R help - Mar 2012 - How to calculate the Deviance for test data based on Cox model

Dear List, 

If I got a Cox model based on training set, then how should I calculate the Cox
log partial likelihood for the test data?
Actually I am trying to calculate the deviance on test dataset to evaluate the
performance of prediction model, the equation is as follows: D =
-2{L(test)[beta_train] - L(test)[0]}. It means using the beta coefficients got
from training set to calculate the likelihood of test data. I know I can got log
likelihood for training model, but how to get it for test data?
Any package or function can do that? Thanks. 		 	   		  
	[[alternative HTML version deleted]]

On Mar 29, 2012, at 6:03 AM, ChiangKevin wrote:
>
>
> Dear List,
>
> If I got a Cox model based on training set, then how should I  
> calculate the Cox log partial likelihood for the test data?
> Actually I am trying to calculate the deviance on test dataset to  
> evaluate the performance of prediction model, the equation is as  
> follows: D = -2{L(test)[beta_train] - L(test)[0]}. It means using  
> the beta coefficients got from training set to calculate the  
> likelihood of test data. I know I can got log likelihood for  
> training model, but how to get it for test data?
Well, that is an equation that has some aspects familiar to many of  
us,  but it is one that does not apply to the statistical problem for  
which you propose to use it. It applies to "nested" model comparisons,
nested both in the sense that the data is the same and the terms are  
shared.
> Any package or function can do that? Thanks. 		 	   		
> 	[[alternative HTML version deleted]]
I think you need to seek statistical consultation.

-- 

David Winsemius, MD
West Hartford, CT

On Thu, Mar 29, 2012 at 11:03 PM, ChiangKevin
<kevinchiang865 at hotmail.com> wrote:
>
>
> Dear List,
>
> If I got a Cox model based on training set, then how should I calculate the
Cox log partial likelihood for the test data?
> Actually I am trying to calculate the deviance on test dataset to evaluate
the performance of prediction model, the equation is as follows: D =
-2{L(test)[beta_train] - L(test)[0]}. It means using the beta coefficients got
from training set to calculate the likelihood of test data. I know I can got log
likelihood for training model, but how to get it for test data?
One way to do it is to get the linear predictors for the test set  (eg
with predict.coxph) and use them as an offset
   coxph(Surv(time,status)~offset(lp), data=testdata)

The loglikelihood reported will be the log partial likelihood
evaluated at the test data and the fitted parameter values

-- 
Thomas Lumley
Professor of Biostatistics
University of Auckland