R help - Dec 2011 - Problem of COX model with time dependent covariate

Hi all,


I am trying to detect association between a covariate and a disease outcome
using R. This covariate shows time-varying effect, I add a time-covariate
interaction item to build Cox model as follows:
COX <-
coxph(as.formula("Surv(TIME,outcome)~eGFR_BASE+eGFR_BASE:TIME"),ori.data);


                                              coef   exp(coef)   e(coef)       
z   p
eGFR_BASE                 6.40     603.5133     0.3702    17.3   0
eGFR_BASE:TIME -3.41      0.0329     0.0772   -44.2   0


But the result seems very different from that got by SPSS: 


                                                    B            	SE    	Wald	  
df    	Sig.	    Exp(B)
eGFR_BASE        	        -13.603	  .548	  616.007	   1	     .000    	.000  
T_COV_*eGFR_BASE	.860 	.115	  55.859  	1	  .000	        2.362   



so my questions are :
1, Is it correct to deal with time dependent covariate in Cox regression using
R? if not, how to do it ?
2, How T_COV_ was calculated in SPSS, and is it the same with the survival time?


Many thanks.
GZ
 		 	   		  
	[[alternative HTML version deleted]]

On Dec 26, 2011, at 3:02 AM, JiangGZ wrote:
>
> Hi all,
>
>
> I am trying to detect association between a covariate and a disease  
> outcome using R. This covariate shows time-varying effect, I add a  
> time-covariate interaction item to build Cox model as follows:
> COX <- coxph(as.formula("Surv(TIME,outcome)~eGFR_BASE 
> +eGFR_BASE:TIME"),ori.data);
>
>
>                           coef   exp(coef)   e(coef)        z   p
> eGFR_BASE                 6.40     603.5133     0.3702    17.3   0
> eGFR_BASE:TIME -3.41      0.0329     0.0772   -44.2   0
>
>
> But the result seems very different from that got by SPSS:
>
>
>                    B    SE    	Wald	    df    	Sig.	    Exp(B)
> eGFR_BASE       -13.603	  .548	  616.007	   1	     .000    	.000
> T_COV_*eGFR_BASE .860 	.115	  55.859  	1	  .000	        2.362
>
>
>
> so my questions are :
> 1, Is it correct to deal with time dependent covariate in Cox  
> regression using R? if not, how to do it ?
The usual way to represent time dependent covariates is to set up  
records with start and stop time and use the Surv(time, time2, event)  
usage. This is clearly described in the help page and there is  
vignette describing time-dependent covariate methods in the package  
documentation.


> 2, How T_COV_ was calculated in SPSS, and is it the same with the  
> survival time?
How could we possibly know how _you_ calculated T_COV? (And this is  
very definitely the wrong mailing list for that question.)

The description in the IBM/SPSS documentation says using the time  
variable on the RHS is used for testing non-proportionality of hazards  
rather than for investigating the association of a covariate with an  
outcome:

http://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=3&ved=0CDcQFjAC&url=http%3A%2F%2Fwww.unileon.es%2Fficheros%2Fservicios%2Finformatica%2Fspss%2Fenglish%2FIBM-SPSS_advanced_statistics.pdf&ei=P2j4TtbDMITi0QG6k-2DAg&usg=AFQjCNH_4hFxKUdl-Ie1VDxEFXRL35nIqw

The SPSS documentation appears designed to lead the naive user into an  
incorrect model formulation, since the only example of use with its  
dialog box shows the user setting up the test for non-proportionality.  
Given that you proposed using a time variable on the RHS in an R  
formula, I am guessing you also do not know how to properly use SPSS  
for the purposes you intend.

-- 
David.
>
>
> Many thanks.
> GZ
> 		 	   		
> 	[[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
David Winsemius, MD
West Hartford, CT

I'll add just a little to what David said. 
The normal meaning of a "time dependent covariate" is a covariate that
changes with time.  For instance in a model that included x="most recent
available blood pressure" the value of x will change at each patient
visit.  You obviously can't get those new values as x=baseline + c*time
unless you are willing to assume a very odd biology for the patient.  I
would think the same about glomular filtration rate (GFR).

A different question is a time dependent coefficient.  That is, the
variable stays the same but it's effect changes over time.  Now in this
case a model that beta(t) = a + bt has some possible merit, but in my
optinion the biology it implies is still quite odd.  So I don't find
that model very interesting, even though some packages impliment it.  In
R there is a more useful approach:
  fit <- coxph(Surv(time, outcome) ~ eGFR.base, ori.data)
  zfit <- cox.zph(fit)
  plot(zfit)

This produces a smoothing spline estimate of beta(t) with confidence
bands.  You can look at the plot and begin to understand the data, not
just create p-values.

Terry T.



On Dec 26, 2011, at 3:02 AM, JiangGZ wrote:
>
> Hi all,
>
>
> I am trying to detect association between a covariate and a disease  
> outcome using R. This covariate shows time-varying effect, I add a  
> time-covariate interaction item to build Cox model as follows:
> COX <- coxph(as.formula("Surv(TIME,outcome)~eGFR_BASE 
> +eGFR_BASE:TIME"),ori.data);
>
>
>                           coef   exp(coef)   e(coef)        z   p
> eGFR_BASE                 6.40     603.5133     0.3702    17.3   0
> eGFR_BASE:TIME -3.41      0.0329     0.0772   -44.2   0
>
>
> But the result seems very different from that got by SPSS: