Skip to content
Prev 244937 / 398502 Next

survival: ridge log-likelihood workaround

You ask some good questions.
Not quite.  The printout contains "se" and "se2".  The first is the
suggestion
of Verwiej and Van Houwlingen and the second is the suggestion of Gray.
The first error estimate is always larger although usually very close to
the
second.  However, I have seen several data sets where the second
collapses and
produces a tiny variance estimate.  Thus the Wald statistic is based on 
the Verweij estimate.
There is literature, but it's been 10 years and I don't remember the
references.  One key issue, not documented I admit, is that for the
difference in LR between two models to be valid at all, they have to
have
the same penalty parameter.  I have almost never compared two penalized
LR to each other, likely because I rarely use ridge().  More often I am
comparing the LR of a penalized model to one without the penalty term, 
for which the issue does not arise.
  The same type of issue will arise, I believe, if one were to fit two
mixed effects models that differed only in the fixed effects, and then
used an LR test.  Comparisons are only valid if two have the same
variance
values, or else the models are not nested.
no
to
I was sure that was somewhere, but I can't find it either.  The manual
page
summary.coxph.object fell off my list somewhere.
  The R^2 value is that of Nagelkirke.  At the time I wrote that code it
was
a simple estimate of R^2.  There have since been several other
proposals, and
a recent review by Schemper suggests that the Nagelkirke method may be
nearly
the worst of the bunch.  Replacing this is another item on my to-do
list.
This is the hard issue.  The "penalized methods" addition to coxph was 
designed so that a new method could be added without any changes at all
to the coxph (or survreg) code.  The primary reason was to allow users
to
add arbitrary methods.  It turned out that no one ever has done so, at
least
none that I have ever heard of, so that particular design goal now looks
silly.
  A consequnce of the design is that the basic printouts for the model
consist of a constant piece + an addition bit supplied by the inserted
method.  There are others: the user has to know whether a particular
penalty plug in is appropriate for survreg, coxph, or both.  This is
true
for pspline and ridge, but most of the parameter combinations for
frailty
only work with coxph.  It is in the documentation for fraily, but has
nevertheless caused many a user to stumble.  
  I think that it should be rewritten, but this is a formidable task.
Note: the "people who developed" the survival package is --- me, and
there
are only so many hours in a day.  As I look ahead to retirement (about
10
years) it clearly is time to enlist more people in this project.
 
Terry Therneau