@paul.buerkner asked if I could open a thread here in The Stan Forums concerning GitHub issue #967 for brms. So here it is. That GitHub issue is a feature request for a PPC for right-censored data in parametric time-to-event models with a response distribution from which sampling is easy (e.g. log-normal or Weibull models). The suggestion is to compare the Kaplan-Meier estimate of the CCDF for the observed data to the posterior CCDFs. Section 4.4 (“Standardised survival probabilities”) of the preprint by Brilleman et al. (2020) describes a similar PPC, however averaging over individual posterior CCDFs. This averaging over individual posterior CCDFs should not be necessary when sampling from the response distribution, which makes the approach suitable for brms’s generic PPC framework. @jonah suggested to integrate such a PPC into the bayesplot package where it may be accessed by other packages as well.
I guess @paul.buerkner’s intention for this thread here was mainly to ask what others think about this approach. Another question from my side is if anyone knows if a similar approach would be possible for other types of censoring other than right censoring (I only know a Kaplan-Meier estimator for right-censored data).
You may find some example code in the corresponding GitHub issue linked above.
EDIT: It turned out that @paul.buerkner’s request for this thread was a misunderstanding; he actually wanted an issue on bayesplot’s GitHub issue tracker (which you may find now here). However, I’m leaving this thread open in case someone stumbles upon it and wants to suggest improvements or leave a comment.