R help - Jul 2009 - negative Somers D from Design package

Dear R help

My problem is very similar to the analysis detailed here.
If we use the mayo dataset provided with the survivalROC package the estimate
for Somer's Dxy is very negative -0.56.
The Nagelkerke R2 is positive though 0.32.
I know there is a difference between explained variation and predictive ability
but I am surprised there is usch a difference given that even a non predictive
model should have Dxy around 0.
Am I doing something wrong or is there an interpretation that makes sense ?

This is with the mayo data so its reproducible but the result with my data is
very similar.
Many thanks in advance

library(survivalROC)
library(Design)
library(survival)
data(mayo)

 Sm <- Surv(mayo$time,mayo$censor)
fm <- cph( Sm ~ mayoscore4    ,mayo,x=T,y=T,surv=T )
validate(fm, B=150,dxy=T)
Iteration 1 ....

        index.orig     training         test      optimism index.corrected   n
Dxy   -0.566027923 -0.566665407 -0.566027923 -0.0006374833    -0.565390440 150
R2     0.325860603  0.327350885  0.325860603  0.0014902826     0.324370320 150
Slope  1.000000000  1.000000000  0.987854765  0.0121452354     0.987854765 150
D      0.093398440  0.095166239  0.093398440  0.0017677983     0.091630642 150
U     -0.001562582 -0.001579618  0.001150175 -0.0027297932     0.001167211 150
Q      0.094961022  0.096745857  0.092248266  0.0044975915     0.090463431 150



Dr Bernard North
Statistical Consultant
Statistical Advisory Service
Advice and Courses on Research Design and Methodology
Imperial College
South Kensington Campus
Room 845, 4th Floor
8 Princes Gardens
London SW7 1NA.

Tel: 020 7594 2034
Fax: 020 7594 1489
Email: bnorth@imperial.ac.uk
Web:  www.ic.ac.uk/stathelp




	[[alternative HTML version deleted]]

For the Cox model Dxy is the rank correlation between predicted log 
hazard and time to event.  As a high hazard means that the time to event 
is short, you need to negate Dxy for your purpose.

Frank


North, Bernard V wrote:
> Dear R help
> 
> My problem is very similar to the analysis detailed here.
> If we use the mayo dataset provided with the survivalROC package the
estimate for Somer's Dxy is very negative -0.56.
> The Nagelkerke R2 is positive though 0.32.
> I know there is a difference between explained variation and predictive
ability but I am surprised there is usch a difference given that even a non
predictive model should have Dxy around 0.
> Am I doing something wrong or is there an interpretation that makes sense ?
> 
> This is with the mayo data so its reproducible but the result with my data
is very similar.
> Many thanks in advance
> 
> library(survivalROC)
> library(Design)
> library(survival)
> data(mayo)
> 
>  Sm <- Surv(mayo$time,mayo$censor)
> fm <- cph( Sm ~ mayoscore4    ,mayo,x=T,y=T,surv=T )
> validate(fm, B=150,dxy=T)
> Iteration 1 ....
> 
>         index.orig     training         test      optimism index.corrected 
n
> Dxy   -0.566027923 -0.566665407 -0.566027923 -0.0006374833    -0.565390440
150
> R2     0.325860603  0.327350885  0.325860603  0.0014902826     0.324370320
150
> Slope  1.000000000  1.000000000  0.987854765  0.0121452354     0.987854765
150
> D      0.093398440  0.095166239  0.093398440  0.0017677983     0.091630642
150
> U     -0.001562582 -0.001579618  0.001150175 -0.0027297932     0.001167211
150
> Q      0.094961022  0.096745857  0.092248266  0.0044975915     0.090463431
150
> 
> 
> 
> Dr Bernard North
> Statistical Consultant
> Statistical Advisory Service
> Advice and Courses on Research Design and Methodology
> Imperial College
> South Kensington Campus
> Room 845, 4th Floor
> 8 Princes Gardens
> London SW7 1NA.
> 
> Tel: 020 7594 2034
> Fax: 020 7594 1489
> Email: bnorth at imperial.ac.uk
> Web:  www.ic.ac.uk/stathelp
> 
> 
> 
> 
> 	[[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.
> 

-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University