Message-ID: <9b53cf$ga2o4j@ironport10.mayo.edu>
Date: 2021-07-12T22:13:53Z
From: Terry Therneau
Subject: syvcoxph and cox.zph for testing the PH assumption
In-Reply-To: <mailman.363590.1.1625997625.38277.r-help@r-project.org>
On 7/11/21 5:00 AM, r-help-request at r-project.org wrote:
> Hello, is it kosher to call cox.zph on a syvcoxph model fit? I see that
> someone proposed a modified version of cox.zph that uses resid(fit,
> 'schoenfeld', **weighted=TRUE**).
>
> https://stats.stackexchange.com/questions/265307/assessing-proportional-hazards-assumption-of-a-cox-model-with-caseweights
> Is that all it takes?
> Thanks,
> Youyi
The cox.zph function does a formal score test. No, it does not account for robust
variance. I hadn't considered that case, but will now think about it. It is quite easy
to show that there is a problem: just give everyone a weight of 100.
The stackexchange conversation was new to me. The solution there won't work with the
current code, which does not make use of resid(). It has been updated to do the proper
score test, the older version of cox.zph, which they modified, used an approximation.
Terry T.