R help - Mar 2013 - Weighted Kaplan-Meier estimates with R

There are two ways to view weights.  One is to treat them as case weights, i.e.,
a weight
of 3 means that there were actually three identical observations in the primary
data,
which were collapsed to a single observation in the data frame to save space. 
This is the
assumption of survfit.  (Most readers of this list will be too young to remember
when
computer memory was so small that we had to use tricks like this.)  The second
assumption
is that the weights are sampling weights and a Horvitz-Thompsen like estimator
should be
used for variance. This is the assumption of the svykm program in the survey
package.  It
appears you want the second behavior.

Terry Therneau


On 03/26/2013 06:00 AM, r-help-request at r-project.org
wrote:
> As part of a research paper, I would like to draw both weighted and
> unweighted Kaplan-Meier estimates, the weight being the ?importance? of the
> each project to the mass of projects whose survival I?m trying to estimate.
>
> I know that the function survfit in the package survival accepts weights
and
> produces confidence intervals. However, I suspect that the confidence
> intervals may not be correct. The reason why I suspect this is that
> depending on how I define the weights, I get very different confidence
> intervals, e.g.
>
> require(survival)
> s<- Surv(c(50,100),c(1,1))
> sf<- survfit(s~1,weights=c(1,2))
> plot(sf)
>
> vs.
>
> require(survival)
> s<- Surv(c(50,100),c(1,1))
> sf<- survfit(s~1,weights=c(100,200))
> plot(sf)
>
> Any suggestions would be more than welcome!
>

Thanks to all of you for your very helpful comments! As Terry suggested,
svykm is what I was looking for.

While testing that I get the same results with the package survey as with
the package survival, I encountered the issue of how to draw survival
curves.  Apparently the implementations in the two packages differ, as I
show below.

I would very much welcome your views, since the tail of the survival curve
has a major impact on the interpretation of my results. In my data, the last
?death? occurs at 2094 days, while the last censoring time is 3297 days.  If
possible, I would like to say something about the probability between 2100
days and 3300 days.

So, my question is that after the last observed death, in the very simple
example below at 883 days, how should one draw the survival curve? The graph
produced by svykm (package survey) ends at 883 days, whereas survfit
(package survival) continues the graph all the way to the last censoring
time, which is at 1022 days.

Please run the code below to see my point. There are no weights. With
weights, I face the same issue.

require(survival)
lungSubSet <- lung
S <- Surv(lung$time,lung$status)
sKm <- survfit(S~1)
##
require(survey)
svyDesignObject<- svydesign(id=~1,weights=~1,data=lungSubSet)
svyKm <- svykm(S~1,design=svyDesignObject,se=T)
##
plot(svyKm,xlim=c(0,1200))
lines(sKm,conf.int=T,mark.time=F,col='green')




--
View this message in context:
http://r.789695.n4.nabble.com/Re-Weighted-Kaplan-Meier-estimates-with-R-tp4662494p4662619.html
Sent from the R help mailing list archive at Nabble.com.