved to charities not modelled. K = $1.

Pain intensities

The gains to hens, shrimps and rodents add up time in the Welfare Footprint Institute's four categories of pain, each weighted by its intensity relative to fully healthy life, whose welfare is 1 AQALY per year. The intensities are uncertain and drawn together so that each category is always more intense than the one below it speculative:

• Hurtful: LN(0.3, 3), median 0.949. Experiences of Pain in this category disrupt the ability of individuals to function optimally (Welfare Footprint Institute). Disability weights for painful conditions are lower, for example Migraine 0·441 (0·294–0·588) (Salomon et al. 2015). But they are capped at 1, the weight of death, and so compress severe states.
• Annoying: hurtful ÷ LN(3, 30); median 0.100, 5th to 95th percentiles 0.0196–0.509. Sufferers can ignore this sensation most of the time (Welfare Footprint Institute).
• Disabling: hurtful × LN(5, 50); median 15.0, 5th to 95th percentiles 2.94–76.4. Disabling pain takes priority over most bids for behavioral execution and prevents most forms of enjoyment or positive welfare (Welfare Footprint Institute). Rethink Priorities' weights put hurtful pain at 0.02–0.2 of disabling pain (shrimp welfare report), the same 5- to 50-fold range as here. The Welfare Footprint Institute finds that the equivalence between Annoying and Disabling pain would be best represented by a ratio of approximately 1 to 94-496 (Short agony or long ache). Here, the median is 150 (5th to 95th percentiles 29.4–764).
• Excruciating: disabling × LN(20, 10 k); covering extreme levels of Pain that are not normally tolerated, even if only for a few seconds (Welfare Footprint Institute). Rethink Priorities puts it at 5–1 k times disabling pain and notes that The biggest uncertainty is how many orders of magnitude worse excruciating pain is than disabling pain (shrimp welfare report). In psychophysics studies cited by the Welfare Footprint Institute, the hardest shock can feel over 3,000 times worse than the mildest, and the worst contact heat was judged 4,000 to 25,000 times more intense than the mildest pain (Short agony or long ache). But the most intense laboratory pain need not be excruciating. The reasoning below supports a heavier upper tail. The median is 6.71 k, the 5th to 95th percentiles 201–224 k, and the mean 65.2 k.

Vasco Grilo's HSI analysis assumes that Hurtful pain is as intense as a practically maximally happy life, that Disabling pain is 10 times as intense as a practically maximally happy life and that Excruciating pain is 100 k times as intense as a practically maximally happy life. These are intensities of 1, 10 and 100 k, taking a practically maximally happy life as fully healthy. With these, 0.864 seconds of excruciating pain cancel 24 hours of such a life, about 8 of which are spent asleep. Cynthia, from the Welfare Footprint Institute, explains that By not tolerated we mean that Excruciating pain can't be sustained for long (e.g., hours, as opposed to minutes) without neurological shutdown, with desperate attempts to escape pain even at the risk of death. She adds that Examples would include severe burning in large areas of the body, dismemberment, or extreme torture (comment). Risking death to escape such pain points to very high ratios. Accepting a 1% risk of losing 40 years of healthy life to escape one minute of it gives up 2.44 days per second, a ratio of 210 k. But such choices are made in agony and may overweight the present. People who jump from burning buildings mostly trade one death for another, which says little about the ratio to healthy life. Choices made in calm conditions point lower. Before anaesthesia, amputations were performed on fully conscious patients who were able to feel the excruciating pain of the surgery, and A skilled surgeon could amputate a leg in under three minutes (History Hit). Patients accepted minutes of dismemberment rather than likely death, which caps the ratio at roughly 1 M–10 M. People who choose death when facing excruciating pain for h hours a day for the rest of their lives only imply a ratio above 24/h, 12.0 for 2 hours a day. Within that range the evidence does not decide, and the scale of these categories is untested speculative. So 100 k is at the 89.7th percentile here. Expected gains are driven by the upper tail: the mean, 65.2 k, is 9.72 times the median. With Vasco Grilo's intensities, GFI leads under model 1 up to n = 0.85, HSI at n = 0.9–1 and GWWC from n = 1.05; under model 2, GWWC leads up to m = 0.8, HSI at m = 0.85–1.1, cage-free campaigns at m = 1.15–1.6 and GWWC from m = 1.65.

People whose deaths are averted

Each child saved gains its remaining years of life, weighted for uncertainty over time, times the QALYs per year of life (births), 35.8 QALYs on average. In GiveWell's model, 73.7% of the deaths averted by nets are of children under 5, with several decades of life ahead, at somewhat less than full health. Net of the births averted, and with descendants across all time weighted for uncertainty over time (births), the gain is 54.4 QALYs per child death averted on average, 61.8% of it from descendants, and 44.7 per death at 5 or older. With other health gains and income gains, GiveWell's grant gains 68.9 QALYs per death averted: 51.9 from the lives saved, 8.13 from other health gains and 8.89 from income gains, with $10.7 k of extra consumption. If the children of the people saved all displaced other people's, the grant would get 0.00884 QALYs per dollar at n = 2, 2.88 times GiveDirectly's Cash for Poverty Relief programme instead of 5.06 times, and GWWC would still lead under model 1 from n = 1.

Hens moved out of cages

The Welfare Footprint Institute estimates that at least 275 hours of disabling pain, 2,313 hours of hurtful pain and 4,645 hours of annoying pain are prevented for each hen kept in an aviary instead of CC during her laying life (Welfare Footprint Institute; CC is conventional cages). These are weighted by the intensities above and multiplied by LN(1, 2.5) for welfare issues they leave out. The Institute writes that Estimates of time in pain should be therefore viewed as the minimum time in pain endured by laying hens, that We did not analyze the rearing phase, and that We did not consider the welfare impacts of the perception of lack of control in cages. No source sizes the omission, and the upper end is a judgement speculative. Vasco Grilo's analysis of corporate campaigns multiplies the Institute's estimates for hens in conventional cages by 1.5 and those for cage-free aviaries by 1.25. The Institute also notes that Approximately 2,000 hours of excruciating pain are expected for every flock of 50,000 laying hens (1-3 hours per individual affected), without attributing it to either system. So it is left out. The share of this reduction in pain that cage-free systems, as implemented, achieve is drawn from one distribution covering 0, Normal(0.9, 0.4) capped at 1. It has mean 0.785 and 5th percentile 0.242. There is a 40.1% chance that cage-free systems achieve all of the reduction, and a 1.23% chance that the share is negative, where cage-free systems are worse. The European Food Safety Authority's 2023 opinion recommends housing all birds in non‐cage systems with easily accessible, elevated platforms. Mortality in aviaries falls with experience to the level of cages (Schuck-Paim et al. 2021). But Keel bone fractures are the most important welfare concern in cage-free aviaries (Welfare Footprint Institute). Aviary mortality was also 11.5% against 4.7% in cages in the US study cited under cage-free campaigns above. A share of −0.1 would add 27.5 hours of disabling pain per hen, against 6.80–34.0 hours from the 6.8-point excess mortality times about 100–500 hours of disabling pain per death, before fractures. The distribution is a judgement speculative. The mean gain is 1.51 AQALYs per laying life (median 0.975, 5th to 95th percentiles 0.156 to 4.62). Leaving out excruciating pain matters: at the mean intensity used here, 0.1 hours of it per hen in either system would be worth 0.744 AQALYs.

Broilers raised to the Better Chicken Commitment

The Welfare Footprint Institute's estimates, which already count the longer lives of slower-growing breeds, are weighted with the intensities above and multiplied by LN(1, 2.5) for welfare issues they leave out, as for hens. That gives 0.254 AQALYs per broiler on average, median 0.123. Hurtful pain can rise as well as fall, but disabling and excruciating pain dominate the gain (broiler campaigns).

Shrimps: the gain is uncertain in sign and size

Vasco Grilo's HSI analysis takes 95% of the slaughter that stunning replaces to be air asphyxiation and 5% ice slurry: All of the shrimps helped transition to electrical stunning, 95 % from air asphyxiation, and 5 % (= 1 - 0.95) from ice slurry. For ice slurry it uses Rethink Priorities' estimate of 0 h in annoying pain, 3.02*10^-4 h in hurtful pain, 0.0239 h in disabling pain, and 0.00604 h in excruciating pain. For air asphyxiation it uses a time in disabling pain equal to the maximum time during which shrimp can remain alive of 30 min, and a time in excruciating pain as a fraction of that in disabling pain equal to that of ice slurry (0.126 h). With his intensities this gives An equivalent additional time of a practically maximally happy life per shrimp helped of 1.37 years. His weights turn this into 0.0426 DALYs averted per shrimp helped (DALYs are disability-adjusted life years), 99.7% of it from excruciating pain during air asphyxiation. The model keeps that structure, but with ice slurry as the usual method. Rethink Priorities' model assumes All shrimp who are slaughtered are slaughtered using ice slurry, bar those on farms with stunners (Rethink Priorities). A review finds that Immersion in an ice slurry bath (thermal shock) is still the most commonly used technique (Blary et al.). On the other side, most farmed shrimps are killed in ice slurry or left to suffocate in air (Shrimp Welfare Project). Aaron Boddy, SWP's chief operations officer, quoted in Vasco Grilo's HSI analysis, describes the usual practice: typically instead the crates of shrimps are “dipped” into ice slurry for a few seconds, so in reality once removed they die of asphyxiation/crushing. He adds that practices can vary significantly by country and production systems (and producer). In a comment, he says it wouldn't surprise me if the number was ~100%. No shares by country were found. So air asphyxiation is a share Beta(mean 0.45, concentration 3) of the slaughter replaced speculative. Air asphyxiation causes LN(0.05, 1.5) hours of disabling pain, as death from oxidative stress only takes a bit longer than 30 minutes for P. vannamei (Rethink Priorities), but shrimps kept cool on ice survive for hours. Excruciating pain is LN(0.03, 0.8) times as long, median 0.155. For air exposure at harvest, Rethink Priorities' mean hours are 0.235 hurtful, 0.0680 disabling and 0.00338 excruciating, a ratio of 0.05, as We, therefore, assign most pain to ‘hurtful’ and ‘disabling’ (Rethink Priorities). Its ice-slurry estimate has a ratio of 0.25. The Welfare Footprint Institute's analysis of air asphyxia in trout splits the time in pain about equally between the two, a ratio of about 1 (Schuck-Paim et al. 2025). All the air time is counted as disabling here, while Rethink Priorities puts most of it in hurtful pain speculative. Rethink Priorities' ice-slurry durations are multiplied by LN(0.7, 12). Its base is a duration of about 4 minutes, 1.81 of them in pain. But brain activity fell to threshold in 4 minutes at -2.5 °C, in 9 minutes at 0 °C and in 28 minutes for 4 of the 5 animals exposed at 2.5-5 °C (Blary et al., on Somerville et al. 2026). Commercial slurries, and storage on ice after a dip, are likely at the warmer end. The stun itself is taken to cause 0.5 seconds of disabling pain, as Vasco Grilo's HSI analysis assumes a time in disabling pain equal to 1.39*10^-4 h (= 0.5/60^2). The intensities are those above.

Whether stunning achieves that gain is unclear. A review of the evidence finds we have very limited data on electrical shrimp stunning that doesn't support a confident conclusion as to whether it's good or bad. It notes that the key studies use a very small sample size (N = 6 for each intervention), and that Electrical stunning without proper ice slurry slaughter poses real potential for causing harm (Animal welfare has an evidence problem). On the other side, a comment on that post by Krzysztof Wojtas of SWP quotes a summary of a University of Stirling study: Electric stunning provides a rapid and effective stun as shown by the lack of response to subsequent cold exposure, though the process triggers a strong muscle contraction and elevated blood lactate. It also quotes a 2020 producer case study that reports >97% stunning efficacy. He adds that SWP uses ice as a kill step after stunning, not in place of it. He also notes that in the Stirling study ice slurry at 2.5–5 °C was slow and unreliable, and that electrical stunning is faster, more consistent, and more monitorable than industry-typical ice slurry stunning. A separate challenge is to the size of the gain. Cornelis Dirk Haupt argues that shrimps probably lack the nerve fibres behind lasting pain, in which case long death is not a long agony. He concludes: So it seems possible that the benefit SWP assumes is far less certain than advertised, possibly near zero. (How do we know shrimp suffer more from asphyxiation than stunning?). Meghan Barrett, an insect neurobiologist, replies that shrimp nociceptors are much more similar to insect class IV (see below) nociceptors than to the fast fibres the argument assumes. She adds that the green crab study it relies on is weak, as This study looked at the effects of exercise on land for a species well-adapted to emersion. Another commenter summarises that shrimp nociceptors most likely ressemble C-types neurons, the kind behind lasting pain. She also warns that electrical shocks that don’t stun before slaughter seem worse than no stunning before slaughter.

These comments pull in both directions without settling the question. The kill step and the published Stirling results make the harmful case less likely where SWP's protocol is followed. But Barrett's warning supports a harm when stuns fail, and the samples remain small. Haupt's argument, which would shrink the gain, is weakened by Barrett's reply, and the lower tails of the slaughter durations above already allow for short pain. So the simulation draws the share of the pain of slaughter averted by stunning, as deployed, from one distribution covering 0, Normal(0.75, 0.4) capped at 1. It has mean 0.685 and 5th percentile 0.0923. There is a 26.6% chance of averting all of the pain of slaughter, and a 3.04% chance that the share is negative, where stunning adds pain. The pain of the stun itself is subtracted in every case. In the peer-reviewed Stirling study, even shrimps not fully stunned and then chilled tail-flipped for a median of 18.5 seconds, against 51–168 seconds in ice alone (Somerville et al. 2026). So a failed stun followed by ice mostly shortens the response rather than adding to it. But All animals tested (n = 10/stunning condition) recovered their righting reflex between 5 s to less than 3 min after a 1-second stun, which makes the ice step essential. The lower part of the distribution covers failed or reversible stuns and the lack of an adequate kill step speculative. In field trials, Immediate stunning efficacy was 100 percent, but 5/72 shrimp (7 percent) tail flipped after stunning (Khan and Rennie, who work for the manufacturer). Electrical stunning alone produces reversible behavioral effects (Blary et al.). After 10 s at 120 V, It took 5 to 10 min before the animals gained some ability to move or swim (Weineck et al. 2018). The mean gain is 0.333 AQALYs per shrimp stunned, but the median is 0.0157 (5th to 95th percentiles 1.57×10⁻⁴ to 0.906): the mean rests on the upper tail of the intensity of excruciating pain.

Rodents and wildlife

• Rodents: the suffering of an anticoagulant death, LN(12, 96) hours of disabling pain and LN(30, 3.6 k) seconds of excruciating pain weighted with the intensities above (mean 2.14 AQALYs, median 0.178, 94.6% of the mean from excruciating pain). A review of rodent control finds times to death typically in the region of four to eight days, with more commonly, an average of one to three days, with a maximum of four to five days of clinical signs. It also finds brodifacoum-poisoned rats that lay prostrate for a mean of 11.4 h prior to death (Mason and Littin 2003). A welfare assessment scores anticoagulants, cholecalciferol and non-toxic cellulose baits (severe to extreme impact for days) (Baker et al. 2022). Not all days of clinical signs are disabling, and haemorrhage into joints and the skull may be excruciating for minutes to an hour. The intervals are judgements speculative. From this is subtracted the suffering of the natural death that replaces it, a share LN(0.05, 1.2) of it, mean 0.390. The share exceeds 1, meaning that the natural death is worse than poisoning, with a probability of 7.25%. Poisoned rats suffer for days, and Animals typically remain conscious during this time (Mason and Littin 2003). Natural deaths are a mix of fast and slow. Of rats killed by cats in Baltimore, more than three-quarters of killed rats were 200 g or less (Glass et al. 2009), mostly young rats killed in seconds to minutes. Starvation, exposure, disease and infected wounds take hours to days. No source quantifies the suffering of natural deaths, and the share is a judgement speculative. This gain, and those below, apply to the share of rodents switched for which fertility control works. The rodents also live on for LN(0.15, 0.7) years at the welfare of wild rats: a wild rat population experiences about 95% mortality, which means that only 5% of rats remain alive after 12 months (ratbehavior.org). That is a constant hazard of 3.00 per year and a remaining life of 0.333 years.
• Wildlife spared poisoning: for the same share, LN(0.01, 0.5) non-target animals per rodent switched, each spared a death like the rodents', less a natural death as for the rodents, weighted as wild small mammals, since most are. Brodifacoum residues: detected in all non-target small mammal taxa during rat control, and BR residues were present in 23% out of 742 samples collected during and after baiting. (Geduhn et al. 2014). Small mammals eat the bait directly: A large proportion (48·6%) of individuals in local populations ate the bait and Local populations of non-target small mammals declined significantly following rodenticidal rat control (Brakes and Smith 2005). Predators are often exposed, but in a French survey of raptors 50% (78/156) were positive for ARs while AR toxicosis was suspected in only 2 of these individuals (US Geological Survey (USGS)). No study gives deaths per rat killed, and the interval is a judgement speculative.

Each is a lognormal with the 5th and 95th percentiles shown, drawn independently. Because a weight is a ratio raised to an exponent, its mean at high exponents exceeds the weight of an organism with median values. The metabolic-rate post's values come from Hoehler et al. (2023).

