R coxph() warning: Loglik converged before variable

Question

I\'m having some trouble using coxph(). I\'ve two categorical variables: Sex and Probable Cause, that I want to use as predictor variables. Sex is just the typical male/female b

Answer 1

When I asked Terry Therneau (author of pkg:survival) about that several years ago he said the test that is being triggered to generate that warning is overly sensitive. Generally the warning is not correct. You can usually just look at your coefficients to see that they are not infinite or even effectively infinite which a coefficient of 20 would be unrealistically large in most instances, but a coefficient less than five would not be.

In your case, however, it seems to be correctly warning you that there may be problems with your data, since you have implausibly large coefficients. A beta coefficient of 2.276e+01 (= 22.7) in an exponential model is just ridiculously high. The estimated relative risk is well over a million! You should be looking at tabular classifications of your data for problems of complete separation. Did any of your control group die, er, have an event?