Right, I (mis?)took the OP to be arguing “reducing salaries wouldn’t have an effect on labour supply, because it is price inelastic”, instead of “reducing salaries wouldn’t have enough of an effect to qualitatively change oversupply.
I’d expect a reduction but not a drastic one. Like I’d predict Open Phil’s applicant pool to drop to 500-600 from 800 if they cut starting salary by $10k-$15k.
This roughly cashes out to an income elasticity of labour (/applicant) supply of 1-2 (i.e. you reduce applicant supply by ~20% by reducing income ~~10%). Although a crisp comparison is hard to find, in the labour market you see figures generally <1, so this expectation slightly goes against the OP, given it suggests EA applicants are more compensation sensitive than typical.
(Obvious CoI/own views, but in my defence I’ve been arguing along these lines long before I had—or expected to have—an EA job.)
I agree ‘EA jobs’ provide substantial non monetary goods, and that ‘supply’ of willing applicants will likely outstrip available positions in ‘EA jobs’. Yet that doesn’t mean ‘supply’ of potential EA employees is (mostly) inelastic to compensation.
In principle, money is handy to all manner of interests one may have, including altruistic ones. Insofar as folks are not purely motivated by altruistic ends (and in such a way they’re indifferent to having more money to give away themselves) you’d expect them to be salary-sensitive. I aver basically everyone in EA is therefore (substantially) salary-sensitive.
In practice, I know of cases (including myself) where compensation played a role in deciding to change job, quit, not apply etc. I also recall on the forum remarks from people running orgs which cannot compensate as generously as others that this hurts recruitment.
So I’m pretty sure if you dropped salaries you would reduce the number of eager applicants (albeit perhaps with greater inelasticity than many other industries). As (I think) you imply, this would be a bad idea: from point of view of an org, controlling overall ‘supply’ of applicants shouldn’t be their priority (rather they set salaries as necessary to attract the most cost effective employees). For the wider community point of view, you’d want to avoid ‘EA underemployment’ in other ways than pushing to distort the labour market.
The inconvenience I had in mind is not in your list, and comprises things in the area of, “Prefer to keep the diet I’m already accustomed to”, “Prefer omnivorous diets on taste etc. grounds to vegan ones”, and so on. I was thinking of an EA who is omnivorous and feels little/no compunction about eating meat (either because they aren’t ‘on board’ with the moral motivation for animal causes in general, or doesn’t find the arguments for veganism persuasive in particular). I think switching to a vegan diet isn’t best described as a minor inconvenience for people like these.
But to be clear, this doesn’t entail any moral obligation whatsoever on the hotel to serve meat—it’s not like they are forcing omnivorous guests to be vegan, but just not cooking them free (non-vegan) food. If a vegan offers me to stay at their house a) for free, b) offers vegan food for free too, c) welcomes me to, if I’m not a fan of vegan food, get my own food to cook at their house whenever I like—which seems basically the counterfactual scenario if I wasn’t staying with them in the first place, and d) explains all of this before I come, they’ve been supererogatory in accommodating me, and it would be absurd for me to say they’ve fallen short in not serving me free omnivorous food which they morally object to.
Yet insofar as ‘free food’ is a selling point of the hotel, ‘free vegan food’ may not be so enticing to omnivorous guests. Obviously the offer is still generous by itself, leave alone combined with free accommodation, but one could imagine it making a difference on the margin to omnivores (especially if they are cost-sensitive).
Thus there’s a trade-off in between these people and vegans who would be put off if the hotel served meat itself (even if vegan options were also provided). It’s plausible to me the best option to pick here (leave alone any other considerations) is the more ‘vegan-friendly’ policy. But this isn’t because the trade-off is in fact illusory because the ‘vegan-friendly’ policy is has minimal/minor costs to omnivores after all.
[Empirically though, this doesn’t seem to amount to all that much given (I understand) the hotel hasn’t been struggling for guests.]
Beyond the ‘silent downvote → anon feedback’ substitution (good, even if ‘public comment’ is even better) substitution, there could also be a ‘public comment --> anon feedback’ one (less good).
That said, I’m in favour of an anon feedback option: I see karma mostly serving as a barometer of community sentiment (so I’m chary of disincentivizing downvotes as this probably impairs resolution). It isn’t a good way of providing feedback to the author (a vote is only a bit or two of information). Text is better—although for me, the main reasons I don’t ‘explain my downvotes’ are mostly time, but occasionally social considerations. An anon option at least removes the latter disincentive.
