Pith. sign in

REVIEW 4 major objections 4 minor 20 references

The concordance index can report near-perfect discrimination while a survival model's probability estimates fail calibration tests by enormous margins, and C-index-only evaluation outside healthcare is not a reliable proxy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-01 12:28 UTC pith:WPTPGC6F

load-bearing objection First real-data test of the C-index critique: the calibration dissociation is real, but the headline meta-verdict rests on one fragile threshold reading. the 4 major comments →

arxiv 2607.19526 v2 pith:WPTPGC6F submitted 2026-07-21 cs.LG

The C-index illusion: discrimination without calibration in published survival models

Rafael da Silva , Danilo Alvares This is my paper
classification cs.LG
keywords concordance indexsurvival analysismodel calibrationcompeting risksproper scoring rulesreproducibilityevaluation metricsnon-clinical applications
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tests, on real published models, whether the widely used concordance index (C-index) can mask serious failures in survival analysis. Reproducing three published survival models outside medicine—hard-drive failure, peer-to-peer credit default, and digital-platform churn—it finds that C-index-only evaluation hides miscalibration, competing-risk bias, and horizon-dependent degradation. A model matching a published C-index of 0.958 fails a formal calibration test at p < 0.001. The paper concludes that single-number discrimination reporting gives misplaced confidence, even when it does not necessarily pick the wrong model.

Core claim

The paper establishes that the illusion warned about by a recent position paper is real and generalizes beyond synthetic data: a survival model can rank risk almost perfectly while assigning probabilities that are systematically wrong. Reproducing three published survival-ML models to tight fidelity, the authors show that all seven reproduced models fail a formal D-Calibration test, including one with C = 0.9595 matching the published 0.958. Treating loan prepayment as ordinary censoring instead of a competing risk biases estimated default probability upward by roughly two percentage points on average and nearly four in the riskiest credit grades. A platform churn model's global C-index sits

What carries the argument

The central machinery is the double-helix ladder: the rule that an evaluation metric is valid only at the same rung of a censoring-assumption hierarchy as the model it judges. The paper applies three instruments from survival analysis—D-Calibration, a formal hypothesis test that checks whether predicted survival quantiles match observed outcomes; IPCW-weighted Integrated Brier Score, a proper scoring rule that penalizes the full predicted probability; and the Aalen-Johansen estimator, which correctly handles competing events. These are applied to reproduced published models while holding the models fixed and varying only the evaluation lens.

Load-bearing premise

The load-bearing premise is that the evaluation instruments—D-Calibration and IPCW-IBS, ported from the position paper's code and validated only for integration fidelity—correctly measure miscalibration on these real datasets; a latent error in that code or dependent censoring inflating apparent miscalibration would weaken the headline dissociation.

What would settle it

Run an independent, from-scratch implementation of D-Calibration and IPCW-IBS on the same reproduced hard-drive failure model and check whether the calibration p-value is still far below 0.001; if it is not, the headline dissociation is an instrumentation artifact. Additionally, apply a copula-adjusted calibration test under assumed dependent censoring to see whether the miscalibration shrinks toward irrelevance.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A published survival model with a high C-index can have probability estimates that a formal calibration test rejects by over a hundred orders of magnitude; calibration checks should accompany discrimination reporting.
  • Treating a competing event such as loan prepayment as ordinary censoring materially inflates estimated default risk, especially in the riskiest segments, which is consequential for lending and capital decisions.
  • A model's global C-index can remain inside a published performance band while its predictive accuracy degrades at longer horizons; horizon-specific proper scores are needed for operational decisions.
  • The failure mode documented is misplaced confidence rather than choosing the wrong model: the ranking-inversion test did not reject, so C-index-only evaluation is more a trust problem than a selection problem.
  • The released evaluation harness makes it straightforward to audit future survival-model results with calibration and proper-scoring metrics rather than discrimination alone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorially: the non-rejection of the ranking-inversion hypothesis should not be read as reassurance, since only two to three models per domain give that test very low power; larger model sets could still reveal metric-driven preference flips.
  • Editorially: extending the same audit to deep survival models would test whether the discrimination-calibration dissociation is specific to Cox and random survival forests or is a more general property of survival ML.
  • Editorially: constructing a copula-adjusted version of D-Calibration—which does not yet exist—would determine how much of the apparent miscalibration is due to dependent censoring rather than genuine model error, and is a natural next step.
  • Editorially: the pattern parallels accuracy under class imbalance in classification; a single aggregate metric becomes blind to the failure mode that matters most, suggesting the supplement-the-metric remedy may be broadly applicable.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper empirically tests whether C-index-only evaluation of survival models misleads outside healthcare. It reproduces three published survival-ML baselines (hard-drive failure, P2P credit default, and Stack Exchange user disengagement), validates an evaluation harness by reproducing the anchor paper's synthetic experiment, and tests five pre-registered hypotheses under Holm-corrected family-wise error control. The headline results are: H2 rejects (a model reproducing published discrimination at C=0.9595 fails D-Calibration at p≈10^-136); H3 rejects under the authors' absolute-count reading of the pre-registered strata rule (naive minus Aalen-Johansen 12-month default CIF bias = 0.0218, exceeding the 0.02 threshold, with D/E/F strata supporting); and H5 rejects (IPCW Brier worsens with horizon while global C-index remains in the reported band). H1 and H4 do not reject. The meta-hypothesis (≥3/5 primary rejections) is therefore declared rejected, but this verdict depends on interpreting the freeze's '≥3 of 5 rating strata' as an absolute count rather than a fraction of the seven observed strata; the authors disclose this ambiguity explicitly.

