Single-center versus multi-center data sets for molecular prognostic modeling: a simulation study