C-index Alone Misleads: Survival Models Need Calibration Validation.

Rafael da Silva, Danilo Alvares· July 23, 2026 View original

Summary

This study empirically demonstrates that relying solely on the C-index for evaluating survival models can be misleading, as models with high discrimination often lack calibration. Researchers reproduced published models across diverse domains and found significant calibration failures, highlighting the need for comprehensive evaluation beyond discrimination metrics.

This research provides empirical evidence supporting the argument that evaluating survival models solely based on discrimination metrics, such as the concordance index (C-index), can lead to systematically misleading conclusions. While the C-index measures how well a model ranks predictions, it fails to account for calibration, which assesses how well predicted probabilities align with actual outcomes. The study reproduced three published survival machine learning models from distinct domains: hard-drive failure, peer-to-peer credit default, and user disengagement on digital platforms. Across these reproductions, the researchers found that models achieving near-perfect published discrimination scores often failed formal calibration tests dramatically. For instance, a model with a C-index of 0.9595 failed calibration at p = 2.6e-136. The study also revealed that treating competing risks incorrectly (e.g., loan prepayment as non-informative censoring) can bias risk estimates. This work underscores that a high C-index can create an "illusion" of accuracy, emphasizing the critical need for comprehensive evaluation that includes calibration and time-dependent accuracy to ensure models are truly reliable and trustworthy.

Why it matters

For data scientists, risk analysts, and product managers relying on survival models, this research is crucial for avoiding misinformed decisions. It highlights that models must be calibrated, not just discriminative, to provide accurate and trustworthy predictions for real-world applications.

How to implement this in your domain

  1. 1Always include calibration checks (e.g., calibration plots, formal tests) in addition to discrimination metrics (C-index) when evaluating survival models.
  2. 2Re-evaluate existing survival models in your domain for calibration issues, especially those with high reported C-index scores.
  3. 3Account for competing risks appropriately in survival analysis to avoid biased probability estimates.
  4. 4Utilize the provided open-source evaluation harness to systematically validate the calibration and time-dependent accuracy of your survival models.

Who benefits

BFSITech (User Retention)Manufacturing (Predictive Maintenance)Healthcare (Prognosis)

Key takeaways

  • High C-index in survival models does not guarantee good calibration.
  • Many published survival models exhibit significant calibration failures.
  • Ignoring calibration leads to misplaced confidence and potentially biased decisions.
  • Comprehensive evaluation, including calibration and competing risks, is essential for reliable models.

Original post by Rafael da Silva, Danilo Alvares

"arXiv:2607.19526v1 Announce Type: new Abstract: "Stop Chasing the C-index when Evaluating Survival Analysis Models" (ICML 2026, Spotlight) argued normatively, on synthetic data, that evaluating survival models by discrimination alone, i.e. the concordance index, produces systemat…"

View on X

Originally posted by Rafael da Silva, Danilo Alvares on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses