[R] Is there an equivalent to predict(..., type="linear") of a Proportional hazard model for a Cox model instead?

Ben Rhelp benrhelp at yahoo.co.uk
Fri Nov 26 20:50:34 CET 2010

 Hi Terry, David, and Thomas,
Thank you for all your emails and the time you to took to clarify my 
misunderstanding on survival analysis. I will need a bit of time to digest all 
this information and to do some more reading.
Best regards,

> From: Terry Therneau
> 1. survreg() does NOT fit a proportional hazards model, a mistake
> repeated  multiple times in your post
> 2. The coxph function operates on the risk  scale: large values of Xbeta
> = large death rates = bad
>    The survreg  operates on the time scale: large values of xbeta =
> longer liftetime =  good.
> 3. predict(fit, type='risk') = exp(predict(fit, type='linear')) in  a Cox
> model returns an estimate of the relative risk for each subject.   That
> is, his/her predicted death rate as compared to the others in  the
> sample.  It has no units of "years" or "days" or anything  else.  The
> predicted survival TIME for a subject is something else  entirely.
>   predict(fit, type='response') in a survreg model does  give predicted
> survvival times.  
>   If you really want to  understand the interrelationships of these
> things more deeply I think you  need some textbook time.  Read the book
> by Kalbfleisch and Prentice for  accelerated failure time models, or even
> better Escobar and Meeker which  comes from the industrial reliability
> view.  For predicted survival from  a Cox model see Chapter 10 of
> Therneau and Grambsch.  The answers to  your specific questions would be
> a document rather than an  email.
> Terry Therneau

More information about the R-help mailing list