stintreg

[Les Boden]

Hi
I have survival data, all of which is (not surprisingly) right censored. For some of my observations, I have an unobserved interval followed after some time by an observed event. This is a situation in which I cannot tell whether the event would have happened in period [t1, t2] before it happened at t3, which I observe. The outcome variable is like the first occurrence of a curable illness. Even though I observe it occurring at t3, it could have occurred first in period [t1,t2]. Can stintreg handle this situation? If so, what would the syntax be?
Thanks.
Les Boden

[Stephen Jenkins]

Survival analysis modelling is about modelling the "time to event", where "time" refers to the survival time since the time (date) at which first became at risk of the relevant event. So, the key question is: when do you think people first become at risk in your particular context, and then what is the length of time until last observed (the number may be an exact length or grouped into bins)? And, then, do you observe these pieces of information in your data set?
An additional issue: you say that all obs are "right censored", which means that no obs had experienced the relevant event at the date/time at which last observed. This severely constrains your ability to fit many models. With no events observed, how can one estimate a baseline hazard, for instance?

[Les Boden]

Sorry, all non-event observations are right-censored. My bad.
The issue is that there are some subjects for whom there is an interval [t1, t2] after the exposure event at t0 but before an observed outcome event at t3 in which we do not observe the subject. The outcome could have occurred during [t1, t2], but if it did we would not observe it. So I would like to treat this interval as censored. Is there a way to do that in Stata?

[Stephen Jenkins]

Les, you appear to have a missing data problem concerning the period t2 to t3.You could treat survival times up to t2 as right-censored -- one way of proceeding.

For the obs observed at t3 and beyond, you don't know when they became at risk of (re)experiencing the event of interest -- it might be in [t2, t3] or before t2 (one single continuous spell) -- which brings back to my earlier comment at #2. You can only proceed, I think, by employing additional information, a.k.a. assumptions, to deal with the missing data issue. You could explore, for instance, what the model estimates looked like if there were no events in the interval. Or what they looked like if you assumed all had an event in [t2,t3] and imputed that date randomly (that would give you start dates for spells at risk of re-experiencing the event). And so on.