Re: st: Main effect for time-varying covariate

[Nicole Boyle <nicboyle@gmail.com>]

Thanks so much for your input, Adam.

With regards to the current discussion on this thread, I agree that
competing risks regression, like many other aspects of biostatistics,
appears to have several arguably "correct" approaches. Every time I
feel like I've settled for one approach, I get pulled in another
direction.

I also agree with you that death is a competing risk here (my study
does concern BMT patients, by the way). You described the situation
very clearly and concisely, in a way that I'm still trying to achieve.

> Again, there is a big conceptual difference between modeling
> biological effect of covariate on one type of event versus predicting
> real-life outcomes.

Agreed. The cause-specific Cox regression's modeling of biological
effects, coupled with its assumption of infection's "independence"
from death, are the two reasons I'm wary of cause-specific Cox for
this study. Fine-Gray feels preferable here, even in light of its
multiple records per subject assumptions for TVC modeling (a minor
assumption breach, IMHO, compared to Cox).

I feel like I need to choose an approach. I'd be happy to just run
both Cox and Fine-Gray and report everything, but I have four models
with many categorical variables.

Ultimately, I'd rather simply report Fine-Gray models and CIFs, even
given its assumptions on TVC modeling post-death, and simply state
something along the lines of "cause-specific Cox models were also run,
and no appreciable differences were observed." This type of vague
reporting is done in many ways (e.g. test results for PH assumption
are never explicitly reported, although testing for the PH assumption
is a fundamental step in proportional hazards modeling). However, it
feels a bit strange to vaguely mention Cox model results instead of
explicitly reporting them, since spotting similarities or differences
between the Cox and Fine-Gray results is more of a subjective task
than an objective task (i.e. there isn't a test to compare these
regressions).

Not sure if we should move this conversation onto a new thread.

Nicole

On Sun, Sep 15, 2013 at 1:25 PM, Steve Samuels <sjsamuels@gmail.com> wrote:
> Thanks, Adam.
>
> Before seeing your post, I reached the same conclusion:
> that infection can be analyzed, not death. This answers Nichole's
> question about how much effort to devote to death: None!
>
> I'll respond to her other issues in a later post.
>
>
>
>
> On Sep 15, 2013, at 2:06 PM, Adam Olszewski wrote:
>
> Hi,
> This discussion is has now gone into a fairly nuanced issues of
> interpretation of cause-specific-hazard and competing-risk survival
> models, and one should note, after Latouche/Fine, that "there appears
> to be no final consensus on how to analyze competing risks endpoints"
> (J Clin Epidemiol. 2013 Jun;66(6):648-53). Therefore it is best left
> to further research rather than any arbitrary statements. With that
> said, ...
>
> On Sun, Sep 15, 2013 at 12:35 PM, Steve Samuels <sjsamuels@gmail.com> wrote:
>> Nicole:
>> I didn't see that nuance. It changes the picture, because it means that you do not have a competing risks problem. For risks to be competing, only one can be observed  (Kalbfleisch & Prentice, 2002, p. 248). In your case, infection and death can both be observed. The proper analysis, therefore, is a standard, single-response, survival model.
>
> This is not correct, and there is a number of detailed analyses of BMT
> and other data that discuss it in detail. The competing risk problem
> still exists and the CIF / Fine-Gray approach is appropriate, as long
> as it is the infection that is the endpoint of interest. In this
> scenario, only one event can be observed, because patients are
> censored at the time of infection and the data on their further deaths
> become irrelevant for the model (and, in fact, invisible through
> -stset- assignment). There is an obvious informative censoring issue
> because of the asymetry: no infections after death. Using a net
> survival estimator would overestimate the risk of infection, which may
> introduce minimal or major bias depending on the risk of each event in
> the population. The same reasoning is widely applied in other
> scenarios of cancer epidemiology, where cancer recurrence and death
> (of any cause) are treated as competing events, even though people
> obviously die after cancer recurrence (yet they do not recur after
> death). Again, there is a big conceptual difference between modeling
> biological effect of covariate on one type of event versus predicting
> real-life outcomes.
> AO
>
> *
> *   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/