Pith. sign in

REVIEW 4 major objections 7 minor 10 references

Federated Causal Inference in Healthcare: Methods, Challenges, and Applications

T0 review · 4 major / 7 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper argues that a FedProx-style proximal update achieves zero asymptotic bias and pooled-level variance for Cox and Aalen-Johansen treatment-effect estimators under site-level heterogeneity.

desk verdict The survey taxonomy is useful, but the claimed survival-theory results are unsupported and almost certainly wrong—FedAvg is not biased under a shared Cox coefficient. read the letter →

arxiv 2505.02238 v1 pith:IRSFKVLG submitted 2025-05-04 cs.LG

Haoyang Li , Jie Xu , Kyra Gan , Fei Wang , Chengxi Zang This is my paper · ORCID
classification cs.LG MSC 62F1262N0162N02
keywords federatedlearningcausalinferencetreatmenteffectestimationsurvivalanalysisCoxproportionalhazardsmodelAalen-Johansenestimatordataheterogeneityhealthcare
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is a survey-and-theory contribution to federated causal inference in healthcare. It classifies existing methods into weight-based and optimization-based approaches, then supplies new asymptotic bias and variance comparisons for time-to-event outcomes under site heterogeneity. Its headline result is that a FedProx-style regularized federated estimator matches pooled analysis, with zero asymptotic bias and pooled-level variance, for both Cox and Aalen-Johansen estimators, while naive averaging and meta-analysis are both biased and less efficient. If that result holds, multi-site treatment-effect studies could obtain centralized accuracy without sharing patient-level data.

What carries the argument

The central mechanism is the FedProx proximal objective: each site minimizes its local Cox or Aalen-Johansen loss $\mathcal{L}_i(\theta)$ plus a quadratic penalty $\frac{\mu}{2}\|\theta-\theta_{\mathrm{global}}\|^2$ that keeps local updates close to the current global parameter. The proximal term is what the paper credits for pulling site-specific estimators toward a shared coefficient vector and for making the aggregated estimator consistent for the global target, despite differences in baseline hazards and covariate distributions across sites.

What would settle it

Simulate two sites with different baseline hazard functions and covariate distributions but a common true hazard ratio; run the paper's Algorithm 1 with a Cox local loss and compare the FedProx coefficient to the pooled maximum partial likelihood estimate as the sample size $n$ grows. If FedProx bias does not shrink to zero while pooled bias does, or if its variance remains above the pooled variance, the claimed equality in Eqs. (40)-(41) fails. A complementary check would compare FedProx Aalen-Johansen cumulative incidence curves with the pooled Aalen-Johansen curve under site-specific competing risks.

Watch

Extended reading notes

Core claim

The paper's central claim is that when each site has its own baseline hazard and covariate distribution but shares a common treatment-coefficient vector, a FedProx-style proximal regularized federated estimator is asymptotically unbiased and achieves the same variance as an analysis of all pooled data, for both Cox proportional hazards and Aalen-Johansen competing-risk estimators. The paper states this as Eqs. (40)-(41): $\mathrm{Bias}_{\mathrm{pooled}}=\mathrm{Bias}_{\mathrm{FedProx}}=0<\mathrm{Bias}_{\mathrm{FedAvg}},\mathrm{Bias}_{\mathrm{Meta\text{-}fixed}},\mathrm{Bias}_{\mathrm{Meta\text{-}random}}$ and $\mathrm{Var}_{\mathrm{pooled}}=\mathrm{Var}_{\mathrm{FedProx}}<\mathrm{Var}_{\mathrm{FedAvg}},\mathrm{Var}_{\mathrm{Meta\text{-}fixed}},\mathrm{Var}_{\mathrm{Meta\text{-}random}}$. This matters because it promises communication-efficient, privacy-preserving multi-site survival analysis with centralized-quality estimates.

Load-bearing premise

The load-bearing premise is that minimizers of the FedProx proximal objective converge to the true shared coefficient vector and global cumulative incidence functions even when each site has its own baseline hazard and covariate distribution; the paper asserts this in Theorems 1 and 2 without proof, and if local estimators are inconsistent under such heterogeneity, adding a norm penalty cannot repair them.

Editorial extensions

If this is right

  • If the claim holds, a communication-efficient FedProx-style algorithm can replace centralized data pooling for Cox and Aalen-Johansen treatment-effect estimation without asymptotic loss.
  • Under site-level heterogeneity, naive FedAvg and inverse-variance meta-analysis are dominated in both bias and variance, so they should not be the default aggregation in survival settings.
  • The result implies that the proximal regularization coefficient acts as a tuning parameter controlling the bias-variance trade-off; the paper's comparisons assume this coefficient is chosen suitably.
  • A single federated estimation framework covers binary, continuous, and time-to-event outcomes, since the local losses can be linear, logistic, Cox, or Aalen-Johansen.
  • Privacy-preserving federated networks of hospitals could achieve pooled-level statistical efficiency while keeping patient-level data in place.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the asymptotic claim holds, the FedProx estimator could serve as a finite-sample benchmark in multi-site trial emulations, but the paper does not provide a convergence-rate or confidence-interval construction.
  • Editorial inference: the same proximal argument might extend to time-varying exposures or recurrent-event models, because the local-loss-plus-quadratic-penalty structure does not depend on a specific hazard form.
  • Editorial inference: a direct testable extension would be to derive the fixed point of the proximal update for Aalen-Johansen cumulative incidence functions and check whether it equals the pooled estimator; the paper leaves the mechanism as an assertion.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. This manuscript is a review of federated causal inference methods for binary/continuous and time-to-event outcomes, together with a claimed theoretical analysis of asymptotic bias and variance under data heterogeneity. The authors organize methods into weight-based and optimization-based categories, describe FedProx, personalized federated learning, peer-to-peer communication, and model decomposition frameworks, review software tools and applications, and conclude in Eqs. (40)-(41) that FedProx-style regularization achieves zero asymptotic bias and pooled-level asymptotic variance for Cox proportional hazards and Aalen-Johansen estimators under heterogeneous site-level baseline hazards and covariate distributions.

Significance. If the central theoretical claim (Theorems 1-2, Eqs. (40)-(41)) were correct, it would be a substantial contribution: it would show that communication-efficient proximal federated estimation matches centralized pooled analysis in survival settings. The paper also usefully summarizes the Khellaf et al. theory for ATE estimation and surveys an emerging software ecosystem. However, the central time-to-event theorem is presented as a table without proof, and the claimed ordering is contradicted by standard partial-likelihood asymptotics under the authors' own model. The review material is broad but largely descriptive, and the theoretical contribution, which is the advertised novelty, is not established.

major comments (4)
  1. [Theoretical Foundations of Federated Estimators for Time-to-Event Outcomes, Theorems 1-2 (Tables 2-3, Eqs. (38)-(39))] The central results are asserted without proof. The abstract claims the paper derives asymptotic bias and variance, but the time-to-event section provides only Definitions 1-4 followed directly by two tables with garbled entries. No regularity conditions, identifiability assumptions, or derivations are supplied, so the claimed zero-bias and variance equalities are not established. A theorem presented as a table with corrupted symbols is not a theorem as it stands.
  2. [Eq. (32) and Table 2 (Cox model)] Under the model in Eq. (32), with site-specific baseline hazard h_0i(t) and a shared coefficient beta_0, the local maximum partial likelihood estimator beta_i is consistent for beta_0 under standard independent censoring conditions, because the local partial likelihood score has expectation zero at beta_0 by the martingale property. Consequently, gamma_i in Eq. (33) is zero and the FedAvg estimator in Eq. (19), a convex combination of consistent local estimators, also has zero asymptotic bias. The nonzero FedAvg bias row in Table 2 and the strict ordering Bias_FedProx < Bias_FedAvg in Eq. (40) are therefore incompatible with the model as stated; the authors would need to add an explicit source of local inconsistency and prove that FedProx remains consistent under that alternative.
  3. [Theorem 2/Table 3 (Aalen-Johansen estimator)] Definition 2 and the heterogeneous-setting paragraph define site-specific true CIFs F_{k,i}(t). If these differ across sites, no estimator computed from local data can be unbiased for an arbitrary external global target F_{k,0}(t) without either defining that target as a known mixture of the site-specific CIFs or imposing an explicit transportability/importance-sampling assumption. The proximal penalty in Eq. (20) does not reweight the populations, so it cannot remove the difference between the local estimands and an unspecified global target. The zero-bias claim in Eq. (41) requires an explicit definition of the global estimand and a positivity condition, neither of which is supplied.
  4. [Eqs. (38)-(39) and (40)-(41)] The asymptotic variance entries in Tables 2 and 3 are garbled to the point of being unreadable, for example '1-#&!!.!(+)' in Table 3. Even setting legibility aside, the manuscript never states which matrices or variance functional the entries denote, nor does it define the pooled estimator in the heterogeneous survival setting. As a result, the claimed variance ordering Varpooled = VarFedProx < VarFedAvg etc. cannot be verified or falsified from the text, and the central comparison is not mathematically meaningful.
minor comments (7)
  1. [Section 'Considering Local Estimator' and earlier ATE section] The equation number (1) is used twice: once for the ATE definition and once for the local estimator in the Khellaf et al. summary; all equations should be renumbered consistently.
  2. [Algorithm 1 and Eq. (20)] Algorithm 1 uses z for the proximal coefficient while Eq. (20) uses mu; unify the notation for the FedProx regularization strength.
  3. [Eq. (7)] The kernel bandwidth h in Eq. (7) is never defined in terms of a data-dependent choice, and the FedProx coefficient is not given any selection guidance; the paper should at least discuss sensitivity to these tuning parameters.
  4. [Reference list] Many references are given only as 'Google Search' entries or bare URLs without full bibliographic information, including refs. 4, 54, 56, 61, 63, 67, 73, 75-77, 79-81, 83, 95, 99, 101, 105, and 110; these are not citable in a journal publication.
  5. [Figure numbering] Figure 2 appears twice with different content, once as the graphical model for heterogeneous distributions and once as a characteristics comparison of meta-analysis, FedAvg, and FedProx; renumber the figures.
  6. [Software Tools section] The sentence 'Among them, FLAME and dsSurvival provide the most direct support for causal inference tasks' is unsupported because FLAME does not appear in Table 4.
  7. [Proofreading] The manuscript contains several typos and garbled glyphs, including 'dost not hold', 'contributions are three folds', and the use of '!' for treatment and '(' for outcome in the binary/continuous outcome section; these should be corrected carefully.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity: the survival theorems are asserted without derivation; self-citations are background and not load-bearing.

full rationale

The paper's central time-to-event claims are the asymptotic bias/variance tables (Theorems 1–2, Eqs. (38)–(41)), but these are presented as table entries with no proof. A circularity finding requires a specific reduction showing that a claimed output equals an input by construction, or that a fitted parameter is renamed as a prediction. No such reduction is exhibited. The FedProx objective in Eq. (20) is defined as a proximal-penalized local loss, yet the zero-bias and pooled-variance rows for FedProx in Tables 2–3 are not derived from that objective; no equations connect the proximal update to the claimed bias and variance. The Aalen–Johansen global target F_{k,0}(t) in Eq. (36) is not formally defined, but the tables do not use that definition to force the FedProx row. For binary/continuous outcomes, the results are explicitly attributed to the external source Khellaf et al. [28], so they are not self-citations. The paper contains several self-citations by the authors, but they appear in background and application sections (OOD generalization, graph representation learning, trial design) and are not used to support Eqs. (38)–(41). Thus the central survival results are unsupported assertions rather than circular derivations; concerns about their correctness, such as the local Cox MLE being consistent for a shared coefficient under Eq. (32), are rigor/correctness issues, not circularity. Score 1 reflects the presence of minor non-load-bearing self-citations and an asserted rather than derived central result, without any circular step.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. Its central estimates rely on standard causal and statistical assumptions plus an unproved convergence assumption for FedProx. The included free parameters are user-chosen hyperparameters whose values are not specified in the theorems.

free parameters (2)
  • FedProx proximal coefficient mu (z in Algorithm 1) = not specified
    The survival theorem tables report a single FedProx entry without stating the value of the regularization coefficient. The result either depends on mu, in which case the theorem is incomplete, or must hold for all mu, which needs proof.
  • Kernel bandwidth h in Eq (7) = not specified
    The kernel-based weighting method described in Eq (7) depends on a bandwidth h, but the paper gives no selection rule. This hyperparameter is not central to the survival theorems.
assumptions (4)
  • domain assumption Each site can fit a full-rank OLS model (Condition 1) or the pooled data is full-rank (Condition 2).
    Used in the binary-outcome theoretical section, attributed to Khellaf et al.; no relaxation or proof is given here.
  • domain assumption A common Cox coefficient vector exists across sites while baseline hazards may vary.
    Definition 1, Eq (32). The survival theorems assume this shared beta; if beta differs by site, the estimand is not well-defined.
  • domain assumption Non-informative right censoring and standard Aalen-Johansen convergence conditions hold.
    Theorem 2 preamble; needed for the asymptotic variance formulas q_i(X).
  • ad hoc to paper The FedProx proximal objective converges to an estimator consistent for the shared global target under heterogeneity.
    Eq (20) and Algorithm 1. This convergence-consistency premise is asserted without proof and is the load-bearing step for Theorems 1 and 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Federated Causal Inference in Healthcare: Methods, Challenges, and Applications." pith.science (2026). https://pith.science/paper/IRSFKVLG

@misc{pith2026250502238,
  author       = {Pith},
  title        = {Pith review of: Federated Causal Inference in Healthcare: Methods, Challenges, and Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRSFKVLG}},
  note         = {Machine review of arXiv:2505.02238}
}
read the original abstract

Federated causal inference enables multi-site treatment effect estimation without sharing individual-level data, offering a privacy-preserving solution for real-world evidence generation. However, data heterogeneity across sites, manifested in differences in covariate, treatment, and outcome, poses significant challenges for unbiased and efficient estimation. In this paper, we present a comprehensive review and theoretical analysis of federated causal effect estimation across both binary/continuous and time-to-event outcomes. We classify existing methods into weight-based strategies and optimization-based frameworks and further discuss extensions including personalized models, peer-to-peer communication, and model decomposition. For time-to-event outcomes, we examine federated Cox and Aalen-Johansen models, deriving asymptotic bias and variance under heterogeneity. Our analysis reveals that FedProx-style regularization achieves near-optimal bias-variance trade-offs compared to naive averaging and meta-analysis. We review related software tools and conclude by outlining opportunities, challenges, and future directions for scalable, fair, and trustworthy federated causal inference in distributed healthcare systems.

Figures

Figures reproduced from arXiv: 2505.02238 by the authors.

Figure 1
Figure 1. Federated causal inference methods framework. Federated Estimation framework In federated causal inference, the design of estimation frameworks plays a critical role in addressing statistical, computational, and communication challenges—especially under heterogeneity. Beyond simple averaging-based approaches, several classes of federated learning (FL) algorithms have emerged, offering various strategies for aggregat… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 5 canonical work pages

  1. [4]

    Federated learning for predicting clinical outcomes in patients with COVID-19 - Google Search. 5. Lee, E. H. et al. An international study presenting a federated learning AI platform for pediatric brain tumors. Nat Commun 15, 7615 (2024). 6. Nguyen, A. et al. Deep Federated Learning for Autonomous Driving. Preprint at https://doi.org/10.48550/arXiv.2110.0...

    work page Pith review arXiv doi:10.48550/arxiv.2110.05754 2024

  2. [14]

    Zhang, Z. et al. Spectral Invariant Learning for Dynamic Graphs under Distribution Shifts. in (2023). 15. Li, H., Zhang, Z., Wang, X. & Zhu, W. Invariant Node Representation Learning under Distribution Shifts with Multiple Latent Environments. ACM Trans. Inf. Syst. 42, 26:1-26:30 (2023). 16. Li, H. et al. Disentangled Graph Self-supervised Learning for Ou...

    arXiv doi:10.48550/arxiv.2307.10616 2023

  3. [29]

    Meurisse, M. et al. Real-World Comparative Effectiveness of SARS-CoV-2 Primary Vaccination Campaigns Against SARS-CoV-2 Infections: A Federated Observational Study Emulating a Target Trial in Three Nations. SSRN Scholarly Paper at https://doi.org/10.2139/ssrn.4869107 (2024). 30. MSR119 Does Federated Analytics Preserve Statistical and Scientific Value of ...

    work page doi:10.2139/ssrn.4869107 2024

  4. [36]

    & Fischer, J

    Mitrovska, A., Safari, P., Ritter, K., Shariati, B. & Fischer, J. K. Secure federated learning for Alzheimer’s disease detection. Front Aging Neurosci 16, 1324032 (2024). 37. Lee, D. Y. et al. Privacy-Preserving Federated Model Predicting Bipolar Transition in Patients With Depression: Prediction Model Development Study. J Med Internet Res 25, e46165 (202...

    arXiv 2024

  5. [51]

    Kidd, B., Wang, K., Xu, Y. & Ni, Y. Federated Learning for Sparse Bayesian Models with Applications to Electronic Health Records and Genomics. Pac Symp Biocomput 28, 484–495 (2023). 52. Enhancing Treatment Effect Prediction on Privacy-Protected Data: An Honest Post-Processing Approach - Working Paper - Faculty & Research - Harvard Business School. https:/...

    work page Pith review arXiv doi:10.48550/arxiv.2311.04375 2023

  6. [59]

    https://dl.acm.org/doi/10.1145/3534678.3539242

    DICE: Domain-attack Invariant Causal Learning for Improved Data Privacy Protection and Adversarial Robustness | Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. https://dl.acm.org/doi/10.1145/3534678.3539242. 60. Terrail, J. O. du et al. FedECA: A Federated External Control Arm Method for Causal Inference with Time-To-...

    arXiv 2025

  7. [67]

    The Application of Causal Inference Algorithms in Federated Recommender Systems - Google Search. 68. Almodóvar, A., Parras, J. & Zazo, S. Propensity Weighted federated learning for treatment effect estimation in distributed imbalanced environments. Computers in Biology and Medicine 178, 108779 (2024). 69. Meurisse, M. et al. Federated causal inference bas...

    arXiv doi:10.48550/arxiv.1712.04912 2024

  8. [76]

    Learning personalized causally invariant representations for heterogeneous federated clients - Google Search. 77. Mitigating data heterogeneity in federated learning with data augmentation - Google Search. 78. Guo, X. et al. Sample Quality Heterogeneity-aware Federated Causal Discovery through Adaptive Variable Space Selection. in vol. 5 4071–4079 (2024)....

    work page Pith review arXiv doi:10.48550/arxiv.2301.00346 2024

Show all 10 references
  1. [94]

    & Ohno-Machado, L

    Bonomi, L., Jiang, X. & Ohno-Machado, L. Protecting patient privacy in survival analyses. J Am Med Inform Assoc 27, 366–375 (2019). 95. Fair and Interpretable Pseudo Value-Based Deep Learning Models for Federated Survival Analysis - Google Search. 96. JMIR AI - Privacy-Preserv...

    2019

  2. [103]

    & Fritz, M

    Rahimian, S., Kerkouche, R., Kurth, I. & Fritz, M. Practical Challenges in Differentially-Private Federated Survival Analysis of Medical Data. Preprint at https://doi.org/10.48550/arXiv.2202.03758 (2022). 104. Hernán, M. A., Brumback, B. & Robins, J. M. Marginal structural mod...

    2022 doi arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.