Open mayrandy opened 8 years ago
Once the general issue regarding the derivatives is resolved, could we inlclude this also in other families?
We could use link and inverse link functions along the lines of countreg (https://r-forge.r-project.org/scm/viewvc.php/pkg/R/?root=countreg).
Ok, that was easier than expected. Should work now with a981957 for as.families()
.
However, @hofnerb is right - we should go through the existing orginal gamboostLSS families and make it possible for users to define other links. In boost-R/mboost#34 I did use stats::make.link()
which automatically provides all necessary functions including d mu / d eta via link$mu.eta(f)
.
@mayrandy Is it possible to include the required changes for the gamboostLSS families in the next release in a few days? If so, you should implement the changes in the master branch. Thanks.
It should be possible to use different link functions in families that we import from
gamlss
viaas.families()
, currently we are just using the default.The link function is crucial for
ngradient,
as ingamlss
they use the derivatives d l / d mu (in case of the first parameter of the distribution). We need d l / d eta_mu, therefore we also need d mu / d eta_mu which depends on the link.