art of rally
08.03.2026 22:53
👍 4
🔁 0
💬 0
📌 0
art of rally
I lucked into an RA position because I was a poli sci student who could code in Python in 2009 - and the rest is history, but do wish I had gotten more of the econometrics exposure as an undergrad.
A very familiar SFS experience!
Curious whether the LLMs that are taking surveys now exhibit more demand effects than actual humans!
I can’t not hear “Markwayne Mullin” in the same cadence as “Darkwing Duck”
We are seeking submissions for a special issue of the Forum, "Political Science and/or Fiction" (yes, the "and/or" is a pun). See attached for details. Submissions due 6/30/26.
the problem with aughts critiques of games like CoD is that they all assumed that the single player campaigns were the most culturally influential forms of the medium
For teaching, I’m on team “the CI is just the inverted hypothesis test” - not a lie and connects to a thought experiment that’s easier to talk about - www.the100.ci/2024/12/05/w...
I’m kind of more optimistic about agents doing *custom software* eliminating the need for these wrapper packages. You can give an agent a methods paper and it’ll implement the method - you can read the code yourself.
The software stuff worries me because one of the biggest fail cases I’ve seen in published papers is authors not knowing the weird quirks of what their estimation software does and making mistakes. I’d be concerned now that the *maintainers* also don’t know!
I’m increasingly moving towards “the text vs. genetic version of the entity in Annihilation”
I prefer this way of thinking as well - the extension beyond like Wikipedia is that these tools are surprisingly good at **interpolation** in the "information space" - and it turns out a lot of knowledge work takes the form of "gap filling"
"just branding" (neutral/exhausted)
And like - it's just branding in the "credibility revolution" designs too.
Quasi-experimental is just branding IMO.
Counterpoint: This is actually correct in terms of what the underlying selection-on-observables assumption actually means.
The problem is that
a) They didn't control for all the differences
b) They don't have a coherent and well-defined intervention.
“Wait, our journals are going to be flooded with submissions and we don’t have the capacity or resources to distinguish what’s actually good?”
*astronaut meme*
🚀 New working paper - joint with S. Oberhänsli:
arxiv.org/abs/2602.23877
We propose a DiD approach to mediation analysis that evaluates direct, indirect, & dynamic treatment effects under conditional parallel trends, using double machine learning for flexible, data-driven covariate control.
Back on the first point, I think perhaps it's a good thing that PIs who have been treating their grad students as postdocs as work horses could have their AI now. Perhaps they should stop working w research assistants for the net benefits of everyone involved and the good of the research culture.
The things people offer as examples of LLM disruption in academia often strike me as insights about the way things already work. E.g. “Now we can mass-produce mediocre papers” or “Now the value of an RA is not that they do the grunt work but that they bring a different perspective.”
Yeah, re-reading I think that’s right/implied. Apologies for being a bit on edge :)
When I add the headers and ask it to install and run the package, it does so, but gets *extremely insistent* that the manual method it wrote is fine and replicates the package exactly and so this is a waste of time.
Fun quirk of the LLM bot is that it will often manually code up whatever method I’m asking for b/c the package couldn’t compile since the headers aren’t in the container (Raspberry Pi)
I don’t have clear intuition about whether cellphone penetration would modify treatment effects (density as a correlate?) and my priors are tech-pessimistic, but PT is honestly more fragile than we’d like.
No comment on the actual paper, but 2 is a slightly wrong intuition. “Common shock” interpretation in DiD isn’t just about exposure but also effect homogeneity. Even if areas similarly exposed, different Fx would -> violate PT.
It’s not clear to me that’s the audience of your blog, but alright.
I just think we’re better off just *doing the thing* instead of shadowboxing in some weird post-Twitter internet tech culture war.
I love working papers and blogs! Most of my best papers are unpublished working papers :)
No, I’m saying that *these* sorts of AI navel gazing blog posts add very little to the conversation. Like - it’s a useful tool, use it and put the research wherever. The meta-discourse is just exhausting.