This is a very high level question, hope it's not too general for this forum:

Is it possible to have a greater risk (hazard) for a factor variable in a COXPH regression for one level against another one, while at the same time when plotting a survival (Kaplan Meier) curve, we have the opposite (i.e. the same factor level with the higher risk also having a better survival probability on the curve than the other level).

Note1: proportionnal assumption is met and p-value are significant.

Note2: the 2 variable are very similar, either be the survival curve or the risk levels (small differences in both case).

Thank you.

  • $\begingroup$ I think your situation is impossible,a better "survival" probability means lower risk of "event".There might be something wrong with your code or you need to understand the meaning of "event" and "censor",. $\endgroup$ – Deep North Feb 3 '16 at 3:16
  • $\begingroup$ Hello, thanks for your reply. Yes I understand it's "impossible", hence my question. I'll keep digging, wish I could share my data. If I find a solution I'll post it here. $\endgroup$ – Guillaume Legoy Feb 3 '16 at 13:52

To everyone it may help, I solved my issue.

My mistake was a bad understanding of how to check for a violation of the proportionnal assumption. I hastily concluded it was good by looking at q-plots of my continuous and categorical variable (the p-value of the proportionnal test didn't apply in my case since my sample is very large), but since these plots were very cluttered, it couldn't help me on that matter.

However I learned, thanks to the below link that one can check for proportionnal hazard violation in a different way by also checking Kaplan Meier curves, and some of them in my case are crossing each other, making invalid the cox regression results.


Thank you.


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.