Thanks for this post, Phil! I work on some related issues at Open Philanthropy, and we’re always grateful for thoughtful engagement like this. What follows here are my own views, not necessarily Open Philanthropy’s.
Overall, I agree with your skepticism of arguments that use simple models to make the case for levered investments. But I disagree that we should model our philanthropic opportunity set as having steeper-than-log diminishing returns, for 2 reasons: one is about aggregating across sub-causes, and another is about the timeline over which we should think about fluctuations in non-EA spending.
I made similar comments on your draft of this piece, and you adapted the writeup in a very reasonable way. But I think there’s still daylight between your model and mine, which is why I’m repeating myself here. Thanks for the opportunity to comment on the draft, and for considering these critiques as carefully as you have.
Aggregating across sub-causes
As you note in appendix C, the spending/impact curvature of the overall philanthropic utility function can be much flatter than the curvature of any specific cause. (I can’t tell how much that consideration flows through to your overall recommendations. I think it should affect your bottom line a lot.) It’s important to note that curvature gets flatter as you aggregate across sub-causes to the cause level, and not just across causes to the level of the philanthropic utility function – and sub-causes can be a lot narrower than you might expect.
For instance, imagine that GiveDirectly recipient households each have a money/utility curvature parameter (“eta”) of 1, but they vary in their baseline level of consumption, and in the costs of getting money to them. Further, assume GiveDirectly is good at targeting its dollars to most efficiently generate utility. So the first million goes to a very poor region with low transaction costs, then the next million is spread across (a) the previously mentioned region plus (b) the next-poorest region with low transactions costs. And so on. If you model this out, you’ll find that the curvature of GiveDirectly’s overall opportunity set can have an eta of much less than 1, depending on your assumptions.
The same should be true of other causes. GiveWell’s malaria funding targets the most cost-effective regions first, then their second million is spread across (a) improving administration/frequency/etc in the same community as the first million, and also (b) expanding to slightly less cost-effective regions. And they face an inefficient baseline, since (a) countries vary in their level of development and in how well they provide basic healthcare, and (b) other funders (like the Global Fund) don’t allocate their money efficiently due to their own constraints. So whatever the spending-impact curvature might be of fighting malaria in one community (and for the record, I don’t see why it should be anchored to 1.5 or even 1), you should expect the overall malaria opportunity set to have flatter curvature than that. (And of course you should expect the overall philanthropic utility function to be even flatter, since it aggregates across causes.)
Time horizon for changes in non-EA spending
I think you’re right to note the distinction between an “all else equal” philanthropic utility function, which holds constant some assumed level of spending by other actors, and an adjusted philanthropic utility function which considers how surprises in others’ spending are correlated with surprises in the resources available to us. I might call these the “conditional” vs “marginal” utility functions, analogous to how we can refer to “conditional” vs “marginal” probability distributions.
However, I don’t think this adjustment will make a huge difference, since a lot of the non-EA spending doesn’t seem especially exposed to asset-market fluctuations in a year-to-year way. I think you might be right over long timelines — for any assets that are intended to fund philanthropy many decades from now, this adjustment could make an important difference to investment decisions. But I think that should be a small share of EA assets.
Here are some non-EA resources that go toward meeting the needs of the rural poor in LMICs: LMIC government spending; foreign aid; the output of subsistence agriculture. (Subsistence agriculture is ~50% of calories in rural Africa.) I don’t think any of those are nearly as exposed to yearly fluctuations in global asset markets as a 60⁄40 portfolio would be. If this non-EA set of resources is taking less risk than is justified by the curvature of the utility function then shouldn’t EAs take more risk to balance that out? (I think the answer to that question is no, because I don’t trust models like these to advise us on how much risk to take.)
Over a timeline of decades rather than years, I agree that the non-EA resources available to the global poor are probably correlated with EA assets. See e.g. Table 6 of Ravallion and Chen 1997. (Though note that individual LMICs will experience growth rates that aren’t perfectly correlated with EA assets – for example, US equity markets did better in the 80s and 90s than 70s and 2000s, but LMIC growth was higher in the 70s and 2000s. Also note that EA’s philanthropic opportunities aren’t perfectly correlated with LMIC poverty – see for example the 2019 paper “Most of Africa’s nutritionally deprived women and children are not found in poor households”, not to mention farm animal welfare or longtermism.) But I don’t think that fact is very action-relevant to us, since most of EA’s current assets are targeted toward spending in the next few decades.
To see this, consider: If you planned for your endowment to stick around perpetually, you could spend at the pace of asset growth and keep a constant level of assets. For example, if you expected 7% asset returns, you’d spend roughly 7% annually. That means you’d expect that by 2043 you’ve spent 75% of your 2023 assets, but the remaining 25% have grown enough to keep your overall asset value constant. (25% * 1.07^20 ≈ 100%)
So from the perspective of 2023, maybe 25% of your assets are “allocated” to spending that is multiple decades from now. Therefore, maybe 25% of your assets should follow your advice here, and pursue something like a 60⁄40 portfolio. That’s if you plan for your endowment to exist perpetually (in contrast, Cari Tuna and Dustin Moskovitz want to spend down their assets within their lifetimes), and if you buy the rest of your arguments about e.g. utility curvatures.
