`i()` does not accept two categorical terms

[DanSimonet]

In R fixest I am able to interact two factors while setting the referent for one.

library(fixest)
mtcars$cyl <- factor(mtcars$cyl)
mtcars$gear <- factor(mtcars$gear)

m1 <- feols(
  mpg ~ i(cyl, gear, ref = 4),
  data = mtcars
)

etable(m1)

Result appears as follows with cycle = 4 set as referent for comparison to the combined groupings.

Coefficient	Estimate (Std. Error)
Constant	26.66*** (1.026)
cyl = 6 × gear = 3	–6.914* (2.616)
cyl = 6 × gear = 4	–6.914** (1.987)
cyl = 6 × gear = 5	–6.964. (3.554)
cyl = 8 × gear = 3	–11.61*** (1.420)
cyl = 8 × gear = 5	–11.26*** (2.616)

However, when running the same syntax in pyfixest

import pyfixest as pf
import statsmodels.api as sm
mtcars = sm.datasets.get_rdataset("mtcars", "datasets").data

mtcars["cyl"]  = mtcars["cyl"].astype("category")
mtcars["gear"] = mtcars["gear"].astype("category")

m1 = pf.feols(
    fml = "mpg ~ i(cyl, gear, ref = 4)", 
    data = mtcars
)

pf.etable(m1, cat_template="{variable}-{value}")

I continue to get the following error even if cyl is category and gear is numeric.

ValueError: 
                        The second variable in the i() syntax cannot be of type "category" or "object", but
                        but it is of type category.

I have seen something akin to this issue referenced in the change log and would love a possible solution to reproduce fixest in its entirety. We make heavy use of the i() functionality for a variety of group comparisons and would like to port over to pyfixest to share pythonic versions with colleagues.

[s3alfisc]

Hi Dan, thanks!

The stated behavior clearly looks like a bug, I'll take a look (possible already today).

On interacting two categoricals - this should not be to difficult to add for the i() operator, I'll try to tackle it over the next week.

Please let me know if you need any other features or stumble over any other quirks - I'm always excited to help with pyfixest adoption =)

[DanSimonet]

It's a great package and I really like the extensions (e.g., gelbach decomposition). Even with R's fixest one of the quirks is working around the unique syntax and how base R parses unique symbols (e.g., ^, |). And for better or worse our team is now using all the fixest quirks 😆.

I can say our use case is in effect trying to run gelbach decomposition on interacted terms + inclusion of fixed effects which is not as trivial an issues as anticipated. But anyways appreciate the response and look forward to any updates.

[s3alfisc]

I can say our use case is in effect trying to run gelbach decomposition on interacted terms + inclusion of fixed effects which is not as trivial an issues as anticipated. But anyways appreciate the response and look forward to any updates.

Hi @DanSimonet , haven't gotten around to this yet (still been busy with the quantreg implementation which is now merged). Note that for the moment, you could interact two factors by using formulaic's C() syntax, i.e. do

C(var1):C(var2) 

and you should also be able to set reference levels via the contr.treatment argument to C, similar to what you can do in R (e.g. see here).

[s3alfisc]

Btw, also great to hear that the Gelbach Decomposition is being used by someone, so far I thought I might be the only one to ever really use it! 😄