iiinova — where your digital self keeps evolving

It appears the Scots and the Irish (or at least the northerners) still use it, which may explain his word choice and maybe this excerpt

030

Cory McCartan @corymccartan.com · 6h

It goes way back to old English I believe. Heaney uses it in the opening lines of his Beowulf translation. Have never seen it used in a modern context.

160

Reposted by Cory McCartan

Diana B. Greenwald @dbgreenwald.bsky.social · 28/09/2026

Attaching some random screenshots but could’ve done this across hundreds of city blocks or stretches of road.

Satellite image of complete destruction of buildings in southern Gaza.

022

Reposted by Cory McCartan

Max Kennerly @maxkennerly.bsky.social · 20/09/2026

"This election is about the threat of trans kids" is the worst possible messaging for Dems yet so many candidates are leaning into it. No voter will flip R->D over it. Low-info voters read this as "you're focused on bullshit that doesn't affect me" while the base reads this as "you're an asshole."

12457561514

Reposted by Cory McCartan

Jed Kolko @jedkolko.bsky.social · 17/09/2026

The 2025 American Community Survey should have come out today. But it has been delayed because of unusual political encroachment at Census. ACS drives business and policy decisions, especially in local communities, and is the primary input to Census’s immigration estimates.

118280

Reposted by Cory McCartan

Bobby Kogan @bbkogan.bsky.social · 17/09/2026

Have been incredibly busy these past two months, but finally really turning my attention back to my upcoming appropriations paper. Some of the data are not 100% done yet, but I want to focus on how much domestic discretionary funding has been cut in the last few years. Posted only on Bluesky.

24717

Reposted by Cory McCartan

Cyrus Samii @cdsamii.bsky.social · 16/09/2026

Half of our seminars should be pre analysis plans and the other half results. Imagine if people had an up front stake in your research, and imagine getting great input before you finalize your analysis strategy. It should be the norm.

1122

Cory McCartan @corymccartan.com · 16/09/2026

Yep, the skewed ones are awful either way—though they are quite skewed! I think often the thing we apply the CLT to (e.g. likelihood scores) is rarely that skewed. Even the 95th sample quantile of these distributions is far closer to normal. Just thinking through what this all means in practice

020

Cory McCartan @corymccartan.com · 16/09/2026

Cool! Seems to me the fairer comparison for coverage is to use t intervals, because the undercoverage there isn't about non-normality of the mean but rather the need to estimate the variance? With t intervals, most of these hit 92-3%+ coverage with fairly small n, except the super skewed ones

110

Cory McCartan @corymccartan.com · 15/09/2026

NYT poll of LV shows D+9 generic ballot ±3.2. Plugging that into my simple congressional swing model (tinyurl.com/cmchousemodel) gives 241D–194R, with 94% chance of D win. Effect of redistricting in that environment is R+4 seats, with TX and FL doing the heavy lifting, and NC slightly backfiring

010

Reposted by Cory McCartan

Mark Joseph Stern @mjsdc.bsky.social · 15/09/2026

Others have said this but it’s truly insane that Brett Kavanaugh, lover of the major questions doctrine, thinks there’s a good chance USPS can use a vague grant of authority over mail processing to radically alter absentee voting procedures in all 50 states without congressional approval. Come on.

1268383

Cory McCartan @corymccartan.com · 15/09/2026

From Alito dissent on mail ballots (L). How can this possibly be squared with the "major questions doctrine"!?!? (R from WV v. EPA)

010

Reposted by Cory McCartan

jamelle @jamellebouie.net · 06/09/2026

the other strange thing about this is that the 2024 election result was exceptionally narrow and yet commentators left and right talk about it like it was 1984

361443127

Reposted by Cory McCartan

Anton Strezhnev @astrezh.bsky.social · 02/09/2026

This is cool but I think an over-use of RD - CEF really looks like it curves at the cut point. But why do we expect a jump? Impact is probably higher for internal locations - so the relevant treatment intensity isn’t at 0.

161

Cory McCartan @corymccartan.com · 01/09/2026

He literally goes around with this slide. The same 'proof' would imply that no stationary AR(1) process with unbounded increments can exist. Error probability is not constant!

24310

Reposted by Cory McCartan

jamelle @jamellebouie.net · 01/09/2026

you could basically predict the direction public opinion under trump if you took the view that a) the 2024 election was about inflation and b) "wokeness" wasn't especially salient nor was it especially unpopular.

2384486

Reposted by Cory McCartan

Amanda Weiss @amandaweiss.bsky.social · 27/08/2026

Ofc I say this and personally put WAY too much effort into reviews. And then I die inside when my papers spend six months under review or get torpedoed because someone was not on board with my type of paper existing or are evaluated by someone who doesn't know how math works. So there's that...

