I realize I'm late to the party here, but while writing the wrapping code a few comments about MixedLM while I can.
I don't see the set_random method that's mentioned in the release notes about this. Was this dropped? Changed? We've stayed away from procedural style, call this method and the result of the call of the next method will change AFAIK. I'd like to stay away from this for as long as possible. I realize that in some situations it may make sense though.
Naming: ranef -> random_effects? We also don't have to continue to use bse for new models.
Just looking at an example not thinking about this, right now it looks like the return of the ranef method is a dictionary with 1d scalar array values. Should we just use a pandas.Series here regardless of wrapper? Likewise for the covariance of the random effects. Will this always be a single value 2d array? If so, same change proposed.
I see that it's not necessarily 1d scalar but I think the comment still holds. I'd almost always want a DataFrame over a dict for a map like this, though maybe this should just be in the wrapper code.
nparams -> k_params
set_random has been dropped and been replaced with an argument in __init__
(no method calling to add arguments anymore)
You are not too late to the party, because the party is not over yet, i.e. needs more review. (I only reviewed parts of it.)
Ok, that's good. I'll fix it in my release branch notes. I might be able to do a partial clean up with the wrapper change. Mainly just how it interoperates with our existing data-handling names/missing/etc. I think there's a similar issue for groups with GEE (?). AFAICT, this needs a once over in that area too, though I'm probably not going to add any tests that take me more than 5-10 minutes.
changed in PR #2008