Date:         Wed, 25 Jun 2008 12:39:43 -0700
Reply-To:     Shawn Haskell <shawn.haskell@STATE.VT.US>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Shawn Haskell <shawn.haskell@STATE.VT.US>
Organization: http://groups.google.com
Subject:      Re: Handling dropouts in phreg
Comments: To: sas-l@uga.edu
Content-Type: text/plain; charset=ISO-8859-1

On Jun 24, 4:51 pm, p...@COLUMBIA.EDU (Pamela Nadash) wrote:
> My study looks at all individuals who enroll in a program during a certain
> time period x to y (18 months).  There are three outcomes: they can live,
> die, or drop out (~13% of sample).  This is the code I used:
>
> proc phreg data=compdata;
>    model LOS_UT*death(0) = program age gender alone asian black
> hispanic cog copd chf cvd dbm arth hbp dep_anx comorb75 comorb10 inc_bowel
> iadlHI iadlLO adl75 adl0 cgnone blacklos cogadl cogiadl / rl;
>   blacklos= black*los_ut;
>  Title 'Cox Model Predicting Days to Death';
> run;
>
> Where los_ut is the number of days an individual is enrolled in the
> program -- ie days to death, if the individual dies.  Talking this over
> with my dissertation committe, I was told that I should have handled
> dropouts differently.  (I didn't exclude them.  I have run the analysis
> where I deleted dropouts and the results were pretty similar.)  I've tried
> to find literature on this but failed.  Would be grateful for advice.
>
> Also, it was said that I should have adjusted for time enrolled in the
> program.  But I can't understand this comment, because the los_ut is the
> relevant variable, and it's already in the model on the LHS.  The person
> who made this comment is a very smart statistician, so he must have meant
> something I'm not quite getting.  Any ideas?  I can ask him but I'd rather
> try to figure it out first (with your kind help).
>
> Many thanks,
> Pamela Nadash

Pamela, your prof may be eluding to time-dependent covariate effects
that would violate the proportional hazards assumption - if they
exist.  The meaning of the PH assumption is not always intuitive, and
it may not always be obvious which covariates (you do have a lot of
them) are likely to violate the assumption and which would not. Thus,
test for it, and remediate with some sort of extended or stratified
Cox model if necessary.  After you have a good understanding of the Ph
assumption, it could be an important finding that "such and such"
covariates do or do not violate the assupmtion.  There are a lot fo
survival texts out there, and I found Kleinbaum and Klein's text
"survival analysis: a self-learning guide" to be very helpful for me,
with the added pleasure of some SAS code.  between a text and the SAS
documentation you should find ample info on right-censorship, which is
part of most survival anlayses. hope that helps.  Shawn