[R] predicting cumulative hazard for coxph using predict

Terry Therneau therneau at mayo.edu
Tue Feb 24 17:23:32 CET 2009


> I am estimating the following coxph function with stratification and 
> frailty, where each person had multiple events.

> m <- coxph(Surv(dtime1,status1) ~ gender +cage +uplf+ strata(enum)+ 
>               frailty(id), xmodel)

>I want to predict the cumulative hazard for each person for their next event at
> a given time in future.What is the best way to compute it? 

 There are two issues.  The first is that what you are asking for -- it is not 
at all clear to me exactly how this should be computed, and as a consequence the 
coxph routines/documentation are not clear.  When there is a random effect in 
the model, should the baseline hazard be computed before or after integrating 
out the random effect?  How should one define the cumulative hazard for a 
subject?  These are hard questions, or at least I've found them so.
 
 What is currently done: The basline hazard is computed treating each of the 
fitted random effects (one per id) as it if were a known fixed value.  The 
predict function treats the random effects as just another coefficient, which is 
clearly the wrong thing to do wrt a new subject who was not in the original data 
set. 
   If you think of the frailty term as a shrinkage estimator, with no particular 
connection to random effects, then what is currently done makes reasonable 
sense.  The "random effects" are just more coefficients in the model.   
   
   Terry Therneau




More information about the R-help mailing list