Nick, here’s an attempt at clarification on the point you’re questioning:

Imagine that you have two problems, Problem X and Problem Y. Problem X is much more tractable to us – we have a 10% chance of solving it if we try – but less impactful in itself (solving the problem would directly save 1000 lives). Problem Y is less tractable – we have an 0.1% chance of solving it if we try – but more impactful (solving the problem would directly save a million lives). Finally, imagine that each life saved represents a person who has a 1% chance of attempting to solve one of the two problems (conditional on the problem’s still being a problem) and similar odds to us of solving each problem conditional on attempting to solve it.

Looking only at direct effects, we’d estimate 100 expected-lives-saved for working on Problem X, vs. a thousand expected-lives-saved for working on Problem Y – working on Problem Y is better by a factor of 10. Incorporating indirect effects, however – the fact that solving Problem X would itself lead to an ~1% chance of solving Problem Y – working on Problem X is better by about a factor of 10.

This phenomenon becomes more pronounced if we start to incorporate the possibility that the people we’re empowering may be better positioned to solve other problems than we are.

While this is obviously a very simplified example, I think it illustrates that working on the problem you’re best positioned to solve can be better than working on another problem, no matter how much more important the other problem is, when flow-through effects are accounted for.

As far as how often this sort of thing is likely to obtain in practice, I agree with you that it depends on how broadly one defines “problems” and that it’s unlikely to obtain for the types of problems you listed, which are both narrow and (related to the fact that we’re interested in them) seem to pass a reasonable threshold of likely/intuitive tractability. As a side note, I think it is more likely to be an issue for “how should I use my talent?” than “how should I use my dollars?” because differences in tractability are so much more pronounced.

Toby, I agree that marginal value of wealth to utility is diminishing at an individual level, and I agree with not wanting to focus myopically on the role of financial wealth. I tried to make this post about the broad concept of “development and empowerment.”

I also agree that this post is far from a conclusive argument about the causal effects, or the magnitude of effects, of development. I definitely believe that $1 added to U.S. per-capita GDP does far less good than $1 added to a developing country’s per-capita GDP, which in turn does less good than $1 given to an effective developing-world health charity. This post should be read as an argument that the comparison is muddier than it appears at first glance, not that there are no such differences.

I don’t know that I agree that “Attempts to increase growth in poor countries have been the main failure story in the history of aid.” I’d say the evidence is inconclusive, taking into account that countries with more obstacles to growth tend to attract more aid. I certainly agree that the evidence for the success of health interventions is stronger than evidence for the success of other interventions. That’s a reason that I tend to prefer health interventions, especially since I think health improvements are roughly likely to flow through to greater prospects for economic development as income improvements are to flow through to greater health.

1) You focus on indirect financial effects rather than other indirect effects. This is probably a useful way to look at it, but note that your argument for the causal benefits of wealth creation is not that strong. Basically there is correlation between wealth and good things in people’s lives, but this doesn’t show the causality. I’m pretty sure there is two-way causality, but I believe this because of general life experience, not from this correlation information.

2) It is generally believed that there are very steeply decreasing marginal benefits of consumption (logarithmic, or more steeply diminishing than that). Thus even exponential economic boosts would be believed to have only linear benefits, which you will want to take into account.

See e.g.
http://www.economist.com/blogs/graphicdetail/2013/05/daily-chart-0

3) Attempts to increase growth in poor countries have been the main failure story in the history of aid, whereas attempts to increase health have been massively successful.

4) Even if wealth creation has caused much of the increase in wellbeing, if exogenous wealth creation would cause a similar increase in poor countries, and if this will also hold true beyond the current levels of wealth in rich countries, it still doesn’t follow that these effects will be large compared to the direct effects of interventions. They could be tiny. e.g. there is $70 tr world GDP, which implies a stock of value of something like $1,400 tr. Doubling this won’t produce close to double the current value due to the diminishing returns, so it seems prima facie like very large amounts of untargetted wealth creation are needed to improve wellbeing much, and that this may be very expensive compared to other ways of increasing wellbeing. I’m not sure if this is right, but I’m just trying to point out that showing economics flow-through effects are good is a long way short of showing that they are large relative to other effects.

This does seem to be conceivable to me, but I have somewhat different intuitions about how often this is likely to be a practical issue for causes that GiveWell and GiveWell supporters take seriously. I get the idea that I could be in one of the following situations:

(A) There is no RFMF for intervention X addressing problem Y. If I don?t fund it, Rich Person Z will. If I do fund it, Rich Person Z will instead spend the money on something worthless.

(B) Intervention X just doesn’t help address problem Y, or exacerbates it.

I agree that, in these cases, funding bednets is a better way to address problem Y than funding intervention X, even if problem Y is extremely narrow (in the sense that all small share of all efforts directly address problem Y). But it seems that, in practice, interventions are more likely to help address the problems they are directed at than they are to exacerbate them, and it can be hard to estimate RFMF dynamics. In light of this, it seems unlikely that you’ll find that funding bednets is a more effective way of addressing publication bias than, say, funding a random cluster of the proposals that people worried about the issue have thrown at you in conversations. And I have similar intuitions for asteroid tracking, geo-engineering research, pre-registration for RCTs, and disease surveillance.

I think that the more broadly you define problem Y, the more likely it is that the situation you describe will obtain. If you define Y as “global catastrophic risk,” I am skeptical but see where you are coming from. If you define Y as “the future doesn’t go as well as it could have,” I find it plausible that the situation you describe could obtain in practice, but that seems like a different issue from what you’re describing in the quoted passage.

I get the sense that we may not on the same page on this issue. If this is the case, I would appreciate it if you could explain your reasoning in greater depth.

One observation you seem to be offering as evidence for this claim is:
(2) Making progress on one problem often indirectly leads to progress on other problems.

It is not totally straightforward to me how (2) is supposed to support (1). It seems to me that, other things being equal, if solving problem X would directly benefit people alive today 10 times as much as solving problem Y, I should expect 10 times larger flow-through effects from solving problem X. So when I account for progress on other problems, it doesn’t necessarily change my bottom line on X vs. Y.

As you point out, certain types of people may be better-positioned to help others, and helping them may foreseeably lead to greater flow-through effects per unit of good accomplished directly (i.e., have a higher “flow-through ratio”). So if problem Y primarily affected people who were a stronger position to contribute to society or help others, addressing problem Y could be more promising than it appears at first. But this doesn’t seem to mean that we should give less weight to problem size, in comparison with tractability and room for more funding. Instead, it seems to mean that we need to give more weight to the flow-through ratio.

One way I might (very crudely) try to put it is that we need to move from the conceptual equation:
promisingness of cause/opportunity = tractability * RFMF * problem size * other stuff
to the equation:
promisingness of cause/opportunity = tractability * RFMF * problem size * flow-through ratio * other stuff
But that this doesn’t change how much weight we’re giving to problem size in comparison with tractability and RFMF.

I do agree with your point that flow-through effects are hard to predict, and that this increases the size of one’s Bayesian adjustment to cost-effectiveness estimates. This may be a way in which RFMF should get more weight, once one accounts for flow-through effects.

(See http://www.cato.org/publications/white-paper/drug-decriminalization-portugal-lessons-creating-fair-successful-drug-policies)