Epistemic state: very uncertain. This is a BOTEC estimate of working in community building. I would appreciate any feedback.
Overview
In this post, I present a quantitative estimate of the effectiveness and impact of running various community building activities from the perspective of an average community builder per year.
I provide a method for quantification of tangible impacts from community-building in effective altruism, as well as some quantitative estimates of impact.
1-1s are very important for engaging new EAs. Community builders should incorporate elements of 1-1 in their own work as much as possible.
Focusing on existing EAs and amplifying their impact is one of the most impactful things that a community builder can work toward.
The impacts of community-building are not insignificant and lead to substantial tangible benefits even in the short-term.
In estimating the effects of community building, second-order effects are of paramount importance; they often triple (or in some cases close to 10x) pure first-order effects.
A community builder working across different community building areas may have a “cross-pollination effect” that increases impact (see the Aside).
Motivation and strategy
In this report, I attempt to quantitatively estimate the average impact of an EA community builder in a year. I begin by identifying relevant community-building activities and separating each activity into a particular bucket. These activities are identified as follows:
organizing a school group and introductory fellowships for their school (e.g. EA at Oxford)
organizing a local group for professionals and introductory fellowships for their professional group (e.g. EA UK or more specific effective altruism groups for specific community niches)
organizing events, which I divide into three types below. Each type of event trades-off engagement for efficiency
3.1. 1-1 discussions — very high engagement which leads to best outcomes, but organizing efficiency is low (e.g. career consultation with 80,000 Hours)
3.2. small-scaled events — moderate engagement potential which leads to moderate outcomes; less efficiency than organizing large-scaled events (e.g. a small-scaled speaker event)
3.3. large-scaled events — low to moderate engagement potential but the efficiency of organizing is likely high (e.g. EAGx)
I am to quantify the impactof an average community builder per year if they help organize the activities above by estimating the increase in the number of engaged EAs per year. I broadly consider engaged EAs to be those who give themselves an engagement score of 4 or 5 on the EA Survey.
I then estimated the expected increase of relevant figures per engaged EA as a result of their participation in community building activities. In particular, I attempt to estimate the impact of community building on two relevant metrics:
increase in the amount of money donated to EA causes or EA-aligned charities as a direct result of community building activities
increase in the number of “effective hours worked on 80,000 Hours cause areas per week” — I loosely define an “effective hour” as an hour that makes some progress towards contributing to a cause area. Community building may accelerate the progress of engaged EAs in upskilling/being productive in cause areas
How do I expect this estimate to be helpful? I broadly expect community builders to use such a quantitative model to estimate their impact — they can measure their impact as the number of engaged EAs that their activities attract, and then multiply this by the metrics above. For instance, someone who works on local group community-building could find their expected impact by:
expected impact per year=increase in $ donated/yr×new engaged EAs from local group building
As an aside, I also present a framework for finding the expected impact of a community builder who spends time across different activities (e.g. 30% working on school community-building, 70% organizing events) in the final section.
I run through the methodology below and present the results (and my takeaways below). Lastly, I list some of the caveats and limitations of this quantitative estimation exercise.
Methodology
(1) Running a university group to increase the number of engaged EAs
I hope to estimate, on average, the impact that a community builder has in increasing the number of new EAs at each university per year (yearlynewstudentEAs). I consider this as getting someone to join the committee or engage in EA events beyond the EA introductory fellowship, and to rate themselves as “engaged EAs” according to the EA survey.
I take the simple assumption that fellowships are the only source by which people decide to commit to effective altruism and estimate the number of fellowships run (numberoffellowshipsrun), the cohort size for each fellowship (fellowshipcohortsize), and the conversion rate for each intro fellowship (introfellowshipconversion); that is, the percentage of students that remain with the effective altruism society following their fellowship.
Note: a fellowship is not the only way to becoming a committed EA. “Introductory fellowship” here is used as a proxy for the number of people reached, but it could easily be replaced by organizing events to expose students to EA ideas, speaking to students in classes, etc. to derive an estimation.
I also estimate a Shapley value (shapley_communitybuilding) to estimate the individual contribution of a school community builder. There are two possible (lognormal 5th and 95th percentile) ranges: one for the team members on an EA community building team and one for the leader(s) of the team.
// 1. Running a university group (new engaged EAs per year)
numberoffellowshipsrun = 3 to 7
fellowshipcohortsize = 5 to 30
introfellowshipconversion = beta(9,11) // approx 0.3 to 0.6, conversion rate to "engaged EA" each intro fellowship
shapley_communitybuilding = mx( // ranges between approx 0.1 and 0.4
[beta(2,10), beta(20,30)],
[0.8, 0.2] // Shapley value for team members and leaders
)
yearlynewengagedstudentEAs =
numberoffellowshipsrun *
fellowshipcohortsize *
introfellowshipconversion *
shapley_communitybuilding
Caveat: it seems fair to assume a very wide spread of outcomes depending on the size of the colleges and the extent of the interest demonstrated by the college students.
(2) Running a group for local professionals to increase the number of engaged EAs
I estimate the number of increased “engaged EAs” that a local group community builder can bring in per year (yearlynewprofessionalEAs).
The estimation strategy is very similar to estimating the impact of running a student group above. numberoffelloships, fellowshipcohortsize and introfellowshipconversion are very similar to above but have slightly smaller ranges because I expect professionals to be busier, for local community builders to have slightly less support and less capacity than student group community builders on average, etc.
// 2. Running a local group (new engaged EAs per year)
numberoffellowshipsrun = 1 to 6 // a bit less than the student group as less capacity
fellowshipcohortsize = 5 to 30 // same as student group
introfellowshipconversion = beta(7,13) // approx 0.2 to 0.5, lower than student group as professionals more busy etc.
yearlynewengagedprofessionalEAs =
numberoffellowshipsrun *
fellowshipcohortsize *
introfellowshipconversion *
shapley_communitybuilding
(3) Organizing events to increase the number of engaged EAs
I now estimate the impact of building on the existing EA community. Specifically, I hope to account for the number of potential EAs that shift from unengaged to engaged as a result of their participation in community building events. (In estimating the statistics below, I refer to previous 80,000 Hours research.)
I start by making some assumptions about the number of people in the EA community overall (numberofEAs), which I take to be around 11k as of 2022.
