Skip to content

standard error of survfit.coxph()

1 message · Terry Therneau

#
1. The computations "behind the scenes" produce the variance of the cumulative hazard. 
This is true for both an ordinary Kaplan-Meier and a Cox model.  Transformations to other 
scales are done using simple Taylor series.

   H = cumulative hazard = log(S);  S=survival
   var(H) = var(log(S))  = the starting point
   S = exp(log(S)), so  var(S) is approx [deriv of exp(x)]^2 * var(log(S)) = S^2 var(H)
   var(log(log(S)) is approx (1/S^2) var(H)

2. At the time it was written, summary.survfit was used only for printing out the survival 
curve at selected times, and the audience for the printout wanted std(S).   True, that was 
20 years ago, but I don't recall anyone ever asking for summary to do anything else.  Your 
request is not a bad idea.
   Note however that the primary impact of using log(S) or S or log(log(S)) scale is is on 
the confidence intervals, and they do appear per request in the summary output.

Terry T.
On 06/28/2014 05:00 AM, r-help-request at r-project.org wrote: