tainted survival data I'm trying to salvage some time-to-event data that have been tainted by scheduled maintenance. The failure, in this case is a replacement. It occurs that through scheduled maintenance multiple assets are replaced while the assets weren't at failure yet. Each asset has its own properties at failure. Is it possible to salvage these data by generating the failure time based on some kind of distribution? Or should I leave these data out of the modeling process?
 A: In event history models (aka time to event models) there is frequently the concept of right censoring, meaning the event was never observed for a particular object being studied. In the logic of event history models, such objects (be they engines at risk of seizing, hard drives at risk of failure, people at risk of dying, etc.) should contribute information towards risk of event while they are observed (i.e. inform the appropriate denominator event hazard in a given time), and only be eliminated from contributing information after the time which they were last observed.
You do not say what kind of event history analysis you are using, or want to use. Giving details as to the kind of model and properties of your data will support more informative answers. That said, there's quite a variety of such models, and their terms sometimes change meaning from one model type to the next (e.g., "hazard" in discrete-time event history models vs. continuous time event history models). I am an epidemiologist, so Kaplan-Meier, Cox proportional hazards, and discrete-time logit hazard (or probit hazard, or complementary log-log hazard) models will be more familiar to me. All of these event history models—and, I would hazard to guess (no pun intended) event history models with which I have not yet been acquainted—incorporate right-censoring as a standard in their baseline applications, no problem.
As asked in your question, I do not know if all or some right-censored objects were censored due to scheduled maintenance, or if only some were. If only some were, and their scheduling for maintenance was anything other than random, you may want to include a "scheduled for maintenance" variable to capture some of the variation in event hazard due to the reasons for scheduling (better still if there are records indicating the nature of those reasons, as constructing variables like "maintenance scheduled due to anticipated failure" and "maintenance scheduled due to routine annual replacement" would be informative).
