5
$\begingroup$

In a survival study with informative censoring (for example, studying the effects of cigarettes on mortality and smokers are more likely to be Lost to Follow Up). This causes the censored data to be from a different distribution that the non-censored data.

Is there an accepted and easily implementable solution to this problem? It turns out in my current study we know the assumption of non-informative censoring is violated and it is having a large impact on results, so I need to fix the issue in some way.

My first idea was to create an unbiased estimate of the actual event-time for the censored data, but I'm not sure how I would accomplish that.

Improve this question
$\endgroup$
2
  • $\begingroup$ Did you ever find an easily implementable solution to this? I am currently asking myself the same question. I believe this can be solved using marginal structural models: insights.ovid.com/pubmed?pmid=10955409 however this does not fit into the 'easily implementable' category. $\endgroup$
    – AP30
    Commented Jan 9, 2019 at 16:34
  • $\begingroup$ @AP30 After quite a bit of trial and error, we ended up ignoring the informative censoring issue - we still got sensible results, I believe Cox PH actually handles informative censoring reasonably well in some cases (i.e. the censoring may be informed, but the data to the right of what is censored is represented well by other records in the data). It's not scientific, but worked well enough for me. You might look at this: stats.stackexchange.com/questions/162104/… $\endgroup$
    – Caleb
    Commented Feb 11, 2019 at 8:26

0

Reset to default

Your Answer

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

Browse other questions tagged or ask your own question.