Dear Antoine,
Thanks for your good point.
However, I would say that if Elaine's dataset is composed of <spells short
enough to have failed during the observation window> only, a selection bias
issue may arise.
Kind Regards,
Carlo
-----Messaggio originale-----
Da: [email protected]
[mailto:[email protected]] Per conto di Antoine Terracol
Inviato: venerdì 23 ottobre 2009 12.06
A: [email protected]
Oggetto: Re: st: R: Problem with Left Truncation
Hi,
Carlo Lazzaro wrote:
Dear Elaine,
Please find beneath the following point-to-point comments about your
query:
<We are doing survival analysis, but unlike other dataset, our dataset
only includes observations that have failed.>
I would not be concerned about all failure=1; how long patient takes to
failure (failure time (tn)- risk onset (t0)) it's the relevant issue.
I haven't done the math, but my intuition is the following.
If Elaine's dataset can be considered as a random sample from the
population, then she can use -stset- and proceed as usual.
I can think of one case where it would not be true: Consider the case
where individuals cannot be at risk before a certain calendar date t0
(say, because she is studying spells in a social program that did not
exist before). If Elaine's observation window is relatively (compared to
the average real duration in the state under study) short, and begins
relatively (again, compared to the real durations) close to t0, then
Elaine's sample will only contain spells short enough to have failed
during the observation window. In this cas, the coefficients will be
biased (although I'm not quite sure if there is a way to handle that
with Stata).
On the other hand, if the spells could have started at any given date
before the observation window (or long enough before the start of the
observation window, relatively to the real durations), then I think her
sample will be random and can be analysed as usual.
HTH,
Antoine
