arxiv: 2605.11362 · v1 · submitted 2026-05-12 · 💻 cs.LG · cs.AI· stat.AP· stat.ML

Recognition: 2 theorem links

· Lean Theorem

Causal Fairness for Survival Analysis

Drago Plecko

Authors on Pith no claims yet

Pith reviewed 2026-05-13 02:36 UTC · model grok-4.3

classification 💻 cs.LG cs.AIstat.APstat.ML

keywords causal fairnesssurvival analysistime-to-eventdisparity decompositiongraphical modelsnon-parametric estimationICU outcomes

0 comments

The pith

A causal framework decomposes survival disparities into direct, indirect, and spurious pathways to explain their origins and evolution.

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

The paper develops a causal approach to fairness in survival analysis that goes beyond statistical measures by separating the mechanisms behind group differences in time-to-event outcomes. It models the problem with a graphical structure that captures censoring and confounding, recovers conditional survival functions from the data, and uses the Causal Reduction Theorem to break the disparities into specific causal contributions. This decomposition reveals how direct effects, effects through mediators, and spurious associations each shape observed gaps and how those contributions shift as time progresses. The method is illustrated by tracing racial differences in post-admission survival within intensive care unit records.

Core claim

The central claim is that disparities in survival data can be decomposed into direct, indirect, and spurious causal contributions by formalizing assumptions about censoring and lack of confounding in a graphical model, recovering the conditional survival function given covariates, applying the Causal Reduction Theorem to reframe the problem for pathway analysis, and performing non-parametric estimation of the resulting effects.

What carries the argument

The Causal Reduction Theorem, which reframes the survival fairness problem into a form that permits explicit decomposition of disparities along direct, indirect, and spurious pathways.

Load-bearing premise

The graphical model assumptions about censoring being independent of the outcome given covariates and the absence of unmeasured confounding are sufficient to recover the conditional survival functions needed for decomposition.

What would settle it

A simulation experiment in which known direct, indirect, and spurious effects are injected into synthetic survival data and the method's estimated pathway contributions fail to recover the injected values.

Figures

Figures reproduced from arXiv: 2605.11362 by Drago Plecko.

**Figure 2.** Figure 2: Standard Fairness Models for different settings. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Case study results: survival curves, disparity metrics, and causal decompositions. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Case study results: survival curves, disparity metrics, and causal decompositions. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Counterfactual graph of the SFM used in the proof of Prop. 1. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

read the original abstract

In the data-driven era, large-scale datasets are routinely collected and analyzed using machine learning (ML) and artificial intelligence (AI) to inform decisions in high-stakes domains such as healthcare, employment, and criminal justice, raising concerns about the fairness behavior of these systems. Existing works in fair ML cover tasks such as bias detection, fair prediction, and fair decision-making, but largely focus on static settings. At the same time, fairness in temporal contexts, particularly survival/time-to-event (TTE) analysis, remains relatively underexplored, with current approaches to fair survival analysis adopting statistical fairness definitions, which, even with unlimited data, cannot disentangle the causal mechanisms that generate disparities. To address this gap, we develop a causal framework for fairness in TTE analysis, enabling the decomposition of disparities in survival into contributions from direct, indirect, and spurious pathways. This provides a human-understandable explanation of why disparities arise and how they evolve over time. Our non-parametric approach proceeds in four steps: (1) formalizing the necessary assumptions about censoring and lack of confounding using a graphical model; (2) recovering the conditional survival function given covariates; (3) applying the Causal Reduction Theorem to reframe the problem in a form amenable to causal pathway decomposition; (4) estimating the effects efficiently. Finally, our approach is used to analyze the temporal evolution of racial disparities in outcome after admission to an intensive care unit (ICU).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a clean causal decomposition of fairness disparities in survival data via the Causal Reduction Theorem, but it stands or falls on graphical assumptions about censoring and no unmeasured confounding that are tough to check in practice.

