[R] Explaining Survival difference between Stata and R

[Göran Broström]

On Thu, May 13, 2004 at 02:10:35AM -0500, Paul Johnson wrote:
> Dear Goran (and others)
> 
> I did not know about the "eha" package, but reading the docs, I see many 
> things I've been looking for, including the parametric hazard model with 
> Weibull baseline.  Thanks for the tip, and the package.

You can find that in 'survival' too. See ?survreg. Watch up for
non-standard parametrization, though.

> I still don't quite understand your point about the reason that coxph 
> crashes.  Why does the difference of scale between the variables cause 
> trouble?  I understand that a redundant set of variables would cause 
> singularity, but do not see why differing scales for variables would 
> cause that.  Does the explanation lie in the optimization algorithm 
> itself?  I'll read the eha package coxreg source code. I bet that will 
> show me what's going on in your fix, anyway.

Nothing special is going on there. I use only standard linpack routines
(found in R) for solving linear equations, and matrix inversion at the
end. I think 'coxph' has its own optimizing routines.

Peter D. explained why scaling can be a problem.

Göran
[...]
-- 
 Göran Broström                    tel: +46 90 786 5223
 Department of Statistics          fax: +46 90 786 6614
 Umeå University                   http://www.stat.umu.se/egna/gb/
 SE-90187 Umeå, Sweden             e-mail: gb at stat.umu.se