RE: st: Proportionality check: stcrreg with an established tvc

["Boyle, Nicole M" <nboyle@fhcrc.org>]

Hi AO,

Wonderful, thanks for your advice! I appreciate it.

Follow-up question: For deeming the "appropriateness" of other time function interactions,
does one also eyeball the results of different outputs in the same manner you described?
I've been considering using texp(ln(_t)), only because this is the only way tvc was taught
to me, but that's obviously not a good enough reason to transform a time function.

Thanks,
Nicole
________________________________________
From: owner-statalist@hsphsun2.harvard.edu [owner-statalist@hsphsun2.harvard.edu] on behalf of Adam Olszewski [adam.olszewski@gmail.com]
Sent: Friday, August 16, 2013 3:07 PM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: Proportionality check: stcrreg with an established tvc

Hi Nicole,
It is actually enough to inspect the results of the stcrreg alone. If
the hypothesis that TVC coefficient for your other variables can be
safely rejected based on the P-values, then this is all you need. You
may, for additional sensitivity, evaluate tvc with the texp() function
to assure that interactions with other time functions are not more
appropriate (e.g. texp(_t^2) or texp(ln(_t)) ).
I am not sure if there is much value of doing a "general" test of
proportionality of all variables at the same time, you would probably
want more to know which variable violates the PH assumption and how
important it is for your estimands of interest to include this effect
in the model.
AO

On Fri, Aug 16, 2013 at 5:55 PM, Boyle, Nicole M <nboyle@fhcrc.org> wrote:
> Hello all,
>
> I'm fitting a model in stcrreg and am trying to check for proportional hazards via the time dependent
> covariate method. I already have a tvc in my model, and I don't know how to properly check for
> proportionality of the other non-time varying covariates.
>
> Here's the model I want to use in my study; I've included "grade" as a tvc, as grade's time-dependence is
> relevant to our study:
>
>         stcrreg myel i.grft i.prop i.year age grade, compete(failure==2) tvc(grade)
>
>
> Here's the model I'd like to use to check for proportionality, where all vars are now tvc:
>
>         stcrreg myel i.grft i.prop i.year age grade, compete(failure==2) tvc(myel i.grft i.prop i.year age grade)
>
> Then followed by the postestimation Wald test, which will test if any of the tvcs in the model are significant:
>
>         test [tvc]
>
>
> My problem:
> I don't want to leave tvc(grade) out of this test, since proportionality of the other covariates
> may change in the presence/absence of tvc(grade). However, keeping tvc(grade) in this test may also result
> in a significant Wald test, and therefore an artifactual proportional hazards violation if tvc(grade) is in fact
> significant.
>
> Should I instead use this postestimation command...
>
>         test [tvc], mtest
>
> ...which will test for the presence any time-dependence AS WELL AS individual time-dependence (per each
>  covariate)? Or should I leave "grade" completely out of the proportionality check model?
>
>
> Thanks a million,
> Nicole
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/

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