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
gam/penalized splines, polynomials, segmented regression, ... are not supported by patsy.
We don't get access to the stateful transform.
One possibility to include those is to use exog_other = [list of instances] to combine standard linear terms with more complex terms, where instance is one of the supported cases.
This would mean that we implement something similar to patsy's stateful transformations but restricted to a few cases where we need a special fit method.
Patsy's stateful transform are only relevant to create the design matrix but do not affect the estimation method, e.g. standard unpenalized splines. In contrast, we need to know additional properties of the transformation, penalization, knot search or similar.
This has the advantage that we can implement specific fit and predict methods.
(OLSAbsorb is a bit similar. We can either just take the absorbed exog and ddof and calculate the rest, or we can absorb inside the model, then we can also get additional information about the absorbed categoricals. A difference is that aborb is more like weights, it affects all exog, and can be added as a simple argument containing the factors.)
The text was updated successfully, but these errors were encountered:
(just another semi-random idea for a general API)
gam/penalized splines, polynomials, segmented regression, ... are not supported by patsy.
We don't get access to the stateful transform.
One possibility to include those is to use
exog_other = [list of instances]
to combine standard linear terms with more complex terms, where instance is one of the supported cases.This would mean that we implement something similar to patsy's stateful transformations but restricted to a few cases where we need a special fit method.
Patsy's stateful transform are only relevant to create the design matrix but do not affect the estimation method, e.g. standard unpenalized splines. In contrast, we need to know additional properties of the transformation, penalization, knot search or similar.
This has the advantage that we can implement specific fit and predict methods.
(OLSAbsorb is a bit similar. We can either just take the absorbed exog and ddof and calculate the rest, or we can absorb inside the model, then we can also get additional information about the absorbed categoricals. A difference is that aborb is more like weights, it affects all exog, and can be added as a simple argument containing the factors.)
The text was updated successfully, but these errors were encountered: