What do RP's tools tell us about giving $100m to AW or GHD?

By Hayley Clatterbuck, arvomm, Bob Fischer, Derek Shiller, David_Moss @ 2024-10-07T20:41 (+167)

Intro

Suppose you have $100M to give away. You are drawn to the many important opportunities to reduce animal suffering or address pressing issues in global health and development. Your choice about how to allocate the funds could depend on considerations like these:

• Moral values: How much moral weight do you assign to various non-human species? Are you focused exclusively on hedonic considerations, like reducing suffering? Or do you have other relevant values, such as autonomy?
• Cost-effectiveness estimates: Species-discounting aside, how many DALYs/$ do the best projects in the area achieve? How fast do returns diminish in these areas?
• Decision-theoretic values: How do you feel about risk-taking? Are you willing to tolerate a substantial probability that projects will fail? What about non-trivial chances of projects backfiring?
• Second-order effects: Will giving to one cause set benefit any of your other values? Are there speculative benefits that might flow from giving to one cause or another? How, for instance, might your giving encourage or discourage other actors to give?

It’s no understatement to say: These questions are difficult. Still, it is possible to investigate clusters of these questions systematically, and we have provided three tools that do so: the Cross-Cause Cost-Effectiveness Model, the Portfolio Builder Tool, and the Moral Parliament Tool. 

Here, we give an overview of how our tools can be used to address the decision at hand and highlight some key insights that we have gained from using them. We hope that others will use our tools and share any insights that they have! 

There are several reasons why we won’t try to give a specific answer about how best to allocate resources between AW and GHD. First, the tools’ outputs depend on user-configured inputs about which we are (and everyone should be) highly uncertain. Second, we do not presuppose that all money must be allocated to one cause area or another; indeed, our tools often recommend diverse allocations. Third, while we think that each tool is a useful aid in decision-making, the tools are designed to highlight different issues; so, we don’t presume to know how to combine their judgments into an overall conclusion about what any particular actor should do, all things considered. Nevertheless, we have identified a few important patterns and cruxes.

Cross-Cause Cost-Effectiveness Model

How it works

• CCM tool
• Intro post
• Explainer video

The Cross-Cause Cost-Effectiveness Model is actually a cluster of different models used to predict the expected cost-effectiveness of various projects across the Animal Welfare, Global Health and Development, and Existential Risk spaces (the latter of which we will ignore here), given user-supplied judgments about key parameters. One of the functions of the model is to track uncertainty. The expected value of a project is derived from the many possible outcomes that could occur, which are explored through Monte Carlo simulations. The CCM also evaluates projects in light of alternative, risk-sensitive decision procedures.

Users can explore the debate topic by investigating individual animal- and GHD-related projects, assessing them for cost-effectiveness and probabilities of success, failure, and backfire. The tool is somewhat limited for present purposes since it evaluates individual projects within cause areas (rather than entire cause areas) and does not permit direct comparison of projects. It does include a diverse array of animal and GHD projects involving different beneficiaries, funding structures, and levels of effectiveness.

Example simulation of a hypothetical intervention to improve shrimp welfare by lowering ammonia concentrations.

What it says

Unsurprisingly given the diversity of projects considered, the CCM doesn’t say that animal projects are uniformly better than GHD ones or vice versa. There are better and worse projects within both spaces. Broadly speaking, however, the most promising animal welfare interventions have a much higher expected value than the leading GHD interventions with a somewhat higher level of uncertainty. For example, compare one simulation of a Cage-Free Chicken campaign vs. a good GHD intervention^[1]:

CCM’s cost-effectiveness estimate for Cage-Free Chicken Campaign (using default parameters)

CCM’s cost-effectiveness estimate for Good GHD Intervention (using default parameters)

The CCM uses RP’s moral weights, which some users might find too high. However, it’s worth noting that animal welfare projects remain competitive with GHD projects with moral weights that are 1-2 orders of magnitude lower. For example, if we set the upper bound of chicken welfare range at 1/10 of the tool’s default setting, the estimated cost-effectiveness is still quite high:

Should we conclude from these results that it would be better to spend an extra $100M on animal welfare than on global health? We think this would be too hasty. First, the default parameters of the CCM were chosen to be plausible but some are not heavily vetted and many are controversial. Users should evaluate what the CCM says given their own beliefs. Second, expected value estimates are not the whole picture for someone concerned about the risk of backfire or failure, so users should consult the alternative weightings as well. Lastly, the value of a very large expenditure in an entire cause area can not be derived solely from the value of any individual project (as our next tool well illustrates).

Portfolio Builder Tool

How it works

• Portfolio Builder Tool
• Intro post
• Explainer videos

The Portfolio Builder tool finds the optimal way to allocate money across cause areas given the user’s budget, empirical assumptions, and attitudes toward risk. It also evaluates proposed allocations in light of various decision theories. The tool can be configured through user answers to a set of guiding questions or it can be directly configured by selecting Other (custom setting).

As a default, the Tool compares three cause areas: GHD, Animal Welfare, and Existential Risk. To directly compare the first two cause areas, users can “zero out” the Existential Risk through the custom settings. There are multiple ways to do so, but the easiest is to click “Other (custom setting)” in the quiz, choose the simple linear model, and set the factor to 0 (meaning no payoff) and/or probability of zero effect to 1 (meaning no effect).

What it says

One of the key inputs is the expected cost-effectiveness of the first $1000 of spending. We can use prior estimates (including those delivered by the CCM) to give ballpark numbers. Effective Animal Welfare projects tend to have much higher estimated cost-effectiveness than effective GHD projects. For example, Open Philanthropy estimates that AMF delivers roughly 25 DALYs/ $1k. Depending on the moral weights that we assign to chickens, corporate cage-free campaigns may deliver between 70 (low moral weights) to 1000 DALYs/$1k (RP’s moral weights).

        Given these cost-effectiveness estimates, the Portfolio Builder tends to recommend giving all or most money to Animal Welfare (see configuration here). This result is robust under quite a few differences in the other parameters of the model, including:

• Both high and low animal moral weights
• Differences in the probability of success: for example, AW is favored when its projects have a 25% chance of success vs. a 99% chance of success for GHD
• Risk attitudes: moderate levels of risk aversion don’t tend to change the results (except when the probability of success for animals is very low)

However, there is one parameter that has a very significant impact on overall allocations: the rate of diminishing returns in each sector. For example, if we assume that returns diminish very slowly for GHD and slowly for AW, the portfolio now recommends allocating nearly all of the budget to GHD:

This result is both unsurprising and under-appreciated. At the scale of $100 million^[2], we need projects that can absorb lots of funding or a long list of very effective smaller projects. Many people lack access to information about the practicalities of philanthropic grantmaking and it’s easy to lose sight of it in abstract cause-prioritization debates. But at the scale of giving we are talking about, it may be the most important factor to consider.

Moral Parliament Tool

How it works

• Moral Parliament Tool
• Intro post
• Explainer videos

The Parliament tool allows users to represent their moral uncertainty via a moral parliament, consisting of different worldviews in proportion to the user’s credence in them. The parliament then decides on an allocation of resources across various philanthropic projects, including a selection of hypothetical GHD (e.g. Tuberculosis Initiative, Better Bangladesh) and Animal Welfare projects (e.g. Lawyers for Chickens, Shrimp Welfare International). The tool includes various allocation methods that the parliament can use to arrive at a decision (e.g. approval voting, Nash bargaining, etc.).

        The default projects that the Parliament tool includes are hypothetical, but users can change project settings to reflect their judgments about actual philanthropic projects. In order to foreground the effects of moral uncertainty, economic considerations are backgrounded: cost-effectiveness is partially captured by the Scale of projects, and allocation methods are sensitive to projects’ diminishing marginal returns (both of which can be manipulated by users).^[3] 

What it says

As with the above tools, the Parliament will only deliver a judgment once a user has input their own particular assumptions. However, we can discern a few general patterns.

Parliament composition matters

Worldviews encompass beliefs about who matters (e.g. chickens, future humans), what matters (e.g. happiness, justice), and how we should try to get what matters (e.g. avoiding the worst, seeking the best). The Parliament tool characterizes popular ethical worldviews and then makes predictions about how much each worldview will value each project. The predictions of individual worldviews are (and were designed to be) intuitive:

• Utilitarians (of all stripes) tend to assign significantly higher value to top AW projects than to top GHD projects, due to the significant weight they assign to animals and the larger scale of animal projects (because of the number of affected individuals)

• Welfare consequentialists (of all stripes) tend to judge the best AW and GHD projects to be equally good; compared to utilitarians, they assign higher moral weights to humans^[4], and this tends to balance out the larger scale of animal projects

• Egalitarians, who give equal weight to all individuals regardless of species, are the most animal-friendly
• Other worldviews tend to be more GHD-friendly, due to their lower moral weights for animals and higher value placed on distinctively human goods.

• At the extreme, the versions of Kantianism and Nietzscheanism represented in the parliament place no value on animals^[5]

• Less obviously, worldviews that place a lot of emphasis on presently-existing individuals and less on individuals in the future tend to favor GHD causes, since most AW projects will benefit animals that do not yet exist

These differences manifest in parliaments made up of these worldviews. In broad strokes, parliaments with equal representations of all worldviews tend to favor GHD-heavy allocations. Consequentialist parliaments tend to be more animal-friendly. Beyond these general results, there is considerable and surprising diversity across different methods of aggregating over uncertainty.

Allocation strategy matters

The tool includes a variety of methods for aggregating the judgments of worldviews into overall allocations. A parliament’s recommendations vary significantly depending on which method is used. To see the difference they make, we can focus on three:

• Maximize Expected Choiceworthiness (MEC): selects the allocation that maximizes the sum of the value across worldviews, weighted by their representation in the parliament
• Approval Voting: worldviews vote to approve or disapprove of each possible allocation of the budget and the allocation with the greatest number of approval votes is selected
• Moral Marketplace: each parliamentarian receives a slice of the budget to allocate as they each see fit; then, each's chosen allocation is combined into one shared portfolio.  

MEC tends to allocate most of the budget to animal causes, even for parliaments containing lots of non-animal friendly worldviews. For example, a parliament with equal representation of all worldviews gives nearly all the budget to shrimp under MEC:

The reason is that the worldviews that like Shrimp Welfare really like it, so the average value of the project across worldviews is very high.

        In contrast, approval voting tends to favor projects that are liked by many worldviews. This tends to favor GHD causes, since all the worldviews we consider care about relieving human suffering. For example, the equal representation parliament gives the highest approval to the following budget (with AW projects in blue, GHD projects in pink):

        Lastly, Moral Marketplace gives each worldview a share of the budget proportional to the number of representatives it has (a stand-in for the credence in that worldview). Unsurprisingly, this method yields significant diversification since different worldviews have different favorite projects, and this will tilt toward AW or GHD depending on the composition of the parliament. For example, the equal representation parliament allocates as follows:

Bumping up the number of utilitarians in the parliament to 50% causes the Moral Marketplace allocation to be noticeably but not tremendously more animal-leaning:

How you distribute money between animal and GHD causes will depend on the worldviews you entertain and how you navigate worldview uncertainty (with significant interaction effects between these two considerations).

Conclusions

Taken together, the tools suggest the following broad lessons for making large-scale Animal Welfare vs. Global Health and Development cause prioritization decisions:

• Given the large number of individuals affected, good AW projects tend to be more (often much more) cost-effective than good GHD projects.
• Judgments of relative cost-effectiveness depend on the moral weights that we assign to animals, though AW projects remain more cost-effective over a very wide range of moral weights.
• If you are uncertain about how much animals matter—if you are uncertain over worldviews—then how you allocate your resources will depend on how you arbitrate moral uncertainty.
  • Some methods, such as MEC, will tell you to give almost everything to AW even when you assign considerable credence to views on which animals matter very little.
  • Other methods recommend that you diversify across AW and GHD projects.
• At the scale of $100M, the diminishing marginal returns of the two cause areas play an extremely important role in determining allocations.

To conclude, we want to step back to consider how we should approach questions like the one posed in this debate.

First, tools like ours have two purposes. They provide guidance to individuals about what they ought to do, given their views. They also allow us to see how decisions depend on particular inputs: what the crux views are that make a big difference to outcomes; how sensitive conclusions are to changes in inputs; etc. This allows us to highlight uncertainties and show when and how they matter. Facing up to our uncertainties allows us to structure our ignorance and redouble our investigations where they are most needed. We hope that our tools will be of use to people as they answer the debate question, and we are eager to hear of any additional insights that come from their use.  

Second, cause prioritization questions—especially those about how to make zero-sum choices within fixed budgets—are important. Our own thinking has been informed both by “from first principles” work (e.g., on the merits of particular approaches to decision-making under uncertainty) and by “reverse engineering” work (e.g., using our tools to identify the sets of assumptions that support particular allocations). It’s valuable to identify the many factors that bear on cause prioritization and be explicit about our stances with respect to them.

That being said, our hope should be to move toward a world where cause prioritization questions are less relevant. As our exploration of diminishing returns demonstrates, it is essential to explore ways to build our capacity to do good across all cause areas. It is also important that we strategize about how to increase the pool of resources devoted to doing good for the world’s most vulnerable individuals. We are grateful to those who do that work.

The CCM, Portfolio Builder Tool, and Moral Parliament Tool are projects of the Worldview Investigation Team at Rethink Priorities. For acknowledgements of contributors to each tool, please visit the intro posts linked above. Thank you to Willem Sleegers for helpful feedback on this post. If you like our work, please consider subscribing to our newsletter. You can explore our completed public work here.

1. ^{^}

    Note that because these cost-effectiveness estimates are generated from Monte Carlo simulations over uncertain outcomes, there will be variation across uses of the model. Users will probably not get precisely the outcome produced here.

2. ^{^}

    Note that at lower spending levels, more of the budget will be allocated to AW causes. For example, if we ask the model to allocate $1 million, it recommends giving 77% to AW even with its faster diminishing returns.

3. ^{^}

    Scale can be changed by editing individual projects. Diminishing marginal returns can be manipulated under the “Edit settings” button on the allocation results page. 

4. ^{^}

    This is because welfare consequentialists think that things other than pain and pleasure contribute to welfare, and humans plausibly have access to more of these sources of welfare.

5. ^{^}

    Of course, the proper interpretation and implications of worldviews are matters of significant debate. For example, many Kantians do think that animals are moral patients. The worldviews in the tool are inspired by common moral theories, but we do not claim that they perfectly capture anyone’s views. Worldviews can be configured to better capture users’ judgments.