151

Reposted by Cory McCartan

AFL-CIO @aflcio.org · 25/08/2026

America’s unions are proud to endorse @sbworkersunited.org’s boycott of stand-alone corporate Starbucks stores. Workers are fed up with unchecked corporate greed, and we’re standing with them in their fight for a fair contract. No contract? NO COFFEE! ✊ www.boycottstarbucks.net

Boycott Starbucks. No Contract? No coffee. Period. BoycottStarbucks.net

315377

Cory McCartan @corymccartan.com · 24/08/2026

FWIW, much of the "don't worry about it" take on data centers was premised on a more ordinary process + community relationship. Clear now with e.g. building fossil fuel plants there are real local externalities (not just water!), and fine to be angry when govts sign NDAs, etc. to avoid scrutiny

020

Cory McCartan @corymccartan.com · 24/08/2026

Data center opposition seems a pretty rational strategy to me even though I may not be totally on board. If you are feeling threatened + angry at new tech, why not contest it on turf you can historically win on & do so locally (NIMBYism) rather than directly try to fight a new technology (hard)?

110

Reposted by Cory McCartan

Ben Williamson @benpatrickwill.bsky.social · 18/08/2026

What's most upsetting me as a journal editor is that many papers we receive where AI use is disclosed for editing and language refinement actually appear to have been ruined rather than improved by AI. We get what seem like promising underpinning studies, but totally obscured by over-complex text.

8518116

Cory McCartan @corymccartan.com · 16/08/2026

Very concerning the NYT killed this excellent piece at the behest of the scammy firms it exposes

0286

Reposted by Cory McCartan

Melody Huang @melodyyhuang.bsky.social · 13/08/2026

Really excited to see that my paper with Erin Hartman on assessing non-ignorable nonresponse in survey weighting is (finally) out! doi.org/10.1093/poq/...

1141

Cory McCartan @corymccartan.com · 12/08/2026

was confused, checked their website. total votes mistranscribed as total write-ins

000

Reposted by Cory McCartan

Etche_homo @etche.bsky.social · 07/08/2026

Corroborated by my own experience with other Big Publisher journal titles Yet we're more impressed on evaluation committees by 1st-author candidates for [things] who made it through the gauntlet. Let's vow to check that impulse & turn to our professional society titles when possible & independent.

001

Cory McCartan @corymccartan.com · 07/08/2026

FWIW, I strongly do not recommend publishing with NHB. A very slow review process and a ton of time-consuming administrative steps post-review and post-acceptance, including micromanaging of language & outdated statistical rules that especially don't make sense for methods work. And OA is $$$!

041

Cory McCartan @corymccartan.com · 07/08/2026

Basic idea is to look at who doesn't get to elect their chosen candidate, but would have under a fair counterfactual plan. That difference = harm, which can be aggregated by & compared across different groups.

131

Cory McCartan @corymccartan.com · 07/08/2026

Out yesterday at Nature Human Behaviour is my paper with @chriskenny.bsky.social on a unified way to measure individual & group harms in redistricting! Due to extortionate APC fees (~$13k) we did not publish open access, but you can read for free at rdcu.be/fyd95 !

rdcu.be

Individual and differential harm in redistricting

Nature Human Behaviour - McCartan and Kenny develop a framework for quantifying the individual-level impacts of redistricting in the USA, showing its applicability for identifying partisan...

1101

Reposted by Cory McCartan

Kevin Morris @kevintmorris.bsky.social · 06/08/2026

Happy birthday to the Voting Rights Act. Complicated feelings on this particular anniversary.

3376

Reposted by Cory McCartan

Ryan Enos @ryanenos.bsky.social · 05/08/2026

Don't posit a theory that is only interesting because it implies X causes variation in Y, test for a statistical relationship between X and Y, but then say you are not interpreting causally. Of course you are interpreting it causally. That was the whole point of what you just did.

25310

Reposted by Cory McCartan

Dan Simpson @danpsimpson.bsky.social · 04/08/2026

Mechanically, what this means is that it’s assuming fairly smooth interactions between covariates. So if you want flexible models of interaction you have to build them in yourself. Which is, after all, what you’d expect with BART.

041

Reposted by Cory McCartan

Dan Simpson @danpsimpson.bsky.social · 04/08/2026

This is definitely interesting. It’s probably worth noting in high dimensions that mixed sobolev spaces are much smoother than usual sobolev spaces. So think of this kernel as closer to a squared exponential than a Matern-1. But still rough enough that the paths aren’t analytic.

271

Reposted by Cory McCartan

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

Fascinating and clear paper by @corymccartan.com and @melodyyhuang.bsky.social, greatly enhancing our understanding of how Bayesian Additive Regression Trees (BART) works and why it is so effective. A must-read for my fellow BART enthusiasts. #statssky #causalinference

13911

Cory McCartan @corymccartan.com · 03/08/2026

We have lots more in the paper, including connections to existing BART literature, random features lit, and seemingly unrelated methods like HAL and HAR (which achieve the same rates) Comments welcome! arxiv.org/abs/2607.28844

arxiv.org

Seeing the Forest for the Trees: The Gaussian Process Limit of BART

Bayesian Additive Regression Trees (BART) have shown state-of-the-art performance in both prediction and causal inference problems. Previous theoretical work has attempted to explain BART's superior p...

