I would like to look at survival in time to event data. So individuals either have an event, or are censored. My problem is the sensitivity for detecting an event differs between arms. I.e. in the intervention, 10% of the events are mislabeled as a censor. In the control, 15%. I do not know the true sensitivity and specificity for detecting the event, just rough estimates and that they are almost certainly different between arms.

Is there a method for dealing with this?

My main problem is that I do not know when an outcome is missing - I just see it as a censored event mislabeled. If the outcome was missing outright (i.e. the outcome was a missing lab value), I could try something like a PMM model imputing data with different levels of bias.

Any help would be gratefully appreciated!

  • $\begingroup$ To me this sounds like informative censoring (ie. being part of certain subsample of your study is associated with becoming censored). Could you add the details of how this could have come to occur (were treated individuals seen more often after the intervention, are reasons for dropout known, etc.)? $\endgroup$ – IWS Feb 20 '17 at 13:15
  • $\begingroup$ Hi @IWS, thanks for your prompt reply. An example would be a propensity score matched cohort. So one arm is a clinical trial where events are actively sought by study nurses querying the patient, their file and their treating physician, while the comparison arm is a registry where events come from linkage. So in both cases there are some events incorrectly labelled as censoring, but the frequency of the mislabeling differs between arms. $\endgroup$ – James Black Feb 20 '17 at 14:30
  • 1
    $\begingroup$ If you truly think there is a risk of informative censoring, then you might want to look into 'Inverse Probability of Censoring Weighting', in which you inversely weight by the estimated probability to be censored, based on covariates of the subject. $\endgroup$ – IWS Feb 20 '17 at 14:59
  • $\begingroup$ Is there some information you can get on how often events are misclassified as censoring or do you have some probabilistic assessment of whether cases are an event or which cases have uncertain stayus? Otherwise, I suspect this will be extremely hard (impossible? ) to deal with. $\endgroup$ – Björn Feb 21 '17 at 6:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.