Significance. If taken at face value, the paper is a valuable existence proof and a methodological contribution: it shows that high discrimination can coexist with formally rejected calibration, that treating a competing event as non-informative censoring creates a directional and practically relevant probability bias, and that horizon-specific degradation of a proper score can be masked by a stable global C-index. The paper's strengths are its pre-registration, its unusually explicit disclosure of a posteriori decisions, its careful reproduction of all three baselines (Domain 1 C-index reproduced within 0.0015), and the release of a reusable harness and notebook. The main weakness is that the formal meta-verdict is fragile: it hinges on an ambiguous pre-registration denominator, and the evaluation instrument is validated only by integration fidelity with the anchor's code, not by independent reimplementation. These limitations are disclosed, but they make the headline 'three of five reject' and the unconditional meta-verdict stronger than the evidence supports. The substantive H2 and H5 findings, and the monotonic H3 pattern, remain informative even if the meta-claim is reframed.

major comments (4)
  1. [§4.5, §5.3, Appendix A (H3/Hmeta)] The central meta-verdict is not robust to the pre-registration denominator ambiguity. The freeze specifies '≥3 of 5 rating strata,' written before inspecting Bondora, which contains seven strata (AA–F). The manuscript applies the absolute count ≥3, satisfied by D/E/F, and rejects H3. But a fractional reading of '≥3 of 5' as '≥60%,' i.e., ≥5 of 7, is equally defensible and would not reject H3. H3's Holm-adjusted p is 0.0448, and the effect (0.0218) exceeds the 0.02 threshold by only 0.0018. Without H3, only H2 and H5 reject, so the pre-registered meta-hypothesis gate of ≥3/5 is not met. The authors disclose both readings, but the abstract and §4.8 state 'three of five reject' and 'the meta-hypothesis is rejected' without making this conditionality the primary framing. Because the paper's central claim is the meta-verdict, this is load-bearing. The revision should present the verdict as ex
  2. [§3.4, §5.3] The instrument validation establishes fidelity of integration, not independent correctness. The authors state that a latent bug in the anchor's D-Calibration/IPCW-IBS implementation would be inherited by their harness. Since H2's headline p<0.001 is computed with this ported implementation, the claim that the instrument is 'validated' is weaker than an independent verification. I am not asserting that a bug exists, but the correctness risk is load-bearing for H2. The authors should either cross-check D-Calibration and IPCW-IBS against an independent implementation, or explicitly state in the results that the calibration p-values come from an integration-validated but not independently reimplemented tool, and discuss how a systematic implementation error would affect the conclusions. The enormous observed p margin makes a full reversal implausible, but the audit's methodological standard
  3. [§4.3, Table 7, §5.1] The conclusion that C-index-only evaluation 'does not fail as a comparison problem' overstates what H1 can support. In Domain 2, the observed Kendall's τ is -1.000 but the bootstrap CI is [-1.000, 1.000], which is completely uninformative; with only two models, the test has no power to distinguish inversion from noise. The paper acknowledges this, yet §5.1 and the conclusion rely on H1's non-rejection to argue that the failure mode is 'misplaced confidence' rather than 'choosing the wrong model.' The correct summary is that H1 is inconclusive in Domain 2 and consistent with agreement in Domains 1 and 3. This distinction should be reflected in the framing of the paper's main message.
  4. [§4.6, §5.1] The post-hoc cluster ablation complicates the 'not a trivial shortcut artifact' interpretation of H2. The pre-registered leave-one-out null stands, but Cluster A (SMART 9/240/241, age/usage proxies) jointly drops C by 0.0920 with non-overlapping CIs, indicating that a substantial share of the reproduced discrimination is concentrated in a small correlated feature group. The manuscript reports this transparently, but the abstract's statement that the calibration failure is 'not a trivial shortcut artifact' should be qualified: no single attribute is responsible, but a small age/usage cluster is. This does not undermine the H2 dissociation, but it narrows the claim about how distributed the discrimination actually is.
minor comments (4)
  1. [Table 3] The notation '≥3/7 strata' in Table 3 is ambiguous: it can be read as an absolute count (3 of 7) or as a fraction (3/7 ≈ 43%). Harmonize it with the Appendix A freeze text, which says '≥3 of 5,' and clarify in the table footnote that the applied rule is the absolute count.
  2. [Figure 6] The x-axis label in Figure 6 renders as 'Kendall's ¿' instead of Kendall's τ, likely an encoding issue. Please fix.
  3. [Abstract / §6] 'Over a hundred orders of magnitude' is imprecise; p=2.60×10^-136 means 135 orders of magnitude below 1. State the p-value or '≈10^-136' instead.
  4. [§3.3] The reproduction tolerance was fixed after observing that all three domains cleared it. This is disclosed, but the admission criterion should be described as 'post-hoc but non-load-bearing' in the main text, not only in the appendix, to avoid readers missing that it was not pre-registered.