I also make assumptions for the proportion of engaged members that participate in 1-to-1 events or larger-scaled events (engagedmembersproportion_1to1 and engagedmembersproportion_events respectively).
Here, I assume that more engaged EAs are more likely to participate in events such as EA Global or EAGx, whereas non-engaged EAs are more likely to attend large-scale events and are more likely to go to 1-1 meetings such as careers consultation with 80,000 Hours.
1-1 meetings also seem more differentially useful for non-engaged EAs to provide them with advice etc. (when compared to the utility for engaged EAs).
However, I am uncertain whether this is a correct assumption to make.
// 3. building on existing community
numberofEAs = normal(11k,500) // from 80k data, extrapolated to 2022 and biased slightly upwards
engagedmembersproportion_1to1 = beta(1,2) // approx 0 to 0.75 % of engaged members in 1-1s
engagedmembersproportion_events = beta(6,2) // approx 0.5 to 0.95, % of engaged members at events
I then estimate the number of newly engaged EAs coming from a community builder working on developing events. Throughout the analysis, I segregate the three event types (1-1, small events, large events). I estimate:
how many participants a community builder can facilitate/interact with per year (totalparticipation_events_(event type), where event type={1to1,small,large}). For small-scale and large-scaled events, this requires predicting the number of events a community builder can realistically contribute towards per year and the number of participants for each event.
a Shapley value for each type of event, where I assume that more effort on behalf of the individual community builder is required for 1-1 events than for small events, and more effort for small events than for large events. I don’t put “1” as the Shapley value for 1-1 events because community builders on 1-1 calls/discussions may also have help for
an engagement factor (engagementfactor_(event type), where event type={1to1,small,large}), a variable representing how engaged the discussion/event is for the participant. 1-1s often provide the most value and require the most engagement, so they have the highest engagement factors. This is used as a proxy for the probability that an unengaged EA becomes engaged as a result of the event.
This ultimately results in the outcome newengagedEAs_(event type), where event type={1to1,small,large}, representing the number of newly engaged EAs per year resulting from the individual community builder’s effort to organize events.
// 3.1. 1 to 1 events — unengaged to engaged EAs
totalparticipants_events_1to1 = mx(
[10 to 100, 100 to 300, 300 to 500],
[0.5, 0.3, 0.2]
)
shapley_events_1to1 = beta(17,3) // approx 0.7 to 0.95, might have some help from others who prepare resources
engagementfactor_1to1 = beta(13,7) // approx 0.4 to 0.8
yearlynewengagedEAs_1to1 = totalparticipants_events_1to1 *
(1 - engagedmembersproportion_1to1) *
engagementfactor_1to1 *
shapley_events_1to1
// 3.2. small events — unengaged to engaged EAs
events_small = mx(
[1 to 10, 10 to 20],
[0.75, 0.25]
) // number of small events organized per year
events_smalloutreachparticipation = beta(2,150) // approx 0 to 0.03, proportion of EA community participating in small events
totalparticipants_events_small = events_small * events_smalloutreachparticipation * numberofEAs
shapley_events_small = beta(10,90) // approx 0.05 to 0.15
engagementfactor_small = beta(9,11) // approx 0.3 to 0.6
yearlynewengagedEAs_smallevents = totalparticipants_events_small *
(1 - engagedmembersproportion_events) *
engagementfactor_small *
shapley_events_small
// 3.3. large events — unengaged to engaged EAs
events_large = tR(1 to 3,1)
events_largeoutreachparticipation = beta(2,16) // approx 0.05 to 0.3
totalparticipants_events_large = events_large * events_largeoutreachparticipation * numberofEAs
shapley_events_large = beta(10,130) // approx 0.05 to 0.1
engagementfactor_large = beta(7,13) // approx 0.2 to 0.5
yearlynewengagedEAs_largeevents = totalparticipants_events_large *
(1 - engagedmembersproportion_events) *
engagementfactor_large *
shapley_events_large
Improving the estimates: second-order effects
Something important to estimate is the multiplier effects of community-building. Increasing the number of engaged EAs implies that the newly-engaged EA may also act as an ambassador for the EA movement, which brings in even more engaged EAs in the future. This snowballing effect seems particularly impactful if the newly-engaged EA brings in many people who bring in even more people in the future; the growth in the number of people could be exponential. (You may be skeptical of the inclusion of second-order effects; I discuss further in the caveats and limitations section below.)
I estimate the second-order effect of community building (second_order_effect) as follows:
I assume that most EAs bring in more EAs. Recruiting a newly engaged EA would lead to the recruitment of more than one engaged EA, because the newly engaged EA might promote the movement. I represent this using newEAs_positivemultiplier: per year, how many new EAs does the newly engaged EA (e.g. via word of mouth) get other engaged EAs to join?
Some EAs may lead to the recruitment of less than one engaged EA per year. This is represented by the variable newEAs_negativemultiplier.
The proportion of EAs that have this effect is represented byproportion_positivemultiplier, which I assume to be slightly more than half.
I assume that each EA brings in new engaged EAs over x number of years, with the number of new engaged EAs they bring in tapering off over time. This is estimated by the
I also include a Shapley value shapley_second_order to discount the impact of the community builder on the second-order effects because there is variability between a community builder’s action and whether this extends into second-order effects.
// Adding second-order effects
newEAs_positivemultiplier = tL(1.1 to 1.6, 1.0001)
proportion_positivemultiplier = beta(35,35) // approx 0.4 to 0.6
newEAs_negativemultiplier = tL(0.4 to 0.9, 0.9999)
second_order_years = 1 to 5
shapley_second_order = beta(11,9) // approx 0.4 to 0.7; you are not fully responsible for the second-order effect
second_order_effect =
(proportion_positivemultiplier *
(newEAs_positivemultiplier^second_order_years - 1)/(newEAs_positivemultiplier - 1)) +
((1 - proportion_positivemultiplier) *
(1 - newEAs_negativemultiplier^second_order_years)/(1 - newEAs_negativemultiplier)) *
shapley_second_order
second_order_newstudentEAs = yearlynewengagedstudentEAs * second_order_effect
second_order_newprofessionalEAs = yearlynewengagedprofessionalEAs * second_order_effect
second_order_1to1 = yearlynewengagedEAs_1to1 * second_order_effect
second_order_smallevents = yearlynewengagedEAs_smallevents * second_order_effect
second_order_largeevents = yearlynewengagedEAs_largeevents * second_order_effect
I then multiply the second-order effect by the yearly newly engaged EA estimates from (1), (2), and (3) above to arrive at second-order estimates for interventions to increase the number of engaged EAs per year. Results can be found below.
I now turn to quantitatively estimating the tangible outcomes. I focus on two outcomes: the amount of money donated to EA causes or EA-aligned charities, and the “effective” number of hours worked on 80k cause areas.
(4) Increasing (in the short-term) the amount of money donated to EA causes or EA-aligned charities
I attempt to estimate the increase in donations for EA by organizing a local group for professionals. This estimate is only relevant to section (2) above where I estimated the number of newly engaged EAs.
I first estimate the proportion of newly engaged professionals EA that donate. As a reference point, Rethink Priorities found that approx 55% of those completing the EA survey made a charitable donation. I use a normal distribution with a mean of 0.5 and s.d. 0.1 to model the proportion of newly engaged professional EAs that donate. (The mean is slightly less than 0.55 because newly engaged EAs might be a bit more uncertain about donating.) This is represented by newEAprofessionalsdonating.
I also estimate the amount that a professional would donate annually professionalsannualdonation (conditional on them donating). I can then estimate the donationpernewprofessional (donation per professional on average) and multiply this with the number of new professionals that the EA community builder successfully engages (from my estimate in section (2)).
// 4. getting people to E2G or donate (professionals only)
newEAprofessionalsdonating = tL(normal(0.5,0.1),0) // from 2020 EA survey data
professionalsannualdonation = mx(
[100 to 1k, 1k to 10k, 10k to 100k, 100k to 1M],
[0.33, 0.33, 0.33, 0.001] // can remove (100k to 1M) option — greatly reduces estimates
)
donationpernewprofessional = newEAprofessionalsdonating *
professionalsannualdonation
increaseindonations = yearlynewengagedprofessionalEAs * donationpernewprofessional
second_order_increaseindonations = second_order_newprofessionalEAs * donationpernewprofessional
For commentary on this result, please see the results section. A very important caveat here: EA students ultimately become professionals and likely donate to EA causes/EA-aligned charities. This model does not account for this, so community builders should definitely not treat this estimate as definitive in any sense!
(5) Increasing the “effective” number of hours worked on 80,000 Hours cause areas
I attempt to estimate the “effective” number of hours worked by an engaged EA on an 80k cause area per week, and how community building could nudge that number upwards.
Why “effective” hours? I introduce the “effective hours” metric to account for how much time is spent by an individual to work on 80k cause areas, and how effective their time is spent (i.e. how much of their time is spent towards solving the problem). I see effective hours as a function of effectiveness, personal fit, and time spent:
Therefore, the number of “effective hours” worked is intended as a proxy for additional existential risk reduced because of community building activities. A community builder could (i) increase the effectiveness of someone working on a cause area by introducing them to fellow EAs, (ii) increase the personal fit by helping EAs focus on areas that they have a competitive advantage in, (iii) increase the time spent by connecting the engaged EA to communities for accountability, etc.
I estimate the difference in effective hours worked per week on 80k cause areas (diff_effectiveweeklyhoursoncauseareas) as follows:
I split into the cases of student group organization, professional group organization, and “enhancement work” from organizing 1-1s or events.
In the case of student EA groups/professional EA groups, I consider the probability that a student/professional changes their career to work more on cause areas (i.e. decide to work on cause areas) after they become an engaged EA. For students, I estimate studentworksmoreoncauseareas, and likewise for professionals, I estimate professionalworksmoreoncauseareas. I think that the variableprofessionalworksmoreoncauseareas should be lower than studentworksmoreoncauseareas, because the costs of switching/committing to work more on cause areas are higher for professionals than for students.
Of the students/professionals that decide to work on cause areas more, I estimate the difference made by the community builder by taking into account the existing effective hours/week worked on cause areas (assumed to be less than 5 hours/week) and multiplying this by a scaling factor (which has multiple distributions — see the code below). This scaling factor represents the significance and various degrees of efficacy of the EA group’s events/activities organized by the community builder on the student/professional on changing their career.
I use a very wide range for the scaling factor here (up to approx 32x) because the initial nudge made by the community builder could be extremely significant (analogous to helping someone completely shift their career to be much more impactful).
Therefore, diff_effectiveweeklyhoursoncauseareas_student and diff_effectiveweeklyhoursoncauseareas_professional can identify the “nudge” effect of the community builder for people to work more on cause areas as a result of student/professional group organization respectively.
// 5. getting people to work on cause areas
// engaged EA students deciding to switch careers from student group intervention
studentworksmoreoncauseareas = beta(10,15) // approx 0.25 to 0.55
diff_effectiveweeklyhoursoncauseareas_student = studentworksmoreoncauseareas *
tR((1 to 3),5) * // existing hours worked
tL(mx(
[1 to 2, 2 to 4, 4 to 8, 8 to 16, 16 to 32],
[0.3, 0.25, 0.2, 0.15, 0.1]), 0.9) // scaling factor for work
// engaged EA students deciding to switch careers from local professional group intervention
professionalworkmoreoncauseareas = beta(4,15) // approx 0.1 to 0.4
diff_effectiveweeklyhoursoncauseareas_professional = professionalworkmoreoncauseareas *
tR((1 to 3),5) * // existing hours worked
tL(mx(
[1 to 2, 2 to 4, 4 to 8, 8 to 16, 16 to 32],
[0.3, 0.25, 0.2, 0.15, 0.1]), 0.9) // scaling factor for work
I also estimate the difference in effective hours worked per week on 80k cause areas for EAs that are already engaged. These EAs may participate in 1-1s or small/large events which would scale their impact further. To estimate this:
I use a mixture distribution to model the existing effective hours/week worked by the EA on cause areas (effectivelyweeklyhoursoncauseareas_existing).
I then estimate scaling factors for a 1-1 and a small/large event (how much the community builder can multiply the existing work of the EA). I use larger scaling factors for 1-1 discussions because they are likely much more efficient for getting connected to resources, learning about tailored individual options etc. This is given by scalingfactorforEAwork.
There is also the possibility of the community builder reducing the impact of an EA, which would lead to a scaling factor of less than 1. However, I consider this possibility to be minimal (and near-negligible) for the average community builder.
