Seeing that this lib is built on top of statsmodels my first response was that I'll just keep using statsmodels. However the simplicity of the diagnostic plots actually seems really nice value add. Showcasing those plots right in your docs might make it more attractive to people checking out your project for the first time.
These are some things I've included in the library which aren't implemented in Statsmodels:
- Breusch-Pagan studentized test of heteroskedasticity (available in R)
- Standardized / beta coefficients (still an open feature request in https://github.com/statsmodels/statsmodels/issues/3857 )
- Leverage vs residuals squared plot (there's an influence plot but not something similar to Stata's lvr2plot)
Even the most common metric I use for assessing models – root MSE – isn't stored in the Statsmodels object summary. To assess an OLS model in Statsmodels I'd find I do so much repetitive code, yet in Stata the commands are fairly succinct.
Other things I also added to make encoding of variables easier:
- InteractionEncoder
- DummyEncoder (to cover different ways of treating missing values)
The more I thought about these missing features, the more I thought they can be wrapped up in a more coherent way. :-)