I have generated a L1 penalized Cox model using the penalized package in R. I used the optL1() function to generate the Breslow object (see below): fit <- optL1(surv.obj, penalized = ..., etc) In the reference manual, it says the fit$predictions are the cross-validated predictions for the left out samples. Does this mean these are predictions are based off the fold that was left out of the model building (what one usually thinks of as cross-validated) or is this just saying that these predictions are based off the optimal lambda which is established by cross-validation and are thus just refitting the data? I noticed that predict(fit$fullfit) and fit$predictions give two different values if that helps. [[alternative HTML version deleted]]