s-news
[Top] [All Lists]

Re: Time dependent covariate assignment in Cox models

To: s-news@lists.biostat.wustl.edu, kieke.burney@mcrf.mfldclin.edu
Subject: Re: Time dependent covariate assignment in Cox models
From: Terry Therneau <therneau@mayo.edu>
Date: Mon, 15 Oct 2007 17:08:14 -0500 (CDT)
Reply-to: Terry Therneau <therneau@mayo.edu>
> Is the general convention to assign a value of 'yes' to the time dependent
> covariate on day A or day A+1? 

  The general convention is on day A+1.
  This avoids several problems that can occur with assignment on day A.  An 
example is increase in PSA (prostate specific antigen) and the declaration of 
progression of prostate cancer.  If the measurement and the declaration occur 
on 
the same visit, it may well be that the measurement does not really predict 
cancer, it predicts that the physician will decide that there is progression. 
(And in any case, is it really interesting to be able to predict a medical 
outcome that will happen tomorrow?)
   In our group, we often move things to day A+3, due to the fact that a 
patient's visit may span multiple days.
   A counter-example would be where the covariate is administration of a 
treatment, and one patient died of an allergic reaction a few hours after 
recieving the drug.  One would certainly count this death on the treatment arm. 
 
But instantaneous cause-effect outcomes like this are unusual.  
  
  
        Terry Therneau

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