RE: st: Mixed continuous and interval censored time-to-event analysis

["MacLennan, Graeme" <g.maclennan@abdn.ac.uk>]

Steve, many thanks for this comprehensive advice.  The failure is failure of knee prosthesis defined by a revision of the original operation.  As this is in the UK secondary care setting your point on interval censoring is very pertinent. I shall consult with my surgical colleagues on waiting times for these types of operations and use to create sensible intervals. (I might even be able to infer this from respondents that trigger failure on questionnaire and then go on to have revision operation).

Many thanks again,
Regards.
Graeme.



The following commands can take mixture of interval censoring and uncensored data( the proper term for what you are calling "continuous").
-intreg-
-intcens- (SSD)
-stpm- (SSC)


They work, because each assumes a parametric model. For parametric models, the likelihood contributions of different types of observations
(uncensored, left-censored, right-censored, interval-censored, late
entry) are well-defined. The likelihood analysis of parametric models is
covered in every text, and you can find some good ones in the Stata
Manual references to -streg-.

The lag between the actual failure event and hospital detection means
that the hospital events are interval-censored. To ignore the lag, you
must have strong evidence that it is "short". A better approach, still
more "exact" in comparison to questionnaire-based detections, is to treat
the hospital-based admissions as interval censored, but with interval lower
endpoints based on theory or on empirical knowledge.

Another issue: if failure is associated with hospitalization,
then the hospital-detected events are a biased sample of all events.

Steve

On Oct 3, 2012, at 10:26 AM, MacLennan, Graeme wrote:

Dear Statalist, I have data on time to an event, the event is "failure" in a randomised controlled trial.  Information on failure is collected through two channels.  Firstly, annual questionnaires where failure is defined as being below a certain cut-off on a self-reported outcome measure, although reported annually this failure will have occurred at some point between the last non-failure questionnaire the failure questionnaire, I consider this to be interval censored time-to-event data.  Secondly failure data is captured through routine data sources on hospital readmissions, and as such is a more exact representation of failure time (putting aside any concerns one might have about lag between time of failure and admission to hospital), and I consider this to be continuous time-to-event data.

A clear strategy is to aggregate the continuous data up to interval censored data and use appropriate methods, but this seems like a waste of information to me.  However, after some initial digging about I can't find any pointers, so my question is do the list members know of any literature on this, particularly with Stata in mind?

Regards, Graeme.


The University of Aberdeen is a charity registered in Scotland, No SC013683.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


The University of Aberdeen is a charity registered in Scotland, No SC013683.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/