read the letter

The core contribution is a four-step non-parametric procedure that decomposes disparities in time-to-event outcomes into direct, indirect, and spurious pathways. It starts with a graphical model to encode assumptions on censoring and confounding, recovers the conditional survival function, applies the Causal Reduction Theorem to enable the decomposition, and then estimates the pieces. They demonstrate it on ICU data tracking how racial disparities in survival evolve over time. That is new relative to the statistical fairness definitions that dominate prior survival work, and the structure is logically coherent on its own terms. The use of an established theorem to reframe the problem is a reasonable move that keeps the method non-parametric where possible. Credit for addressing a real gap in high-stakes temporal settings like healthcare. The main limitation is that steps two and three only identify the target quantities if the graphical model correctly rules out informative censoring and all backdoor paths. In observational ICU records, unmeasured severity often affects both race proxies, treatment, dropout, and survival, so the recovered survival function may not correspond to the interventional effects needed for the decomposition. The paper states the assumptions but the provided material shows no sensitivity checks, partial identification bounds, or simulation results under violated censoring. Estimation error propagation is also left implicit. This is useful for researchers already working on causal fairness extensions or survival models who want a pathway-based explanation rather than a single fairness metric. A reader comfortable with graphical causal models and willing to add robustness checks themselves will get value. It is coherent enough and addresses a genuine gap, so it deserves a serious referee who can press on the identification assumptions and ask for empirical validation of the decomposition under realistic violations.

Referee Report

2 major / 2 minor

Summary. The paper claims to develop a causal framework for fairness in survival/time-to-event analysis that decomposes observed disparities into direct, indirect, and spurious pathway contributions. It proceeds non-parametrically in four steps: (1) formalizing censoring and no-confounding assumptions via a graphical model, (2) recovering the conditional survival function S(t|X), (3) invoking the Causal Reduction Theorem to enable pathway decomposition, and (4) efficient estimation, with an application to temporal racial disparities in ICU data.

Significance. If the identification assumptions hold, the work meaningfully extends causal fairness methods to temporal settings by supplying human-interpretable, time-evolving explanations of disparity mechanisms that purely statistical fairness definitions cannot provide. The explicit use of the Causal Reduction Theorem together with a non-parametric pipeline is a strength, as is the concrete ICU demonstration that shows how the decomposition can be computed in practice.

major comments (2)

[Abstract and §2–3] Abstract and four-step outline (§2–3): the central claim that the graphical model renders censoring conditionally independent of event time and blocks all backdoor paths is required for step (2) to recover an identifiable S(t|X) and for step (3) to apply the Causal Reduction Theorem. The manuscript provides no sensitivity analysis, partial-identification bounds, or robustness checks against informative censoring or unmeasured confounding, which are load-bearing for observational TTE data.
[§5] Results and estimation section (§5): the reported temporal evolution of racial disparities contains no error bars, bootstrap intervals, or cross-validation diagnostics on the decomposed effects, leaving the quantitative claims about pathway contributions without uncertainty quantification.

minor comments (2)

[Notation] Notation: distinguish more clearly between the observed conditional survival function and the interventional quantities that appear after the Causal Reduction Theorem is applied.
[Figure 1] The graphical model figure would benefit from explicit labels on the direct, indirect, and spurious paths to improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback and for recognizing the significance of extending causal fairness methods to survival analysis. We address the two major comments point by point below and commit to revisions that strengthen the manuscript's robustness and empirical rigor.

read point-by-point responses

Referee: Abstract and four-step outline (§2–3): the central claim that the graphical model renders censoring conditionally independent of event time and blocks all backdoor paths is required for step (2) to recover an identifiable S(t|X) and for step (3) to apply the Causal Reduction Theorem. The manuscript provides no sensitivity analysis, partial-identification bounds, or robustness checks against informative censoring or unmeasured confounding, which are load-bearing for observational TTE data.

Authors: We agree that these assumptions are foundational and that the lack of sensitivity analysis represents a limitation in the current version. Our framework is explicitly non-parametric and relies on the graphical model for identification. In the revised manuscript, we will add a new subsection in the discussion that explores the implications of violating the censoring and no-confounding assumptions, including simple sensitivity analyses (e.g., varying the strength of potential unmeasured confounders) and partial identification bounds for the pathway contributions where analytically tractable. revision: yes
Referee: Results and estimation section (§5): the reported temporal evolution of racial disparities contains no error bars, bootstrap intervals, or cross-validation diagnostics on the decomposed effects, leaving the quantitative claims about pathway contributions without uncertainty quantification.

