Teacher for 7 years; about to start working as a Researcher at CEARCH: https://exploratory-altruism.org/
I’ll be constructing cost-effectiveness analyses of various cause areas, identifying the most promising opportunities for impactful work.
Stan Pinsent
Karma: 283
I think Ghandi’s point nods to the British Empire’s policy of heavily taxing salt as a way of extracting wealth from the Indian population. For a time this meant that salt became very expensive for poor people and many probably died early deaths linked to lack of salt.
However, I don’t think anyone would suggest taxing salt at that level again! Like any food tax, the health benefits of a salt tax would have to be weighed against the costs of making food more expensive. You certainly wouldn’t want it so high that poor people don’t get enough of it.
Thanks again!
I think I have been trying to portray the point-estimate/interval-estimate trade-off as a difficult decision, but probably interval estimates are the obvious choice in most cases.
So I’ve re-done the “Should we always use interval estimates?” section to be less about pros/cons and more about exploring the importance of communicating uncertainty in your results. I have used the Ord example you mentioned.
Thanks for your feedback, Vasco. It’s led me to make extensive changes to the post:
More analysis on the pros/cons of modelling with distributions. I argue that sometimes it’s good that the crudeness of point-estimate work reflects the crudeness of the evidence available. Interval-estimate work is more honest about uncertainty, but runs the risk of encouraging overconfidence in the final distribution.
I include the lognormal mean in my analysis of means. You have convinced me that the sensitivity of lognormal means to heavy right tails is a strength, not a weakness! But the lognormal mean appears to be sensitive to the size of the confidence interval you use to calculate it—which means subjective methods are required to pick the size, introducing bias.
Overall I agree that interval estimation is better suited to the Drake equation than to GiveWell CEAs. But I’d summarise my reasons as follows:
The Drake Equation really seeks to ask “how likely is it that we have intelligent alien neighbours?”, but point-estimate methods answer the question “what is the expected number of intelligent alien neighbours?”. With such high variability the expected number is virtually useless, but the distribution of this number allows us to estimate the number of alien neighbours. GiveWell CEAs probably have much less variation and hence a point-estimate answer is relatively more useful
Reliable research on the numbers that go into the Drake equation often doesn’t exist, so it’s not too bad to “make up” interval estimates to go into it. We know much more about the charities GiveWell studies, so made-up distributions (even those informed by reliable point-estimates) are much less permissible.
Thanks again, and do let me know what you think!
My attempt to summarize why the model predicts that preventing famine in China and other countries will have a negative effect on the future:
Much of the value of the future hinges on whether values become predominantly democratic or antidemocratic
The more prevalent antidemocratic values (or populations) are after a global disaster, the likelier it is that such values will become predominant
Hence preventing deaths in antidemocratic countries can have a negative effect on the future.
Or as the author puts it in a discussion linked above:
To be blunt for the sake of transparency, in this model, the future would improve if the real GDP of China, Egypt, India, Iran, and Russia dropped to 0, as long as that did not significantly affect the level of democracy and real GDP of democratic countries. However, null real GDP would imply widespread starvation, which is obviously pretty bad! I am confused about this, because I also believe worse values are associated with a worse future. For example, they arguably lead to higher chances of global totalitarianism or great power war.
I agree with the author that the conclusion is confusing. Even concerning.
I’d suggest that the conclusion is out-of-sync with how most people feel about saving lives in poor, undemocratic countries. We typically don’t hesitate to tackle neglected tropical diseases on the basis that doing so boosts the populations of dictatorships.
Perhaps it can be captured by ensuring we compare counterfactual impacts.
For an urgent, “now or never” cause, we can be confident that any impact we make wouldn’t have happened otherwise.
For something non-urgent, there is a chance that if we leave it, somebody else could solve it or it could go away naturally. Hence we should discount the expected value of working on this (or in other words we should recognise that the counterfactual impact of working on non-urgent causes, which is what really matters, is lower than the apparent impact).