// enhancement of work from events
effectiveweeklyhoursoncauseareas_existing = mx(
[0 to 10, 10 to 20, 20 to 40, 40 to 60, 60 to 100],
[0.3, 0.2, 0.25, 0.2, 0.05]
)
scalingfactorforEAwork_1to1 = tL(mx(
[1 to 1.5, 1.5 to 2, 2 to 4, 4 to 8],
[0.45, 0.35, 0.15, 0.05]), 0.9)
scalingfactorforEAwork_event = tL(mx(
[1 to 1.5, 1.5 to 2, 2 to 4, 4 to 8],
[0.5, 0.4, 0.09, 0.01]), 0.9) // similar for small and large events
diff_effectiveweeklyhoursoncauseareas_1to1 = effectiveweeklyhoursoncauseareas_existing *
(scalingfactorforEAwork_1to1 - 1)
diff_effectiveweeklyhoursoncauseareas_event = effectiveweeklyhoursoncauseareas_existing *
(scalingfactorforEAwork_event - 1)
From these assumptions, I find the diff_effectivelyweeklyhoursoncauseareas for 1-1s and professionals, which is given by the product of the existing effective hours/week and the (scaling factor−1).
Results
Summary of results
I summarize the 50th percentile outcomes below. The 25th and 75th percentile outcomes are provided in parentheses. Please note that these results are the expected impact per community builder per year on average. All estimates are approximate and use a sample size n=1000.
Newly engaged EAs per year
from one community builder organizing 1-1 events for one year
second-order effects (new EAs get more EAs to join) and first-order effects: 12 new engaged EAs/yr (25th: 4.5/yr, 75th: 32/yr)
Expected increase in donations (in the short-term) from one community builderorganizing a professional group
first-order effects only: increase in $2,800/yr (25th: $410/yr, 75th: $16k/yr)
second-order effects (new professional EAs get more professional EAs to join) and first-order effects: increase in $7,900/yr (25th: $1,200/yr, 75th: $49k/yr)
Expected difference in effective weekly hours worked on 80k cause areas
following organizing a 1-1 event for individuals already contributing to cause areas: increase in 8.7 effective hours per week per individual (25th: 3.1, 75th: 22)
following a small-scaled/large-scaled event for individuals already contributing to cause areas: increase in 7.7 effective hours per week per individual (25th: 2.7, 75th: 22)
from organizing a student group across 1 year: increase in 2.4 effective hours per week per student (25th: 1.2, 75th: 5.7)
from organizing a professionals group across 1 year: increase in 1.3 effective hours per week per professional (25th: 1.2, 75th: 5.7)
For detailed results and distributional analysis (and to play around with the numbers), please see the Squiggle notebook.[1]
What do the results show?
I present some of my preliminary thoughts about this quantitative estimation exercise:
1-1s are extremely effective. The differential of a community builder organizing 1-1 events seems a lot more effective than contributing to helping with the organization of a large-scale/small-scaled event. It is perhaps the case that community builders should work towards finding more opportunities for 1-1 engagement (e.g. incorporating 1-1 elements between community builders and students in a student group).
Focusing on existing EAs seems useful. The results reflect that focusing on EAs already making an impact and working to support them might have a much higher impact than attracting new students/professionals. As seen from the estimates, community builders who enhance the impact of individuals who are already working on causes by connecting them to other EAs, suggesting internships and resources etc. may be a very effective move for enhancing the movement overall.
The impact of community building is not insignificant. An average community builder working on building up professional communities can raise $7.9k per year for EA-related causes (second-order estimate), with the amount increasing significantly assuming a 75th percentile (or more extreme) community-building outcome. For reference, Givewell’s cost-effectiveness analysis of Deworm the World in Kenya suggests that each philanthropic dollar spent leads to an increase of 0.11 “units of value”. Therefore, raising $7.9k leads to an increase of ~869 units of value (equivalent to preventing approximately 10.5 deaths of an individual 5 or older from malaria per year). Increasing the number of effective hours worked also seems very important for pushing ahead on important EA cause areas, especially if many of the 80k cause areas are talent-constrained.
Second-order effects are important, particularly for community building. Second-order effects (sometimes known as “multiplier effects”) are pivotal to fully capture the impact of community building. They can often triple (or more) the estimated first-order effects.
Aside: how would a community builder go about envisioning their impact?
We have estimated the number of engaged EAs that a community builder can recruit per year. The estimates above correspond to my thoughts about what a full-time community-builder would be able to achieve per year if they were only devoted to that one activity. Of course, community builders can mix between activities (e.g. spend 30% of the time working on a university group and 70% of the time organizing 1-1 events.) If a community builder wants to estimate their impact, I propose using something like the following formula:
Impact=λ×∑activities% time spent on activity×impact of activity
Impact is the sum of all activities that a community builder helps organize multiplied by the time spent on the activity and a coefficient λ. I see λ as a coefficient which measures cross-pollination effects from participating in different activities, which would increase the efficacy of an individual’s community building effects (which would increase λ) or a distraction effect from not focusing attention on one thing (which would decrease λ).
For instance, if I wanted to measure the (first-order) number of engaged EAs that a 50th percentile community builder can attract per year with a 30-70 mix between organizing a uni group and 1-1 events, I can arrive at engaged EAs per year≈λ×((0.3×4.4)+(0.7×31))=23.02λ EAs per year. If participating in multiple activities gives the community-builder a strong cross-pollination effect, I could assume λ=1.1, which would mean that the community-builder can attract approx 23.02×1.1≈25.3 EAs per year.
Caveats and limitations
I list some caveats and potential limitations of my estimates below.
Subjective. The estimates are largely based on my assumptions and prior credences as of 2022. You may have different credences based on your own information or think that my assumptions are systematically biased, could be too optimistic, or too pessimistic.
Outcomes may not be useful. I’m very specifically focused on the increase in donations and increase in effective time spent on 80k cause areas. It’s plausible that there are more important outcomes or more relevant outcomes for the community builder than the outcomes that I have analyzed.
Unspecific outcomes. The outcomes that have been analyzed are too unspecific and could be meaningless (e.g. the concept of an “effective hour” worked on a cause area may not be useful). Outcomes such as “effective hours” are also subjective and may not equate directly to progress on solving the problem (e.g. more time spent on AI alignment does not necessarily equate to more progress made on AI). It could be the case that effective time worked on cause areas has greatly diminishing returns, which is not accounted for within this model.
