MNLogit tests are broken

[vincentarelbundock]

Multinomial logit models do not match R.

At first glance, it looks like things work, and there are some tests commented out here:

https://github.com/vincentarelbundock/pymarginaleffects/blob/main/tests/test_statsmodels_mnlogit.py

Things to check:

• Do the point estimates match?
• Do the standard errors match?
• If the SEs don't match, is it because of marginaleffects, or do the fitted models themselves have different covariance matrices? In R you can inspect it with vcov(model). in Python it's usually model.cov_params().

[vincentarelbundock]

@artiom-matvei you could take a look at this

[artiom-matvei]

@vincentarelbundock It seems like the support for mlogit is deprecated in R. Just double checking before continuing on this issue. https://github.com/vincentarelbundock/marginaleffects/blob/74ef4b3d5a81c83efa3aeff728b9f8eea5681293/NEWS.md?plain=1#L28

[vincentarelbundock]

Right, the equivalent should be multinom() from the nnet package. Examples:

https://github.com/vincentarelbundock/marginaleffects/blob/74ef4b3d5a81c83efa3aeff728b9f8eea5681293/inst/tinytest/test-pkg-nnet.R#L8