[R] Hazard model with long-term survivor (cure model)

[David Winsemius]

"=?UTF-8?B?5a6L5pe25q2M?=" <shigesong at gmail.com> wrote in
news:5abc11d80802180008w51fa2139jcedae7efe068c67c at mail.gmail.com: 

> Dear All,
> 
> Are there R packages that can estimate survival model with long-term
> survivors? This is sometimes known as "cure" model or
> "split-population" model. Thanks.

The usual Cox model would certainly allow the analysis of observations 
with long-term survivors, but I am wondering if you want some sort of 
parametric model or one which compares a cohort's survival to some sort 
of external standard population expected survival. Therneau and Gramsch's 
text has a chapter on working with such expected survival estimates and 
the survival package can probably be considered an R atandard. If you set 
up a parametric Weibull model with a decreasing hazard, you get a cure, 
at least asymptotically. I think Harrell's Design package can conjure up 
accelerated time models that include the Weibull. 

Although the SRAB cancer methodologists assert that: "Neither SAS nor 
Splus can be used to fit survival models to relative survival data." (1)
I would be very dubious regarding such a claim. See for instance:

<http://www.mf.uni-lj.si/ibmi/biostat-center/predtiski/CMPB_Pohar_Stare_relsurv.pdf>
<http://www.ncbi.nlm.nih.gov/pubmed/15848272>
<http://cran.r-project.org/web/packages/relsurv/relsurv.pdf>

-- 
David Winsemius

1) <http://srab.cancer.gov/cansurv/models.html>