Authors: We concur that uncertainty quantification is essential for interpreting the empirical results. The original manuscript prioritized presenting the decomposition methodology and qualitative trends in the ICU data. For the revision, we will recompute the estimates with bootstrap resampling to provide 95% confidence intervals for each pathway contribution over time and include cross-validation results for the non-parametric estimators used in Section 5. revision: yes

Circularity Check

0 steps flagged

Derivation relies on external Causal Reduction Theorem and graphical model assumptions; no internal reduction to fitted quantities

full rationale

The four-step non-parametric approach formalizes censoring/no-confounding assumptions via a graphical model, recovers the conditional survival function S(t|X), invokes the Causal Reduction Theorem to enable pathway decomposition, and estimates effects. No equation or step equates a claimed prediction to a fitted parameter by construction, nor renames a known result. The theorem and identification results are treated as external inputs rather than derived within the paper, so the central fairness decomposition does not collapse to self-definition or self-citation load-bearing. This yields only a minor score for possible incidental self-citations that do not carry the main claim.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard causal identification assumptions and the Causal Reduction Theorem; no new free parameters or invented entities are introduced in the described approach.

axioms (1)

domain assumption Assumptions about censoring and lack of confounding hold as formalized in the graphical model.
Step (1) invokes these assumptions to enable recovery of the conditional survival function.

pith-pipeline@v0.9.0 · 5554 in / 1261 out tokens · 48125 ms · 2026-05-13T02:36:06.513279+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 1 (Reduced Standard Fairness Model): ... the structural mechanism of the random variable Φ is then given by f_Φ(x,z,w)=ϕ(P(T|X=x,Z=z,W=w)). Therefore, the variable Φ is a deterministic function of X,Z,W...
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Our non-parametric approach proceeds in four steps: (1) formalizing the necessary assumptions about censoring and lack of confounding using a graphical model; (2) recovering the conditional survival function...

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages

[1]

O. Aalen. Nonparametric inference for a family of counting processes.The Annals of Statistics, pages 701–726, 1978

work page 1978
[2]

Angwin, J

J. Angwin, J. Larson, S. Mattu, and L. Kirchner. Machine bias: There’s soft- ware used across the country to predict future criminals. and it’s biased against blacks.ProPublica, 5 2016. URL https://www.propublica.org/article/ machine-bias-risk-assessments-in-criminal-sentencing

work page 2016
[3]

Bang and J

H. Bang and J. M. Robins. Doubly robust estimation in missing data and causal inference models.Biometrics, 61(4):962–973, 2005

work page 2005
[4]

F. D. Blau and L. M. Kahn. The gender earnings gap: learning from international comparisons. The American Economic Review, 82(2):533–538, 1992

work page 1992
[5]

F. D. Blau and L. M. Kahn. The gender wage gap: Extent, trends, and explanations.Journal of economic literature, 55(3):789–865, 2017

work page 2017
[6]

Braekers and N

R. Braekers and N. Veraverbeke. A copula-graphic estimator for the conditional survival function under dependent censoring.Canadian Journal of Statistics, 33(3):429–447, 2005

work page 2005
[7]

Brennan, W

T. Brennan, W. Dieterich, and B. Ehret. Evaluating the predictive validity of the compas risk and needs assessment system.Criminal Justice and Behavior, 36(1):21–40, 2009

work page 2009
[8]

Buolamwini and T

J. Buolamwini and T. Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In S. A. Friedler and C. Wilson, editors,Proceedings of the 1st Conference on Fairness, Accountability and Transparency, volume 81 ofProceedings of Machine Learning Research, pages 77–91, NY , USA, 2018

work page 2018
[9]

Chen and C

T. Chen and C. Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016

work page 2016
[10]

Chernozhukov, D

V . Chernozhukov, D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, W. Newey, and J. Robins. Double/debiased machine learning for treatment and structural parameters, 2018

work page 2018
[11]