Circularity Check

0 steps flagged

No circularity: the paper is an empirical audit against external published models; disclosed limitations are validity risks, not circular reductions.

full rationale

The paper's derivation chain is an audit, not a construction that folds its conclusion into its inputs. The three baselines are external published papers (Ahmed & Green, Bone-Winkel & Reichenbach, Abedi Firouzjaei), and the headline quantities—Harrell C, D-Calibration p-values, naive vs. Aalen-Johansen cumulative incidence, and horizon-specific IPCW Brier scores—are computed from reproduced models on public data. No fitted parameter is renamed as a prediction: the reproduced C-index (0.9595) is matched to the published value and then independently evaluated for calibration; the calibration failure is a measured property of the fitted model, not a consequence of the reproduction target. H3's bias (0.0218 vs. 0.02 threshold) is also a measurement, and neither the measurement nor its confidence interval depends on the pre-registration denominator; the disclosed '>=3 of 5' vs. '>=3 of 7' ambiguity affects only the counting rule for the meta-verdict, and the paper explicitly reports both readings rather than silently choosing the favorable one. The instrument validation against the anchor paper's synthetic experiment is explicitly limited to 'fidelity of integration, not independent correctness,' and the paper flags that a latent bug in the anchor's code would be inherited—this is a validity limitation, not a circular step, and it is weighed as such here. The reproduction admission criterion was declared a posteriori, but the paper verifies that it governs baseline admission only and does not alter any hypothesis outcome. There is no self-citation chain, no imported uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known result as a new derivation. The close calls are pre-registration fragility and instrument-validation strength, not definitional circularity, so the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The audit contributes no new theoretical entities. Its conclusions depend on hand-chosen decision thresholds (especially H3's), on faithful reproduction given under-specified baseline texts, and on the correctness of ported metric code that was validated only by matching the anchor's own outputs.

free parameters (6)
  • H3 CIF bias threshold = 0.02
    Pre-registered threshold; observed bias 0.0218 clears it by only 0.0018. Central to H3 rejection.
  • H3 strata-count rule = >=3 (absolute) of 7 observed strata
    Freeze text '>=3 of 5' written expecting 5 strata; applied as absolute count. A fractional reading (>=5 of 7) would not reject H3 and would drop the meta-verdict to 2/5.
  • H4 delta-C threshold = 0.03 with non-overlapping CIs
    Chosen in freeze; determines H4 non-rejection. Cluster ablation crosses it (ΔC≈0.09) but LOO does not.
  • H5 C-index band = [0.66, 0.76]
    Pre-registered; only one of three D3 models qualifies, and H5 rejection rides on that single model.
  • Domain 1 Cox penalizer = 0.01
    Baseline paper silent on regularization; chosen by us. Affects reproduced C/HRs mildly (Table 12 L5).
  • Domain 1 SMART snapshot = last observed day
    Baseline silent on first/last/mean raw SMART; our choice may shift C/HRs vs the private pipeline (Table 12 L3).
axioms (6)
  • domain assumption The three reproduced baselines are faithful enough to the published models to stand in for them
    Used throughout Phase A/B; under-specified public details (private repos, unstated hyperparameters, SMART snapshot timing) require our choices (Appendix B Tables 12, 14, 15), so the audited models may differ from the original fits.
  • domain assumption D-Calibration (Haider et al.) correctly detects miscalibration on these data
    Central to H2 and the 7/7 failure pattern; assumes non-informative censoring, acknowledged in §5.3 as potentially inflating apparent miscalibration under dependent censoring.
  • domain assumption The ported anchor evaluation code is correct
    Instrument validation (§3.4) is integration-only; a latent bug in the anchor's implementation would be inherited.
  • domain assumption Standard survival metrics (Harrell C, IPCW-IBS) are valid under each domain's censoring mechanism
    Phase B holds models fixed and varies the lens; the sensitivity sweep (§4.9) shows metric magnitudes change with assumed censoring dependence, so this assumption is load-bearing.
  • standard math Bootstrap CIs with B=1000 adequately represent sampling variability
    Used for H1, H3, H4, H5 decision rules; paired bootstrap with fixed seed 42.
  • standard math Two-level Holm-Bonferroni controls FWER at alpha=0.05
    Used for family-wise verdict; conservativeness noted in §3.7.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of The C-index illusion: discrimination without calibration in published survival models." pith.science (2026). https://pith.science/paper/WPTPGC6F

@misc{pith2026260719526,
  author       = {Pith},
  title        = {Pith review of: The C-index illusion: discrimination without calibration in published survival models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPTPGC6F}},
  note         = {Machine review of arXiv:2607.19526}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent work has argued normatively, on synthetic data, that evaluating survival models by discrimination alone (concordance index) yields systematically misleading model comparisons, because the metric ignores calibration and time-dependent accuracy. Whether this matters for real, published, non-clinical models has not been tested. We reproduce three published survival-ML models across three structurally distinct domains -- hard-drive failure prediction, peer-to-peer credit default, and user disengagement on digital platforms -- validate our instrument against the anchor paper's own synthetic experiment, and test five pre-registered hypotheses under a Holm-corrected family-wise error rate. Three of five reject (though one pre-registered threshold clears by a narrow margin). A model reproducing the published literature's discrimination almost exactly (C = 0.9595 vs. 0.958 reported) fails a formal calibration test at p < 0.001; a broad feature-ablation search finds no single attribute responsible for its discrimination, so the calibration failure is not a trivial shortcut artifact. A lender's estimated default risk is biased upward by roughly two percentage points, growing to nearly four in the riskiest segment, when loan prepayment is treated as non-informative censoring rather than a competing risk. A platform's churn model shows probability estimates that degrade with the horizon even as global discrimination stays within the pre-registered C-index band. A direct test of whether metric choice inverts model preference does not reject, though with limited power given two to three models per domain; the failure mode we document is better characterized as misplaced confidence in a chosen model than as choosing the wrong one. We release a pre-registered evaluation harness with full code and an annotated notebook, so these results can be verified independently and the audit extended.

Figures

Figures reproduced from arXiv: 2607.19526 by Danilo Alvares, Rafael da Silva.

Figure 1
Figure 1. Figure 1: Overview of the study pipeline. Three published baselines and the anchor’s synthetic experiment [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Discrimination (C-index) versus proper-score (IPCW-IBS) rankings, by model and domain. Points [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: D-Calibration histogram for the Domain 1 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Rating n ∆ (naive − AJ) 95% CI Supports H3 AA 176 0.0016 [0.0004, 0.0037] No A 114 0.0114 [0.0045, 0.0202] No B 636 0.0078 [0.0052, 0.0110] No C 1,373 0.0158 [0.0131, 0.0194] No D 1,785 0.0210 [0.0182, 0.0243] Yes E 2,031 0.0300 [0.0264, 0.0346] Yes F 342 0.0390 [0.0258, 0.0530] Yes 12 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Bias in twelve-month cumulative incidence of default, by credit rating (safest to riskiest, left to [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: IPCW Brier score at three prediction horizons, for all three Domain 3 (Stack Exchange) models. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity of CG-IPCW evaluation metrics to the assumed strength of censoring dependence [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 9 canonical work pages

  1. [4]