The s-risk people I’m familiar with are mostly interested in worst-case s-risk scenarios that involve vast populations of sentient beings over vast time periods. It’s hard to form estimates for the scale of such scenarios, and so the importance is difficult to grasp. I don’t think estimating the cost-effectiveness of working on these s-risks would be as simple as measuring in suffering-units instead of QALYs.
Tobias Baumann, for example, mentions in his book and a recent podcast that possibly the most important s-risk work we can do now is simply preparing to be ready in some future time when we will actually be able to do useful stuff. That includes things like “improving institutional decision-making” and probably also moral circle expansion work like curtailing factory farming.
I think Baumann also said somewhere that he can be reluctant to mention specific scenarios too much because it may lead to complacent feeling that we have dealt with the threats: in reality, the greatest s-risk danger is probably something we don’t even know about yet.
I hope the above is a fair representation of Baumann’s and others’ views. I mostly agree with them, although it is a bit shady not to be able to specify what the greatest concerns are.
I could do a very basic cause-area sense-check of the form:
The greatest s-risks involve huge populations
SO
They probably occur in an interstellar civilisation
AND
Are likely to involve artificial minds (which could probably exist at a far greater density than people)
HENCE
Work on avoiding the worst s-risks is likely to involve influencing whether/how we become a spacefaring civilisation and whether/how we develop and use sentient minds.
Thanks. I’ll read up on the power law dist and at the very least put a disclaimer in: I’m only checking which is better out of normal/lognormal.
Thanks for the reply and sorry for the long delay! I decided to dive in and write a post about it.
I check when using distributions is much better than point-estimates: it’s when the ratio between upper/lower confidence bounds is high—in situations of high uncertainty like the probability-of-life example you mentioned.
I test your intuition that using lognormal is usually better than normal (and end up agreeing with you)
I check whether the lognormal distribution can be used to find a more reliable mean of two point estimates, but conclude that it’s no good
Making better estimates with scarce information
Great summary. You must have an incredibly organised system for keeping track of your reading and what you take from each post!
I suspect this has given me most of the benefit of hours of unguided reading at a fraction of the time cost.
Well said. I share @Henry Howard ’s reservations about WELLBYs, but I would argue that even if WELLBY comparisons are near-meaningless between New Yorkers and Sentinelese, they are probably much more meaningful when comparing one individual’s wellbeing before and after treatment, or even comparing control and intervention groups drawn randomly from the same population.
I think it’s a great idea. My intuition is that you ought to exaggerate what makes your work different from the existing EA canon. For example, you might want to be much more accessible than the works put out by moral philosophers. To this end, I suggest partnering with someone with a track record publishing pop-science, self-help or similar.
It doesn’t mean you have to water down EA ideas. But it would probably mean distilling the essence of EA thought into a few clear principles, which can then use to illustrate why EA leads to various conclusions.
For example (off the top of my head) your principles might be:
The outcomes are what matters (consequentialism)
Do your best with the information available (Bayesian thinking)
Then, in your chapter on personal consumption choices, you can show why (to borrow Geoffrey Miller’s example) transitioning from chicken to grass-fed beef, with an offset donation to Vegan Outreach (a bewildering choice to most people) stems from the principles.
In short, you should aim to be accessible, but not one of those books that you have to flick back through to find the answers to each of life’s questions. Readers should be left with a clear and lasting grounding in the basics.
While I’m writing about setting yourself apart from the EA canon, I feel I should point out the obvious—women, especially mothers, are not well-represented among EA authors. If you can find some way to productively collaborate with Julia, you should.
On purely pragmatic grounds, having a female name on the cover will affect the readership
Poor representation is plausibly linked to increased attrition (I see the higher rates of attrition among women studying math as an example of this)
Parenthood, which is essentially a lifelong commitment to favour your child over others, urgently needs discussing in the context of EA, especially by those who are both EA and parents.
This is great! Your super-simple code is helping me learn the basics of Squiggle. Thanks for including it.
I currently use point estimates for cost-effectiveness analyses but you are convincing me that using distributions would be better. A couple of thoughts/questions:
When is a point-estimate approach much worse than a distribution approach? For example, what is it about the Fermi paradox that leads to such differing results between the methods?
