2

Assume we model the probability of disease incidence. When an individual's predicted probability of incident disease, or absolute risk, is greater than a certain threshold, we start preventive measures. Therefore, model calibration is most important near the threshold, and for predictions far away from the threshold we don't care too much about precision.

Is there a model that can be tuned for more precise prediction in a pre-defined probability range?

I have time-to-event data and am using the Cox model, although I understand it's not too well-suited for prediction.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
miura
  • 3,364
  • 3
  • 21
  • 27
  • 1
    if there is a one single set threshold, then this is just a discrimination issue. if you have a range of reasonable thresholds I don't think there is an answer. Vickers 2006, has some ways of visualizing/comparing models but not tuning them. – charles Dec 25 '13 at 16:05
  • 1
    Sounds like a problem that calls simply for trial and error. I don't know of any named method that would specifically accomplish this task. Checking for nonlinear relationships ought to give you better chances. – rolando2 Dec 26 '13 at 10:32

0 Answers0