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: Main effect for time-varying covariate


From   Nicole Boyle <[email protected]>
To   [email protected]
Subject   Re: st: Main effect for time-varying covariate
Date   Wed, 4 Sep 2013 13:37:46 -0700

On Tue, Sep 3, 2013 at 4:22 AM, Phil Clayton
<[email protected]> wrote:
> I probably wouldn't describe the Fine-Gray model as "protecting" anything.

You're right, that's an incorrect descriptor. Thanks for clarifying
those definitions.

> I went to a talk by Jason Fine last year and he gave the following general advice:
> - use a Cox model for each of the competing outcomes (in your case infection & death)
> - use a Fine-Gray model for each of the competing outcomes
> - present all of those results

Thanks for the advice! What's the utility of presenting model results
for the outcome of death if death is not an outcome of interest in my
study? Feel free to direct me to a paper if you'd like.

I also agree with your comment about how the assumption necessary for
a binary time-varying covariate to function under -stcrreg- is
unreasonable (for my data). I think I'll only model through -stcox- as
a result, but the advice about running both  Fine-Gray and Cox models
is good to know for future analyses.

Also, conditional landmark analysis is an interesting idea; however,
I'm not sure if it's right for this situation. I want to measure the
effect of, not simply condition on, this particular risk factor.
Additionally, I have detailed information on time of risk factor onset
per each individual, and I'd be losing information if I chose a single
reference point in time.

Nicole

On Tue, Sep 3, 2013 at 4:22 AM, Phil Clayton
<[email protected]> wrote:
> Nicole,
>
> I probably wouldn't describe the Fine-Gray model as "protecting" anything. It models the cumulative incidence function in the presence of competing risks, whereas a Cox model models the cause-specific hazard (ie what would happen if the competing event didn't occur).
>
> I went to a talk by Jason Fine last year and he gave the following general advice:
> - use a Cox model for each of the competing outcomes (in your case infection & death)
> - use a Fine-Gray model for each of the competing outcomes
> - present all of those results
>
> I must say, in my experience the Cox and Fine-Gray models (without time-varying covariates) give very similar results most of the time.
>
> Your binary time-varying covariate is a problem unless you can assume that it wouldn't have changed in people who died. It doesn't sound like that's a very reasonable assumption in your dataset, but you would know better than us.
>
> Therefore I think if it were me I'd follow Steve's advice and model the cause-specific hazard using -stcox-
>
> You could also try a conditional landmark analysis (see the paper that Adam pointed you to, Giobbie-Hurder et al, J. Clin. Oncol. 2013 Aug 10;31(23):2963–9). For example, if most infections and death are occurring after 3 months, but most rejections are occurring before 3 months, you could start your analysis time at 3 months and model the competing events using rejection status at 3 months as your baseline covariate. This would allow you to use -stcrreg- without introducing the TVC bias.
>
> Phil
>
> On 01/09/2013, at 9:51 AM, Nicole Boyle <[email protected]> wrote:
>
>> (Just to briefly mention, I meant to address the beginning of my
>> last response to Adam, but it seems that magic [?] deleted
>> that portion of my response once posted. Thanks again, Adam.)
>>
>>
>> Phil:
>> Thanks for crphplot! I'm going to fiddle with it this weekend.
>>
>> I'm reading the "Multiple records per subject" section you've recommended,
>> and I'm very glad you've advised this. According to the section, it seems like
>> competing risks STILL have the opportunity to bias the study, even within
>> the context of a competing risks regression, if the inherent assumption of
>> stcrreg (the value of a subject's time-varying covariate at time of failure via
>> competing risk remains fixed after this failure) is an invalid assumption.
>>
>> Therefore, with the Fine and Gray model, modeling a categorical variable
>> as time-varying PRECLUDES that same variable from being protected by
>> the model's inherent competing risks environment. Is this accurate?
>>
>> In other words, take my study of post-transplant patients:
>>    * Event of interest = post-transplant infection
>>    * Time-varying factor = onset of an irreversible post-transplant
>> complication
>>    * Competing risk = death
>>
>> With regards to these parameters, the Fine and Gray model does account
>> for the competition of death vs. infection, but does NOT account for
>> the possible
>> competition of death vs. post-transplant complication.
>>
>> POTENTIAL PROBLEM?:
>> Looking at my data now, about 50% of those subjects who "exit" observation
>> via competing risk (death) also had the post-transplant complication.
>> _If_ my understanding of the issue described in the manual excerpt is correct,
>> then I need to make sure that the other 50% of those subjects who died prior
>> to having this post-transplant complication would NOT have had this complication
>> if alive and given the opportunity.
>>
>> QUESTIONS:
>> (1) Does my understanding of the issue check out? And if so...
>> (2) Any possible remedies, or is this simply a model assumption that
>> must be noted?
>> (3) Does this issue also apply to cause-specific hazards (stcompet)?
>>     stcompet literature:
>> http://www.stata-journal.com/sjpdf.html?articlenum=st0059
>>
>> Thank you so much!
>> Nicole
>>
>> On Fri, Aug 30, 2013 at 8:25 PM, Phil Clayton
>> <[email protected]> wrote:
>>> Here's an example. No doubt it could be improved. Much of the code is borrowed from stphtest.ado.
>>>
>>> Phil
>>>
>>> -------- program crphplot --------
>>> program define crphplot, rclass
>>>        version 11
>>>        syntax varname(fv), *
>>>
>>>        capture assert e(cmd)=="stcrreg"
>>>        if _rc {
>>>                di as error "crphplot can only be used after stcrreg"
>>>                error 498
>>>        }
>>>
>>>        * convert factor variable notation
>>>        _ms_extract_varlist `varlist'
>>>        local varlist "`r(varlist)'"
>>>
>>>        * calculate schoenfeld-like residuals
>>>        tempname b
>>>        mat `b' = e(b)
>>>        local dim = colsof(`b')
>>>        forval i = 1/`dim' {
>>>                tempvar sch`i'
>>>                local schvars `schvars' `sch`i''
>>>        }
>>>        qui predict double `schvars' if e(sample), schoenfeld
>>>        local n : word count `schvars'
>>>
>>>        * subtitle will use variable label if it exists
>>>        local varlab: variable label `varlist'
>>>        if "`varlab'"!="" local subtitle subtitle("`varlab'")
>>>
>>>        * lowess plot of the relevant variable vs time
>>>        forvalues i=1/`n' {
>>>                local lbl: variable label `sch`i''
>>>                local lbl=substr("`lbl'", 23, .)
>>>                if "`lbl'"=="`varlist'" ///
>>>                        lowess `sch`i'' _t, mean noweight ///
>>>                        title("Test of proportional subhazards assumption") ///
>>>                        xtitle(Time) `subtitle' `options'
>>>        }
>>> end
>>> -------- end program --------
>>>
>>> -------- example analysis --------
>>> webuse hypoxia, clear
>>> stset dftime, failure(failtype==1)
>>> stcrreg ifp tumsize pelnode, compete(failtype==2)
>>> crphplot ifp
>>> -------- end example --------
>>>
>>>
>>> On 31/08/2013, at 11:30 AM, Phil Clayton <[email protected]> wrote:
>>>
>>>> It wouldn't be hard to program a wrapper for generating and plotting the Schoenfeld-like residuals. That's essentially what -estat phtest, plot()- does (take a look at -viewsource stphtest.ado-)
>>>
>>>
>>> *
>>> *   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/

*
*   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/


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