s-news
[Top] [All Lists]

Re: Kaplan Meier - confidence interval for the median

To: Chris.Barker@cvt.com
Subject: Re: Kaplan Meier - confidence interval for the median
From: Terry Therneau <therneau@mayo.edu>
Date: Tue, 23 Sep 2008 08:35:35 -0500 (CDT)
Cc: s-news@lists.biostat.wustl.edu
Reply-to: Terry Therneau <therneau@mayo.edu>
 >
 > I'd like to find out whether or not the method, used in SPLUS,  for
 >computing the confidence interval for the median of a Kaplan Meier
 >estimate is the method of
 >Brookmeyer, Crowley, "A confidence Interval for the Median Survival
 >Time", Biometrics, 38, 29-41, 1982.
 >
 > Interestingly, for the simple dataset of , (10, 15, 23, 30, 35, 52,
 >100+) (and + indicates censoring)
 > SPLUS gives, the median and CI as 30 (15, NA), (R reports the upper
 >limit as inf) 
 >
 > while SAS gives 30 (15,52)
 >
 >  Chris Barker, Ph.D.
 >Consulting Statistician
 >   CV Therapeutics
 >

   I don't have a copy of that issue of Biometrics so can't answer the question 
directly, but I can tell you what survfit does do.  I'm moderately sure that it 
is the method found in said paper.
   
   Given a set of pointwise confidence bands for the survival curve S(t), draw 
a 
horizontal line on the graph at S = .5.  The median survival is where that line 
intesects the curve S(t), the upper and lower confidence intervals are where it 
intersects the lower and upper pointwise bands, respectively.  If it does not 
intersect, the corresponding end of the interval is unknown -- how to notate 
that to the user is a matter of personal preference.
   
   Beware!  The quality of this procedure depends on the quality of the 
original 
pointwise bands.  There are several methods to generate the bands, and lots of 
papers comparing them.
    empirical likelihood (Thomas and Grunkmeier): excellent
    bootstrap methods: okay to excellent, depending on details of how its done
    transformed CI, log, logit, log-log, etc: very good.  Different papers rank
        different ones of these first in the list.
    plain,  S(t) +- 1.96* std: poor
    
There is no excuse for a modern package to use the last of these, given the 
literature.  Guess which one is the default in SAS?  

        Terry Therneau
        


<Prev in Thread] Current Thread [Next in Thread>