Take Control of Censoring, Generate Real-World Like Synthetic Data

Benchmarking survival models under varying censoring regimes requires data where censoring is both realistic and controllable. We propose a semi-synthetic framework that, starting from a real survival dataset, generates derived datasets with an exact user-specified censoring rate among initially uncensored subjects, through covariate-informed selection of subjects to censor and subject-specific censoring-time sampling. We evaluate realism with Wasserstein distance to a reproducible reference distribution and assess preservation of event--censoring dependence using copula-based Kendall's τ on synthetic settings with observable (T,C). Across six public datasets and censoring rates from 10% to 90%, our method better matches the reference distribution and preserves dependence compared to parametric and random baselines.