I think I get the idea:
Suppose (heaven forbid) a close relative has cancer, and there’s a new therapy which fractionally improves survival. The NHS doesn’t provide it on cost-effectiveness grounds. If you look around and see the NHS often provides treatment it previously ruled out if enough public sympathy can be aroused, you might be inclined try to do the same. If instead you see it is pretty steadfast (“We base our allocation on ethical principles, and only change this when we find we’ve made a mistake in applying them”), you might not be—or at least change your strategy to show the decision the NHS has made for your relative is unjust rather than unpopular.
None of this requires you to be acting in bad faith looking for ways of extorting the government—you’re just trying to do everything you can for a loved one (the motivation for pharmaceutical companies that sponsor patient advocacy groups may be less unalloyed). Yet (ideally) the government wants to encourage protest that highlights a policy mistake, and discourage those for when it has done the right thing for its population, but is against the interests of a powerful/photogenic/popular constituency. ‘Caving in’ to the latter type pushes in the wrong direction.
(That said, back in EA-land, I think a lot things that are ‘PR risks’ for EA look bad because they are bad (e.g. in fact mistaken, morally abhorrent, etc.), and so although PR considerations aren’t sufficient to want to discourage something, they can further augment concern.)
Related: David Manheim’s writing on network theory and scaling organisations.
A bit of both:
I’d like to see more forecasting skills/literacy ‘in the water’ of the EA community, in the same way statistical literacy is commonplace. A lot of EA is about making the world go better, and so a lot of (implicit) forecasting is done when deciding what to do. I’d generally recommend most people consider things like opening a Metaculus account, reading superforecasting, etc.
This doesn’t mean everyone should be spending (e.g.) 3 hours a day on this, given the usual story about opportunity costs. But I think (per the question topic) there’s also a benefit of a few people highly developing this skill (again, a bit like stats: it’s generally harder to design and conduct statistical analysis than to critique one already done, but you’d want some folks in EA who can do the former).
This is more a ‘skill I’d like to see more of in the EA community’, rather than a career track. It seems a generally valuable skill set for a lot of EA work, and having some people develop expertise/very high performance in it (e.g. becoming a superforecaster) looks beneficial to me.
[Not one of the downvoters]
The leading rationale of “Learn a trade --> use it for EA projects that need it” looks weak to me:
There’s not a large enough density of ‘EA’ work in any given place to take up more than a small fraction of a tradepersons activity. So this upside should be discounted by (substantial) time to learn the trade, and then most of one’s ‘full time job’ as (say) an electrician will not be spent on EA work.
It looks pretty unlikely to have ‘nomadic’ tradespeople travelling between EA hubs, as the added cost of flights etc. suggest it might be more efficient just to try and secure good tradespeople by (e.g.) offering above market rates.
As you say, it could be a good option for some due to good earning power (especially for those with less academic backgrounds, cf. kbog’s guide) but the leading rationale doesn’t seem substantial reason to slant recommendations (e.g. if you could earn X as a plumber, but 1.1X in something else, the fact they could occasionally help out for EA projects shouldn’t outweigh this.
[I didn’t downvote.] I fear the story is that this is something of a ‘hot button’ issue, and people in either ‘camp’ have sensitivities about publicly speaking out on one side or the other for fear of how others in the opposing ‘camp’ may react towards them. (The authors of this document are anonymous; previous conversations on this area in this forum have had detractors also use anon accounts or make remarks along the lines of, ‘I strongly disagree with this, but I don’t want to elaborate further’). Hence why people who might be opposed to this (for whatever reason) preferring anonymous (albeit less-informative) feedback via downvoting.
There are naturally less charitable explanations along the lines of tribalism, brigadeing, etc. etc.
Thanks for your reply, and the tweaks to the post. However:
[I] decided to keep the discussion short because the regression seemed to offer very limited practical significance (as you pointed out). Had I decided to give it more weight in my analysis then it certainly would be appropriate to offer a fuller explanation. Nonetheless, I should have been clearer about the limited usefulness of the regression, and noted it as the reason for the short discussion.
I think the regression having little practical significance makes it the most useful part of the analysis: it illustrates the variation in the dependent variable is poorly explained by all/any of the variables investigated, that many of the associations found by bivariate assessment vanish when controlling for others, and gives better estimates of the effect size (and thus relative importance) of those which still exert statistical effect. Noting, essentially, “But the regression analyses implies a lot of the associations we previously noted are either confounded or trivial, and even when we take all the variables together we can’t predict welcomeness much better than taking the average” at the end buries the lede.
A worked example. The summary notes, “EAs in local groups, in particular, view the movement as more welcoming than those not in local groups” (my emphasis). If you look at the t-test between members and nonmembers there’s a difference of ~ 0.25 ‘likert levels’, which is one of the larger effect sizes reported.
