This strategy is quite amenable to be run on the GPU - for "hard" problems where the alternating projections algorithm struggles to converge, running the demeaning on a 𝗚𝗣𝗨 can provide orders of magnitude of speedups.

Also works with "negative" effects:

Vis method for decomposition now merged to main, feedback welcome!

My attempt - how to tell you I work in industry without telling you ;) what do you think?

Currently revamping the default output view for the Gelbach Decomposition. Goal is to help both academics but also data scientists in industry to communicate results easily. Are the column and panel headers immediately clear? Any better way to summarize decomp results?

Re fast routines for the quantile regression process - we follow Chernozhukov et al (2019) and their algo's are quite an improvement over "naive" for loops over quantiles:

𝐏𝐲𝐅𝐢𝐱𝐞𝐬𝐭 0.30.1 is on 𝐏𝐲𝐏𝐢 and I am extremely happy with this release, mostly because of two 𝐧𝐞𝐰 𝐟𝐞𝐚𝐭𝐮𝐫𝐞𝐬: a Rust back end for all performance critical parts and quantile regression.

I've started to work on a quantile regression implementation for pyfixest. There's a basic PR that implements the Frisch-Newton solver in numpy + a quantreg interface. Still plenty of things to do (standard errors, performance improvements), so please let me know if you want to contribute!

Fully saturated model support -> you get pretty plots like this out of the box =)

Thanks to @apoorvalal.com, we now have some benchmarks for PyFixest on the GPU: 2-3x improvements for large problems via JAX relative to numba!

We also released pyfixest 0.28.0, mostly with small bug fixes, but it also contains JAX support and GLMs (no fixed effects yet) as experimental features.

Something cool for the next pyfixest release 😊

And multiple testing with pretty power curves at the end of the nb: py-econometrics.github.io/pyfixest/mul...

For Gelbach's decomposition: py-econometrics.github.io/pyfixest/reg...

We have a new notebook that motivates multiple testing corrections and shows how to use them via pyfixest. At the end, it produces this pretty figure: py-econometrics.github.io/pyfixest/mul...

We’re about to finish an implementation of @gelbach.bsky.social‘s regression decomposition for the next pyfixest release.

But there‘s a problem! We only match Stata‘s b1x2 to the 3-4th digits after the comma.

Is this a bug, or just numerical error?

A very simple simulation suggest that it works:

I’ve released a new version of pyfixest - it ships some performance improvements for CRV1 standard errors. I’ve also set up some benchmarks, and I was pleased to see that pyfixest looks quite competitive:

Now on PyPi: pyfixest with a more flexible API for did2s():