I have data across 8 different time-points that is both left, right, and interval censored (also randomly censored) that I am trying to use to conduct a survival analysis. The left-censoring should not be a problem because I can adjust respondents and set T1 as each respondents' first response (meaning that there will be a lot of right-censoring but I think this is okay).

My issue is that a lot of respondents come and go throughout the dataset (i.e., they may respond at time 1 and 2, not respond at time 3 and 4, and reappear for time 5 and 6, etc.) and I wasn't sure what the best way is to deal with this data... should I use data imputation to propagate those with missing values between valid responses? For context, my data is trying to identify whether smokers 'survive' as smokers or quit smoking (i.e., failure).

Additionally, if anyone has experience with this type of issue particularly when using R, I would greatly appreciate suggestions on what packages/functions I should be using.

I hope this makes sense and I have provided sufficient information. Thanks very much in advance.

  • $\begingroup$ How are you dealing with the possibility that someone stops smoking (the "event/failure" that you specify) and then resumes? Are you allowing for multiple "events" per individual, or modeling the return-to-smoking "event"? $\endgroup$
    – EdM
    Feb 1, 2023 at 17:42
  • $\begingroup$ That's the thing I'm wondering... I'm new to survival analysis so I'm not sure what's the better option. I could just do the one "event" per individual and once I get a non-response, count that as the last time point? Or I could do multiple "events" as well... Any suggestions? $\endgroup$ Feb 1, 2023 at 17:47

1 Answer 1


Presumably all these individuals entered the study as "smokers." As it's all too possible for an individual to become a "former smoker" and then return to being a "relapsed smoker," this requires a multi-state model rather than a single-event-type survival model. You will have to use your understanding of the subject matter to decide whether you want to model this with two interchangeable states ("smoker" and "former smoker") or to separate out states of "initial smoker" from "relapsed smoker."

What you have is panel data, with individuals evaluated repeatedly at a handful of times. The msm package in R is a respected choice for multi-state modeling of panel data, although I don't have direct experience with it. A vignette explains how to use the software and discusses issues like the implications of different reasons for missing data.

  • $\begingroup$ Thank you very much for the insight, I really appreciate it! $\endgroup$ Feb 11, 2023 at 14:14

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.