I'm trying to understand a paper that claims to have identified a gene expression signature that can distinguish primary from metastatic tumors. The authors stratify their data into patients with and without the gene signature and then plot Kaplan Meier survival curves and compute p-values from the log rank test to show that the signature has diagnostic utility.
I had a few questions about this method as opposed to using a Cox Regression model
Is a significant log rank test sufficient to say the gene signature is predictive of metastasis? Wouldn't a AUC-ROC plot better show the quality of predictions made with the signature?
Would a Cox model with relevant clinical covariates, such as age and tumor size etc, better evaluate the importance of the gene signature as the Kaplan Meier plots don't tell us how much new information the gene signature provides?
When trying to evaluate the diagnostic utility of a variable, when is it appropriate to use the approach the authors take (Kaplan Meier stratifying by groups with a log rank test) vs building a baseline and a full Cox model and comparing their predictive scores?
In summary, I find myself wondering when (if ever) a Kaplan Meier plot is useful to evaluate diagnostic utility. Kaplan Meier's lack of ability to handle multiple covariates and its restriction to a single binary variables seems to make it inferior to the Cox model in most ways and makes me doubt these figures provide a good representation of the gene signature's diagnostic utility.