[R] Binary outcome with non-absorbing outcome state

[Thomas Lumley]

On Fri, 29 Jul 2005, John Sorkin wrote:

> I am trying to model data in which subjects are followed through time to
> determine if they fall, or do not fall. Some of the subjects fall once,
> some fall several times. Follow-up time varies from subject to subject.
> I know how to model time to the first fall (e.g. Cox Proportional
> Hazards, Kaplan-Meir analyses, etc.) but I am not sure how I can model
> the data if I include the data for those subjects who fall more than
> once.

Various people have already given references that deal with marginal Cox 
models. I'd second Frank Harrell's recommendation of Therneau & Grambsch.
  Computationally this is very straightforward: each person has multiple 
records corresponding to the times between events, and in a Cox model you 
add
    +cluster(id)
to the model formula to get the right standard errors, where id is unique 
identifier for individuals.

The difficult part is deciding which person-time to compare: eg should 
someone who has recently had a second event at time 500 be compared to 
other people who have recently had a second event, other people who have 
recently had any sort of event, other people at time 500, etc.

Another possibility is frailty models, the analogue of generalized linear 
mixed models.  As with GLMMs, even fitting these is tricky and statistical 
theory isn't that well-developed. The survival package does have an 
implementation, though.


 	-thomas