Re: st: survival analysis query regarding stsplit, tvc or using enter and exit options

[Joerg Luedicke <joerg.luedicke@gmail.com>]

On Thu, Apr 14, 2011 at 7:44 AM, Melissa Wright <WrightM10@cardiff.ac.uk> wrote:
> Hi
> I'm working on a breast cancer survival project and am looking at the
> effect of grade of tumour, size of tumour, nodal status, screening status
> etc on survival.
> The proportional hazards assumption is not met for my multivariate Cox
> regression.  There are several significant interactions between variables.
>  Including interaction terms in the model or stratifying by grade of
> tumour etc doesn't help.
>
> Fitting tvc terms showed that many variables have an interaction with
> time.  For example:-
> xi:stcox i.agegroup i.size i.grade i.screencat i.nodes i.diagperiod9609,
> tvc(i.grade) texp(_t>365.25)
>
> I fitted a linear relationship with time, but also allowed for a different
> effect after 1 year and 5 years.
>
> Based on the results I thought it would be sensible and straightforward to
> stratify by time.  I have presented 3 models,
>
> 1st year follow up
> stset ftime, failure (dead==1) exit(time 365.25)
> xi:stcox  i.agegroup i.size i.grade i.screencat i.nodes i.diagperiod9609
> 1-5 years follow up
> stset ftime, failure (dead==1) enter(time 365.25) exit(time 1826.25)
> xi:stcox  i.agegroup i.size i.grade i.screencat i.nodes i.diagperiod9609
> 5+ years follow up
> stset ftime, failure (dead==1) enter(time 1826.25)
> xi:stcox  i.agegroup i.size i.grade i.screencat i.nodes i.diagperiod9609
>
> Assumptions are met for all the above models.  I'm not confident that this
> is a legitimate technique however, or if there is a more widely used
> method of dealing with this problem.  Stsplit doesn't seem to be
> appropriate in this instance, unless I'm misunderstanding it's use.
>
These are merely my 0.02$ as I am not into clinical studies etc. so
there may be specialists who have better advice. But, that being said,
running separate models while conditioning on time looks rather
inefficient and misleading to me. For one thing, you are now looking
at survival within three completely different samples. For instance,
in your last model, you only look at women who survived for 5 years
after biopsy/surgery (or whatever it is). This group is (most likely)
totally different from those who did not make it to that point. You do
not get an overall picture for the relation between your covariates
and survival that way. Maybe a frailty model would help or otherwise a
piecewise approach with period specific effects. Isn't there some
default in the clinical literature how to deal with that problem? I
would imagine this being a common problem there.

J.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/