If I suspect a lognormal distribution, is there an easy way to turn discrete estimates into a aggregate estimate? Suppose I want to estimate how many eggs my pet toad will lay. One source suggests it’s 50, another says 5000. The lognormal seems like the right model here, but it takes me several minutes with pen, paper and calculator to estimate the mean of the underlying lognormal distribution. Ideally I’d like a ‘mean’ that approximates this for me, but neither arithmetic, geometric nor harmonic means seem appropriate.
I welcome help from anyone!!
Great post, thanks for sharing. To provide a bit of context could you sketch out what an increased emphasis on long-term brand marketing might look like for EA orgs?
For example, what might it look like for:
A research org looking to get more from small donations in response to the withdrawal of a major donor
A startup charity embarking on a 5-year expansion plan in which talent acquisition will be a major bottleneck
A “shopfront” EA org that counts inducting new people to EA ideas among its goals.
“Public Health England has estimated that 85% of the salt people ingest is already in food at the point of purchase and consumers only add the other 15% during cooking or at the table.”
Anecdotally, people add less salt to the food they cook themselves than is added to processed foods in the factory. If this is true, even if people do ‘top up’ salt levels in their food, they will likely still end up with less salt.
Plus there’s salt in all sorts of crazy things like breakfast cereal. If that gets reduced, I don’t think people are going to start salting their fruit loops.
I’d love to do that! Just sent you a message.
Its embarassing because my methods are so low-tech. I literally went to the Twitter feeds of each org and copied the 10 most recent original tweets into a spreadsheet. Hence the call for someone with data-scraping skills at the end.
My background is: I’m a teacher, run a podcast on Education, looking to transition to a more impactful career, have a degree in Mathematics. Any research skills I have are self-taught.
I don’t have a specific vision for a project at the moment but would very much be interested in doing a larger study on EA and Twitter. I’d be interested in projects that i) allow EAs and orgs to better form social media strategies and ii) build my research skills & credibility.
Thanks for pointing that out. I also get a sense that people are getting more traction than brands in EA twitter. Before my initial social media study I followed very few EA orgs on Twitter and I wasn’t getting exposed to new ones, whereas prominent individuals were popping up.
Perhaps orgs should be trying that out. I expect some friction—people who already use social media in a personal capacity may want to keep it separate from their job, while others are consciously off the platforms and actively against spending time on them. Maybe people could just get secondary, job-aligned accounts.
I skimmed the piece on axiological asymmetries that you linked and am quite puzzled that you seem to start with the assumption of symmetry and look for evidence against it. I would expect asymmetry to be the more intuitive, therefore default, position. As the piece says
At just the first-order level, people tend to assume that (the worst) pain is worse than (the best) pleasure is pleasurable. The agonizing ends for non-human animals in factory farms and in the wild seem far worse than the best sort of life they could realize would be good. [...] it’s hard to find any organisms that risk the worst pains for the greatest pleasures and vice versa.
I would expect that a difference in magnitude between the best pleasure and worst possible is the most obvious explanation, but the piece concludes that these judgments are “far more plausibly explained by various cognitive biases”.
As far as I can tell this would suggest that either:
Someone who has recently experienced or is currently experiencing intense suffering (and therefore has a better understanding of the stakes) would be more willing to take the kind of roulette gamble described in the piece. This seems unlikely.
People’s assessments of hedonic states are deeply unreliable even if they have recent experience of the states in question. I don’t like this much because it means we have to fall back on physiological evidence for human pleasure/suffering, which, as shown by the mayonnaise example, can’t give us the full picture.
On a slightly separate note, I played around with the BOTEC to check the claim that assuming symmetry doesn’t change the numbers much and I was convinced. The extreme suffering-focused assumption (where perfect health is merely neutral) resulted in double the welfare gain of the symmetric assumption (when the increase in welfare as a percentage of the animals’ negative welfare range is held constant).
My main question on this last point is: why use “percentage of the animals’ negative welfare range” when “percentage of the animals’ total welfare range” seems more relevant and would not vary at all across different (a)symmetry assumptions?
Can we see it yet?