1 of 1

Followup with Thomas Ptashnik

Further scoping, access, PhD partner

Thomas Ptashnik is a Psychology PhD student interested in working on this with us. He is using the SOEP-Core data and familiar with SEM/Latent variable methods.

We have gained access to the relevant data

Here's the link to the Fehr appendix that contains the survey items they created (starting at Appendix B on page 33).\

Some salient example content:

These items correspond to the SOEP-IS surveys, which can be found here (use item names, like Q132, to search quickly
2017: https://paneldata.org/soep-is/inst/soep-is-2017-f
2018: https://paneldata.org/soep-is/inst/soep-is-2018-f
These links also mention that individuals with preexisting data access can apply for expanded access. I [Thomas] have access to SOEP-core version 36 (1984-2020 surveys),..

DR: Some interesting content (at a quick peek)

From 2017...

Q380: What you value in your work likerts ... includes "Having much influence" and "Socially responsible and important work"

Q160: Optimism/pessimism about the future

Q162: ... bunch of Likerts on "attitudes towards life and the future" (e.g., 'The options that I have in life are determined by social circumstances.')

From 2019

... they seem to collect genetic data

A proposed project

Does the Fehr/SOEP data provide valuable 'outcome measures' of EA and effective giving support?

I think we might see positive responses to the Fehr et al questions and donation choices as ‘necessary but not sufficient' for people to become effective givers or even EAs. If (especially in spite of the de-biasing) people still don’t support international redistribution, international orgs, and don’t opt to give from the lottery earnings to the global poor person … I think they are very unlikely to be susceptible to an EA or effective giving (e.g., GiveWell) appeal. (See further discussion and debate on this below). (But, as a check on this, it might be good to try to ask these same questions on a sample of actual EA’s and effective givers, and a comparison group!. #surveyexperiments)

Two projects on the same data

I envision two related projects on the same data: 1. Building a 'portable' model for prediction to aid targeting and 2. Building a 'deeper' model to aid understanding

I’m hoping that looking for predictors of (or ‘coherent factors explaining’) these responses in the SOEP data would prove useful for organizations like GWWC to consider ‘which groups to target in doing outreach’ (and perhaps especially ‘which groups to rule out’)

I hope we can do a sort of ‘leak-proof validated predictive ML model for this’
perhaps especially relevant for the German/EU context

Thomas: After talking it over with some colleagues, I think this approach is our best bet in terms of developing something with practical utility that still has a chance of being published in an academic journal. This is not my area of expertise, but if I remember correctly you have some R code already written. So I should quickly be able to put something together.

2. An (exploratory model) to help understand key factors that might be driving EA-adjacent attitudes and behaviors, offering insight into ‘what drives people towards or away from this mindset’.

Here we could engage the richer set of SOEP variables and consider latent factors

Anonymous colleague; caveats on 'the two goals'

if one simply wants to target people for giving to some specific EA-aligned cause in terms of a donation. In that case of the hypothetical African Christian women are likely to give, and it doesn't matter so much how they get to that decision. Quite a different set of metrics is desired (the kind of things we are trying to get at) if one is trying to actually select/find 'effective altruists'[RT2] if one simply wants to target people for giving to some specific EA-aligned cause in terms of a donation. In that case of the hypothetical African Christian women are likely to give, and it doesn't matter so much how they get to that decision. Quite a different set of metrics is desired (the kind of things we are trying to get at) if one is trying to actually select/find 'effective altruists'

Red team

But I'm less sure about: ..."would prove useful for orgs like GWWC to consider ‘which groups to target in doing outreach’ (and perhaps especially ‘which groups to rule out’)"

[suppose] you measure something like 'interested in giving to people in poverty in Africa' (or, at best, cosmopolitanism), and you find that the people highest in this are [Classical music fans], but the people most interested in EA stuff are [Techno ravers]. I think there are lots of reasons why this might occur. It could be that interest in EA is a combination of cosmopolitanism + interest in maximising effectiveness, but differences in the latter swamp the former. (If so the reasoning would at least be along the right lines, but would potentially be very practically misleading to GWWC)...

But I think what could be going on could be even worse, i.e.:

The measures measure something like 'not being so parochial that you won't give to a non-German charity', which is (ex hypothes) a necessary condition, but so minimal it's not really informing us about the much more demanding thing
... it measures something more specific/narrow that may be orthogonal or even antagonistic to EA (e.g. interest in overseas charity/poverty specifically [even if it doesn't maximise effectiveness]). Thought experiment: how would a libertarian-leaning AI-safety concerned German EA respond to the questions?

[still, this] seems worthwhile... I'd just be very tentative about inferring anything about what GWWC should do etc

Red team analogy

(I think of this case as a bit like studying interest in Marxism by asking about whether people are interested in helping the poor (or some such) In one sense you might think of this as a necessary condition / people who don't have any concern for this are not likely to be interested in Marxism. OTOH you'll probably mostly be picking up the 99% of people who are interested in helping the poor but not interested in the much more niche / slightly weirder thing that is also closely related to helping the poor, but is also associated with slightly counterintuitive views like 'donating to the poor is not good, you need to be concerned with [systemic change and global revolution / AI safety] etc.)

Red team:

[red team]
I guess it will be interesting to find out through your analysis:
Are these measures predicted by plain altruism + cosmopolitanism (which a priori we might say are more likely to be connected to EA)
Or are these measures predicted by egalitarianism + belief we should repay the third world / belief the rich should help the poor (which seem like they may be less closely connected with EA)*
*of course EAs are overwhelmingly liberal/egalitarian, but liberal/egalitarians are overwhelmingly not EA, which I think is an important complication"

DR and TP response to red team

Good points, and I even think “global redistribution” might rub some actual EAs the wrong way, as well as many EAs rejecting the 'repay our collective guilt' aspect.

Still, GWWC and TLYCS are pushing more for behaviors (esp. giving) than for intellectual alignment with EA. They are also pushing the traditional global poverty part of the EA agenda. I suspect the Fehr/Soep measures will pick up people more receptive to this than to longtermist 'avant garde' EA.
- Thomas: This is the main point to highlight. We probably need to limit our generalizability to the people-oriented neartermist worldview bucket. As the comments above note, I'm not sure this worldview necessarily maps onto the longtermist individual concerned about, say, AI safety risk. However, as you point out, there is still utility in focusing on understanding individuals that have this worldview for GWWC and other EA orgs, and this worldview (according to the EA survey) is currently the largest in the community.
  - DR: Agreed, but we probably need to make sure not to water it down too much; ideally we would retain some notion of 'the importance of prioritization and cost-effectiveness' in the worldview we are targeting
- TP: As you mentioned, it would be interesting to replicate this survey with explicitly EA endorsing individuals. Particularly, in seeing how well the ML model can predict cohorts that fall into the three different worldview buckets.
  - DR: yes but the model that predicts "EA/global poverty supporting types within a general population may be unlikely to predict groups *among explicit EA's*" ... still, the comparison could be interesting (and we've done a bit of this already with the EA survey)
- TP: Also, as a long-term idea, it could be useful to consider developing more EA-oriented items for SOEP-IS (the survey Fehr and colleagues used) that take into account all the issues listed here.
  - DR: that would be great!

RT2: Is there any way you can think of to get at EA more like a style of thinking/justification of choices as opposed to possibly the highly context-dependent choices are themselves? Some kind of relevant psychometric things are probably possible e.g., need for cognition or something similar RT1:

One option create or use measures of maximising + cosmopolitanism + altruism (or of maximising cosmopolitan altruism) ... maybe we are getting at 'EA style of thinking'. And if we can show that these more abstract measures are connected to behavioural or otherwise more concrete measures of EA inclination (whether that's decisions/choices, signing up for mailing list or something else) then it does seem reasonable to think of these as capturing EA inclination.
The risk otherwise is that theoretically we think these 3 things correspond to EA thinking... and actually they don't ...
Consider NFC, IRT, Rationality Quotient etc. as predictors of EA-inclination \

DR: My conception was maximizing + cosmopolitanism + _altruism + willing-to-sacrifice/non-competitiveness …_I think many people think “I should work to help humanity” but also think ‘yeah but I’ll be a sucker if I give to charity while my neighbor gets a new swimming pool and Hawaii holiday…’That’s where “willing-to-sacrifice/non-competitiveness” comes in, in my mind. (It needs a better name?)I think this last trait more important for effective giving than for EA-intellectual-engagement… and it may not be important at all for the latter.

Thomas: In psychology, altruism captures this notion. Prosociality is a concept of helping others but allows for self-concern, while altruism is distinguished by a purer form of selflessness (I have a paper under review that goes into detail about this, which I can privately share....argh, the closed doors of academia). Fortunately, altruism is widely studied and there are even a few items that capture it in the SOEP dataset. \

Value of incentivized measures here

(DR ideas)

IMO it would be nice to have some meaningful behavioral (incentivized) measures on top of the ‘psych’ ones. The ‘donation to the very poor’ measure in Fehr et al gets at this a bit … although its a pretty small probablistic sacrifice. And I suspect it measures all three of the above except maximizing. And I don’t think these things are all separable, so I think that the fact that it measures ‘altruism and willing to sacrifice in a cosmopolitan-relevant context’ is good.
It would also be pretty nice to have a behavioral/incentivized measure of ‘maximizing in an altruistic context’ …If Fehr ea had asked them to (e.g.) allocate giving among a German poor person, an African poor person, and themselves, this might have been a decent measure.
(We have this choice in some other contexts though … not as rich data but maybe worth digging into). Why might that choice have been better (in some ways) than a hypothetical choice? Because I imagine in a hypothetical choice some people would be like “OK they obviously want me to say support the poor person in Africa, and I see the maximization arguments, so, fine.'But when it involves real money, and even their own money, I expect that for some people, other motives will outweigh the ‘maximizing motive’…“wait, I’d rather keep the money than give it to an African who will waste it”“wait, if this is real, I’d rather help someone local”.

