iiinova — where your digital self keeps evolving

If you work on very large and very sparse fixed effects problems as in worker-firm panels, or doctor-patient relations, we'd love to learn if it works for you / speeds up your regressions!

001

There is also long-form documentation of the new algorithmic strategy github.com/py-econometr... plus a new vignette in which we try to explain why our initial "vanilla MAP" algorithm struggled so much on "hard" fixed effects problems: pyfixest.org/explanation/...

100

You can find the repo here: github.com/py-econometr...

100

You can access the new algo via the `demeaner_backend = "rust-cgs"` option of `pf.feols()`, and we have published it in a standalone rust crate (including Python bindings. R bindings are WIP and we could use some help! Even more with Stata).

101

Fixed effects are 'sparse' if each observation belongs to only a few groups. In a worker firm panel, this means that most workers only work at a single firm and (almost) never change their employer.

100

Alex @s3alfisc.bsky.social · 28/02/2026

I am very excited that PyFixest 0.50.0 is on PyPi, including a new graph-based solver for demeaning that makes fixed effects estimation in PyFixest significantly faster for "sparse" fixed effects structures.

1157

This looks really good! Would you be open to me adding a pyfixest fitting skill (😇)? Also, have you found a good way to make the docs more LLM accessible, i.e. via providing md twins for each html page / providing an llms.txt / do you think it might be valuable? github.com/NVIDIA/sphin...

100

Alex @s3alfisc.bsky.social · 25/02/2026

MacKinnon Nielsen Webb have also suggested a "fast" cluster jackknife - I had implemented it a long time ago in `summclust`. Description of the ago here: s3alfisc.github.io/summclust/ar...

130

Alex @s3alfisc.bsky.social · 22/02/2026

In Kopenhagen schmeckt der 6 Euro Kaffee aber immerhin ...

000

Alex @s3alfisc.bsky.social · 21/02/2026

PyFixest Sprint We will be doing a 3-day PyFixest sprint in two weeks - are there any features / regression-based methods widely useful to applied researchers you'd like us to explore/ add?

Alex @s3alfisc.bsky.social · 19/02/2026

Also shamelessly plugging @mchow.com as I am wondering if this potentially already available with quartodoc?

000

Alex @s3alfisc.bsky.social · 19/02/2026

@jacobtomlinson.dev has been working on sphinx llm - core idea: provide a .md file for each page of your docs. Does this really help LLMs navigate the documentation? And what about an llms.txt? If yes, what are best practices? github.com/NVIDIA/sphin...

github.com

GitHub - NVIDIA/sphinx-llm: LLM extensions for Sphinx Documentation

LLM extensions for Sphinx Documentation. Contribute to NVIDIA/sphinx-llm development by creating an account on GitHub.

100

Alex @s3alfisc.bsky.social · 18/02/2026

For me it might be "most fixest installs on CI via r2u" 😀

Alex @s3alfisc.bsky.social · 18/02/2026

This is very funny, is it possible to send kudos after burning through a lot of tokens? 😂

(Ok, FixedEffectsModels.jl does pretty awesome as well. And we still have quite some catchup work to do with pyfixest =) )

000

If you fit hdfe regressions with complex fixed effects structures, fixest is unquestionably the goat in terms of performance.

100

Plus, the performance, which you get via a highly optimized MAP algorithm (with local IT acceleration, global smoothing, and a range of internal tricks and heuristics, see here for lrberge.github.io/fixest/refer...) and all the perf improvements that multiple estimation syntax allows.

I could of course go on here... for more ideas in API design, read the paper!

100

- Then of course you get regression tables out of the box with etable. - And you can "bring your own vcov's" functionality with easy plug ins. - We get a new operator for interacting variables with i(var1, var2) (very under appreciated imo!) - Last, my favorite feature: multiple estimation syntax!

100

To point out just a few innovations on the API side that imo set the standard for regression packages. You get - separation of estimation and inference - and you can adjust inference post-estimation "on the fly" - a lot of handy post-estimation processing: coef, se, iplot, predict, and much more ..

100

Alex @s3alfisc.bsky.social · 28/01/2026

Strong agree. And do read the paper! It illustrates very nicely how much thought and creativity has gone into designing fixest's API and performance optimizations which are really the two things that make fixest as great.

240

The neat thing is that this idea generalizes to regression and is worked out in the paper by Wong et al listed in the dbreg docs.

030

Alex @s3alfisc.bsky.social · 28/01/2026

You are left with a 2x2 "compressed" data matrix of "sufficient statistics" and "weights", which here is just the number of treated and untreated observations. You can now compute an ATE (as sum_Y(1) / w(1) - sum_Y(0) / w(0)) and also SEs.

140

Alex @s3alfisc.bsky.social · 28/01/2026

Core idea: say you want to compute the impact of an RCT on Y. You only need: - the sum of Y in test and control sum_Y(D) - the sum of squares of Y in test and control - and store the number of treated and non treated units. w(D) You can get all of it in a simple group by and sum evaluation.

120

Reposted by Alex

Grant McDermott @gmcd.bsky.social · 26/01/2026

Things are grim. But in more frivolous news... @jamesbrandecon.bsky.social and I have been chipping away at `dbreg`, a 📦 for running big regression models on database backends. For the right kinds of problems, the speed-ups are near magical. Website: grantmcdermott.com/dbreg/ #rstats [1/2]

grantmcdermott.com

dbreg

46815

Alex @s3alfisc.bsky.social · 06/01/2026

Cool! I have many question: Is this in Python and do you have a GitHub repo and would you be up to contribute to moderndid? 😀

Alex @s3alfisc.bsky.social · 06/01/2026

I'll take a close look, thanks!

010

Alex @s3alfisc.bsky.social · 05/01/2026

Very flattering (and exciting!) to see that Claude Code, even though prompted to use other packages, decided that it was a good choice to fit a TWFE regression in Python via PyFixest =)

060

Alex @s3alfisc.bsky.social · 04/01/2026

I also like the `event_study` API @kylefbutts.bsky.social introduced (and that we also have in pyfixest), though I think this is mostly a functional wrapper around the att_gt class?

140

Alex @s3alfisc.bsky.social · 04/01/2026

I've been wondering if pyfixest syntax is ideal for the new DiD estimators - what moderndid does seems to make very good sense to me - first there is a model class call, and then users can specify aggregation levels of ATTs. This is more or less the same API that the R did package uses:

131

If you want to get started with OSS and have questions, please reach out! I'm happy to help wherever I can =)

100

As all packages are developed by volunteers, please consider helping out with writing code or documentation - the hurdle for contributing to OSS has never been lower. Plus, you might learn a thing or two about software engineering best practices, which will be a useful skill to have in the future!

- mlsynth for synth controls (github.com/jgreathouse9...) - moderndid & PyFixest for DiD - rdrobust (which needs better documentation and maybe a cleaner API) github.com/rdpackages/r... - @apoorvalal.com 's torchonometrics for metrics on the GPU: github.com/apoorvalal/t...

Btw - the Python open source applied metrics tooling is growing rapidly. We now have: - PyFixest for HDFE regressions (github.com/py-econometr...) - marginaleffects for marginal effects (github.com/vincentarelb...) - maketables for regression tables (github.com/py-econometr...)

131

If you are aware of any other high quality implementations of "modern" estimators in Python, please let us know!

111

From PyFixest, we plan to contribute - DiD2s a la Gardner - Local Projections as in @arindube.bsky.social - Fully Interacted Estimators as in Sun&Abraham Plus, hopefully, the imputation estimator from github.com/jsr-p