S. Chiappa. Path-specific counterfactual fairness. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 7801–7808, 2019

work page 2019
[12]

Creager, D

E. Creager, D. Madras, T. Pitassi, and R. Zemel. Causal modeling for fairness in dynamical systems. InInternational conference on machine learning, pages 2185–2195. PMLR, 2020

work page 2020
[13]

Datta, M

A. Datta, M. C. Tschantz, and A. Datta. Automated experiments on ad privacy settings: A tale of opacity, choice, and discrimination.Proceedings on Privacy Enhancing Technologies, 2015 (1):92–112, Apr. 2015. doi: 10.1515/popets-2015-0007

work page doi:10.1515/popets-2015-0007 2015
[14]

J. P. Fine and R. J. Gray. A proportional hazards model for the subdistribution of a competing risk.Journal of the American statistical association, 94(446):496–509, 1999

work page 1999
[15]

I. O. Gallegos, R. A. Rossi, J. Barrow, M. M. Tanjim, S. Kim, F. Dernoncourt, T. Yu, R. Zhang, and N. K. Ahmed. Bias and fairness in large language models: A survey.Computational Linguistics, 50(3):1097–1179, 2024

work page 2024
[16]

Hu and G

S. Hu and G. H. Chen. Fairness in survival analysis with distributionally robust optimization. Journal of machine learning research, 25(246):1–85, 2024

work page 2024
[17]

Ishwaran, U

H. Ishwaran, U. B. Kogalur, E. H. Blackstone, and M. S. Lauer. Random survival forests.The Annals of Applied Statistics, 2(3):841–860, Sept. 2008. doi: 10.1214/08-AOAS169

work page doi:10.1214/08-aoas169 2008
[18]

E. L. Kaplan and P. Meier. Nonparametric estimation from incomplete observations.Journal of the American statistical association, 53(282):457–481, 1958

work page 1958
[19]

A. E. Khandani, A. J. Kim, and A. W. Lo. Consumer credit-risk models via machine-learning algorithms.Journal of Banking & Finance, 34(11):2767–2787, 2010. 10

work page 2010
[20]

J. P. Klein and M. L. Moeschberger.Survival Analysis. Statistics for Biology and Health. Springer, New York, NY , 1 edition, September 1999

work page 1999
[21]

W. A. Knaus, D. P. Wagner, E. A. Draper, J. E. Zimmerman, M. Bergner, P. G. Bastos, C. A. Sirio, D. J. Murphy, T. Lotring, A. Damiano, et al. The apache iii prognostic system: risk prediction of hospital mortality for critically iii hospitalized adults.Chest, 100(6):1619–1636, 1991

work page 1991
[22]

M. J. Kusner, J. Loftus, C. Russell, and R. Silva. Counterfactual fairness.Advances in neural information processing systems, 30, 2017

work page 2017
[23]

Lange and J

T. Lange and J. V . Hansen. Direct and indirect effects in a survival context.Epidemiology, 22 (4):575–581, 2011

work page 2011
[24]

M. Liu, Y . Ning, H. Wang, C. Hong, M. Engelhard, D. S. Bitterman, W. G. La Cava, and N. Liu. Equitable survival prediction: A fairness-aware survival modeling (fasm) approach.arXiv preprint arXiv:2510.20629, 2025

work page arXiv 2025
[25]

J. F. Mahoney and J. M. Mohen. Method and system for loan origination and underwriting, Oct. 23 2007. US Patent 7,287,008

work page 2007
[26]

S. K. McGowan, K. A. Sarigiannis, S. C. Fox, M. A. Gottlieb, and E. Chen. Racial disparities in icu outcomes: a systematic review.Critical care medicine, 50(1):1–20, 2022

work page 2022
[27]

R. Nabi, D. Malinsky, and I. Shpitser. Learning optimal fair policies. InInternational conference on machine learning, pages 4674–4682. PMLR, 2019

work page 2019
[28]

W. Nelson. Theory and applications of hazard plotting for censored failure data.Technometrics, 14(4):945–966, 1972

work page 1972
[29]