In AQALYs per organism-year: 1 is a year of fully healthy life for that organism. Groups are drawn independently. P(negative) is for each group's draw, before the deviation for each land class or biome. For soil organisms these are the levels on natural land. Farmland shifts them (below). Whether wild animals, and soil animals above all, have positive or negative lives is contested. Horta argues that r-selected animals often die at young ages by brutal means such as starvation and predation (Faunalytics summary). Browning and Veit note that it has become common to depict animals in the wild as existing in a state dominated by suffering and question the evidence for it (Positive wild animal welfare). Zach Groff, presenting his paper with Yew-Kwang Ng, says that When the error is fixed, the balance of suffering and enjoyment can go either way (Does suffering dominate enjoyment in the animal kingdom?). Deaths set the means. The life of an animal that is not dying is worth at most about +1 AQALY a year, while the suffering of dying is paid once per animal. In a stable population, deaths per animal-year are 1 ÷ life expectancy. Brian Tomasik argues that the pain of death is probably the dominant factor for considering wild-insect welfare (Crop cultivation and wild animals). Wild Animal Initiative cautions that we do not know where to draw the line between net-positive and net-negative welfare for any species (Welfare expectancy). At the median intensities used here, 10 seconds of excruciating pain per death costs about 0.0776 AQALYs a year for a nematode living 10 days, and about 0.00425 for a mite or ant worker living half a year speculative. So groups with faster turnover get more negative means, while dormancy, which dilutes both good and bad states, and quick deaths keep them close to zero. The means for soil organisms are small next to their spreads, and the sign of each group's welfare stays close to a coin flip. The spreads are wide because small fractions of life in severe pain go a long way on this scale. With the median intensity of excruciating pain used here, 6.71 k times that of healthy life (pain intensities), an animal that spends 0.01% of its life in excruciating pain loses 0.671 AQALYs per year. For soil arthropods, a Normal spread of 0.7 around a mean of 0 gives a 15.3% chance that welfare is beyond ±1 AQALY per year. The levels for farmed animals are set to be consistent with the hours of pain estimated by the Welfare Footprint Institute and Rethink Priorities, weighted by the pain intensities used here.

The distributions are guesses made for this analysis from the sources above speculative.

Welfare by land class

Each soil group's welfare in a land class is its draw above plus a deviation for that class speculative:

Cropland is tilled, sprayed, harvested and left bare, and favours short-lived opportunists over long-lived species. Cropland holds 18.4 k mites per m² in Rosenberg et al.'s global data, against 45.8 k–99.6 k in tropical and temperate grasslands and forests (land effects). Intensive agriculture is making soil food webs less diverse and composed of smaller bodied organisms (Tsiafouli et al. 2015). Of the effects of pesticides on soil invertebrates, 70.5% of tested parameters showed negative effects (Gunstone et al. 2021). Also, reduced tillage intensity or no‐tillage increases the total density of springtails (+35%), mites (+23%) (Betancur-Corredor et al. 2022). Grazing harms soil life less: effects are strongest under heavy, continuous grazing (Tian et al.). This evidence is about numbers and community composition, not welfare per animal. But cropland shifts each group towards faster life histories, which means more deaths per animal-year. Where Australian native pastures were replaced by cultivated, regularly fertilised ones, total nematode abundance increased under cultivation, primarily due to the proliferation of bacterivorous nematodes (Yeates and King 1997, via a review), opportunists with short lives. With more acute deaths from tillage and pesticides, this suggests that soil animals on cropland fare worse than on natural land, and on pasture somewhat less so. Fertiliser, residues and irrigation could ease starvation and desiccation, which keeps the expected shifts small speculative. The spread of each deviation is half that of the group's welfare, since the conditions of life can change welfare by much of its plausible range. In Vasco Grilo's estimates for chickens, welfare per living time is -2.27 AQALYs per year for a broiler in a conventional scenario against -0.161 in a reformed one. It is -1.69 for a hen in a conventional cage against -0.333 in a cage-free aviary. The expected shifts matter even so: conversion to cropland cuts nematode numbers by 23.1% on average, and conversion to pasture changes them by −1.21%. So a shift of −0.0180 for cropland nematodes cancels the effect of their lower numbers. With no shift for farmland, under model 1 GWWC leads at every exponent.

