Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: FW: stcrreg: when the proportional hazards assumption fails


From   Steve Samuels <[email protected]>
To   [email protected]
Subject   Re: st: FW: stcrreg: when the proportional hazards assumption fails
Date   Fri, 22 Oct 2010 08:43:44 -0400

--

Zoe:

So, num_cancers has time-varying constants  for categories 2 and 3 vs
0, but not for  1 vs 0.  I had something different in mind.


1. First run stcreg with the tvc() specified for num_cancers 2 vs 0, 3
vs 0, and texp() at whatever you specified, (You will still want
i.num_cancers in the other part of the model).  Call it model A.

2. Then generate the predicted CIFs holding num_cancers at values 0,
1, 2, 3  and other covariates at single specified values.

3. Next run stcreg four times: if num_cancers = 0, 1, 2, 3. It's still
possible that for the main-effects, num_cancers = 0 and num_cancers= 1
are different, so you want  different runs for num_cancers= 0 and
num_cancers=1

4. Generate the  predicted  CIF plots for model and compare to
corresponding plot from model A.  -stcurve- will generate results, so
that you can plot each pair together.  And, yes, just eyeball the
results.

I'll be away from my computer for the next couple of days, so won't be
able to respond further for awhile.

Good luck.

Steve

Steven J. Samuels
[email protected]


On Fri, Oct 22, 2010 at 3:57 AM, Zoe Hyde <[email protected]> wrote:
> Thanks, Steve.
>
> Sorry, there are four levels to the ordinal variable - I was
> forgetting the reference category.
>
> Regarding your suggestion, do you mean something like this:
>
>
> stset d_event, failure(compete==2) origin(d_dob) entry(d_clinicdate)
> id(id) scale(365.25)
>
> stcrreg i.lh_quintile i.numcancers prevcvd age whr hyp dyslipid i.smoker
> diabetes if numcancers == 0 | numcancers == 1, compete(compete==1)
> stcurve, cif at1(lh_quintile=0) at2(lh_quintile=1) at3(lh_quintile=2)
> at4(lh_quintile=3) at5(lh_quintile=4)
>
> stcrreg i.lh_quintile i.numcancers prevcvd age whr hyp dyslipid i.smoker
> diabetes if numcancers == 0 | numcancers == 2, compete(compete==1)
> stcurve, cif at1(lh_quintile=0) at2(lh_quintile=1) at3(lh_quintile=2)
> at4(lh_quintile=3) at5(lh_quintile=4)
>
> stcrreg i.lh_quintile i.numcancers prevcvd age whr hyp dyslipid i.smoker
> diabetes if numcancers == 0 | numcancers == 3, compete(compete==1)
> stcurve, cif at1(lh_quintile=0) at2(lh_quintile=1) at3(lh_quintile=2)
> at4(lh_quintile=3) at5(lh_quintile=4)
>
>
> ...and then just eyeballing the results?  The curves look
> pretty much identical.
>
>
> Zoe.
>
>
>>On Thu, Oct 21, 2010 at 04:13 PM, Steve Samuels <[email protected]>
> wrote:
>>Zoe-
>>
>>Ah, I see what you mean. The tvc() coefficients provide evidence of
>>non-proportionality, but might not provide the correct model. With
>>regular Cox, we'd stratify by categories of the offending variable, as
>>you say, but that's not available here. -stcompadj- (from SSC) also
>>does not provide a stratified analysis.
>>
>>One possibility: run the model in the two (three?) subgroups of your
>>ordinal variable that violate proportionality. Compare the separate
>>cumulative incidence curves to that predicted by -stcrreg- or
>>-stcompadj-. Perhaps they are close, and you have a good model after
>>all.
>>
>>Otherwise, store the estimates of coefficients of the variables common
>>to all the models and compute weighted averages, weighting by the
>>inverses of the estimated variances.  I know this is easier said than
>>done!
>>
>>Steve
>>
>>Steven J. Samuels
>>[email protected]
>>18 Cantine's Island
>>Saugerties NY 12477
>>USA
>>Voice: 845-246-0774
>>Fax:    206-202-4783
>>
>>
>>
>>On Thu, Oct 21, 2010 at 10:06 AM, Steve Samuels <[email protected]>
> wrote:
>>> Zoe:
>>>
>>> I don't see that you have a problem. You seem to have a fairly
>>> complete model if you include the ordinal variable with the tvc() and
>>> texp() commands, perhaps omitting the non-significant indicator. As
>>> the Stata 11 Manual states on p 214, it is the coefficients which are
>>> time varying.
>>>
>>> One issue: a three-level variable would have only two indicators, not
>>> three. Showing your code and results, as the FAQ request, would
> really
>>> help avoid this kind of misunderstanding.
>>>
>>> Steve
>>>
>>> Steven J. Samuels
>>> [email protected]
>>> 18 Cantine's Island
>>> Saugerties NY 12477
>>> USA
>>> Voice: 845-246-0774
>>> Fax:    206-202-4783
>>>
>>> On Thu, Oct 21, 2010 at 4:45 AM, Zoe Hyde <[email protected]>
> wrote:
>>>> Hello All,
>>>>
>>>> I am wondering what options are available when the proportional
> hazards assumption
>>>> doesn't hold in a competing-risks regression.  The assumption holds
> for my main
>>>> independent variable of interest, but not for another (ordinal)
> variable that I'd
>>>> like to adjust for; fitting it as a time-varying covariate gives a
> significant
>>>> result for 2 of its 3 levels.
>>>>
>>>> I could get around this by stratifying by this variable in a
> standard Cox model,
>>>> but this doesn't seem to be supported (yet) by stcrreg.
>>>>
>>>> Are there any alternatives?
>>>>
>>>>
>>>> Regards,
>>>>
>>>> Zoe.
>>>>
>>>>
>>>> Western Australian Centre for Health and Ageing (M570)
>>>> University of Western Australia
>>>> 35 Stirling Highway, Crawley 6009
>>>> Western Australia
>>>>
>>>> Courier address:
>>>> Level 6, Ainslie House, Royal Perth Hospital
>>>> 48 Murray Street, Perth 6000
>
> *
> *   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/
>

*
*   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/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index