You might also think that the impact might be a lot larger for students who decide to work on AI safety rather than other cause areas and treat the conversion to “priority cause areas” (e.g. AI safety/biosecurity) as a more important metric.
The concept of an engaged EA is meaningless. You may not think that there is a clear delineation between an “unengaged EA” and an “engaged EA”, or that the difference is not binary and should be further nuanced. The quantitative estimate that I have provided above does not suitably capture the nuances of this sliding scale.
Short-term. One could also consider that the thinking behind this quantitative estimate could be considered very short-term, and the relevant outcomes that I have put forward today may be very different in the future as the EA movement grows further. This particularly seems relevant for the estimates for increase in donations, because the estimates do not account for the donations of students in the future and only account for the increase in donations when a local professional group is organized that moves people from being an unengaged professional EA to an engaged professional EA.
The number of EAs does not matter. You might think that getting more people into effective altruism doesn’t matter. For instance,some new EAs may quickly lose interest or some may be very dedicated and help recruit even more students next year, so the statistic itself is not useful for making estimates or analysis.
No account for increasing/diminishing marginal returns. The model does not account for increasing/diminishing marginal returns (e.g. an EA going to 500 1-1s per year would probably derive much less value than an EA going to five 1-1s per year). It only accounts for the average effect of organizing a particular activity.
Second-order effects are poorly estimated. The second-order effects may be very understated, particularly in the long term. It also does not consider further nth-order effects. As a result, it would be a good idea to treat this model as negatively biased. Alternatively, one could think that the second-order effects need to be further discounted because the community-builder might not actually be accountable for the second-order effects. A potential point of improvement could be to further reduce the Shapley values for the second-order effects so that the second-order effects are further discounted.
High variance. The quantitative estimates vary too much to be useful (i.e. there is too much uncertainty). We might need to gather more quantitative data to be more well-informed about outcome of community building.
Poor external validity. The model may not be useful because the reasoning is not valid across different contexts (e.g. some groups may not run introductory fellowships etc.)
Conclusion
I have provided a quantitative estimate of the impact that an average community builder can expect by working on further community building activities. I hope that this report is useful for those who are considering community building activities or for community builders who are thinking about ways to further enhance their impact.
For the graphs on the right of the Squiggle notebook, several outcomes have two graphs assigned to them. The first graph (“0”) represents the first-order effect only, while the second graph (“1″) represents both first-order and second-order effect.
Estimating the impact of community building work
Epistemic state: very uncertain. This is a BOTEC estimate of working in community building. I would appreciate any feedback.
Overview
In this post, I present a quantitative estimate of the effectiveness and impact of running various community building activities from the perspective of an average community builder per year.
This is a submission to NunoSempere’s $5k challenge to quantify the impact of 80,000 hours’ top career paths. I used Squiggle for modelling and estimation; my Squiggle code is linked here. Comments or feedback are very welcome.
Executive summary
I provide a method for quantification of tangible impacts from community-building in effective altruism, as well as some quantitative estimates of impact.
1-1s are very important for engaging new EAs. Community builders should incorporate elements of 1-1 in their own work as much as possible.
Focusing on existing EAs and amplifying their impact is one of the most impactful things that a community builder can work toward.
The impacts of community-building are not insignificant and lead to substantial tangible benefits even in the short-term.
In estimating the effects of community building, second-order effects are of paramount importance; they often triple (or in some cases close to 10x) pure first-order effects.
A community builder working across different community building areas may have a “cross-pollination effect” that increases impact (see the Aside).
Motivation and strategy
In this report, I attempt to quantitatively estimate the average impact of an EA community builder in a year. I begin by identifying relevant community-building activities and separating each activity into a particular bucket. These activities are identified as follows:
organizing a school group and introductory fellowships for their school (e.g. EA at Oxford)
organizing a local group for professionals and introductory fellowships for their professional group (e.g. EA UK or more specific effective altruism groups for specific community niches)
organizing events, which I divide into three types below. Each type of event trades-off engagement for efficiency
3.1. 1-1 discussions — very high engagement which leads to best outcomes, but organizing efficiency is low (e.g. career consultation with 80,000 Hours)
3.2. small-scaled events — moderate engagement potential which leads to moderate outcomes; less efficiency than organizing large-scaled events (e.g. a small-scaled speaker event)
3.3. large-scaled events — low to moderate engagement potential but the efficiency of organizing is likely high (e.g. EAGx)
I am to quantify the impact of an average community builder per year if they help organize the activities above by estimating the increase in the number of engaged EAs per year. I broadly consider engaged EAs to be those who give themselves an engagement score of 4 or 5 on the EA Survey.
I then estimated the expected increase of relevant figures per engaged EA as a result of their participation in community building activities. In particular, I attempt to estimate the impact of community building on two relevant metrics:
increase in the amount of money donated to EA causes or EA-aligned charities as a direct result of community building activities
increase in the number of “effective hours worked on 80,000 Hours cause areas per week” — I loosely define an “effective hour” as an hour that makes some progress towards contributing to a cause area. Community building may accelerate the progress of engaged EAs in upskilling/being productive in cause areas
How do I expect this estimate to be helpful? I broadly expect community builders to use such a quantitative model to estimate their impact — they can measure their impact as the number of engaged EAs that their activities attract, and then multiply this by the metrics above. For instance, someone who works on local group community-building could find their expected impact by:
expected impact per year=increase in $ donated/yr×new engaged EAs from local group buildingAs an aside, I also present a framework for finding the expected impact of a community builder who spends time across different activities (e.g. 30% working on school community-building, 70% organizing events) in the final section.
I run through the methodology below and present the results (and my takeaways below). Lastly, I list some of the caveats and limitations of this quantitative estimation exercise.
Methodology
(1) Running a university group to increase the number of engaged EAs
I hope to estimate, on average, the impact that a community builder has in increasing the number of new EAs at each university per year (yearlynewstudentEAs). I consider this as getting someone to join the committee or engage in EA events beyond the EA introductory fellowship, and to rate themselves as “engaged EAs” according to the EA survey.
I take the simple assumption that fellowships are the only source by which people decide to commit to effective altruism and estimate the number of fellowships run (numberoffellowshipsrun), the cohort size for each fellowship (fellowshipcohortsize), and the conversion rate for each intro fellowship (introfellowshipconversion); that is, the percentage of students that remain with the effective altruism society following their fellowship.
Note: a fellowship is not the only way to becoming a committed EA. “Introductory fellowship” here is used as a proxy for the number of people reached, but it could easily be replaced by organizing events to expose students to EA ideas, speaking to students in classes, etc. to derive an estimation.
I also estimate a Shapley value (shapley_communitybuilding) to estimate the individual contribution of a school community builder. There are two possible (lognormal 5th and 95th percentile) ranges: one for the team members on an EA community building team and one for the leader(s) of the team.
Caveat: it seems fair to assume a very wide spread of outcomes depending on the size of the colleges and the extent of the interest demonstrated by the college students.
(2) Running a group for local professionals to increase the number of engaged EAs
I estimate the number of increased “engaged EAs” that a local group community builder can bring in per year (yearlynewprofessionalEAs).
The estimation strategy is very similar to estimating the impact of running a student group above. numberoffelloships, fellowshipcohortsize and introfellowshipconversion are very similar to above but have slightly smaller ranges because I expect professionals to be busier, for local community builders to have slightly less support and less capacity than student group community builders on average, etc.
(3) Organizing events to increase the number of engaged EAs
I now estimate the impact of building on the existing EA community. Specifically, I hope to account for the number of potential EAs that shift from unengaged to engaged as a result of their participation in community building events. (In estimating the statistics below, I refer to previous 80,000 Hours research.)
I start by making some assumptions about the number of people in the EA community overall (numberofEAs), which I take to be around 11k as of 2022.
I also make assumptions for the proportion of engaged members that participate in 1-to-1 events or larger-scaled events (engagedmembersproportion_1to1 and engagedmembersproportion_events respectively).
I refer to the 80k estimates: ~35% of the EA community (approximate 2.6k people) were “committed” EA members in 2021. As engaged EAs form a significant proportion of the EA community, it is realistic that a lot of engaged EAs would participate in EA events, especially social events for EAs or large-scale EA conferences.
Here, I assume that more engaged EAs are more likely to participate in events such as EA Global or EAGx, whereas non-engaged EAs are more likely to attend large-scale events and are more likely to go to 1-1 meetings such as careers consultation with 80,000 Hours.
1-1 meetings also seem more differentially useful for non-engaged EAs to provide them with advice etc. (when compared to the utility for engaged EAs).
However, I am uncertain whether this is a correct assumption to make.
I then estimate the number of newly engaged EAs coming from a community builder working on developing events. Throughout the analysis, I segregate the three event types (1-1, small events, large events). I estimate:
how many participants a community builder can facilitate/interact with per year (totalparticipation_events_(event type), where event type={1to1,small,large}). For small-scale and large-scaled events, this requires predicting the number of events a community builder can realistically contribute towards per year and the number of participants for each event.
a Shapley value for each type of event, where I assume that more effort on behalf of the individual community builder is required for 1-1 events than for small events, and more effort for small events than for large events. I don’t put “1” as the Shapley value for 1-1 events because community builders on 1-1 calls/discussions may also have help for
an engagement factor (engagementfactor_(event type), where event type={1to1,small,large}), a variable representing how engaged the discussion/event is for the participant. 1-1s often provide the most value and require the most engagement, so they have the highest engagement factors. This is used as a proxy for the probability that an unengaged EA becomes engaged as a result of the event.
This ultimately results in the outcome newengagedEAs_(event type), where event type={1to1,small,large}, representing the number of newly engaged EAs per year resulting from the individual community builder’s effort to organize events.
Improving the estimates: second-order effects
Something important to estimate is the multiplier effects of community-building. Increasing the number of engaged EAs implies that the newly-engaged EA may also act as an ambassador for the EA movement, which brings in even more engaged EAs in the future. This snowballing effect seems particularly impactful if the newly-engaged EA brings in many people who bring in even more people in the future; the growth in the number of people could be exponential. (You may be skeptical of the inclusion of second-order effects; I discuss further in the caveats and limitations section below.)
I estimate the second-order effect of community building (second_order_effect) as follows:
I assume that most EAs bring in more EAs. Recruiting a newly engaged EA would lead to the recruitment of more than one engaged EA, because the newly engaged EA might promote the movement. I represent this using newEAs_positivemultiplier: per year, how many new EAs does the newly engaged EA (e.g. via word of mouth) get other engaged EAs to join?
Some EAs may lead to the recruitment of less than one engaged EA per year. This is represented by the variable newEAs_negativemultiplier.
The proportion of EAs that have this effect is represented byproportion_positivemultiplier, which I assume to be slightly more than half.
I assume that each EA brings in new engaged EAs over x number of years, with the number of new engaged EAs they bring in tapering off over time. This is estimated by the
I also include a Shapley value shapley_second_order to discount the impact of the community builder on the second-order effects because there is variability between a community builder’s action and whether this extends into second-order effects.
I then multiply the second-order effect by the yearly newly engaged EA estimates from (1), (2), and (3) above to arrive at second-order estimates for interventions to increase the number of engaged EAs per year. Results can be found below.
I now turn to quantitatively estimating the tangible outcomes. I focus on two outcomes: the amount of money donated to EA causes or EA-aligned charities, and the “effective” number of hours worked on 80k cause areas.
(4) Increasing (in the short-term) the amount of money donated to EA causes or EA-aligned charities
I attempt to estimate the increase in donations for EA by organizing a local group for professionals. This estimate is only relevant to section (2) above where I estimated the number of newly engaged EAs.
I first estimate the proportion of newly engaged professionals EA that donate. As a reference point, Rethink Priorities found that approx 55% of those completing the EA survey made a charitable donation. I use a normal distribution with a mean of 0.5 and s.d. 0.1 to model the proportion of newly engaged professional EAs that donate. (The mean is slightly less than 0.55 because newly engaged EAs might be a bit more uncertain about donating.) This is represented by newEAprofessionalsdonating.
I also estimate the amount that a professional would donate annually professionalsannualdonation (conditional on them donating). I can then estimate the donationpernewprofessional (donation per professional on average) and multiply this with the number of new professionals that the EA community builder successfully engages (from my estimate in section (2)).
