A Semi-Synthetic Social and Clinical Survival Dataset with Structured Dependence and Informative Censoring

This repository provides a semi-synthetic survival dataset designed for methodological research in survival analysis under structured dependence, informative censoring, and latent heterogeneity. The dataset is constructed by combining empirically grounded covariate structures with synthetically generated time-to-event and censoring processes, preserving realistic dependence patterns while ensuring full reproducibility and data privacy. The covariate space and hierarchical identifiers are calibrated to reflect properties commonly observed in real-world observational data, including multilevel clustering and shared latent risk across individuals. Event times and censoring times are then generated through controlled stochastic mechanisms that allow systematic violations of the independent censoring assumption. No row in the dataset corresponds to a real individual, and no real outcomes are included. The dataset is intended exclusively for methodological development, benchmarking, replication, and teaching, particularly in contexts involving hierarchical survival models, network-aware survival methods, and machine learning approaches. Methods: Semi-Synthetic Data-Generating Process The construction of the dataset proceeds in two stages. First, the covariate space and hierarchical identifiers are calibrated to reflect empirical properties commonly observed in applied survival data, including marginal distributions, correlation structure, and multilevel grouping. These components define the structural backbone of the dataset. Second, event times are generated from a hazard function that depends on observed covariates and latent group-level effects. Censoring times are generated from a separate stochastic process that is also conditionally dependent on covariates and latent structure, inducing informative censoring. This design allows explicit control over the strength of dependence and the degree of censoring informativeness. The resulting dataset consists of individual-level observations nested within higher-level units, with survival outcomes that reflect realistic follow-up dynamics while remaining fully reproducible.