Is loglik = -2 Log Likelihood for comparing Cox PH models in survival package?


I'm new to R and trying to do some model building using Cox proportional hazards modelling. When I've done this in the past (using SAS phreg) I've compared the -2 Log Likelihood between two different models to see what happens to it when I add in new variables.

I'm having trouble finding that in the survival package, or finding something that would help me understand and compare the fit of my models. If I use the summary() function I get loglik (which says its the partial likelihood), but I'd expect the -2 log likelihood to be a positive number.

What am I missing?

I've looked at this previous topic (p-value of cox regression model in survival package “coxph()”) and the wiki it referenced (, as well as the vignette on the survival CRAN site, but to no avail.

Thank you!

Yes, unless you are making a parametric assumption about your outcome (say Weibull), you would have to use the partial likelihood and there are various approximations to compute this.

logLik.coxph is pretty well tested and, IIRC, the package author was involved with the SAS proc too. This document is a bit old but discusses the R and SAS output and calculations. There's also a vignette about the package's validation. Sections 3.x of that document have examples worked by hand that show negative log-(partial)likelihood estimates.

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.