Bottom line
Overall, I end up skeptical of claims that we should act as if our philanthropic opportunity set has steeper-than-logarithmic diminishing marginal returns, for both empirical and theoretical reasons – though I agree with your general conservatism about investment risk, and your deference to the wisdom of longstanding practice in portfolio management.
On empirical grounds: OP has tried to estimate empirically the spending/impact curvature of a big philanthropic opportunity set – the GiveWell top charities – and ended up with an eta parameter of roughly 0.38. (I.e. each 10% increase in annual spending means marginal cost-effectiveness decreases by 3.8%.) Your main critique of that estimate seems to be that it’s “conditional” rather than “marginal” as defined above – but I think it’s very unlikely that your proposed adjustment will bring it from 0.38 to something above 1, for the reasons I gave in the “time horizon” section of this comment.
And then on theoretical grounds: even if each household’s utility function has steeper-than-log diminishing returns, by the time you aggregate across households to the sub-cause level, and then aggregate across sub-causes to the cause level, and then aggregate across causes to the level of the overall philanthropic opportunity set, you’ll end up with a much shallower opportunity set in aggregate; as I argued in the “aggregating across sub-causes” section of this comment.
I think the answer to that question is no, because I don’t trust models like these to advise us on how much risk to take.
How would you prefer to decide how much risk to take?
OP has tried to estimate empirically the spending/impact curvature of a big philanthropic opportunity set – the GiveWell top charities – and ended up with an eta parameter of roughly 0.38.
I would love to see more on this if there’s any public writeups or data.
Hi Peter, thanks again for your comments on the draft! I think it improved it a lot. And sorry for the late reply here—just got back from vacation.
I agree that the cause variety point includes what you might call “sub-cause variety” (indeed, I changed the title of that bit from “cause area variety” to “cause variety” for that reason). I also agree that it’s a really substantial consideration: one of several that can single-handedly swing the conclusion. I hope you/others find the simple model of Appendix C helpful for starting to quantify just how substantial it is. My own current feeing is that it’s more substantial than I thought when I first started thinking about this question, though not enough to unambiguously outweigh countervailing considerations, like the seemingly unusually high beta of EA-style philanthropic funding.
I also agree that the long-run correlations between asset returns and the consumption of the global poor seems like an important variable to look into more insofar as we’re thinking about the global poverty context, and that it could turn out to be weak enough that using an effective eta<1 is warranted even if we’re operating on a long time horizon.
Thanks for this post, Phil! I work on some related issues at Open Philanthropy, and we’re always grateful for thoughtful engagement like this. What follows here are my own views, not necessarily Open Philanthropy’s.
Overall, I agree with your skepticism of arguments that use simple models to make the case for levered investments. But I disagree that we should model our philanthropic opportunity set as having steeper-than-log diminishing returns, for 2 reasons: one is about aggregating across sub-causes, and another is about the timeline over which we should think about fluctuations in non-EA spending.
I made similar comments on your draft of this piece, and you adapted the writeup in a very reasonable way. But I think there’s still daylight between your model and mine, which is why I’m repeating myself here. Thanks for the opportunity to comment on the draft, and for considering these critiques as carefully as you have.
Aggregating across sub-causes
As you note in appendix C, the spending/impact curvature of the overall philanthropic utility function can be much flatter than the curvature of any specific cause. (I can’t tell how much that consideration flows through to your overall recommendations. I think it should affect your bottom line a lot.) It’s important to note that curvature gets flatter as you aggregate across sub-causes to the cause level, and not just across causes to the level of the philanthropic utility function – and sub-causes can be a lot narrower than you might expect.
For instance, imagine that GiveDirectly recipient households each have a money/utility curvature parameter (“eta”) of 1, but they vary in their baseline level of consumption, and in the costs of getting money to them. Further, assume GiveDirectly is good at targeting its dollars to most efficiently generate utility. So the first million goes to a very poor region with low transaction costs, then the next million is spread across (a) the previously mentioned region plus (b) the next-poorest region with low transactions costs. And so on. If you model this out, you’ll find that the curvature of GiveDirectly’s overall opportunity set can have an eta of much less than 1, depending on your assumptions.
The same should be true of other causes. GiveWell’s malaria funding targets the most cost-effective regions first, then their second million is spread across (a) improving administration/frequency/etc in the same community as the first million, and also (b) expanding to slightly less cost-effective regions. And they face an inefficient baseline, since (a) countries vary in their level of development and in how well they provide basic healthcare, and (b) other funders (like the Global Fund) don’t allocate their money efficiently due to their own constraints. So whatever the spending-impact curvature might be of fighting malaria in one community (and for the record, I don’t see why it should be anchored to 1.5 or even 1), you should expect the overall malaria opportunity set to have flatter curvature than that. (And of course you should expect the overall philanthropic utility function to be even flatter, since it aggregates across causes.)
