Date: Mon, 20 Aug 2007 17:13:56 -0500
Reply-To: Mary <mlhoward@avalon.net>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Mary <mlhoward@AVALON.NET>
Subject: Re: proportional hazards regression
Content-Type: text/plain; format=flowed; charset="iso-8859-1";
reply-type=original
Cornel,
Proportional Hazards is measuring the time to an event. It seems like your
data is saying that people who have a more recent negative comorbid outcome
are likely to have the event.
Consider the group that had an event, say a heart failure or diabetes, 10
years ago, but haven't had anything in the past 3 years. Recognize that
that group is not ALL of the people who had that comorbid outcome 10 years
ago, only the people who LIVED, since proportional hazards does not model
those who died.
Perhaps those who lived 10 years ago radically changed their lifestyle or
had other factors that made them live over those that died.
But of the group that has had an outcome within the past 3 years, some of
that group will not live 10 years, and thus if we repeated the study in 10
years those persons who died would not be included in the proportional
hazards of surviving.
Think Hope, Arkansas: Mike Huckabee and Bill Clinton (both Governors of
Arkansas) both radically changed their lifestyles following cardiac
problems/diabetes (lost weight, exercised, etc.). They are alive whereas
people less motivated or able to change may have already died. Thus the
"survivors" are more likely to be included in your analysis from the group
who had co-morbidities 10 years ago than those who did not survive.
An alternative to getting at this would be to do other methods in addition
to proportional hazards, such as logistic regression, that include the
people who did not survive from an comorbid outcome 10 years ago.
-Mary
----- Original Message -----
From: "Cornel Lencar" <clencar@INTERCHANGE.UBC.CA>
To: <SAS-L@LISTSERV.UGA.EDU>
Sent: Monday, August 20, 2007 3:31 PM
Subject: proportional hazards regression
> Hi,
>
> I am running a roportional hazards model on a cohort of ~ 870000 subjects
> between 45 and 85 years old. I am looking at the risk of having acute
> coronary outcomes function of sex, age class and presence of comorbidities
> in the past. The start of follow up is january 1, 1999, and the end of F/U
> is december 2002. Co-morbidities are acute coronary outcomes, chronic
> coronary outcomes, chronic heart failure, hypertension, diabetes, all
> determined on the period 1991, dec 31, 1998.
>
> So far so good. The results considering co-morbidities as 1=Yes, 0=No
> only, plus age and sex look as expected, with males having higher risk, an
> increase in risk with age and an increase in risk with the presence of co-
> morbidities (i.e. HR=3.1 in one example).
>
> However, if I split the subjects with comorbidities in two, (ones that had
> co-morbidities less than 3 years prior to the outcome of interest in the
> follow-up period and the ones that had co-morbidities earlier than that -
> more than 3 years prior to the outcome of interest), I get some
> inexplicable results:
> HR=322.7 (co-morbidities < 3 years prior)
> HR=101.77 (co-morbidities > 3 years prior);
> all the other covariates (sex, age) being the same.
>
> A cross tabulation between co-morbidities and outcome of interest looks
> like this:
>
> Co-morbid(0)/ACS(0) = 555468
> Co-morbid(1)/ACS(0) = 283577
> Co-morbid(0)/ACS(1) = 9597
> Co-morbid(1)/ACS(1) = 21425 or if split by time 1484(<3yr) & 19941(>3yr);
>
> What can cause such great hazard ratios?
>
> Sincerely,
>
> Cornel Lencar
>
| 