Yet we presumably care about how much of this difference can be attributed to local groups. If the story is “EAs in local groups find EA more welcoming because they skew (say) male and young”, it seems better to focus attention on these things instead. Regression isn’t a magic wand to remove confounding (cf.), but it tends to be better than not doing it at all (which is essentially what is being done when you test association between a single variable and the outcome).
As I noted before, the ‘effect size’ of local group membership when controlling for other variables is still statistically significant, but utterly trivial. Again: it is ~ 1/1000th of a likert level; the upper bound of the 95% confidence interval would only be ~ 2/1000th of a likert level. By comparison, the effect of gender or year of involvement are two orders of magnitude greater. It seems better in the conclusion to highlight results like these, rather than results the analysis demonstrates have no meaningful effect when other variables are controlled for.
A few more minor things:
(Which I forgot earlier). If you are willing to use means, you probably can use standard errors/confidence intervals, which may help in the ‘this group looks different, but small group size’ points.
Bonferroni makes a rod for your back given it is conservative (cf.); an alternative approach is false discovery rate control instead of family wise error rate control. Although minor, if you are going to use this to get your adjusted significance threshold, this should be mentioned early, and the result which ‘doesn’t make the cut’ should be simply be reported as non-significant.
It is generally a bad idea to lump categories together (e.g. countries, cause areas) for regression as this loses information (and so statistical power). One of the challenges of regression analysis is garden of forking path issues (even post-hoc—some coefficients ‘pop into’ and out of statistical significance depending on which model is used, and once I’ve seen one, I’m not sure how much to discount subsequent ones). It is here where an analysis plan which pre-specifies this is very valuable.
FWIW: I think I know of another example along these lines, although only second hand.
Thanks for this—the presentation of results is admirably clear. Yet I have two worries:
1) Statistics: I think the statistical methods are frequently missing the mark. Sometimes this is a minor quibble; other times more substantial:
a) The dependent variable (welcomeness—assessed by typical Likert scale) is ordinal data i.e. ‘very welcoming’ > welcoming > neither etc). The write-up often treats this statistically either as categorical data (e.g. chi2) or interval data (e.g. t-test, the use of ‘mean welcomeness’ throughout). Doing the latter is generally fine (the data looks pretty well-behaved, t-tests are pretty robust, and I recall controversy about when to use non-parametric tests). Doing the former isn’t.
chi2 tests against the null of (in essence) the proportion in each ‘row’ of a table is the same between columns: it treats the ordered scale as a set of 5 categories (e.g. like countries, ethnicities, etc.). Statistical significance for this is not specific for ‘more or less welcoming’: two groups with identical ‘mean welcomeness’ yet with a different distribution across levels could ‘pass statistical significance’ by chi2. Tests for ‘ranked dependent by categorical independent’ data exist (e.g. Kruskall-Wallis) and should be used instead.
Further, chi2 assumes the independent variable is categorical too. Usually it is (e.g. where you heard about EA) but sometimes it isn’t (e.g. age, year of joining, ?political views). For similar reasons to the above, a significant chi2 result doesn’t demonstrate a (monotonic) relationship between welcomeness and time in EA. There are statistical tests for trend which can be used instead.
Still further, chi2 (ditto K-W) is an ‘omnibus’ test: it tells you your data is surprising given the null, but not what is driving the surprise. Thus statistical significance ‘on the test’ doesn’t indicate whether particular differences (whether highlighted in the write-up or otherwise) are statistically significant.
b) The write-up also seems to be switching between the descriptive and the inferential in an unclear way. Some remarks on the data are accompanied with statistical tests (implying an inference from the sample to the population), whilst similar remarks are not: compare the section on ‘time joining EA’ (where there are a couple of tests to support a ‘longer in EA—finding it more welcoming’), versus age (which notes a variety of differences between age groups, but no statistical tests).
My impression is the better course is the former, and so differences being highlighted to the readers interest should be accompanied by whether these differences are statistically significant. This uniform approach also avoids ‘garden of forking path’ worries (e.g. ‘Did you not report p values for the age section because you didn’t test, or because they weren’t significant?’)
c) The ordered regression is comfortably the ‘highest yield’ bit of statistics performed, as it is appropriate to the data, often more sensitive (e.g. lumping the data into two groups by time in EA and t-testing is inferior technique to regression), and helps answer questions of confounding sometimes alluded to in the text (“Welcoming seems to go up with X, but down with Y, which is weird because X and Y correlate”), but uniformly important (“People in local groups find EA more welcoming—but could that driven by other variables between those within and without local groups?“)
It deserves a much fuller explanation (e.g. how did ‘country’ and ‘top priority cause’ become single variables with a single regression coefficient—is the ‘lumping together’ implied in the text post-hoc? How was variable selection/model choice decided? Model 1 lacks only ‘top priority cause’, so assumedly ‘adding in political spectrum didn’t improve explanatory power’ is a typo?). When its results vary with the univarible analysis, I would prefer the former over the latter. That fb membership, career shifting (in model 2), career type, and politics aren’t significant predictors means their relationship to welcomingness, if, even if statistically significant, probably confounding rather than true association.
It is unfortunate some of these are highlighted in the summary and conclusion, even more so when a crucial negative result from the regression is relatively unsung. The ~3% R^2 and very small coefficients (with the arguable exception of sex) implies very limited practical significance: almost all the variation in whether an EA finds EA welcoming or not is not predicted by the factors investigated; although EAs in local groups find EA more welcoming, this effect—albeit statistically significant—is (if I interpret the regression right) around 0.1% of a single likert level.
2) Selection bias: A perennial challenge to the survey is issues of selection bias. Although happily noted frequently in discussion, I still feel it is underweighed: I think it is huge enough to make the results all but uninterpretable.
Facially, one would expect those who find EA less welcoming are less likely to join. We probably wouldn’t think that how welcoming people already in EA think it is would be informative to how good it is at welcoming people into EA (caricatured example: I wouldn’t be that surprised if members of something like the KKK found it generally welcoming). As mentioned in the ‘politics’ section, the relative population size seems a far better metric (although the baseline hard to establish) to which welcomingness adds very little.
Crucially, selection bias imposes some nigh-inscrutable but potentially sign-inverting considerations to any policy ‘upshot’. A less welcoming subgroup could be cause for concern, but alternatively cause for celebration: perhaps this subgroup offers other ‘pull factors’ that mean people who find EA is less welcoming nonetheless join and stick around within it (and vice versa: maybe subgroups whose members find EA very welcoming do so because they indirectly filter out everyone who doesn’t). Akin to Wald and the bombers in WW2, it is crucial to work out which. But I don’t think we can here.
I think the reason the OP had a high fraction of ‘long’ processes had more to do with him being a strong applicant who would get through a lot of the early filters. I don’t think a typical ‘EA org’ hiring round passes ~50% of its applicants to a work test.
This doesn’t detract from your other points re. the length in absolute terms. (The descriptions from OP and others read uncomfortably reminiscent of more senior academic hiring, with lots of people getting burned competing for really attractive jobs). There may be some fundamental trade-offs (the standard argument about ‘*really* important to get the right person, so we want to spent a lot of time assessing plausible candidates to pick the right one, false negatives at intermediate stages cost more than false positives, etc. etc.’), but an easy improvement (mentioned elsewhere) is to communicate as best as one can the likelihood of success (perhaps broken down by stage) so applicants can make a better-informed decision.
If you’re Open Phil, you can hedge yourself against the risk that your worldview might be wrong by diversifying. But the rest of us are just going to have to figure out which worldview is actually right.
Minor/Meta aside: I don’t think ‘hedging’ or diversification is the best way to look at this, whether one is an individual or a mega-funder.
On standard consequentialist doctrine, one wants to weigh things up ‘from the point of view of the universe’, and be indifferent as to ‘who is doing the work’. Given this, it looks better to act in the way which best rebalances the humanity-wide portfolio of moral effort, rather than a more narrow optimisation of ‘the EA community’, ‘OPs grants’, or ones own effort.
This rephrases the ‘neglectedness’ consideration. Yet I think people don’t often think enough about conditioning on the current humanity-wide portfolio, or see their effort as being a part of this wider whole, and this can mislead into moral paralysis (and, perhaps, insufficient extremising). If I have to ‘decide what worldview is actually right’, I’m screwed: many of my uncertainties I’d expect to be resilient to a lifetime of careful study. Yet I have better prospects of reasonably believing that “This issue is credibly important enough that (all things considered, pace all relevant uncertainties) in an ideal world humankind would address X people to work on this—given in fact there are Y, Y << X, perhaps I should be amongst them.”
This is a better sketch for why I work on longtermism, rather than overall confidence in my ‘longtermist worldview’. This doesn’t make worldview questions irrelevant (there are lot of issues where the sketch above applies, and relative importance will be one of the ingredients that goes in the mix of divining which one to take), but it means I’m fairly sanguine about perennial uncertainty. My work is minuscule part of the already-highly-diversified corporate effort of humankind, and the tacit coordination strategy of people like me acting on our best guess of the optimal portfolio looks robustly good (a community like EA may allow better ones), even if (as I hope and somewhat expect) my own efforts transpire to have little value.
