...
If you suspect that there is a problem with proportionality, then you should investigate that first.
I'd use -estat phtest, log detail- to test the proportionality assumption on the full model with other relevant covariates. The dichotomous variable may fail the proportionality test in a model by itself, but this might disappear when you condition on other variables.
The phtest is, however, only a significance test and its significance or otherwise is going to depend on how many events are observed during follow-up since that will affect the power of the test. So, even if the test is not significant, non-proportionality may still be problematic.
A graph of the Schoenfeld residuals overlayed with a lowess smoother will enable you to see what's going on in the first 48 hours.
If, after this, it still appears that the effect of the dichotomous variable is changing over time, one option would be to use spline to model this effect over time. There is an example of this in:
Gray RJ (1992). Flexible Methods for Analyzing Survival-Data Using Splines, with Applications to Breast-Cancer Prognosis. J Am Stat Assoc 87(420): 942-951.
______________________________________________
Kieran McCaul MPH PhD
WA Centre for Health & Ageing (M573)
University of Western Australia
Level 6, Ainslie House
48 Murray St
Perth 6000
Phone: (08) 9224-2701
Fax: (08) 9224 8009
email: [email protected]
http://myprofile.cos.com/mccaul
http://www.researcherid.com/rid/B-8751-2008
______________________________________________
If you live to be one hundred, you've got it made.
Very few people die past that age - George Burns
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of moleps islon
Sent: Monday, 31 August 2009 1:38 AM
To: [email protected]
Subject: st: Survival analysis
Dear listers,
I´m investigating the effect of a dichotomous variable on survival.
Looking at the graphs they are parallell after the first 48 hrs, but
there is a difference within the first 48 hrs. Is it reasonable to
present the cox regression analysis with all the patients and then
discard the patients dead within 48 hrs and redo the analysis-(this
gives a significant effect in the first analysis and non-significant
in the latter as suspected from the graphs)? Or is there a more
analytical way to find the exact time from which the two groups no
longer differ?
Regards,
M
*
* 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/
