0
$\begingroup$

Im currently working on a data set and I can not get my statistics to add up. It is a survival analysis and I'm using Kaplan-Meier and Cox proportional-Hazards regression. I have used STATA for analysis. My cox regression indicates no increased nor decreased risk of survival. I then test the PH assumption and obtain a p-value of 0.4, så assumtions are met. However, when I graph the PH assumption, the lines are crossing. What am I doing wrong? I can see my log likelihood is -914, what does that exactly tell me about the model? enter image description here enter image description here enter image description here

$\endgroup$

1 Answer 1

0
$\begingroup$

In reverse order:

The log-likelihood is compared against that of a null model (no covariates) to get the overall LR chi2 estimate of model significance. As you note, on that basis the overall model doesn't meet the usual standard of significance at p < 0.05. I wouldn't give up on this project on that basis, however, as there does seem to be evidence that logco is associated with outcome. Further study seems called for. Don't confuse statistical significance with practical significance, in either direction.

You aren't doing anything wrong in terms of proportional hazards. The lines crossing in the plots, based on groupings by age group, evidently weren't bad enough to influence the global test of proportional hazards.

A couple of things to note on that. First, you show the Global test for proportional hazards, but it might be possible for a test specifically on age to have had a different result. Second, evaluating proportional hazards by grouping a continuous predictor like age for Kaplan-Meier plots or log(-log(survival)) plots can be tricky. As I understand what you've done, those plots don't take other predictors into account. With continuous predictors, I prefer to look at plots of scaled Schoenfeld residuals from the full model over time instead.

$\endgroup$

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.