Gabor's Data PyFixest Wishlist

[s3alfisc]

I plan to tackle some of these over the holidays:

• no prediction interval -> LOW EFFORT if no fixed effects supported, as PR prepared by @greenguy33, more effort with fixed effects
• R2 for WLS -> LOW EFFORT Issue
• logit support -> HIGH EFFORT. Issue: link. PR in progress here
• probit support -> HIGH EFFORT. PR in progress here
• HAC SEs -> MEDIAN EFFORT Issue
• locally defined functions cannot be used inside the formula (lspline) -> FEASIBILITY?
• ARIMA VAR -> WONT DO

[adamvig96]

Hi @s3alfisc,

I created a poc for context capture aka locally defined functions cannot be used inside the formula (lspline) here: adamvig96@96fa4fc

I think it would be good to have this in pyfixest, apart from the fact that we'd need it in the Gabor's DA book codebase – eg. statsmodels have this functionality. This is just a POC as I'm not familiar with the pyfixest codebase, but it worked for me in some case studies I tested.

FYI @gbekes, we need this for logit and probit too as splines are used extensively there.

[s3alfisc]

Hi @adamvig96 , to make sure I understand the problem correctly - you need formulaic to parse a custom data transform (in this case a spline) and need to provide the extra context via the feols / fepois APIs? If yes I very much agree that this would be a very useful feature to have!

If yes, I took a quick initial look at your code and I think it looks quite good already - I don't think there'd be much to add except maybe some tests (happy to help with that =) ). If you open a PR, we can get started with getting this into pyfixest =)

[adamvig96]

Yes, I think that's an accurate description of the issue. I opened #769 and #770.

[s3alfisc]

Thanks!

[s3alfisc]

I will merge logit + probit tomorrow (I am encountering a strange bug that I currently cannot explain), will then continue with R2 for WLS. Then the prediction intervals, last HAC.

[s3alfisc]

We've just merged R2 for WLS (and within and adjusted R2 too)! PR here, thanks to @AronNemeth. I am also finalizing the prediction standard error PR , I should finish it by tomorrow evening. Then only HAC will be left =)

[s3alfisc]

... and we've merged the prediction errors PR. @adamvig96 @gbekes

[DanielBarabas]

Hi, @s3alfisc, in chapter 17 of the book we would like to show the AMEs of a logit. In the documentation you gave an example using the marginaleffects package that I tried to replicate (in the "Quick reminder about probability models" part of the notebook) but the avg_slopes function simply gives back the original logit coefficients, not the AMEs, as if it was an OLS regression. Have you encountered anything similar? Or do you have any ideas that we chould try?

On another note: in our conda environment we could only use marginaleffects==0.0.11 because of a dependency issue regarding numpy (numba needs numpy<1.27, that comes with pyfixest, while marginaleffects=0.1.0 needs numpy>2.0.0), maybe this causes the issue?

[s3alfisc]

Hi @DanielBarabas , this can indeed be the case!

To create the docs, I am using marginaleffects 0.0.14:

marginaleffects            0.0.14                                     147.1 KiB   pypi   marginaleffects-0.0.14-py3-none-any.whl

But I am also running higher numpy and numba versions?

numba                      0.61.0                                     11.7 MiB    pypi   numba-0.61.0-cp312-cp312-win_amd64.whl
numpy                      2.1.3                                      40.2 MiB    pypi   numpy-2.1.3-cp312-cp312-win_amd64.whl

This is all under pyfixest 0.28.0 and the docs environment. I suppose you are running the same version as you are installing from pypi?

Which version of numba do you have installed?

[DanielBarabas]

Thanks, I will check the numba bc that might be it (when i get to my laptop)