Question regarding convergence and check_conv_feols functionality

[Skonar90]

Thank you very much for package.

I have a problem with the convergence of the demeaning algorithm using feols. My data is huge (~ 500 million observations) and I am estimating a rather complex model using fixed effects and varying slopes. At a certain point, when I include more fixed effects I receive the following message (I have already decreased the tolerance as it takes roughly 4 hours to estimate this model):

First, I was wondering why the algorithm stopped iterating onwards, despite not reaching the maximum default value of 10,000. Then I found your very useful function check_conv_feols() (https://lrberge.github.io/fixest/reference/check_conv_feols.html) to check whether all first-order conditions are close to 0. Here I do not understand why there seems to be a problem, because all of the maximum absolute values of the first-order conditions are below the tolerance value of (e-04)

Do I missunderstand the check_conv_feols() function? How can I check what leads to the problem of convergence?

On line 1750 the code (https://github.com/lrberge/fixest/blob/6b7967350be4ea9672b4a13ee2febf1b7d8bfcc3/R/estimation.R) for raising the warning can be found. There you use a different function, namely check_conv(). How is this related?

Thank you again for this great package!

[lrberge]

Hi: thanks for reporting. At first sight, the warning message does not match the one of the current version. Are you using the last version?

[Skonar90]

Hi, I think thats the latest build which I am running:

[lrberge]

Thanks. This looks odd indeed: I need some time to investigate.

[Skonar90]

Great, thanks!