Time horizon for changes in non-EA spending
I think you’re right to note the distinction between an “all else equal” philanthropic utility function, which holds constant some assumed level of spending by other actors, and an adjusted philanthropic utility function which considers how surprises in others’ spending are correlated with surprises in the resources available to us. I might call these the “conditional” vs “marginal” utility functions, analogous to how we can refer to “conditional” vs “marginal” probability distributions.
However, I don’t think this adjustment will make a huge difference, since a lot of the non-EA spending doesn’t seem especially exposed to asset-market fluctuations in a year-to-year way. I think you might be right over long timelines — for any assets that are intended to fund philanthropy many decades from now, this adjustment could make an important difference to investment decisions. But I think that should be a small share of EA assets.
Here are some non-EA resources that go toward meeting the needs of the rural poor in LMICs: LMIC government spending; foreign aid; the output of subsistence agriculture. (Subsistence agriculture is ~50% of calories in rural Africa.) I don’t think any of those are nearly as exposed to yearly fluctuations in global asset markets as a 60⁄40 portfolio would be. If this non-EA set of resources is taking less risk than is justified by the curvature of the utility function then shouldn’t EAs take more risk to balance that out? (I think the answer to that question is no, because I don’t trust models like these to advise us on how much risk to take.)
Over a timeline of decades rather than years, I agree that the non-EA resources available to the global poor are probably correlated with EA assets. See e.g. Table 6 of Ravallion and Chen 1997. (Though note that individual LMICs will experience growth rates that aren’t perfectly correlated with EA assets – for example, US equity markets did better in the 80s and 90s than 70s and 2000s, but LMIC growth was higher in the 70s and 2000s. Also note that EA’s philanthropic opportunities aren’t perfectly correlated with LMIC poverty – see for example the 2019 paper “Most of Africa’s nutritionally deprived women and children are not found in poor households”, not to mention farm animal welfare or longtermism.) But I don’t think that fact is very action-relevant to us, since most of EA’s current assets are targeted toward spending in the next few decades.
To see this, consider: If you planned for your endowment to stick around perpetually, you could spend at the pace of asset growth and keep a constant level of assets. For example, if you expected 7% asset returns, you’d spend roughly 7% annually. That means you’d expect that by 2043 you’ve spent 75% of your 2023 assets, but the remaining 25% have grown enough to keep your overall asset value constant. (25% * 1.07^20 ≈ 100%)
So from the perspective of 2023, maybe 25% of your assets are “allocated” to spending that is multiple decades from now. Therefore, maybe 25% of your assets should follow your advice here, and pursue something like a 60⁄40 portfolio. That’s if you plan for your endowment to exist perpetually (in contrast, Cari Tuna and Dustin Moskovitz want to spend down their assets within their lifetimes), and if you buy the rest of your arguments about e.g. utility curvatures.
Bottom line
Overall, I end up skeptical of claims that we should act as if our philanthropic opportunity set has steeper-than-logarithmic diminishing marginal returns, for both empirical and theoretical reasons – though I agree with your general conservatism about investment risk, and your deference to the wisdom of longstanding practice in portfolio management.
On empirical grounds: OP has tried to estimate empirically the spending/impact curvature of a big philanthropic opportunity set – the GiveWell top charities – and ended up with an eta parameter of roughly 0.38. (I.e. each 10% increase in annual spending means marginal cost-effectiveness decreases by 3.8%.) Your main critique of that estimate seems to be that it’s “conditional” rather than “marginal” as defined above – but I think it’s very unlikely that your proposed adjustment will bring it from 0.38 to something above 1, for the reasons I gave in the “time horizon” section of this comment.
And then on theoretical grounds: even if each household’s utility function has steeper-than-log diminishing returns, by the time you aggregate across households to the sub-cause level, and then aggregate across sub-causes to the cause level, and then aggregate across causes to the level of the overall philanthropic opportunity set, you’ll end up with a much shallower opportunity set in aggregate; as I argued in the “aggregating across sub-causes” section of this comment.
How would you prefer to decide how much risk to take?
I would love to see more on this if there’s any public writeups or data.
Hi Peter, thanks again for your comments on the draft! I think it improved it a lot. And sorry for the late reply here—just got back from vacation.
I agree that the cause variety point includes what you might call “sub-cause variety” (indeed, I changed the title of that bit from “cause area variety” to “cause variety” for that reason). I also agree that it’s a really substantial consideration: one of several that can single-handedly swing the conclusion. I hope you/others find the simple model of Appendix C helpful for starting to quantify just how substantial it is. My own current feeing is that it’s more substantial than I thought when I first started thinking about this question, though not enough to unambiguously outweigh countervailing considerations, like the seemingly unusually high beta of EA-style philanthropic funding.
I also agree that the long-run correlations between asset returns and the consumption of the global poor seems like an important variable to look into more insofar as we’re thinking about the global poverty context, and that it could turn out to be weak enough that using an effective eta<1 is warranted even if we’re operating on a long time horizon.