probitFE and logitFE implementation about fixest

[lrberge]

On giving it a shot: Fantastic!

On staggered did: I was thinking at least GB and de Chaisemartin & d'Hautefeuille. Then maybe Abraham & Sun, Callaway & Sant'Anna, and Imai & Kim, but for these last 3 I didn't go through the theory yet.

[lrberge]

Hello and thanks for the words! :-)

The short and incomplete answer is yes. The complete answer contains a disappointing but: but not before a year or two.

I really do want to include this kind of tools because so far the raw FE logit/probit are, indeed, hardly useful due to the presence biases. This is however very low priority for me because I don't use these models in my day-to-day research work.

To give you a hint of the stuff I want to finish before coming to that:

• finishing a massive overhaul of the internal demeaning algorithm to accommodante closed form solutions for variables with varying slopes (I should end that soon),
• implementing some staggered DiD tools that have been recently developed (mostly because I'll use them in my papers),
• implementing IV estimations,
• implementing marginal effects, in a way that is completely integrated with the exports.

And since I also need to move on to work that is more important career-wise... well, it'll take some time before I come to it!

Way to go?

It would be fantastic if you could implement it. But it looks like a big job to get close to the Stata implementation. Also I can't guarantee to include it in the package directly--since I'm a bit of a control freak, I need to ensure it gets embedded with other functions (print/exports/vcov and maybe others) and that users can't misuse the new function (and it takes time to check :-( ).
Or maybe differently, what do you think?

Other comments

By the way, do you know how to get clustered standard-error? Because we need the score for each observation, is it well defined in the jackknife? (Or we just compute it at the jackknife-point estimates?) I'm asking because I've never thought about it.

If you do it, I can try to have an (incomplete) marginal effects function ready (although that won't be before a few months :-( ).

Just a comment: Tables 1 and 2 of the Stata paper are unfortunately not really convincing since all results really look the same :-( and it seems hard to choose between the different models. Are there rules of thumbs to choose between them?

[apoorvalal]

Agree re: relatively low-priority for most use cases; I find the marginal effects to be pretty comparable, but some referees are morons.
I'll fork the repo and give it a shot; at the very least it looks like a fun little programming challenge. I'm not really sure about clustering in this setting either.

RE: tasks - those all look very useful. What staggered DiD methods are you looking to implement? Goodman-bacon weights should be relatively easily doable. Also is Chaisemartin WFE in the pipeline?

[lrberge]

Just for the update, I won't implement GB nor CH (2020) nor CS (forthcoming), but Sun and Abraham (forthcoming) will be there in the next version.