Could statsmodels dependency be replaced with scipy?

[twiecki]

It seems that some package maintainers that would like to rely on seaborn are hesitant to add more dependencies. I was thus wondering if statsmodels is really required as scipy provides linear regression (via stats.linregress) and KDE.

Did miss anything that's only in statsmodels?

[mwaskom]

Hah, I just moved the KDE from scipy to statsmodels :) The main advantage is the availability of different kernels, and faster computation in the univariate gaussian case.

There are a few things that require statmodels (or at least require either statsmodels or my implementing some statsmodelish things in pure scipy), like logit=True, robust=True, and lowess=True in regplot, coefplot, interactplot, etc.

I have plans (cf #67) to make the statsmodels dependency optional (but strongly recommended) by pushing the import down into the local context where it's needed.

[twiecki]

Ah, missed that issue. That makes sense!