Date:         Fri, 18 Sep 1998 14:26:07 GMT
Reply-To:     Candy Kane <candykane@CANDYLAND.ORG>
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         Candy Kane <candykane@CANDYLAND.ORG>
Organization: http://extra.newsguy.com
Subject:      Re: Testing proportional hazards assumption with PROC PHREG
Content-Type: text/plain; charset=us-ascii

On Thu, 17 Sep 1998 19:41:47 GMT, mgonen@my-dejanews.com wrote:
 
>Dear Colleagues:
>
>I am trying to test the proportional hazards (PH) assumption using a time
>dependent covariate. I am aware of the statistical shortcomings of this
>method, but I want to ask a specific question regarding how SAS does this.
>
>SAS manual suggests the following code:
>
>PROC PHREG DATA=TEST;
>  MODEL TIME*CENSOR(0) = GROUP X;
>  X = GROUP*(LOG(TIME));
>RUN;
>
>(in fact they suggest subtracting the mean from LOG(TIME), but that is
>irrelevant
>for my purposes). Let's call this the first method.
>
>Essentially, this code is supposed to create a time-dependent variable X. If
>the coefficient of X is significant then we conclude PH assumption is
>violated (again I am aware of the shortcomings, but please keep on reading).
>
>I try creating the same variable in a data step before the PHREG:
>
>DATA TEST;
>  SET TEST;
>  X = GROUP*(LOG(TIME));
>RUN;
>
>PROC PHREG DATA=TEST;
>  MODEL TIME*CENSOR(0) = GROUP X;
>RUN;
>
>and I get very very different results. The p-value for X using the first
>method is 0.99 and the second method is 0.0001.
>
>A colleague of mine says that he attended a survival analysis class where the
>instructor specifically said the second approach is incorrect.
>
>I am at a loss here. Does anybody have any ideas of what is going on?
 
The data step method is incorrect. The X variable needs to be
calculated dynamically within the procedure.
 
Jay Weedon.