(considering right-censoring)

Without knowing the experimental design, is there any conclusion/idea about the type of censoring I have just with the data? I was googling and looking around in literature, but I couldn't find anything related to that.

I guess I could at least some measure of independence? But how should I do that and which test would be appropriate? My first idea would be to look into the relationship between Y (time to event) and the censoring flag (1 or 0), but then of course that non-censored individuals have a higher Y. I would have to check if knowing the censored time for an individual gives me any extra information about the real time, but is that possible in my case?

  • $\begingroup$ You are asking multiple very different questions. Maybe splitting it into multiple individual questions will help to provide good answers. $\endgroup$
    – Michael M
    May 19 at 19:24

Hernán and Robins in Chapter 8 of Causal Inference (Section 8.5) discuss an approach to correct for the bias introduced by censoring. It's an extension of your idea of evaluating the censoring data. In principle, it could get around the problem of distinguishing types of right censoring. It does rest on some assumptions, however.

The idea is to form a pseudo-population via weighting uncensored cases by their inverse probabilities (IP) of censoring. Then an uncensored case "accounts in the analysis not only for herself, but also for those like her" in terms of having the same treatment and covariate values. That requires an adequate model of the probability of censoring as a function of treatment status and covariate values.

The following assumptions are also required:

First, the average outcome in the uncensored individuals must equal the unobserved average outcome in the censored individuals with the same values of [treatment and covariates]...causal interpretation of the resulting adjustment for selection bias depends on this untestable exchangeability assumption.

Second, IP weighting requires that all conditional probabilities of being uncensored given the [covariate values] must be greater than zero....

The third condition is consistency, including well-defined interventions... the pseudo-population effect measure is equal to the effect measure had nobody been censored. This effect measure may be relatively well defined when censoring is the result of loss to follow up or non-response, but not when censoring is defined as the occurrence of a competing event.

Although the presentation in Chapter 8 is for outcomes without respect to time, Chapter 17 covers extensions to time-to-event analysis.

  • $\begingroup$ Interestingly, this is very close to how I got to this problem. I am in fact trying to estimate a marginal hazard ratio using IPTW; and in my simulations found some bias in high censoring proportions (with non-informative censoring). My question was: Is IPTW messing with the "non-informativess" of the censoring? How can I test that? (among some other possible problems) I tried using censoring weights x iptw and then throw those into a Cox Regression. However, I only found this solution to work under informative censoring, but didn't help if the censoring was non-informative at the start. $\endgroup$ May 19 at 21:33
  • $\begingroup$ Which makes sense since with informative censoring it is possible to properly model the probability of censoring. $\endgroup$ May 19 at 21:53

Your Answer

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

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