I am dealing with retrospective survival data covering 60 years. Some people start their first spell in 1960, others start their first spell in 2000 (they are never left truncated). People can experience events multiple times which means that they have multiple spells. I am curious on how to deal with spells that begin later in my observation period. I can imagine that they have a greater chance to be right censored. Any advice?

P.S I am treating every episode as independent and I am accounting for this by conducting a multilevel weibull model (so spells nested in people)


1 Answer 1


These are multiple non-fatal events per person, so it seems that with few if any exceptions there will be right censoring after the last recorded event.* After the last recorded event there is at most a lower limit to the time to the next event. That type of censoring isn't necessarily a problem per se, as survival analysis is designed to handle such right-censored survival times.

There can be problems if censoring is informative. To evaluate that, you need to apply your knowledge of the subject matter to your data collection and modeling strategies. See this answer for an introduction, and this answer for a more technical description. This review provides several examples of how and when censoring patterns can lead to problems. If events are truly independent within an individual, then the distributions of event times and censoring times (each expressed relative to the preceding event) should provide no information about each other, consistent with non-informative censoring.

I'm actually more worried about your assumption of independence of events within individuals, as that's a pretty strong requirement. You might want to consider more flexible ways of evaluating your data, for example with plots of cumulated hazard over time (at least to help evaluate the assumption of independence), models in which the number of prior events affects the hazard of a subsequent event, or multi-site models. The vignettes for the R survival package provide a good place to start to look for guidance.

*Unless you are also modeling death in a multi-state model there should be censoring at time of death, as that's a competing event that clearly changes the probability of further non-fatal events.

  • $\begingroup$ Thank you, and concerning your comment about my assumption of independence: My statement might be wrongly worded and I should've added some information about my research. If I study time until my next scientific publication I thought it was ok to model time from previous event. Therefore every episode starts at 0. I add a random effects term in my multilevel model which accounts for unobserved heterogenity and I add the number of previous episodes/events as a covariate to account for some kind of event dependence. $\endgroup$
    – MaiMai
    Apr 15, 2021 at 13:05
  • $\begingroup$ @MaiMai the danger in this particular application is that scientific publications might tend to happen in bursts. For example, an extended period of laboratory work might results in several related publications appearing at about the same time. That burst of publications might then be followed by another period of relative publication silence while a new set of laboratory experiments is undertaken. My sense here might be mistaken, or the departures from independence might not be large enough to matter. So try to evaluate how well your assumption holds. $\endgroup$
    – EdM
    Apr 15, 2021 at 15:09

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.