For commentary on this result, please see the results section. A very important caveat here: EA students ultimately become professionals and likely donate to EA causes/EA-aligned charities. This model does not account for this, so community builders should definitely not treat this estimate as definitive in any sense!
(5) Increasing the “effective” number of hours worked on 80,000 Hours cause areas
I attempt to estimate the “effective” number of hours worked by an engaged EA on an 80k cause area per week, and how community building could nudge that number upwards.
Why “effective” hours? I introduce the “effective hours” metric to account for how much time is spent by an individual to work on 80k cause areas, and how effective their time is spent (i.e. how much of their time is spent towards solving the problem). I see effective hours as a function of effectiveness, personal fit, and time spent:
effective hours=effectiveness×personal fit×time spentTherefore, the number of “effective hours” worked is intended as a proxy for additional existential risk reduced because of community building activities. A community builder could (i) increase the effectiveness of someone working on a cause area by introducing them to fellow EAs, (ii) increase the personal fit by helping EAs focus on areas that they have a competitive advantage in, (iii) increase the time spent by connecting the engaged EA to communities for accountability, etc.
I estimate the difference in effective hours worked per week on 80k cause areas (diff_effectiveweeklyhoursoncauseareas) as follows:
I split into the cases of student group organization, professional group organization, and “enhancement work” from organizing 1-1s or events.
In the case of student EA groups/professional EA groups, I consider the probability that a student/professional changes their career to work more on cause areas (i.e. decide to work on cause areas) after they become an engaged EA. For students, I estimate studentworksmoreoncauseareas, and likewise for professionals, I estimate professionalworksmoreoncauseareas. I think that the variableprofessionalworksmoreoncauseareas should be lower than studentworksmoreoncauseareas, because the costs of switching/committing to work more on cause areas are higher for professionals than for students.
Of the students/professionals that decide to work on cause areas more, I estimate the difference made by the community builder by taking into account the existing effective hours/week worked on cause areas (assumed to be less than 5 hours/week) and multiplying this by a scaling factor (which has multiple distributions — see the code below). This scaling factor represents the significance and various degrees of efficacy of the EA group’s events/activities organized by the community builder on the student/professional on changing their career.
I use a very wide range for the scaling factor here (up to approx 32x) because the initial nudge made by the community builder could be extremely significant (analogous to helping someone completely shift their career to be much more impactful).
Therefore, diff_effectiveweeklyhoursoncauseareas_student and diff_effectiveweeklyhoursoncauseareas_professional can identify the “nudge” effect of the community builder for people to work more on cause areas as a result of student/professional group organization respectively.
I also estimate the difference in effective hours worked per week on 80k cause areas for EAs that are already engaged. These EAs may participate in 1-1s or small/large events which would scale their impact further. To estimate this:
I use a mixture distribution to model the existing effective hours/week worked by the EA on cause areas (effectivelyweeklyhoursoncauseareas_existing).
I then estimate scaling factors for a 1-1 and a small/large event (how much the community builder can multiply the existing work of the EA). I use larger scaling factors for 1-1 discussions because they are likely much more efficient for getting connected to resources, learning about tailored individual options etc. This is given by scalingfactorforEAwork.
There is also the possibility of the community builder reducing the impact of an EA, which would lead to a scaling factor of less than 1. However, I consider this possibility to be minimal (and near-negligible) for the average community builder.
From these assumptions, I find the diff_effectivelyweeklyhoursoncauseareas for 1-1s and professionals, which is given by the product of the existing effective hours/week and the (scaling factor−1).
Results
Summary of results
I summarize the 50th percentile outcomes below. The 25th and 75th percentile outcomes are provided in parentheses. Please note that these results are the expected impact per community builder per year on average. All estimates are approximate and use a sample size n=1000.
Newly engaged EAs per year
from one community builder organizing 1-1 events for one year
first-order effects only: 31 new engaged EAs/yr (25th: 11/yr, 75th: 84/yr)
second-order effects (new EAs get more EAs to join) and first-order effects: 85 new engaged EAs/yr (25th: 24/yr, 75th: 260/yr)
from one community builder organizing small events for one year
first-order effects only: 4.4 new engaged EAs/yr (25th: 1.6/yr, 75th: 13/yr)
second-order effects (new EAs get more EAs to join) and first-order effects: 13 new engaged EAs/yr (25th: 3.4/yr, 75th: 44/yr)
from one community builder organizing large events for one year
first-order effects only: 4 new engaged EAs/yr (25th: 1.7/yr, 75th: 8.8/yr)
second-order effects (new EAs get more EAs to join) and first-order effects: 11 new engaged EAs/yr (25th: 3.8/yr, 75th: 31/yr)
from one community builder organizing a local professional group for one year
first-order effects only: 1.8 new engaged EAs/yr (25th: 0.84/yr, 75th: 3.9/yr)
second-order effects (new EAs get more EAs to join) and first-order effects: 5.2 new engaged EAs/yr (25th: 1.9/yr, 75th: 13/yr)
from one community builder organizing a university student group for one year
first-order effects only: 4.4 new engaged EAs/yr (25th: 2.2/yr, 75th: 8.3/yr)
second-order effects (new EAs get more EAs to join) and first-order effects: 12 new engaged EAs/yr (25th: 4.5/yr, 75th: 32/yr)
Expected increase in donations (in the short-term) from one community builder organizing a professional group
first-order effects only: increase in $2,800/yr (25th: $410/yr, 75th: $16k/yr)
second-order effects (new professional EAs get more professional EAs to join) and first-order effects: increase in $7,900/yr (25th: $1,200/yr, 75th: $49k/yr)
Expected difference in effective weekly hours worked on 80k cause areas
following organizing a 1-1 event for individuals already contributing to cause areas: increase in 8.7 effective hours per week per individual (25th: 3.1, 75th: 22)
following a small-scaled/large-scaled event for individuals already contributing to cause areas: increase in 7.7 effective hours per week per individual (25th: 2.7, 75th: 22)
from organizing a student group across 1 year: increase in 2.4 effective hours per week per student (25th: 1.2, 75th: 5.7)
from organizing a professionals group across 1 year: increase in 1.3 effective hours per week per professional (25th: 1.2, 75th: 5.7)
For detailed results and distributional analysis (and to play around with the numbers), please see the Squiggle notebook.[1]
What do the results show?
