s-news
[Top] [All Lists]

Re: C-statistic comparison

To: Felipe Medeiros <fmedeiros@uol.com.br>
Subject: Re: C-statistic comparison
From: Frank E Harrell Jr <f.harrell@vanderbilt.edu>
Date: Fri, 28 Jan 2005 22:40:09 -0500
Cc: "'Naomi B. Robbins'" <n.robbins@att.net>, "'s-news'" <s-news@wubios.wustl.edu>
In-reply-to: <20050128214518.ADF667BD8@scorpion5.uol.com.br>
References: <20050128214518.ADF667BD8@scorpion5.uol.com.br>
User-agent: Mozilla Thunderbird 0.9 (X11/20041105)
Felipe Medeiros wrote:
Does anybody know of an s-plus function or method to compare two
c-statistics (Harrell's c) obtaining from two different models applied to
the same group of patients (in a survival analysis context)?

Thank you for any help.

Felipe Medeiros.

This is not the comparison most people do (they test for a difference in C indexes) but the rcorrp.cens function is more powerful although it tests a different hypothesis. Note that neither rcorrp.cens nor rcorr.cens are meant to be used when there is a chance of overfitting (e.g., if evaluating on the training data, and the ratio of events to the total of all CANDIDATE preditors is less than 20). However if both models have the same amount of overfitting (had the same CANDIDATE degrees of freedom, no matter how many parameters were in the final model) then rcorrp.cens may still work.

Frank Harrell

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

<Prev in Thread] Current Thread [Next in Thread>