Political advocacy is an important lever for reducing existential risk. One way to make political change happen is to support candidates for Congress.
In October, Eric Neyman wrote Consider donating to Alex Bores, author of the RAISE Act. He created a cost-effectiveness analysis to estimate how donations to Bores's campaign change his probability of winning the election. It's excellent that he did that—it's exactly the sort of thing that we need people to be doing.
We also need more people to check other people's cost-effectiveness estimates. To that end, in this post I will check Eric's work.
I'm not going to talk about who Alex Bores is, why you might want to donate to his campaign, or who might not want to donate. For that, see Eric's post.
The basic structure of Eric's model:
The model specifically looks at the primary for New York's 12th Congressional district. It doesn't look at the general election because the district is deep blue and whoever wins the Democratic primary will almost certainly win the election.
I reproduced Eric's model using Squiggle.
Before getting into the numbers, my take on this model is that it's very reasonable. Some simplifying assumptions had to be made to make the model tractable, and I fully agree with all of Eric's choices in that regard. When reproducing the model, I only make one small change that (I think) didn't affect the final numbers at all.[1]
Some simplifying assumptions that the model makes:
My only high-level critique is that the original model used point estimates instead of credence intervals for the input parameters. For my Squiggle version, I converted most inputs into credence intervals using my own judgment about each parameter's uncertainty.
(To be fair, doing a cost-effectiveness estimate with credence intervals is a lot more work if you're not using a tool like Squiggle.)
Eric's model has seven input parameters:
(Eric talked about all of these parameters in more detail in his post, although they were split across a few sections.)
I will go through the values Eric gave for each of these parameters and if I have disagreements. Then I will do a sensitivity analysis.
Eric Neyman assumed a typical campaign costs $100 per vote based mainly on "numbers thrown around casually by experts", then multiplied by 3 because New York has higher costs than average.
I spent 15 minutes looking for literature and I found:
Given the information I found Eric's guess of $100 for the average election seems reasonable to me. Raising this to $300 for New York elections sounds about right to me. But I used a wide credence interval, with my 75th percentile estimate being 10x higher than my 25th percentile.
I believe it would be possible to come up with a more confident estimate with another 5–10 hours of work. If I wanted to improve this cost-effectiveness estimate, that's where I'd start.
According to Ballotpedia, the New York 12th District primary elections had about 90,000 voters in 2020, 2022, and 2024. So 90,000 is a reasonable estimate for voter turnout.[7]
Eric modeled the margin of victory as following a uniform distribution from 0-30%, on the assumption that near the beginning of a campaign, it's very hard to predict how close an election will be. I think that's reasonable and that's how I would have done it.
(A 30% margin means that, e.g., the top candidate gets 55% of the vote and the #2 candidate gets 25%.)
Eric described a second model where he treated candidates' votes as following a Dirichlet distribution. This alternative model got approximately the same answer. I didn't attempt to replicate it; I agree that it's more accurate to reality, but I don't think a Dirichlet distribution adds enough value to justify its complexity, so I just modeled the distribution as uniform.
There are currently three candidates in the race; there are two spots in the top two; therefore there's a 2/3 chance that Bores is in the top two. This is a very simple line of reasoning and I have no objection to it.
If your candidate would win without any additional funding, then additional funding doesn't help. Donations only matter if they would lose otherwise.
There's a 50% chance that your candidate is on the losing side, conditional on the election being close.
Eric applied a 10% discount based on the possibility that if Bores raises more funding than expected, then the AI anti-regulation super PAC will donate more money to his opposition. That discount seems too low to me, but I don't have any evidence about what the right number would be. (My model still used a credence interval centered on a 10% discount.)
I think the probability that Bores-funding induces anti-Bores-funding is pretty high, but I also think super PAC spending is less valuable than individual-donor spending due to campaign funding restrictions (as I understand, super PACs can pay for ads, but they can't directly advertise for or against particular candidates).
Eric expects that early campaign fundraising consolidates party support—it makes it easier to get more endorsements, raise more money from funders who don't want to back a losing candidate, etc. He estimates that early funding is 2x as valuable. I didn't do any research on this, but 2x sounds reasonable to me. I converted Eric's point estimate into the 50% credence interval [1.33, 3].
Four of the inputs have relatively narrow credence intervals: voter turnout, probability that your candidate is in the top two, and probability that your candidate is on the losing side.
Margin of victory is based on a coarse assumption of uniform probability, but I don't think there's much value in adding complexity to this parameter.
Two parameters, the opposition fundraising discount and the early fundraising multiplier, are completely made up. These are the #2 and #3 most important parameters (but not necessarily in that order). But I don't actually think the credence intervals are that wide—I don't think their 50% CIs span a factor of 10.
By far the most important parameter is the cost per vote changed. My 50% credence interval for this parameter does span a factor of 10.
That's why I think the best way to improve this model would be to spend more time figuring out the cost per vote changed. The simple version is to come up with a more well-researched number for the cost-effectiveness of campaign spending. A more sophisticated implementation could attempt to model the rate of diminishing returns to spending and apply that to where the Bores campaign is at currently.
Eric gave a 50% credence intervals of "something like [$40k, $170k]" for donations made specifically on October 20. Based on the other things he said, I'd infer that his 50% CI for donations in 2025 (but after October 20) is [$49k, $210k]. To my knowledge, he did not explicitly model credence intervals for input parameters.
My replication finds a 50% CI of [$36k, $380k], which is notably wider, spanning 11x compared to Eric's 4.3x.
This cost-effectiveness model estimates the expected cost to change the outcome of the election. That's not what we ultimately care about. What we really care about is the expected cost to prevent AI extinction via donating to political candidates. That number is much harder to estimate. But it's still nice to have a model that gets you part of the way there.
For a cost-effectiveness to go all the way, it would need to model how representatives affect what AI safety legislation gets passed, and how that legislation decreases x-risk. That's a good question for another day.
Namely, Eric's model estimated the probability that the vote margin falls within 1000 votes, and then used that plus the expected voter turnout to estimate the probability that the margin is within one percentage point. My reproduction used voter turnout to directly estimate the probability that the margin is within one percentage point. ↩︎
Le, T., Onur, I., Sarwar, R., & Yalcin, E. (2024). Money in Politics: How Does It Affect Election Outcomes?. ↩︎
Bombardini, M., & Trebbi, F. (2007). Votes or Money? Theory and Evidence from the US Congress.. ↩︎
Eric also notes that this study looked at general elections, not primaries, which are probably more expensive to influence because there are relatively fewer undecided voters. ↩︎
Gerber, A. S. (2004). Does Campaign Spending Work?. ↩︎
Without having carefully read the paper, I'm more inclined to trust the methodology if it's coming from Poverty Action Lab than if it's coming from some author I've never heard of. ↩︎
One thing that puzzles me is that the 2018 turnout was only 45,000, and the 2016 turnout was 17,000 (!). I don't know why voter turnout changed so much in only four years, and then barely changed for the subsequent four years. I thought perhaps it's because New York changed its districts, but the last redistricting was in 2012. So I have no idea what caused this sudden change in turnout, and I can't rule out that it won't happen again for the 2026 election. ↩︎