This repository contains the code, data, and results from a simulation study on the properness of scoring rules in survival analysis, alongside related experiments and visualizations.
Check the supplementary HTML report, which summarizes the results and includes reproducibility instructions. It also links code outputs to manuscript figures and tables for clarity.
See also here the R script that contains a numerical evaluation that complements the analytical proof of ISBS improperness.
The report includes:
-
High variability for generated survival and censoring distributions [link]
(corresponds to Figure 1 in the main manuscript) -
Appendix D experiment results:
-
Degenerate Model Exploits ISBS Scoring Rule [link]
A minimal model that outperforms established methods under ISBS, illustrating its vulnerability to gaming. -
Real-World Benchmark Evaluation [link]
An exploratory comparison of scoring rules (RCLL, RCLL*, ISBS, C-index, D-calibration) on validation tasks using real-world survival datasets.
Sonabend, R., Zobolas, J., Kopper, P., Burk LMU Munich, L., & Bender, A. (2022). Examining properness in the external validation of survival models with squared and logarithmic losses. https://arxiv.org/abs/2212.05260v3