020

Cory McCartan @corymccartan.com · 03/08/2026

Most importantly, because random tree features are just a (random) basis expansion of your data, you can stick them inside any other model with a linear predictor, including alongside structured terms like fixed/random effects, spatial regression terms, inside hazard models, etc!

100

Cory McCartan @corymccartan.com · 03/08/2026

We also show random tree features have similar uncertainty quantification to full BART, despite not learning the tree structure

100

Cory McCartan @corymccartan.com · 03/08/2026

Enough theory—does this work in practice? We show random tree features are competitive with full BART, random forests, and xgboost, and lie along the Pareto frontier of accuracy & computation

100

Cory McCartan @corymccartan.com · 03/08/2026

Turns out (with some conditions) you only need n^(1/3) trees to achieve the optimal learning rate! Notably this rate is faster than existing rates for BART theory, which suffer from the curse of dimensionality* *not quite this simple, but basically

100

Cory McCartan @corymccartan.com · 03/08/2026

OK, but this is for infinitely many trees, right? What do we do in practice? We propose "random tree features," which are a BART model where the tree structure isn't learned

100

Cory McCartan @corymccartan.com · 03/08/2026

We can characterize the RKHS (function space) that BART's kernel corresponds to. It lies in between a weak class (just requiring 1 derivative; slow learning rates) and a smooth class (p derivatives; fast-ish rates), and has a nice minimax learning rate that depends only logarithmically on dimension!

100

Cory McCartan @corymccartan.com · 03/08/2026

The BART kernel is anisotropic, a.k.a. not rotation invariant, which means intuitively it prioritizes main effects over interactions (more about this below)

110

Cory McCartan @corymccartan.com · 03/08/2026

In fact, we show formally for the first time that as the # trees grows, BART converges to a Gaussian Process with a kernel we can describe in closed form! This means (a) no tree learning happens in the limit, and (b) we can study the kernel to learn things about the types of functions BART learns

100

Cory McCartan @corymccartan.com · 03/08/2026

We first try knocking out different parts of the BART model to see what matters most. Turns out, at least when you have many trees, it's (mostly) not being Bayesian, nor learning the tree structures, nor even the flexible part of trees per se

100

Cory McCartan @corymccartan.com · 03/08/2026

New WP w/@melodyyhuang.bsky.social studying the success of BART models, which regularly win causal inference competitions! We argue that BART should be thought of as a random features approximation to a limiting GP. This view helps understand BART & apply it in more places arxiv.org/abs/2607.28844

Seeing the Forest for the Trees: The Gaussian Process Limit of BART

Cory McCartan & Melody Huang

Abstract:
Bayesian Additive Regression Trees (BART) have shown state-of-the-art performance in both prediction and causal inference problems. Previous theoretical work has attempted to explain BART’s superior performance by establishing posterior contraction rates for standard BART models, but these rates depend strongly on the number of covariates. Here, we take a different approach and study the behavior of BART as the number of trees grows towards infinity. We show that in this regime, BART converges to a Gaussian process (GP) with a particular kernel. The kernel and its corresponding reproducing kernel Hilbert space (RKHS) have favorable inferential properties that help explain BART’s excellent performance. We introduce random tree features as an approximation to this limiting GP, and establish minimax-optimal learning rates for ridge
regression on these random features that depend only logarithmically on dimension. In addition to providing insight into the empirical success of BART, random tree features offer a computational benefit over traditional MCMC estimation. The random-features approximation also allows
practitioners to easily incorporate BART into any model which has a linear predictor, expanding the applicability and flexibility of BART.

192

Reposted by Cory McCartan

ArXiv Paperboy (Stat.ME+Econ.EM) @paperposterbot.bsky.social · 03/08/2026

arXiv📈🤖 Seeing the Forest for the Trees: The Gaussian Process Limit of BART By McCartan, Huang

Bayesian Additive Regression Trees (BART) have shown state-of-the-art performance in both prediction and causal inference problems. Previous theoretical work has attempted to explain BART's superior performance by establishing posterior contraction rates for standard BART models, but these rates depend strongly on the number of covariates. Here, we take a different approach and study the behavior of BART as the number of trees grows towards infinity. We show that in this regime, BART converges to a Gaussian process (GP) with a particular kernel. The kernel and its corresponding reproducing kernel Hilbert space (RKHS) have favorable inferential properties that help explain BART's excellent performance. We introduce *random tree features* as an approximation to this limiting GP, and establish minimax-optimal learning rates for ridge regression on these random features that depend only logarithmically on dimension. In addition to providing insight into the empirical success of BART, random tree features offer a computational benefit over traditional MCMC estimation. The random-features approximation also allows practitioners to easily incorporate BART into any model which has a linear predictor, expanding the applicability and flexibility of BART.

032