GEE with censored data

[Michele Siegel]

Hi Medstats,

I would like advice on the following.

I have a data set of 2000 respondents surveyed at admission into a psych program and every 6 months until discharge. Clients are followed for up to five years. Once they are discharged, they are no longer followed. They may or may not be discharged during the study period.  I plan to analyze predictors of time to discharge during the observation window, using a Cox proportional hazard model. That will tell me predictors of a quicker discharge during the study period. I would also like to do the analyses with Generalized Estimating Equations, which will tell me predictors of discharge, as opposed to time to discharge, during the study period. Can I use GEE when respondents are followed for a varying number of periods, based on their time to discharge?

Thanks very much!

Michele

[Tzippy Shochat]

If I understand correctly, Discharge is the Dependent and six month follow- ups are Independent. If so, a Cox survival model, with time varying covariates. may be appropiate. 

Tzippy Shochat

בתאריך יום ה׳, 10 במאי 2018, 21:24, מאת Michele Siegel ‏<mjsi...@gmail.com>:

--
--
To post a new thread to MedStats, send email to MedS...@googlegroups.com .
MedStats' home page is http://groups.google.com/group/MedStats .
Rules: http://groups.google.com/group/MedStats/web/medstats-rules

---
You received this message because you are subscribed to the Google Groups "MedStats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to medstats+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Raed Bahelah]

I think you can still use GEE in your scenario, but why are using GEE while Cox model (if proportionality assumption holds) can answer your question? I don't see a justification for using Cox for predictors of "time to discharge" vs. GEE for "predictors of discharge"?

--

--
To post a new thread to MedStats, send email to MedS...@googlegroups.com .
MedStats' home page is http://groups.google.com/group/MedStats .
Rules: http://groups.google.com/group/MedStats/web/medstats-rules

---
You received this message because you are subscribed to the Google Groups "MedStats" group.

To unsubscribe from this group and stop receiving emails from it, send an email to medstats+unsubscribe@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

Raed Bahelah, PhD, MPH&TM
Postdoctoral Associate, Department of Epidemiology
Robert Stempel College of Public Health and Social Work
Florida International University, Miami, FL 33199
Email: rbah...@fiu.edu; Office phone: 305-348-7792

[John Whittington]

At 13:10 11/05/2018, Raed Bahelah wrote:
>I think you can still use GEE in your scenario, but why are using
>GEE while Cox model (if proportionality assumption holds) can answer
>your question? I don't see a justification for using Cox for
>predictors of "time to discharge" vs. GEE for "predictors of discharge"?

I'm not convinced that GEE is necessarily the best approach to the
second one, but is not Michelle talking about two different questions
- firstly the time to dischanrge and, secondly, whether or not a
patient was discharged within a 5-year period. For the latter, one
would obviously have to decide what to do about patients who were,
for whatever reason (including death) 'lost to follow-up' before the
5-year period had elapsed.

Kind Regards,


John

----------------------------------------------------------------
Dr John Whittington, Voice: +44 (0) 1296 730225
Mediscience Services Fax: +44 (0) 1296 738893
Twyford Manor, Twyford, E-mail: Joh...@mediscience.co.uk
Buckingham MK18 4EL, UK
----------------------------------------------------------------

[Marc Schwartz]

> On May 11, 2018, at 8:33 AM, John Whittington <Joh...@mediscience.co.uk> wrote:
>
> At 13:10 11/05/2018, Raed Bahelah wrote:
>> I think you can still use GEE in your scenario, but why are using GEE while Cox model (if proportionality assumption holds) can answer your question? I don't see a justification for using Cox for predictors of "time to discharge" vs. GEE for "predictors of discharge"?
>
> I'm not convinced that GEE is necessarily the best approach to the second one, but is not Michelle talking about two different questions - firstly the time to dischanrge and, secondly, whether or not a patient was discharged within a 5-year period. For the latter, one would obviously have to decide what to do about patients who were, for whatever reason (including death) 'lost to follow-up' before the 5-year period had elapsed.
>
> Kind Regards,
>
>
> John

Hi Michele,

As noted, these are two very different questions here, one simply involving a binary endpoint, irrespective of time, the other keyed to the time to the event.

It seems to me that, given the length of the follow up interval in this setting, time to discharge would be a more important question, as opposed to simply whether or not discharge took place at any time.

That is, the length of time a patient is still in the program is likely to be associated with the severity of their underlying disease, co-morbid conditions and things like compliance with treatment (both counseling and pharma), if you are tracking such parameters over time.

Both types of treatment can vary, as patients might shift the type of counseling, especially if their diagnosis changes, and certainly, the pharma regimen can and will change over time.

Some of these are baseline conditions, others, like treatment compliance, treatment regimen and diagnosis, may vary over time.

Since there is no scenario under which multiple discharges, which infers re-admissions, will be tracked in this study design, the only logic, as I see it, for using GEE, would be for dealing with time varying covariates, in the setting of a time independent, binary endpoint.

If you are not tracking time varying covariates at the follow up interactions, then this would distill down to a simple logistic regression, however, still having to deal with censoring, as John noted above.

While there are censored regression scenarios, given that it would appear, to me at least, that time to discharge is a relatively more important question, Cox regression, with time varying covariates, if you are tracking them, would be more apropos here.

If the PH assumptions underlying Cox regression fail, you could consider parametric survival analysis.

Regards,

Marc Schwartz

[Michele Siegel]

Marc, John, Raed, and Tzippy,

Thanks very much. The research question is actually a bit more complicated than what I posted. Clients may have a positive, negative or neutral discharge, or they may not be discharged within the observation window. "Loss to follow-up" would be coded as a negative or neutral discharge, depending on the reason. I definitely plan to run a Cox proportional hazard model with time varying covariates and competing risks (not yet sure whether the software allows me to do both in the same model). I would also fit pairwise models comparing 2 at a time. That will estimate predictors of time to a successful discharge as opposed to predictors of a successful discharge at any point during the observation window.

I plan to use GEE to examine predictors of a successful discharge at any point during the observation window, with time varying covariates. To some extent, it would be a sensitivity analysis of whether predictors of a successful discharge at any point during the observation window differ from predictors of a quicker successful discharge. We are interested in both. Does that make sense?

Thanks again,
Michele