[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: competing risk
Steve Samuels <firstname.lastname@example.org>
Re: st: competing risk
Thu, 8 Jul 2010 11:19:59 -0400
No wonder your question sounded familiar. There was a thread about
this topic two months ago:
I would add:
1. If there is censoring, the standard assumption is that censorship
within the interview is unrelated to outcome. This assumption appears
to be shaky foryour data: Some losses to followup might have occurred
because subjects moved to accept employment. One might guess that this
is more likely if the new employment was open-ended.
2. If there is much censoring, you can make the analysis completely
transparent by adding a fourth outcome of "unknown" to the list and do
the multiple logistic analysis.
4. Unknown (no interview)
Now, when you are considering the probabilities of a final status, you
can run a sensitivity analysis. Allocate those in the unknown status
to one of the other three. This approach would also allow you to
examine predictors of missingness, even if these did not distinguish
among the first three statuses.
3. You can analyze employment status at each interview separately.
This will allow you to incorporate information on people who missed
earlier interviews and will give you some extra information.
4. You should check the sequence of reported employment status.
Changes between interviews might give you some insight about changes
within intervals. You do this by tabulating status at interview K
against status at interview J. You will need to put data in "wide"
format to do this. See -reshape-.
18 Cantine's Island
Saugerties NY 12477
On Wed, Jul 7, 2010 at 3:09 PM, Steve Samuels <email@example.com> wrote:
> Welcome to Statalist!
> See Chapter 9, especially Section 9.3, of Stephen Jenkins's wonderful
> book "Survival Analysis" at
> and his lesson 8 on setting up the analysis with Stata at
> You don't really have discrete survival times, you have "interval
> censored" or grouped times To quote Stephen (Lesson 8): "If one needs
> to use a discrete time model because one has interval-censored data
> (continuous survival times are available only in grouped form), then
> modelling is rather complex, and one needs special programs to
> estimate the models" There is an exception, I think: if you have no
> (or almost no) loss-to-follow up within intervals, then you can use
> the "multinomial logit" model that Stephen describes-just censor those
> "lost" individuals at the start of the interval.
> Questions about data with this structure (probably the same data set!)
> come up from time to time here. I always wonder why these surveys did
> not ask about dates of transitions prior to interview, even on a
> subsample; that would greatly enhance the usefulness of the data.
> Also, the questions always assume implicitly that nobody changed status
> twice in an interval. (e.g. unemployed--> open-ended -> fixed term).
> Good luck!
> Steven Samuels
> 18 Cantine's Island
> Saugerties NY 12477
> Voice: 845-246-0774
> Fax: 206-202-4783
> On Wed, Jul 7, 2010 at 6:02 AM, <firstname.lastname@example.org> wrote:
>> Dear all,
>> I am running a survival analysis with discrete time data using a stock sample
>> (1750 individuals followed up one, three and five years after graduation).
>> The period at risk of being employed may ends in two competing events, namely
>> fixed term and open ended contracts.
>> I am new in using Stata. Could anyone please advise me how to deal with
>> competing risk in a discrete time setting?
>> In a single risk model I have already organized the data so that there is one
>> row per person per each time at risk of being employed in a fixed or open ended
>> Thank you for your attention,
* For searches and help try: