Based on my reading on time-varying survival analysis, I am encountering two different and conflicting sets of advice with regards to time-varying covariates and interpolation.
- The first advice is to avoid basing covariates on future events, which may introduce bias. As example, suppose a subject has two lab measurements 25 at time 0 and 50 at time 2; using counting process notation, the subject would be entered as two time intervals A and B: A. (0,2] 25 died = 0, and B. (2,5] 50 died = 1. Under one interpolation of the values the subject would have 37.5 for A. Based on the above advice, bias (perhaps small?) may be introduced as the value 37.5 is based on a future event.
- The second advice is to go ahead and interpolate, and there are some creative methods such as joint mixed models which seem to do this.
Which advice to take? If it depends, on what situations would it be appropriate to prefer one of the other?