Discussion:
[pystatsmodels] GAM: an update
j***@gmail.com
2018-11-14 03:54:59 UTC
Permalink
https://gist.github.com/josef-pkt/5e73596fab72da427a9f2be59c0fef1b

The core functionality works.
I have now full agreement with R package mgcv for two types of splines,
cyclic cubic splines and B-splines, and their penalization.

Some differences to mgcv are in the significance/wald test and in the
options for penalization parameter selection. Wood/mgcv is either doing
something that I don't understand or something too complicated that I don't
want to implement right now.

Some convenience features like formula usage are still incomplete or not
supported.

It fully support partial linear models.
y = X b + f_1(z_1) + f_2(z_2) + ...
where the f are nonlinear functions approximated by penalized splines.

Extras like var_weights and freq_weights are not yet supported.
But we inherit some goodies like sandwich covariances, that are for example
robust to unspecified heteroscedasticity, which mgcv does not have..

Josef

Loading...