D. Pager. The mark of a criminal record.American journal of sociology, 108(5):937–975, 2003

work page 2003
[30]

Pearl.Causality: Models, Reasoning, and Inference

J. Pearl.Causality: Models, Reasoning, and Inference. Cambridge University Press, New York,

work page
[31]

Pessach and E

D. Pessach and E. Shmueli. A review on fairness in machine learning.ACM Computing Surveys (CSUR), 55(3):1–44, 2022

work page 2022
[32]

T.-H. Pham, J. Chen, S. Lee, Y . Wang, S. Moroi, X. Zhang, and P. Zhang. The boundaries of fair ai in medical image prognosis: A causal perspective.arXiv preprint arXiv:2510.08840, 2025

work page arXiv 2025
[33]

Ple ˇcko and E

D. Ple ˇcko and E. Bareinboim. Causal fairness analysis: A causal toolkit for fair machine learning.Foundations and Trends® in Machine Learning, 17(3):304–589, 2024

work page 2024
[34]

Plecko, P

D. Plecko, P. Secombe, A. Clarke, A. Fiske, S. Toby, D. Duff, D. Pilcher, L. A. Celi, R. Bellomo, and E. Bareinboim. An algorithmic approach for causal health equity: A look at race differentials in intensive care unit (icu) outcomes.arXiv preprint arXiv:2501.05197, 2025

work page arXiv 2025
[35]

M. M. Rahman and S. Purushotham. Fair and interpretable models for survival analysis. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1452–1462, 2022

work page 2022
[36]

Royston and M

P. Royston and M. K. Parmar. Restricted mean survival time: an alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome.BMC medical research methodology, 13(1):152, 2013

work page 2013
[37]

Secombe, J

P. Secombe, J. Millar, E. Litton, S. Chavan, T. Hensman, G. K. Hart, A. Slater, R. Herkes, S. Huckson, and D. V . Pilcher. Thirty years of anzics core: a clinical quality success story. Critical Care and Resuscitation, 25(1):43–46, 2023

work page 2023
[38]

M. Y . Shaheen. Applications of artificial intelligence (ai) in healthcare: A review.ScienceOpen Preprints, 2021

work page 2021
[39]

Shpitser and J

I. Shpitser and J. Pearl. What counterfactuals can be tested. InProceedings of the Twenty-third Conference on Uncertainty in Artificial Intelligence, page 352–359, 2007. 11

work page 2007
[40]

Sonabend, F

R. Sonabend, F. Pfisterer, A. Mishler, M. Schauer, L. Burk, S. Mukherjee, and S. V ollmer. Flexible group fairness metrics for survival analysis.arXiv preprint arXiv:2206.03256, 2022

work page arXiv 2022
[41]

P. J. Stow, G. K. Hart, T. Higlett, C. George, R. Herkes, D. McWilliam, R. Bellomo, A. D. M. Committee, et al. Development and implementation of a high-quality clinical database: the australian and new zealand intensive care society adult patient database.Journal of critical care, 21(2):133–141, 2006

work page 2006
[42]

L. Sweeney. Discrimination in online ad delivery. Technical Report 2208240, SSRN, Jan. 28

work page
[43]

URLhttp://dx.doi.org/10.2139/ssrn.2208240

work page doi:10.2139/ssrn.2208240
[44]

L. T. Sweeney and C. Haney. The influence of race on sentencing: A meta-analytic review of experimental studies.Behavioral Sciences & the Law, 10(2):179–195, 1992

work page 1992
[45]

T. J. VanderWeele. Causal mediation analysis with survival data.Epidemiology, 22(4):582–585, 2011

work page 2011
[46]

Xie and Y

T. Xie and Y . Ge. Fairness in survival analysis: A novel conditional mutual information augmentation approach.arXiv preprint arXiv:2502.02567, 2025

work page arXiv 2025
[47]

Zhang and E

J. Zhang and E. Bareinboim. Fairness in decision-making—the causal explanation formula. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 12 Technical Appendices forCausal F airness for Survival Analysis The source code for reproducing all the experiments can be found in our anonymized code reposi- tory https://anonymous.4...

work page 2018