Analysis Plan, sample, and variables under consideration (01/31/22, Ptashnik)

DR: See sidebar comments

Analysis plan

Lasso regression to identify the most salient cluster [DR: how is this defined?] of predictors for effective giving

I will use k-fold cross-validation to compare a lasso model with ridge regression and OLS to confirm it is the best method for handling our data [DR: 'best in what sense? I recommend the elastic net approach if possible.]

Bayesian and latent lasso

TP: There is now a Bayesian form of lasso, but the R packages to run this analysis are in their infancy and the results between the methods are strikingly similar (Steorts, 2015). So, on the first pass I will just use one of the methods above but may rerun the analysis time-permitting to check my assumption that results won’t change.

Similarly, there is latent lasso regression, but most of our constructs have only one indicator and the R package for this analysis also appears to be at a nascent stage.·

Sample

To start, I’m just considering the 2017 survey and the control group (i.e., those who weren’t notified of their position in the national and global income distribution (~700 individuals). We can expand to the 2018 survey and the treatment group in future analyses using the same method (although some items may not be included across surveys).

Outcome Variable

Q280 and 281 in the SOEP-IS dataset developed by Fehr et al. (2019)

You were paired with another household in Kenya or Uganda. This household belongs to the poorest 10 percent of households worldwide. Now, you have 50 EUR at your disposal and can split this amount between the other household and you in any way you want. If this task is selected for payout, you will receive the amount you decided to keep at the end of the interview. The amount you want to give the other household will be given in full to the other household (without transaction costs) at the end of the field period by Heidelberg University via a charitable organization. In full means that every given euro will be received by the other household 1:1. A leaflet with information about the donations will be given to you after you have made your decision. I ask you to make this decision alone now.”

“How much of the 50 EUR do you want to keep and how much do you want to give the other household?”

2017 survey questions: https://paneldata.org/soep-is/inst/soep-is-2017-f

Variables Under Consideration

Below I list variables below in terms of what the intended construct I’m trying to get at and the proxy measures that are available within the SOEP dataset.

Theoretical rationale for construct from 'charitable giving' review

Theoretical rationale for these constructs comes from the most comprehensive review on predictors of charitable giving I could find (Bekkers & Wiepking, 2007; also see Bekkers & Wiepking, 2011 and Wiepking & Bekkers, 2012 for follow-ups on this review). These reviews seem like a reasonable starting point because they are cross-disciplinary and only consider studies that involve real money to real charitable organizations. There were a surprising number of what I think of as common-sense variables that weren’t included in these reviews that I add in the table below (i.e., those without an asterix).

There were several variables omitted because I did not think they were relevant or other constructs exist that better get at the underlying effect. ...

Home ownership: Appears to just be an indicator of wealth, so using income is preferrable.

Perceived financial position: Bivariate studies (Bennet & Kottasz; Havens et al., 2007) conclude those who perceive their financial situation as more positive are more generous donors. However, Fehr et al. (2019)—which has a more robust design—reports that “we find no evidence that perceived rank in the global income distribution affects support for global redistribution, donations to the global poor, globalization or immigration. If anything, when thinking about these policy preferences, it matters more how one compares to other people nationally than to others around the globe.” Given these findings and the fact that we are using the same data, it is probably sensible to omit this variable. Although studies have found confidence in the economy (Okunade, 1996), so an interesting pivot could be to measure optimism (both domain-specific and general forms).

Place of residence and years of residence: Mixed findings and it appears to be a weak predictor regardless.

Immigration and citizenship status: Better captured by other variables. “Osili and Du (2005) found that immigrants in the United States are less likely to give to charitable organizations and also give less, but that these differences are due to differences in racial background, lower levels of income, and education” (Bekkers & Wiepking, 2007: 15).

Youth participation: Impacts donations through socialization, which is better captured through parental background. It also strengthens social bonds of the children in the community, making them less likely to make effective donations over local causes.

Volunteering: In simple bivariate analysis, volunteers are usually found to donate more to charity. However, differences between volunteers and non-volunteers often vanish in multiple regression analyses controlling for joint determinants of giving and volunteering (Bekkers, 2002, Bekkers, 2006a, Wiepking & Maas, 2006). Given SOEP only asks about time spent volunteering and does not categorize where one volunteers, this variable seems like a blunt tool that is likely to be insignificant.

Awareness of need: A strong predictor of general philanthropy, but Fehr et al. (2019) did not find significant effects for effective giving. DR: I think 'failing to find significant effects' shouldn't be reason to exclude this!

[DR: I think 'previous failire to find significant effects' shouldn't be reason to exclude!]

Variables held constant by the survey design (see Bekkers & Wiepking 2007 for detailed explanation): Solicitation, benefits, reputation, and efficacy.

DR comments:

A very interesting list of features
were these all asked before the charity questions? (I'm worried about reverse causality otherwise)
maybe remove 'unavailable' rows for space\

We should discuss how the fitted model will be used and interpreted ... maybe identifying a few collections of useful subsets: