Date: Tue, 7 Jun 2005 16:15:43 -0300
Reply-To: Hector Maletta <firstname.lastname@example.org>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Hector Maletta <email@example.com>
Subject: Re: logistic regression with a varying followup time
Content-Type: text/plain; charset="US-ASCII"
You can, indeed, use logistic regression with time of exposure as one of the
predictor variables. But bear in ming that this assumes the likelihood of
the event increases logistically with length of exposure, which may or not
be realistic depending on the type of problem. The risk of a new hard disk
crashing is highest at the start and then decreases dramatically; exposure
to a random event increases the risk in a more or less linear fashion, as
more time elapsed means simply more time for the event to occur. In some
specific phenomena de relationship may be logistic: little increase of the
risk with a little time elapsed, then period a larger rate of increase in
the risk, and finally some kind of saturation where risk does not increase
by much with further time elapsed.
On the other hand, your idea of not using survival analysis (Cox regression)
does not seem particularly justified, since yours is precisely the situation
in which such method is supposed to be applied.
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU]
> On Behalf Of firstname.lastname@example.org
> Sent: Tuesday, June 07, 2005 3:53 PM
> To: SPSSX-L@LISTSERV.UGA.EDU
> Subject: logistic regression with a varying followup time
> A statistical model, not an spss question.
> Suppose that new persons enter a study at different times
> such as would the case if a new intake group were enrolled
> each week. At intake a person characteristic is measured. At
> sometime during the followup period, which ends on a specific
> date, an event may occur. Suppose we know the date the event
> occurs--if it occurs. While the best way to analyze this
> would be survival, what difficulties do i get into by using
> logistic regression? Can i reduce those difficulties if i
> include either a) the exposure period (difference between end
> date and enrollment date) or b) time to event occurrence, but
> not both, as a covariate? Or do i create other difficulties
> by doing either?
> Thanks, Gene Maguin