[R] predicting from coxph with pspline

Terry Therneau therneau at mayo.edu
Mon May 12 20:13:15 CEST 2008


 -== begin included message ---------
 Hello.

I get a bit confused by the output from the predict function when used
on an object from coxph in combination with p-spline, e.g. 

fit <- coxph(Surv(time1, time2, status)~pspline(x), Data)
predict(fit, newdata=data.frame(x=1:2))

--------- end included ----------

  Yes, you should be confused.  Coxph still retains a well known problem with S 
models, namely that the prediction is incorrect when there are data-dependent 
transformations in the formula such as ns(), poly() or pspline().  That is, the 
set of basis functions chosen by pspline(x) depends on the range of x; for a new 
data prediction the basis functions are re-calculated, giving results that are 
wrong (unless the new x happens to have the exact same lower and upper limits).
  
  This is on my to-be-fixed list.  Once a few other things are cleared away 
(actually a long list of other things).   These fixes have been applied to lm 
and others for long enough now, coxph needs to catch up.
   
  	Terry Therneau



More information about the R-help mailing list