Imagine a simple Kaplan-Meier analysis where administrative censoring is applied at time=5 years. In other words, I may have follow-up data on persons out further but i am choosing to censor them at 5 years.

  1. Would it make sense to try censoring at 30 days if I was interested in short term survival? This seems to me that I am throwing data away. Is this ok to do?
  2. Is one testing the similarity of curves from 0 to 90 days and the other comparing curves from 0 to 5 years? Would i ever find a significant difference over the shorter time period and then not find a difference when comparing the longer period (or vice versa)?

1 Answer 1


You are correct in your first point, that deliberately censoring at 30 or 90 days is throwing away data. The question is whether those discarded data would help test your hypothesis of interest, which is whether certain variables are associated with short-term survival.

If you are willing to assume that the influences of the variables on survival are the same over time, then in principle you might get better estimates of 30-day or 90-day survival from a model that took all the data into account. That's a strong assumption that gets to your second point. With a Cox model, for example, you wouldn't be assuming that the overall shape of the survival is the same at short versus long times. You would, however, be assuming that the associations of each of the variables with survival (in terms of the instantaneous hazard of death) are constant over time.

There's no way to know in general whether that's true. In practice, short-term survival generally has to do with things like surgical complications and hospital-acquired infections that don't necessarily affect longer-term survival (at least with the same functional form over time). So you can only answer your question by interrogating your own data set.


Your Answer

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

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