Perform Cox PH when KM curves cross

Hi everyone,

I would like to look at pre-transplant criteria (Milan vs UCSF) as predictor of patient’s survival after organ transplantation (and tumour recurrence, separately). I plotted the KM curves, and found that the curves cross each other - can I still do Cox proportional hazards model on this data?


Besides, I would like to ask what are the correct statistical methods for survival data, and how to present them. I understand from past forum posts, that log-rank tests are frowned upon due to it mainly not being able to take in continuous variables (some papers ended up discretizing them using arbitrary cut-offs).

Can I proceed with doing univariate Cox PH on all variables to show their independent association with survival and then use predictor variables of interest (that are clinically significant) to perform multivariable analysis?

The proportion of event is rather small at < 25% - so I think using a lot of variables for the multivariable analysis may not be appropriate? Hence, I thought I should do univariate analysis for all variables using Cox PH first.

Appreciate any advice. Thanks!

There are a number of problems with your strategy especially the non-pre-specification of covariates to adjust for.

Regarding non-PH in your example, confidence bands on the differences between any two of the K-M curves would tell you that you don’t have enough data to make such a statement.

Univariate analysis causes nothing but trouble. See