Re: Calibration and discrimination in survival models

[Frank E Harrell Jr <f.harrell@vanderbilt.edu>]

bruno leal wrote:
> Thank you very much, prof. Harrell.
>  
> I was able to sort that out . I have calculated c and used groupkm to 
> compare expected vs. observed probabilities. 
> However, I was wondering if there is an s-plus function to calculate an 
> overall  goodness-of-fit measure that could be applied in my situation 
> (i.e, having a prespecified model and new data). D'Agostino and Nam have 
> proposed a measure similar to the Hosmer-Lemeshow test and used in 
> D'Agostino et al. Validation of the Framingham Coronary Heart Disease 
> Prediction Scores. JAMA 2001; 286:180-187 and in several other 
> publications. Although the H-L test is limited by an arbitrary selection 
> of the number of partitions, is such a measure available from any S-plus 
> functions?
>  
> Thank you once again
>  
> Bruno Leal.
There have been several papers on goodness of fit in survival data 
published in the literature.  Some of them are referenced in my 
bibliographic database at the bottom of http://biostat.mc.vanderbilt.edu/rms
Frank

> */Frank E Harrell Jr <f.harrell@vanderbilt.edu>/* wrote:
>
>     bruno leal wrote:
>      > Dear S-Plus users,
>      >
>      >
>      >
>      > I have a pre-specified model (i.e., I have the Cox model
>     equation, but I
>      > don't have the original data from where it was derived) for
>     predicting
>      > risk at 5 years for a particular condition. I am interested in
>     checking
>      > the discrimination and calibration of this model on a new
>     dataset. I am
>      > aware that the design library has tools to check discrimination and
>      > calibration and their use seems to be straightforward to check
>      > discrimination and calibration during the steps of model
>     development.
>      > However, I am not sure how to use the functions to test a
>     pre-specified
>      > model on a new data set.
>      >
>      >
>      >
>      > I would appreciate any help I can get.
>      >
>      >
>      >
>      > Bruno Leal
>      >
>      >
>      >
>      >
>      >
>
>     When the model is fully prespecified, including the underlying survival
>     curve and all the regression coefficients, use rcorr.cens to estimate
>     the predictive discrimination and look at groupkm for calibration
>     curves. Also look at the more experimental val.surv function. All of
>     this is under the area of external validation.
>
>     Frank
>
>     -- 
>     Frank E Harrell Jr Professor and Chair School of Medicine
>     Department of Biostatistics Vanderbilt University

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