You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Just an idea, I don't know yet whether this is feasible.
use case:
In penalized estimation, the penalization is often calibrated or predefined for standardized coefficients, standardized data (rescaled exog).
If the user uses a transformation prior to creating the model, then it might still be of interest to have the results, tests and reports for the original scale. That is we fit in one parameterization or scaling and have the Results using a different one.
This is a generalization of fit_constrained and fit_transformed that allows user to use only the second half of the transformation.
Problem with this:
While it is easy to transform params and cov_params, there are many results that directly depend on the attached exog, like fittedvalues, resid, rsquared and similar. The attached exog would have the wrong scaling, if the user rescaled it.
fit_constrained and fit_transformed use an auxiliary model that contains the transformed data, but still has the original data in the main model.
(Doesn't sound good anymore.)
correction
If only exog is transformed but not endog, then fittedvalues and resid would still be independent of linear or affine transformation of exog.
Predict, Margins and others would depend on the original scaling.
The text was updated successfully, but these errors were encountered:
Just an idea, I don't know yet whether this is feasible.
use case:
In penalized estimation, the penalization is often calibrated or predefined for standardized coefficients, standardized data (rescaled exog).
If the user uses a transformation prior to creating the model, then it might still be of interest to have the results, tests and reports for the original scale. That is we
fit
in one parameterization or scaling and have the Results using a different one.This is a generalization of
fit_constrained
andfit_transformed
that allows user to use only the second half of the transformation.Problem with this:
While it is easy to transform params and cov_params, there are many results that directly depend on the attached exog, like fittedvalues, resid, rsquared and similar. The attached exog would have the wrong scaling, if the user rescaled it.
fit_constrained
andfit_transformed
use an auxiliary model that contains the transformed data, but still has the original data in the main model.(Doesn't sound good anymore.)
correction
If only exog is transformed but not endog, then fittedvalues and resid would still be independent of linear or affine transformation of exog.
Predict, Margins and others would depend on the original scaling.
The text was updated successfully, but these errors were encountered: