I am conducting a survival analysis to examine the impact of a set of covariates X on the time-to-event Y. However, I have collected data on participants even after the event has occurred, where the event does not necessarily indicate death. I am unsure whether there is any valuable information in the observations obtained after the event. If there is, how should I handle this type of data in my survival analysis? Are there alternative models that may be more suitable than the Cox proportional hazard model? In other words, should I consider approaches other than survival analysis? It is important to note that the event is permanent, and using a (mixed effects) logistic regression does not account for this, making it unsuitable for my analysis.

I have attempted to search the literature, but I am uncertain which keywords will lead me to relevant papers. Any assistance in identifying appropriate keywords would be greatly appreciated.

  • 1
    $\begingroup$ Whether there is any relevant information after the event depends on the nature of the event and of the data you collect. As a simple example, the event could be casting a vote in an election and the subsequent information could be whom the person voted for. (In recent US elections, one major party's members tended to vote early compared to the other party.) Thus it looks like you are trying to ask a question that is (a) not statistical and (b) requires more information to be answered at all. What information can you share to make it answerable here on CV? $\endgroup$
    – whuber
    Commented Jun 28, 2023 at 20:33

1 Answer 1


The approach depends on your question of interest. If you're trying to develop a causal model for the event time as a function of the covariates, you should exclude observations occurring after the event since they obviously can't cause the event. Likewise, if you're trying to predict survival time, you should exclude such observations because you wouldn't have access to them in the case of a new individual who hasn't yet experienced an event.

The only scenario I can imagine where you might want to include the "extra" covariate measurements is that where the occurrence of the event does not affect the covariate measurements (an unrealistic assumption, I'm guessing). In that case, you might be interested in a joint model for the event times and the longitudinal covariate measurements (e.g., Rizopoulos 2023). Using all the covariate measurements to estimate the longitudinal part of the model would presumably improve the efficiency of the estimators of the effects of the covariates on the event time. But interpretation of these effects is much more complex; I can't say that I'd recommend it based on the info you've provided.

  • $\begingroup$ This is very clear. Thank you! $\endgroup$
    – gap9497
    Commented Jun 29, 2023 at 0:04
  • $\begingroup$ @gap9497 if this answers your question, please mark it as the accepted answer. $\endgroup$
    – EdM
    Commented Jun 29, 2023 at 13:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.