I am performing backward selection for the Cox Proportional Hazards model with several variables (it is required that I do backward selection). One of these variables (let's call it "A") does not meet the proportionality assumption in the full model. So I would like to consider the possibility of stratifying on variable A. How should I perform backward selection in this scenario?

I am considering performing backward selection on the full model while stratifying on A (call the model selected "model B"). I will then compare fit statistics between model B with stratification and model B without stratification. If I see that the stratification is providing only a marginally better fit, then I will perform another backward selection on model B without stratification.

Two questions:

  1. Does this seem reasonable or is there a better way?
  2. How do I formally test if model B without stratification is better than model B with stratification? A likelihood ratio test would not work because the models are not nested and comparing AIC or BIC doesn't make sense because there are the same number of variables in each model.

Edit: any general advice on model selection in COX PH is appreciated as well!

  • $\begingroup$ Could you please expand the question to show how variable A does not meet the assumption of proportionality? I'm guessing you're basing this on residuals, but would be helpful for readers unfamiliar with this difficulty... also this question may be helpful. $\endgroup$
    – dardisco
    Dec 2, 2013 at 0:50

1 Answer 1


Do not perform backward selection. Nor forward. Nor stepwise. Don't do it for Cox or OLS or logistic or any other model. Instead, think about what variables you have, why you have them, what they mean, what theory says about them, how including them affects other variables and so on.

If you must use some automated method, LASSO or LAR have nice properties.

But no automated method is as good as your brain and your knowledge.

If a variable doesn't meet one of the assumptions, consider modifying the variable; or consider another method of analysis. Also, with a large N, the test of proportionality may be overly conservative. You also might consider interactions.

  • $\begingroup$ It is required that I do backward selection. $\endgroup$ Dec 1, 2013 at 23:39
  • 2
    $\begingroup$ Required by whom? And why? Can you show him/her the folly of doing this? $\endgroup$
    – Peter Flom
    Dec 1, 2013 at 23:40
  • $\begingroup$ For privacy concerns, I would rather not say. But, no, I cannot show him/her the folly of doing this. $\endgroup$ Dec 1, 2013 at 23:44
  • $\begingroup$ Perhap you would find it more satisfactory if I asked, how do I test whether or not I need to stratify on a variable when the proportionality assumption is not met? $\endgroup$ Dec 2, 2013 at 0:39
  • $\begingroup$ Try it stratified and not stratified and plot the results against each other. $\endgroup$
    – Peter Flom
    Dec 2, 2013 at 10:57

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.