With respect to Cox models + frailty, and post-fit survival curves. 1. There are two possible survival curves, the conditional curve where we identify which center a subject comes from, and the marginal curve where we have integrated out center and give survival for an "unspecified" individual. I find the first more useful. More importantly to your case, the survival package currently has no code to calculate the second of these. 2. When the number of centers is large the coxph code may have used a sparse approximation to the variance matrix, for speed reasons. In this particular case one cannot use the "newdata" argument. The reason is entirely practical --- the code turned out to be very hard to write. The need for this comes up very rarely, and the work around is to use coxph(....... + frailty(center, sparse=1000, ....) where we set the "sparse computation" threshold to be some number larger than the number of centers, i.e., force non-sparse computation. Terry Therneau