Soil arthropods

Densities by land class are from Table S4 of Rosenberg et al. (2023), unrounded, as listed in the spreadsheet for Vasco Grilo's Expanding agricultural land may increase or decrease the welfare of soil animals. Its note reads Values shared by Yuval Rosenberg on 10 March 2026, and The values rounded to 1 significant digit are in Table S4 of Rosenberg et al. (2023). Every mean and 95% confidence interval below matches it. Each density is a lognormal with the mean shown and the spread of the 95% confidence interval of the mean across sites. Animals per m²:

For springtails and mites, farmland density is instead a multiple of the density of the natural land displaced, since the evidence on farmland is mixed. For springtails, a global synthesis finds minimum densities in tropical forests (5831 individuals m−2) and agricultural ecosystems (3438 individuals m−2 (Potapov et al. 2023). A meta-analysis of intensified against undisturbed land finds a substantial decline in springtails abundance (-34%) (Betancur Corredor et al., at the European Geosciences Union's 2024 General Assembly (EGU24)). For mites, the same meta-analysis finds that Mites show a positive increase in abundance (+129%). In central Amazonia, oribatid mites had the highest density in POA (23,371 ind/m2), a polyculture, against FLO (10,398 ind./m2) in primary forest. Springtails there were clearly higher in the primary forest (FLO), with a value of 3,296 ind/m2 (Franklin et al. 2001). Rosenberg et al.'s global means, by contrast, have far fewer mites on cropland than on natural land. Farmland ÷ natural density is LN(0.3, 1.1) for springtails on cropland, mean 0.621, and LN(0.35, 1.4) on pasture, mean 0.765; for mites, it is LN(0.25, 1.6) on both, mean 0.742 speculative. Rosenberg et al.'s farmland means are kept for ants and termites, and for the dikes of shrimp ponds.

New cropland displaces, in prior mean shares: tropical savanna 0.27, tropical forest 0.15, temperate grassland 0.03, desert 0.03, temperate forest and Mediterranean 0.01 each, boreal 0.005, pasture 0.33 (re-created at the pasture frontier) and recultivated fallow 0.165. New pasture displaces tropical forest 0.38, tropical savanna 0.33, temperate grassland 0.10, desert 0.06, temperate forest and Mediterranean 0.05 each, boreal 0.02 and tundra 0.01. The cropland prior also gives tundra 0.001. Both are Dirichlet priors with total concentration 10, based on Potapov et al. (2022), for whom The other half of cropland expansion (51%) was due to pasture conversion and recultivation of abandoned arable land. By region, In Africa, 79% of the gross cropland gain occurred at the expense of natural vegetation, against 61% in Southeast Asia and 39% in South America (Resource Watch). For pasture, of agriculture-driven tropical deforestation, pasture expansion forms the most important driver, accounting for around half of the deforestation (Pendrill et al. 2022, Nature-based Solutions Initiative). Worldwide, cropland expansion is causing almost 50 percent of global deforestation, followed by livestock grazing, accounting for 38.5 percent. The loss of tropical forests accounted for more than 90 percent of the global deforestation from 2000 to 2018 (FAO). With deforestation of 7.8 M ha a year in 2010–2018, cropland other than oil palm took about 3.32 M ha of forest a year, 0.244 of the 13.6 M ha of yearly gross cropland gain (217.5±37.7Mha (17%) represents new cropland established since 2003, Potapov et al.). The share is less once forest that does not become productive farmland is excluded. Dryland irrigation accounts for 5% of the gain (Potapov et al., Table 2), mostly in deserts. Pasture expansion is concentrated at tropical frontiers: The expansion of pasture land to raise cattle was responsible for 41% of tropical deforestation (Our World in Data). But there are 140 million fewer hectares of pasture today than there were in 2000, with more than two-thirds of all countries now experiencing flat or declining pasture area (Blaustein-Rejto et al. 2019). So extra grazing demand partly slows the abandonment of drylands and temperate rangeland rather than clearing forest, hence the weight on temperate grassland and desert. Pasture displaced by new cropland is re-created at the frontier one for one. The net change in farmland is already scaled by the share that survives land-market responses, and within that net change the land converted must eventually be natural land. The split among biomes within these shares is a judgement speculative. Pond farms hold a fraction Beta(mean 0.12, concentration 5) of the cropland density. In an FAO design for Malaysian ponds, To include the spaces taken up by the dikes, their side slopes and water canals, the average gross area for one unit of family pond is in the order of 4.0 acres (1.6 ha), for 1.2 ha of ponds (FAO). Across the five leading shrimp exporters, The average LWR for the five leading shrimp exporting countries was 1.67, the land-to-water ratio, with embankments 0.032–0.325 of farm area, 0.160 on average (Boyd, Davis and McNevin 2022, Table 1). Dikes at perhaps 0.7 of cropland density, plus pond bottoms dried between cycles, give about 0.12; pond water and anoxic bottoms hold few soil arthropods speculative. New ponds displace, in prior mean shares (Dirichlet, concentration 6), cropland 0.45, salt flats and tidal flats 0.25 (as desert), mangroves 0.15 (as tropical forest, the closest class in the table) and other wetlands and grazing land 0.15 (as tropical savanna). Mangroves were a large source in the past, but The present-day loss of mangrove area to shrimp farming is less than in the past (Boyd et al.). In Southeast Asia, more than 100,000 hectares of mangrove forest were lost between 2000 and 2012, aquaculture amounting to 30 percent of the region's mangrove displacement (Richards and Friess, via Mongabay). That is about 4.30% of the 581 km² of coastal ponds added each year in Vietnam, Indonesia, Thailand and Myanmar over 1990–2015 (Luo et al. 2022). In northeast Brazil, Between 8% ... and 10.5% ... of the total shrimp pond area is from mangrove deforestation, with most ponds on salt flats (Lacerda et al. 2021). Conversion of farmland is documented locally, for instance conversion from agricultural land to aquaculture accounting to an extent of 237.01 hectares in an Indian village (Saxena et al. 2014). The shares of cropland and salt flats are judgements speculative. Mediterranean ants are imputed at 171 per m² (the geometric mean of the other classes). Single-site cells get a ×/÷3 interval.

Soil nematodes

Rosenberg et al. do not cover nematodes. So managed and natural land are compared here in the public aggregated file of the global nematode database (van den Hoogen et al. 2020; data). Its authors note Across all samples, 67.2% originate from natural sites and 32.8% from agricultural or managed sites. A pixel counts as managed if EarthEnv's cultivated-plus-urban share is at least 50%, and natural if under 10%. Densities are per m² over the top 15 cm. The script is nem.py, below.

Site studies are mixed. A meta-analysis finds that The overall abundance of nematodes per 100 g of soil was similar in forest, NGL, DGL, and agricultural ecosystems. (Journal of Nematology meta-analysis; NGL and DGL are natural and disturbed grassland). In Sumatra, Land use did not affect the total abundance of litter- and soil-dwelling nematodes under rubber and oil palm (Krashevska et al. 2019). The simulation draws the cropland ÷ natural ratio from LN(0.4, 1.3), mean 0.769. New cropland is mostly tropical, and weighting the tropical result for forest-like land and the higher savanna ratio by the displacement priors above gives about 0.73. Pasture keeps more of the soil, with wider uncertainty, since grazing can raise or lower numbers: its ratio is LN(0.5, 1.7), mean 0.988 speculative. A meta-analysis of 192 studies finds that Nematodes, on the other hand, experience a negative impact on abundance (-7%) from agricultural management (Betancur Corredor et al., EGU24), close to the global result. Natural density is LN(800 k, 4 M) per m². The global total is 4.4 ± 0.64 × 10^20 nematodes (van den Hoogen et al. 2019), 3.38 M per m² over about 130 trillion m² of ice-free land. The study finds higher abundances in sub-Arctic regions (38% of total) than in temperate (24%) or tropical (21%) regions. Tropical land, where most new farmland is, holds about 2 M per m². So new cropland removes 465 k nematodes per m² on average, and new pasture 23.3 k. Weaknesses: land cover is per pixel, not per sample; urban land is mixed in; sampling depths differ.

Soil bacteria and archaea

Converting natural land cuts microbial biomass by a quarter to over a half. A meta-analysis finds that converting grassland to cropland and abandoned land decreased MBC by 47.80 % and 38.74 %, respectively (Science of the Total Environment meta-analysis; MBC is microbial biomass carbon). A meta-analysis for India finds a smaller loss, 26% of microbial biomass carbon when forest is converted to cultivation (Padbhushan et al. 2022). In the tropics, where most new farmland is, losses are larger. A meta-analysis of tropical land use finds that Soil microbial biomass decreased with forest conversion to agriculture and plantations (Díaz Vallejo et al. 2021), by 58.6% for agriculture (a log response ratio of −0.881). As reported in the author's dissertation, Conversion to pasture did not affect microbial biomass C., a non-significant 16.2% fall. Assuming cell numbers fall with biomass carbon, the share lost to cropland is Beta(mean 0.45, concentration 16), and to pasture Beta(mean 0.1, concentration 10) speculative. New cropland on recultivated fallow changes neither microbes nor nematodes. Natural density is LN(1.2, 4.5)×10¹⁵ cells per m², around the land average of 2.31×10¹⁵. Of Most bacteria and archaea on Earth (1.2 × 10^30 cells), soil (3 × 10^29) (Flemming and Wuertz 2019), the figure the metabolic-rate post uses, over about 130 trillion m² of ice-free land.

Trees

The metabolic-rate post counts 10 trillion trees, 0.246 per m² of the world's 40.6 trillion m² of forest. Counting trees at least 10 cm in diameter, approximately 1.39 trillion exist in tropical and subtropical forests, with 799 billion in tropical moist forest and 318 billion in tropical grasslands (Crowther et al. 2015, Figure 4). That is about 0.04 and 0.016 per m² over about 20 trillion m² each. A survey of dense cerrado counts 1964 indivíduos per hectare with stems at least 5 cm across (University of Brasília), 0.196 per m². Weighted by the displacement priors, these counts of larger trees give about 0.02–0.03 per m² of natural land converted. New farmland removes LN(0.01, 0.12) trees per m² of natural land converted, mean 0.0461, allowing for smaller trees and for the scattered trees pasture keeps; new cropland on recultivated fallow removes none speculative.

Costs per unit of output

QALYs per dollar are computed draw by draw, dividing each draw's welfare change by its cost.

Food and animals

• Correlation between welfare levels and pain intensities. Welfare levels are drawn independently of the intensities, though the levels of farmed animals were set with the intensities in mind. So draws with intense pain raise the gains of hens and shrimps without lowering the welfare of broilers, pigs or soil animals, which favours HSI and cage-free campaigns slightly against School Plates and GFI.
• Effects of the wild-fish share of the replacement basket on wild fish populations.
• Fungi, protists and plants other than trees. Model 2 follows the metabolic-rate post, which covers trees, soil bacteria and archaea. Fungi hold about 12 Gt of carbon and protists about 4 Gt (Bar-On et al. 2018). Soil bacteria hold about 7 Gt, a tenth of all bacteria, since Bacteria include about 90% deep subsurface biomass (mostly in aquifers and below the seafloor). Doubling the microbial term would move the exponent below which soil organisms dominate model 2 for the intervention leading there by only 0.0222.
• Microorganisms in pond sediments, and wild vertebrates on converted land.
• Variation in size within a group: each group has one uncertain typical neuron count and BMR, whereas at exponents above 1 the weight of a group of varied sizes is dominated by its largest members, notably for trees.
• Changes in farm soil management (tillage, fertiliser), which change microbial numbers without changing land class.
• Funging of donations restricted to a programme (using the results).
• Pest damage while rodent fertility control works more slowly, and any effects of interventions on future advocacy or policy.
• For GiveDirectly's programme: livestock bought with the part of the transfers that is invested, wild fish eaten, local price changes, and the different weights of goats, sheep and cattle, all of which count as cattle.
• For GiveWell's grant: stillbirths prevented, which add people, and treatment costs averted, which add consumption (9% and 20% in GiveWell's supplemental adjustment); livestock other than chickens for the extra people; and changes in chickens per person over time.
• For buying beef: the eater's enjoyment of the beef, greenhouse-gas emissions, and antibiotic resistance.
• For School Plates and the Good Food Institute: changes in pupils' and consumers' health and enjoyment, money saved by caterers and consumers and spent on other goods, and effects on eggs and fish. For GFI, effects on dairy are also not modelled (Giving Green's underlying scenarios also shrink dairy, which would decrease farmland more).
• For cage-free campaigns: effects on broilers and on campaigns for other species, and the hens' own soil footprint beyond feed.
• Correlation between the welfare of different soil groups, which are drawn independently.

Download the scripts (.zip) Download the results (mc.json)

1. Download the scripts (mc.py, model.py, nem.py, rankcorr.py, sens.py, sens_flip.py and textnums.py, with a README) and keep them in one folder, or copy them from below.
2. Install the dependencies: pip install numpy scipy pandas.
3. Run python mc.py. It executes model.py up to its output section, runs every intervention under both models, prints a summary and writes mc.json, which holds every number in the charts. The seed is fixed. So the results are reproduced exactly.
4. For the nematode comparison, download data/nematode_aggregated_wCovariateData.csv from the dataset's repository into the folder and run python nem.py.
5. For the rank correlations, run python rankcorr.py. It reruns the simulation with the same draws and writes rank.json.
6. For the sensitivity analyses, run python sens.py and python sens_flip.py. They print the leading intervention at each exponent.
7. For the numbers in the text, run python textnums.py. It recomputes them from the scripts' outputs, checks the qualitative claims, and writes nums.json.