This repository contains the code for the position paper:

Stop Chasing the C-index when Evaluating Survival Analysis Models
Christian Marius Lillelund, Shi-ang Qi, Russell Greiner, and Christian Fischer Pedersen
Accepted at ICML 2026 (Spotlight)

Preprint: arXiv:2506.02075

The paper argues that survival models should be evaluated with metrics whose assumptions match the modeling objective and the censoring mechanism. The experiments in this repository illustrate the ladder hypothesis of model-metric consistency, which shows that model evaluation can introduce significant bias if censoring is not adjusted for.

Repository contents

├── data/                  # Saved experiment outputs
├── dgp.py                 # Weibull data-generating process
├── ladder_hypo.ipynb      # Main experiments
├── plot_metrics.ipynb     # Plotting code for metric-error curves
├── stats.ipynb            # Summary statistics for literature survey
├── utility.py             # Helper functions for splitting, formatting, and evaluation
├── requirements.txt       # Python dependencies
├── LICENSE
└── README.md

Experiments

See this Notebook for controlled experiments of the ladder hypothesis.

Citation

If you find this paper useful in your work, please consider citing it:

@article{lillelund_stop_2026,
  title={Position: Stop Chasing the C-index when Evaluating Survival Analysis Models},
  author={Christian Marius Lillelund and Shi-ang Qi and Russell Greiner and Christian Fischer Pedersen},
  journal={preprint, arXiv:2506.02075},
  year={2026},
}