R help - Jul 2011 - Plotting survival curves from a Cox model with time dependent covariates

Dear all,

Let's assume I have a clinical trial with two treatments and a time to 
event outcome. I am trying to fit a Cox model with a time dependent 
treatment effect and then plot the predicted survival curve for one 
treatment (or both).

library(survival)
test <- 
list(time=runif(100,0,10),event=sample(0:1,100,replace=T),trmt=sample(0:1,100,replace=T))
model1 <- coxph(Surv(time, event) ~ tt(trmt), data=test, tt=function(x, 
t, ...) pspline(x + t))
newdat1 <- data.frame(trmt=1,time=list(0,1,2,3,4,5))
plot(survfit(model1,newdata=newdat1,individual=T), xlab = "Years", 
ylab="Survival")

Where I think I am failing is with how to correctly specify what I want 
the survfit function to do. My understanding on reading the 
documentation for the survival package is that I should use newdata to 
not only specify the treatment, but also timepoints for which I want 
survival estimates and that this is the scenario for which the 
individual=T option can be appropriate. However, I just seem to fail to 
figure out exactly how I should specify this.

It would be greatly appreciated if someone who has done this before or 
knows how to do it could give me a quick (or extensive, of course) hint.

Many thanks,
Bj?rn

PS: Yes, I realise that a Kaplan-Meier plot would do something like the 
above very nicely, but once I get this to work, I am actually looking at 
something a bit more complicated where a KM plot would not help me.

The time-transform (tt() arguments) feature is the most recent addition
to coxph.  Most of the follow-up functions, in particular survfit(fit)
have not yet been updated to deal with such models.  Your message points
out that I need to at least update them to add a "not yet available"
error message.  
  I'm still learning what can be done with the tt() option.  If I assume
that your trmt variable is 0/1, then the code below is a clever way to
look at time dependent treatment effects.  I had not thought of it.

 Terry Therneau

--- begin included message ---

Let's assume I have a clinical trial with two treatments and a time to 
event outcome. I am trying to fit a Cox model with a time dependent 
treatment effect and then plot the predicted survival curve for one 
treatment (or both).

library(survival)
test <- 
list(time=runif(100,0,10),event=sample(0:1,100,replace=T),trmt=sample(0:1,100,replace=T))
model1 <- coxph(Surv(time, event) ~ tt(trmt), data=test, tt=function(x, 
t, ...) pspline(x + t))