    URL https://doi.org/10.1007/s42521-024-00114-3

    doi: 10.1007/s42521-024-00114-3. URL https://doi.org/10.1007/s42521-024-00114-3. Maarten Coemans, Geert Verbeke, Bernd Döhler, Caner Süsal, and Maarten Naesens. Bias by censoring for competing events in survival analysis.BMJ, 378:e071349,

    work page doi:10.1007/s42521-024-00114-3

  2. [5]

    URL https://doi.org/10.1136/bmj-2022-071349

    doi: 10.1136/bmj-2022-071349. URL https://doi.org/10.1136/bmj-2022-071349. Frank Emmert-Streib and Matthias Dehmer. Introduction to survival analysis in practice.Machine Learning and Knowledge Extraction, 1(3):1013–1038,

    work page doi:10.1136/bmj-2022-071349 2022

  3. [9]

    URLhttps:// doi.org/10.1002/(SICI)1097-0258(19990915/30)18:17/18<2529::AID-SIM274>3.0.CO;2-5

    doi: 10.1002/(SICI)1097-0258(19990915/30)18:17/18<2529::AID-SIM274>3.0.CO;2-5. URLhttps:// doi.org/10.1002/(SICI)1097-0258(19990915/30)18:17/18<2529::AID-SIM274>3.0.CO;2-5. Humza Haider, Bret Hoehn, Sarah Davis, and Russell Greiner. Effective ways to build and evaluate individual survival distributions.Journal of Machine Learning Research, 21(85):1–63,

    doi:10.1002/(sici)1097-0258(19990915/30)18:17/18

  4. [13]

    compbiomed.2025.111176

    doi: 10.1016/j. compbiomed.2025.111176. URLhttps://doi.org/10.1016/j.compbiomed.2025.111176. 21 Jaya M. Satagopan, Leah Ben-Porat, Marianne Berwick, Mark Robson, David Kutler, and Arleen D. Auer- bach. A note on competing risks in survival data analysis.British Journal of Cancer, 91(7):1229–1235,

    arXiv 2025

  5. [16]

    URLhttps://doi.org/10

    doi: 10.1136/bmj-2021-069249. URLhttps://doi.org/10. 1136/bmj-2021-069249. Ping Wang, Yan Li, and Chandan K. Reddy. Machine learning for survival analysis: A survey.ACM Computing Surveys, 51(6):110:1–110:36,

    work page doi:10.1136/bmj-2021-069249 2021

  6. [17]

    URLhttps://doi.org/10.1145/ 3214306

    doi: 10.1145/3214306. URLhttps://doi.org/10.1145/ 3214306. David Wissel, Nikita Janakarajan, Aayush Grover, Elisa Toniato, María Rodríguez Martínez, and Valentina Boeva. SurvBoard: Standardized benchmarking for multi-omics cancer survival models.Briefings in Bioinformatics, 26(5):bbaf521,

    doi:10.1145/3214306

  7. [18]

    URLhttps://doi.org/10.1093/bib/ bbaf521

    doi: 10.1093/bib/bbaf521. URLhttps://doi.org/10.1093/bib/ bbaf521. Ewa Wycinka. Competing risk models of default in the presence of early repayments.Econometrics. Ekonometria. Advances in Applied Data Analysis, 23(2):99–118,

    work page doi:10.1093/bib/bbaf521

  8. [19]

    URL https://doi.org/10.15611/eada.2019.2.07

    doi: 10.15611/eada.2019.2.07. URL https://doi.org/10.15611/eada.2019.2.07. A Protocol Freeze Protocol Freeze Record (version2026-07-12.c00.v5.1). Every methodological decision documented here wasfixedbeforeanytestdatawereexamined, exceptwhereexplicitlynoted(C00.4, Subsection3.3). Rejected alternatives are listed for transparency. •Frozen at (UTC):2026-07-...

    work page doi:10.15611/eada.2019.2.07 2019

  9. [20]

    metric":

    Hypotheses (decision rules) H1 – Ranking inversion under metric-assumption misalignment.H 0:Within each domain, C- index ranking equals IPCW-IBS ranking (τK = 1).H 1:τ K <1in at least one domain (math alternative; decision = C00.1).Statistic:Kendallτ K (C-index rank vs. IPCW-IBS rank); subject-level stratified boot- strap CI (primary); rank permutation (s...

    2020

  10. [1982]

    URL https://doi.org/10.1001/jama.1982.03320430047030

    doi: 10.1001/jama.1982.03320430047030. URL https://doi.org/10.1001/jama.1982.03320430047030. Vincent Jeanselme, Brian Tom, and Jessica Barrett. Competing risks: Impact on risk estimation and algorithmic fairness,

    arXiv 1982

  11. [1999]

    10474144

    doi: 10.1080/01621459.1999. 10474144. URLhttps://doi.org/10.1080/01621459.1999.10474144. Halina Frydman and Anna Matuszyk. Random survival forest for competing credit risks.Journal of the Operational Research Society,

    arXiv 1999

  12. [2004]

    URLhttps://doi.org/10.1038/sj.bjc.6602102

    doi: 10.1038/sj.bjc.6602102. URLhttps://doi.org/10.1038/sj.bjc.6602102. Hajime Uno, Tianxi Cai, Michael J. Pencina, Ralph B. D’Agostino, and L. J. Wei. On the C-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data.Statistics in Medicine, 30(10):1105–1117,

    work page doi:10.1038/sj.bjc.6602102

  13. [2005]

    URLhttps://doi.org/10

    doi: 10.1002/sim.2427. URLhttps://doi.org/10. 1002/sim.2427. Georg F. Bone-Winkel and Felix Reichenbach. Improving credit risk assessment in P2P lending with explain- able machine learning survival analysis.Digital Finance,

    doi:10.1002/sim.2427

  14. [2011]

    URLhttps://doi.org/10.1002/sim.4154

    doi: 10.1002/sim.4154. URLhttps://doi.org/10.1002/sim.4154. Nan van Geloven, Daniele Giardiello, Edouard F. Bonneville, Lucy Teece, Chava L. Ramspek, Maarten van Smeden, Kym I. E. Snell, Ben van Calster, Maja Pohar-Perme, Richard D. Riley, Hein Putter, and Ewout W. Steyerberg. Validation of prediction models in the presence of competing risks: A guide thr...

    doi:10.1002/sim.4154

  15. [2019]

    URLhttps://doi.org/ 10.3390/make1030058

    doi: 10.3390/make1030058. URLhttps://doi.org/ 10.3390/make1030058. Jason P. Fine and Robert J. Gray. A proportional hazards model for the subdistribution of a competing risk.Journal of the American Statistical Association, 94(446):496–509,

    work page doi:10.3390/make1030058

  16. [2020]

    URLhttps://doi.org/10

    doi: 10.1080/01605682.2020.1759385. URLhttps://doi.org/10. 1080/01605682.2020.1759385. Erika Graf, Claudia Schmoor, Willi Sauerbrei, and Martin Schumacher. Assessment and comparison of prognostic classification schemes for survival data.Statistics in Medicine, 18(17-18):2529–2545,

    arXiv 2020

  17. [2021]

    Irene Rossi, Federico Sartori, Corrado Rollo, Giovanni Birolo, Piero Fariselli, and Tiziana Sanavia

    URLhttps://arxiv.org/abs/2003.12206. Irene Rossi, Federico Sartori, Corrado Rollo, Giovanni Birolo, Piero Fariselli, and Tiziana Sanavia. Beyond Cox models: Assessing the performance of machine-learning methods in non-proportional hazards and non-linear survival analysis.Computers in Biology and Medicine, 198:111176,

    Pith/arXiv arXiv 2003

  18. [2022]

    URLhttps://doi.org/10.1007/s13278-022-00914-8

    doi: 10.1007/s13278-022-00914-8. URLhttps://doi.org/10.1007/s13278-022-00914-8. 20 Junaid Ahmed and Roger Green. Leveraging survival analysis in cost-aware DeepNet for efficient hard drive failure prediction.Neural Computing and Applications,

    work page doi:10.1007/s13278-022-00914-8

  19. [2024]

    URL https://doi.org/10.1007/s00521-024-10479-6

    doi: 10.1007/s00521-024-10479-6. URL https://doi.org/10.1007/s00521-024-10479-6. Laura Antolini, Patrizia Boracchi, and Elia Biganzoli. A time-dependent discrimination index for survival data.Statistics in Medicine, 24(24):3927–3944,

    work page doi:10.1007/s00521-024-10479-6

  20. [2025]

    Christian Marius Lillelund, Shi-ang Qi, and Russell Greiner

    URLhttps://arxiv.org/abs/2508.05435. Christian Marius Lillelund, Shi-ang Qi, and Russell Greiner. Overcoming dependent censoring in the evalu- ation of survival models, 2025a. URLhttps://arxiv.org/abs/2502.19460. Christian Marius Lillelund, Shi-ang Qi, Russell Greiner, and Christian Fischer Pedersen. Position: Stop chasing the C-index when evaluating surv...

    Pith/arXiv arXiv 2019

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.