Date:         Fri, 19 Apr 1996 14:19:58 GMT
Reply-To:     c.kane@fertile.com
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         Candy Kane <c.kane@FERTILE.COM>
Organization: Growing greener every day
Subject:      Re: Event-history problem

Ted Greenstein <Ted_Greenstein@ncsu.edu> wrote:
 
>I'm analyzing some data on marital stability.  I have two waves of a panel,
>where the waves are separated by about five years.  The unit of analysis is
>the married couple; I start out with about 5,000 couples who were married
>prior to wave 1, and my outcome variable is whether they are still married at
>wave 2.  If their marriages terminated between wave 1 and wave 2, I have data
>on the date; if they were still married at wave 2, I know the date at which
>they were right-censored.
>
>One problem that has arisen is the fact that the couples are of varying
>marital durations prior to wave 1.  That is, some of the couples got married
>30 years prior to wave 1, some 20 years, and so forth.  Obviously, I need to
>take this into account in my model (I've been using a piecewise exponential
>model with LIFEREG).  So far, I've been including an indicator of this prior
>marital duration in the LIFEREG model, and the results seem reasonable.  As
>might be expected, prior marital duration is negatively related to the hazard
>rate of experiencing divorce at all durations.
>
>Does this procedure (controlling for prior marital duration) seem reasonable?
>Or there some trap I'm not seeing?  Any comments or suggestions would be
>welcome.
 
Ted:
 
It's not really clear from your post what your actual research
question is. Are you simply trying to model divorce rate as a function
of marriage duration? Or are there predictors you haven't mentioned?
 
The data you seem to have are these: a date of marriage for 5000
couples, and either a divorce date or a censoring date corresponding
to the earlier of a death date or date of collection of Wave 2.
 
It seems to me that modelling divorce rate here is a straightforward
process of subtracting the marriage date from the
divorce/death/censoring date to generate a response time, and
classifying divorces as events, and other cases as censored at the
earlier of death/Wave 2. You can then use either a parametric method
(proc lifereg) or non-parametric (proc lifetest) survival analysis to
generate your divorce hazard function.
 
Have I missed the point of the project here?!
Jay Weedon.