I am planning on creating a small package that just contains the 'mgcv' constructor/predict functions for domain transformations. There are 3 basis types that I'm going to include:
pi: "parametric interactions", which is the one that you wrote for pcox a few months ago, and I edited. This reduces a bivariate smooth to a univariate smooth, e.g., which assumes a particular form for the interaction.
dt: "domain transformations". This allows the user to enter a transformation function for the domain indices (argvals), e.g., s-t or s/t.
nb: "no basis" is applied, but it still allows for penalizing the coefficients associated with a term with a specified penalty, perhaps a difference penalty. Are there other types of penalties that would make sense to allow, and if so do you have any suggestions for a user-friendly interface to specify different penalties? The difference penalty is the main one that I want to implement.
The reason I want to do this as a separate package is because I want these functions to be available to both refund::lf.vd() and in pcox, and this seemed like a better way than having multiple versions of the same functions. You and I would be the authors of the package. I might need some help with the
dt basis... this is the only one I haven't written yet.
Do you have any suggestions for a name for the package? Right now I'm going with mgcvTransformations.
Once I do this I am going to re-write lf.vd to play nicer with the new pfr().
These have been implemented. For now the functions will live in both refund and pcox as necessary; mgcvTrans will likely be accompanied by a small paper.
Hey Fabian,
I am planning on creating a small package that just contains the 'mgcv' constructor/predict functions for domain transformations. There are 3 basis types that I'm going to include:
pi
: "parametric interactions", which is the one that you wrote forpcox
a few months ago, and I edited. This reduces a bivariate smooth to a univariate smooth, e.g., which assumes a particular form for the interaction.dt
: "domain transformations". This allows the user to enter a transformation function for the domain indices (argvals
), e.g.,s-t
ors/t
.nb
: "no basis" is applied, but it still allows for penalizing the coefficients associated with a term with a specified penalty, perhaps a difference penalty. Are there other types of penalties that would make sense to allow, and if so do you have any suggestions for a user-friendly interface to specify different penalties? The difference penalty is the main one that I want to implement.The reason I want to do this as a separate package is because I want these functions to be available to both
refund::lf.vd()
and inpcox
, and this seemed like a better way than having multiple versions of the same functions. You and I would be the authors of the package. I might need some help with thedt
basis... this is the only one I haven't written yet.Do you have any suggestions for a name for the package? Right now I'm going with
mgcvTransformations
.Once I do this I am going to re-write
lf.vd
to play nicer with the newpfr()
.