[R] Time dependent Cox model

Marco Barbàra jabbadhutt at libero.it
Fri May 21 15:18:22 CEST 2010



> > ... interactions between covariables and time.
> 
>   A model such as "coxph(Surv(ptime, pstat) ~ age + age*ptime, ...."
> is invalid -- it is not at all what you think.  

Actually what i'm trying to fit is 

coxph(Surv(start,end,event)~age+age:start)

to model a time-varying effect \beta(t)=a+b*t.

Conceptually, is not H_0:b=0 the hypothesis tested by cox.zph()?

(i know: i have to study more deeply this topic, my copy of your book
is on the way to my house)

> 
> > Is is somewhat sensible to use cox.zph() to investigate which
> variables need time interaction...
> 
>   The cox.zph function is primarily graphical; I would respond to your
> question with "is it good to look at scatterplots before fitting a
> linear model?"  My answer to this is emphatically yes.  
> 
> Terry Therneau
> 
> 

Thank you,

Marco Barbàra.



More information about the R-help mailing list