Austin, thank you for your prompt response.
Agreed that death is one of the main component of the endpoint and not
a censor mechanism. A composite endpoint of death or rehospitalization
(whichever event occurred first) is what traditionally has been used
as clinical outcomes in heart failure trials. However, by indication
of regulatory agencies (FDA, EMEA) new treatment need to be tested
using "days alive and out of the hospital", which is a sort of the
reciprocal of the previous endpoint. The theoretical advantage of this
endpoint is that it combines mortality, lenght of stay (LOS) of the
index hospital stay, and the burden of subsequent hospital stays into
a single endpoint. Using survival analysis was mi first thought, but I
wasn't sure how to deal with follow-up time. Then, I considered the
option of using "censornb" from Joseph Hilbe, approach that will
ignore death as the main endpoint and treat it as a censor mechanism.
If I understand the first part of your answer, you meant to use stcox
and for the time variable, just to discount the days spent in the
hospitals from the total follow-up. Is that right?
I agreed that competing events may be playing an important role in
obtaining unbiased estimates, particularly when the posibility of
informative censoring may be operating in this situation (lost to FU
related with the treatment or death with rehospitalization as you
mentioned). However, I don't know how to implement such competing risk
approach in this specific setting.
This is the description of censornb: censornb fits a maximum
likelihood censored negative binomial regression of depvar on
indepvars, where depvar is a non-negative count variable. The censor
option is required. If no observations are censored, a censor variable
with all 1's must be specified.
Regards,
Eduardo
On Thu, Aug 6, 2009 at 12:46 PM, Austin Nichols<[email protected]> wrote:
> Eduardo Nunez<[email protected]>:
> Death is not censoring here; it is the true end of risk for increases
> to the duration "days alive and out of the hospital" but the end of
> the observation period is a censoring event. You want a survival
> model here, I think, and you could run it as single risk by
> subtracting days in the hospital from the endpoint, but a asymmetric
> competing-risks multiple-failure model is the underlying model (death
> prevents you from admission to the hospital, but admission to the
> hospital do not prevent death). Do you observe time-varying
> characteristics of patients over time? See also:
> -help st-
> http://www.iser.essex.ac.uk/iser/teaching/module-ec968
> http://home.fsw.vu.nl/m.buis/wp/survival.html
> etc.
>
> On Thu, Aug 6, 2009 at 11:18 AM, Eduardo Nunez<[email protected]> wrote:
>> Stata/SE 11.0 for Windows (64-bit x86-64)
>> Born 13 Jul 2009
>>
>> Hi all,
>>
>> I would appreciate if someone can advise me on the best way to analyse
>> and endpoint that is based on counts with censoring mechanism. It has
>> been suggested in heart failure trials to use as endpoint "days alive
>> and out of the hospital" in order to evaluate new treatment
>> modalities. This composite endpoint is supposed to capture the burden
>> of mortality and hospital stay (for rehospitalization) during the
>> follow-up period. Patients will be censored either by death or lost to
>> follow--up. I was thinking in censored-poisson, although I doubt the
>> day-counts follows a poisson distribution. Moreover, using simple
>> regression analysis won't take into account the censoring mechanisms.
>>
>> Regards,
>>
>> Eduardo Nunez, MD, MPH
>> Servicio de Cardiología, Hospital Clínico Universitario, Universitat
>> de Valencia., Spain
>>
>
> *
> * 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/
