Dear Enzo, Michael and Bobby:
Many thanks for your prompt and very helpful replies, they have
certainly cleared up a few things. Bobby I will try your suggestion of
fitting an empty model with stccreg.
While I can understand how adding startfication terms to the competing
risks modelling can complicate matters substantially, I have to say
that I am a little surprised at the current lack of options
considering how frequently casue-specific survival analysis is
performed.
Thanks again!
Best wishes
Wei
Date: Wed, 14 Oct 2009 20:58:36 +0200
From: Enzo Coviello <[email protected]>
Subject: Re: st: stcrreg: startifications and median survival?
Hi Wei Xun,
> However, I noticed that unlike stcox, stcrreg does not allow any
> stratification terms within the model. Are there any particular
> reasons why this not possible?
A further complication in the competing risks regression with respect to
the usual cause-specific survival model is that we must take into
account the censoring distribution that varies across different strata.
The censoring distribution provides weights used in the estimation of
the competing risks model.
Note also that a -strata- option (i.e. an option allowing different
baseline hazard of the subdistribution for each value of the stratavar)
is not available within the -cmprsk- package too written by Fine and
Gray themselves.
Therefore, I suspect that this topic is not completely explored until now.
> Could I alternatively use stcompet to calculate the cumulative incidence
> curve and estimate it from the point of 50%?
>
This should not be correct.
Note that 1-Cumulative Incidence is not a Survival probability and that
this function (1-CI) is not easily interpretable in practical research.
Best wishes
Enzo
- --
Enzo Coviello
Epidemiology Unit ASL Bari
Piazza V. Emanuele, 14 - 70054 GIOVINAZZO (BA)
Italy
tel +39 080 3357883 - fax +39 080 3357867
mobile 347 5016016
^^^^
Home
via Imbriani, 186 - 70055 MINERVINO MURGE (BA)
tel +39 0883 695055
*
* 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/
Date: Wed, 14 Oct 2009 14:17:25 -0500
From: [email protected] (Roberto G. Gutierrez, StataCorp)
Subject: Re: st: stcrreg: startifications and median survival?
Wei Xun <[email protected]> asks:
> I apologise if this has been asked before but my search in the archives did
> not return any relevant results.
> I am currently trying to estimating the effects of certain exposures on
> mortality from specific cancers where a large proportion of deaths occur
> from other causes. The new command stcrreg in STATA 11 has been most useful.
> However, I noticed that unlike stcox, stcrreg does not allow any
> stratification terms within the model. Are there any particular reasons why
> this not possible? I am aware of the fact that the power will be
> significantly reduced.
To the best of our knowledge, the theoretical aspects of stratification in the
Fine and Gray (1999) regression model have yet to be established in the
literature. Fine and Gray do discuss stratifying to achieve better estimates
of the censoring distribution of the data, a key component in competing-risks
regression, and that is a feature we do plan on adding to -stcrreg-. They do
not, however, discuss stratification in the sense of allowing different
subhazard functions over groups while maintaining unified subhazards ratios,
and we unaware of any other reference that makes this extension.
Offhand I can think of no reason why stratification would not work just as it
does in Cox regression, but because with competing risks you are dealing with
subdistribution functions and not proper probability distribution functions it
is hard to say whether the theory will back that up. More investigation is
needed.
> Also my second problem is that I also would like to calculate the median
> survival time after the adjustment for competing causes. The stci command
> which works with stcox does not extent to stcrreg. Could I alternatively use
> stcompt to calculate the cumulative incidence curve and estimate it from the
> point of 50%?
Yes. Or you could alternatively fit an -stccreg- model with no covariates
(or with covariates for that matter) and use -stcurve- afterward to plot the
cumulative incidence curve. From that curve, you can estimate the 50% point.
Reference:
Fine, J. P. and R. J. Gray. 1999. A proportional hazards model for the
subdistribution of a competing risk. Journal of the American Statistical
Association. 94: 496-509.
- --Bobby
[email protected]
*
* 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/