When generating survival times to simulate Cox proportional hazards models, does it matter to generate them in days or in years?

In theory, I guess it does not matter. But in practice? Is there a preference regarding computational issues?

Thank you! Marco

  • 4
    $\begingroup$ There are no computational issues. Choose a time unit that is readily interpretable. $\endgroup$ – whuber Feb 11 '11 at 15:15
  • 1
    $\begingroup$ I believe it's advantageous to avoid ties when dealing with time as a continuous variable. Therefore I think you'd want to base your results on the finest-grained measurements you can get (if that is part of what your question implies). $\endgroup$ – rolando2 Feb 11 '11 at 18:53
  • $\begingroup$ @Whuber: Ok Ok, thanks! @rolanda2: Yes, you are right. But if I generate survival times in years I keep all decimal points. Thx $\endgroup$ – ocram Feb 12 '11 at 7:08
  • $\begingroup$ @whuber, @rolando2 Make answers of that. $\endgroup$ – user88 Feb 13 '11 at 14:36

There's no computational issues behind using a different time scale. I would make sure the time scale you used is biologically/practically meaningful and plausible to collect. If you're talking about the time-to-failure of a single node in a computer cluster, you might be able to say you'd know the event time down to a fraction of a second. If you're talking about time until HIV seroconversion, you'd be lucky to know the day, and more likely know the week or month. Simulation studies are most useful when they're conducted similarly.

I wouldn't worry to much about ties in your time - they'll happen in real-life too, and its useful for your situation to know how to handle them.

So I'd say, generally, pick a good, interpretable time scale, and retain as much precision as is reasonable in a real-world setting.


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.