I present some of my preliminary thoughts about this quantitative estimation exercise:
1-1s are extremely effective. The differential of a community builder organizing 1-1 events seems a lot more effective than contributing to helping with the organization of a large-scale/small-scaled event. It is perhaps the case that community builders should work towards finding more opportunities for 1-1 engagement (e.g. incorporating 1-1 elements between community builders and students in a student group).
Focusing on existing EAs seems useful. The results reflect that focusing on EAs already making an impact and working to support them might have a much higher impact than attracting new students/professionals. As seen from the estimates, community builders who enhance the impact of individuals who are already working on causes by connecting them to other EAs, suggesting internships and resources etc. may be a very effective move for enhancing the movement overall.
The impact of community building is not insignificant. An average community builder working on building up professional communities can raise $7.9k per year for EA-related causes (second-order estimate), with the amount increasing significantly assuming a 75th percentile (or more extreme) community-building outcome. For reference, Givewell’s cost-effectiveness analysis of Deworm the World in Kenya suggests that each philanthropic dollar spent leads to an increase of 0.11 “units of value”. Therefore, raising $7.9k leads to an increase of ~869 units of value (equivalent to preventing approximately 10.5 deaths of an individual 5 or older from malaria per year). Increasing the number of effective hours worked also seems very important for pushing ahead on important EA cause areas, especially if many of the 80k cause areas are talent-constrained.
Second-order effects are important, particularly for community building. Second-order effects (sometimes known as “multiplier effects”) are pivotal to fully capture the impact of community building. They can often triple (or more) the estimated first-order effects.
Aside: how would a community builder go about envisioning their impact?
We have estimated the number of engaged EAs that a community builder can recruit per year. The estimates above correspond to my thoughts about what a full-time community-builder would be able to achieve per year if they were only devoted to that one activity. Of course, community builders can mix between activities (e.g. spend 30% of the time working on a university group and 70% of the time organizing 1-1 events.) If a community builder wants to estimate their impact, I propose using something like the following formula:
Impact=λ×∑activities% time spent on activity×impact of activityImpact is the sum of all activities that a community builder helps organize multiplied by the time spent on the activity and a coefficient λ. I see λ as a coefficient which measures cross-pollination effects from participating in different activities, which would increase the efficacy of an individual’s community building effects (which would increase λ) or a distraction effect from not focusing attention on one thing (which would decrease λ).
For instance, if I wanted to measure the (first-order) number of engaged EAs that a 50th percentile community builder can attract per year with a 30-70 mix between organizing a uni group and 1-1 events, I can arrive at engaged EAs per year≈λ×((0.3×4.4)+(0.7×31))=23.02λ EAs per year. If participating in multiple activities gives the community-builder a strong cross-pollination effect, I could assume λ=1.1, which would mean that the community-builder can attract approx 23.02×1.1≈25.3 EAs per year.
Caveats and limitations
I list some caveats and potential limitations of my estimates below.
Subjective. The estimates are largely based on my assumptions and prior credences as of 2022. You may have different credences based on your own information or think that my assumptions are systematically biased, could be too optimistic, or too pessimistic.
Outcomes may not be useful. I’m very specifically focused on the increase in donations and increase in effective time spent on 80k cause areas. It’s plausible that there are more important outcomes or more relevant outcomes for the community builder than the outcomes that I have analyzed.
Unspecific outcomes. The outcomes that have been analyzed are too unspecific and could be meaningless (e.g. the concept of an “effective hour” worked on a cause area may not be useful). Outcomes such as “effective hours” are also subjective and may not equate directly to progress on solving the problem (e.g. more time spent on AI alignment does not necessarily equate to more progress made on AI). It could be the case that effective time worked on cause areas has greatly diminishing returns, which is not accounted for within this model.
You might also think that the impact might be a lot larger for students who decide to work on AI safety rather than other cause areas and treat the conversion to “priority cause areas” (e.g. AI safety/biosecurity) as a more important metric.
The concept of an engaged EA is meaningless. You may not think that there is a clear delineation between an “unengaged EA” and an “engaged EA”, or that the difference is not binary and should be further nuanced. The quantitative estimate that I have provided above does not suitably capture the nuances of this sliding scale.
Short-term. One could also consider that the thinking behind this quantitative estimate could be considered very short-term, and the relevant outcomes that I have put forward today may be very different in the future as the EA movement grows further. This particularly seems relevant for the estimates for increase in donations, because the estimates do not account for the donations of students in the future and only account for the increase in donations when a local professional group is organized that moves people from being an unengaged professional EA to an engaged professional EA.
The number of EAs does not matter. You might think that getting more people into effective altruism doesn’t matter. For instance, some new EAs may quickly lose interest or some may be very dedicated and help recruit even more students next year, so the statistic itself is not useful for making estimates or analysis.
No account for increasing/diminishing marginal returns. The model does not account for increasing/diminishing marginal returns (e.g. an EA going to 500 1-1s per year would probably derive much less value than an EA going to five 1-1s per year). It only accounts for the average effect of organizing a particular activity.
Second-order effects are poorly estimated. The second-order effects may be very understated, particularly in the long term. It also does not consider further nth-order effects. As a result, it would be a good idea to treat this model as negatively biased. Alternatively, one could think that the second-order effects need to be further discounted because the community-builder might not actually be accountable for the second-order effects. A potential point of improvement could be to further reduce the Shapley values for the second-order effects so that the second-order effects are further discounted.
High variance. The quantitative estimates vary too much to be useful (i.e. there is too much uncertainty). We might need to gather more quantitative data to be more well-informed about outcome of community building.
Poor external validity. The model may not be useful because the reasoning is not valid across different contexts (e.g. some groups may not run introductory fellowships etc.)
Conclusion
I have provided a quantitative estimate of the impact that an average community builder can expect by working on further community building activities. I hope that this report is useful for those who are considering community building activities or for community builders who are thinking about ways to further enhance their impact.
Sources
Rethink Priorities — EA Survey 2020: demographics
Rethink Priorities — EA Survey 2020: how people got involved in EA
Rethink Priorities — EA Survey 2020: donation data
80,000 Hours careers guide for building effective altruism
80,000 Hours: is effective altruism growing?
GiveWell cost-effective analysis 2021
For the graphs on the right of the Squiggle notebook, several outcomes have two graphs assigned to them. The first graph (“0”) represents the first-order effect only, while the second graph (“1″) represents both first-order and second-order effect.