R help - Mar 2012 - Difference in Kaplan-Meier estimates plus CI

I thought this would be trivial, but I can't find a package or function
that does this.

I'm hoping someone can guide me to one.

Imagine a simple case with two survival curves (e.g. treatment & control).

I just want to calculate the difference in KM estimates at a specific time
point (e.g. 1 year) plus the estimate's 95% CI.  The former is
straightforward, but the estimates not so much.

I know methods exist such as Parzen, Wei, and Ying, but was surprised not
to find a package that included this.

Before I code it up, I thought I'd ask if I was just missing it somewhere.

Thank you
Jason

	[[alternative HTML version deleted]]

On Thu, Mar 8, 2012 at 4:50 AM, Jason Connor <jconnor at alumni.cmu.edu>
wrote:
> I thought this would be trivial, but I can't find a package or function
> that does this.
>
> I'm hoping someone can guide me to one.
>
> Imagine a simple case with two survival curves (e.g. treatment &
control).
>
> I just want to calculate the difference in KM estimates at a specific time
> point (e.g. 1 year) plus the estimate's 95% CI. ?The former is
> straightforward, but the estimates not so much.
>
> I know methods exist such as Parzen, Wei, and Ying, but was surprised not
> to find a package that included this.
>
> Before I code it up, I thought I'd ask if I was just missing it
somewhere.
summary.survfit() in the survival package will give you the point
estimate and standard error, and then combining these into a
difference and confidence interval for the difference is easy.

   -thomas

-- 
Thomas Lumley
Professor of Biostatistics
University of Auckland

Did you try the survival package?



On Wed, 07-Mar-2012 at 10:50AM -0500, Jason Connor wrote:

|> I thought this would be trivial, but I can't find a package or
function
|> that does this.
|> 
|> I'm hoping someone can guide me to one.
|> 
|> Imagine a simple case with two survival curves (e.g. treatment &
control).
|> 
|> I just want to calculate the difference in KM estimates at a specific time
|> point (e.g. 1 year) plus the estimate's 95% CI.  The former is
|> straightforward, but the estimates not so much.
|> 
|> I know methods exist such as Parzen, Wei, and Ying, but was surprised not
|> to find a package that included this.
|> 
|> Before I code it up, I thought I'd ask if I was just missing it
somewhere.
|> 
|> Thank you
|> Jason
|> 
|> 	[[alternative HTML version deleted]]
|> 
|> ______________________________________________
|> R-help at r-project.org mailing list
|> https://stat.ethz.ch/mailman/listinfo/r-help
|> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
|> and provide commented, minimal, self-contained, reproducible code.

-- 
~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.   
   ___    Patrick Connolly   
 {~._.~}                   Great minds discuss ideas    
 _( Y )_  	         Average minds discuss events 
(:_~*~_:)                  Small minds discuss people  
 (_)-(_)  	                      ..... Eleanor Roosevelt
	  
~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.

--begin included message --
I thought this would be trivial, but I can't find a package or function
that does this.

I'm hoping someone can guide me to one.

Imagine a simple case with two survival curves (e.g. treatment &
control).

I just want to calculate the difference in KM estimates at a specific
time
point (e.g. 1 year) plus the estimate's 95% CI.  The former is
straightforward, but the estimates not so much.

--------------  end inclusion ----------

As others have pointed out, use "survfit".  Because the curves are
step
functions, they likely won't have a stored value in the data at exactly
365 days.  The summary() function for survival curves allows you to read
off the curve's values at any arbitrary time point.

A note on standard errors:  S(t) +- std is a terrible confidence
interval.  You will be much more accurate if you use log scale.  (Some
argue for logit or log-log, in truth they work well.)   If n is large
enough, however, you should be ok.

Terry T.

On Fri, Mar 9, 2012 at 3:08 AM, Terry Therneau <therneau at mayo.edu>
wrote:
> A note on standard errors: ?S(t) +- std is a terrible confidence
> interval. ?You will be much more accurate if you use log scale. ?(Some
> argue for logit or log-log, in truth they work well.) ? If n is large
> enough, however, you should be ok.
Very true, but if one really wants a confidence interval for
S_1(t)-S_2(t)  (not for S_1(t)/S_2(t)) then one is pretty much forced
to use the raw probability scale.

    -thomas


-- 
Thomas Lumley
Professor of Biostatistics
University of Auckland