0

I am using a Cox proportional hazards model (PHREG) in SAS. I have used the (t1,t2)*event specification to indicate the age at which an individual came into the risk set and the age at which s/he left as described here.

I am trying to figure out how to correctly test/assess the proportional hazards assumption for my primary predictor. My initial guess is to include an interaction between the predictor of interest and log(t2-t1) but I'd like some confirmation beyond my own gut. Thanks for your help.

Vincent
  • 955
  • 2
  • 15
  • 32
user1499701
  • 143
  • 1
  • 5
  • I could be entirely wrong, but if you have a time dependent variable you don't meet the Proportional Hazards assumption. – Reeza Dec 30 '14 at 04:20
  • Thanks for this comment! I looked into it more and you're right, once there are time-varying variables the model is no longer a proportional hazards model. It's still unclear to me whether the time-fixed variables need to meet the PH assumption under this circumstance, though. – user1499701 Dec 30 '14 at 20:14

1 Answers1

0

You can do that, check graphically and/or use tests.

Here is a nice guide I found on how to do that. What I'd do is use a Harrell test and check that Schoenfeld residuals graph is 'horizontal' but I'm not sure whether you can do it with SAS or not : I never did (I used R) and that's not in the guide.

However, what you propose seems to be fine too (it's the second presented way).

Vincent
  • 955
  • 2
  • 15
  • 32
  • Thanks, I've used that resource in the past but most of it pertains to the more traditional time-scale specification. I'm trying to figure out how to deal with the PH assumption with the (t1,t2)*event specification (i.e. time-varying covariates). – user1499701 Dec 30 '14 at 19:38