The reason I shouldn’t ‘hedge’ but Open Phil should is not so much because they can afford to (given they play with much larger stakes, better resolution on ‘worldview questions’ has much higher value to them than to I), but because the returns to specialisation are plausibly sigmoid over the ‘me to OP’ range. For individuals, there’s increasing marginal returns to specialisation: in the same way we lean against ‘donation splitting’ with money, so too with time (it seems misguided for me to spend—say − 30% on bio, 10% on AI, 20% on global health, etc.) A large funder (even though it still represents a minuscule fraction of the humanity-wide portfolio) may have overlapping marginal return curves between its top picks of (all things considered) most promising things to work on, and it is better placed to realise other ‘portfolio benefits’.
Excellent. This series of interviews with superforecasters is also interesting. [H/T Ozzie]
Thanks. I should say that I didn’t mean to endorse stepwise when I mentioned it (for reasons Gelman and commenters note here), but that I thought it might be something one might have tried given it is the variable selection technique available ‘out of the box’ in programs like STATA or SPSS (it is something I used to use when I started doing work like this, for example).
Although not important here (but maybe helpful for next time), I’d caution against using goodness of fit estimators (e.g. AIC going down, R2 going up) too heavily in assessing the model as one tends to end up with over-fitting. I think the standard recommendations are something like:
Specify a model before looking at the data, and caveat any further explanations as post-hoc. (which sounds like essentially what you did).
Split your data into an exploration and confirmation set, where you play with whatever you like on the former, then use the model you think is best on the latter and report these findings (better, although slightly trickier, are things like k-fold cross validation rather than a single holdout).
LASSO, Ridge regression (or related regularisation methods) if you are going to select predictors ‘hypothesis free’ on your whole data.
(Further aside: Multiple imputation methods for missing data might also be worth contemplating in the future, although it is a tricky judgement call).
Neither of your examples backs up your point.
The 80000 hours article you cite notes in its summary only that:
Giving some money to charity is unlikely to make you less happy, and may well make you happier. (My emphasis)
The GWWC piece reads thus:
Giving 10% of your income to effective charities can make an incredible difference to some of the most deprived people in the world. But what effect will giving have on you? You may be concerned that it will damage your quality of life and make you less happy. This is a perfectly reasonable concern, and there is no shame in wanting to live a full and happy life.The good news is that giving can often make you happier.… (My emphasis)
As I noted in prior discussion, not only do these sources not claim ‘giving effectively will increase your happiness’, I’m not aware of this being claimed by any major EA source. Thus the objection “This line of argument confuses the effect of donating at all with the effect of donating effectively” targets a straw man.
My impression FWIW is that the ‘giving makes you happier’ point wasn’t/isn’t advanced to claim that the optimal portfolio for one’s personal happiness would include (e.g.) 10% of charitable donations (to effective causes), but that doing so isn’t such a ‘hit’ to one’s personal fulfilment as it appears at first glance. This is usually advanced in conjunction with the evidence on diminishing returns to money (i.e. even if you just lost—say − 10% of your income, if you’re a middle class person in a rich country, this isn’t a huge loss to your welfare—and given this evidence on the wellbeing benefits to giving, the impact is likely to be reduced further).
E.g. (and with apologies to the reader for inflicting my juvenilia upon them):
[Still being in the a high global wealth percentile post-giving] partly explains why I don’t feel poorly off or destitute. There are other parts. One is that giving generally makes you happier, and often more happier than buying things for yourself. Another is that I am fortunate in non-monetary respects: my biggest medical problem is dandruff, I have a loving family, a wide and interesting circle of friends, a fulfilling job, an e-reader which I can use to store (and occasionally read) the finest works of western literature, an internet connection I should use for better things than loitering on social media, and so on, and so on, and so on. I am blessed beyond all measure of desert.
So I don’t think that my giving has made me ‘worse off’. If you put a gun to my head and said, “Here’s the money you gave away back. You must spend it solely to further your own happiness”, I probably wouldn’t give it away: I guess a mix of holidays, savings, books, music and trips to the theatre might make me even happier (but who knows? people are bad at affective forecasting). But I’m pretty confident giving has made me happier compared to the case where I never had the money in the first place. So the downside looks like, “By giving, I have made myself even happier from an already very happy baseline, but foregone opportunities to give myself a larger happiness increment still”. This seems a trivial downside at worst, and not worth mentioning across the scales from the upside, which might be several lives saved, or a larger number of lives improved and horrible diseases prevented.