Debiased maximum-likelihood estimators for hazard ratios under kernel-based machine-learning adjustment

Satoshi Asai; Takashi Hayakawa

REVIEW 3 major objections 5 minor 62 references

Debiased maximum-likelihood estimators for hazard ratios under kernel-based machine-learning adjustment

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Hazard ratios become causally interpretable when the Cox baseline hazard is replaced by kernel machine learning, and debiased ML estimators make them reliably estimable.

desk verdict A genuinely new debiased-ML hazard-ratio estimator whose main asymptotic guarantee currently rests on an unproved conjecture. read the letter →

arxiv 2507.17686 v3 pith:FE2DO23F submitted 2025-07-23 stat.ML cs.LG

Takashi Hayakawa , Satoshi Asai This is my paper · ORCID

classification stat.MLcs.LG MSC 62N0262G0562D20

keywords hazardratiodebiasedmachinelearningNeymanorthogonalitykernelmethodscausalinferencesurvivalanalysislatentvariablemodels

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

The paper proposes that the Cox model's unspecified baseline hazard is the main obstacle to giving hazard ratios a causal reading, because it lets the same parameter values describe many different risk-set selection processes. To fix this, it works with an exponential hazard model $h(t|A_t,X_t)=\exp(\theta'A_t+f(X_t))$ that has no baseline hazard, with $f$ learned by kernel machine learning. Under consistency, conditional randomization of treatment, time-homogeneity, and correct specification, the maximum-likelihood parameter $\theta^*$ equals the covariate-adjusted log hazard contrast between two counterfactual treatment schedules that branch at time $t$ (Proposition 1). Since kernel ML estimators of $\theta$ are not root-$n$ consistent, the paper builds two Neyman-orthogonal score families and a cross-fitted debiased ML estimator, proving convergence under an assumption the authors flag as a conjecture. Simulations with nonlinear confounding and with an unobserved risk factor show the debiased estimates hit the true hazard ratios with minimal bias, unlike naive ML.

What carries the argument

The load-bearing object is the exponential hazard model without a baseline hazard, $h(t|A_t,X_t)=\exp(\theta'A_t+f(X_t))$, where $f$ lies in a reproducing-kernel Hilbert space built from multiple kernels. The likelihood's Hessian operator $H$ (with blocks $H_{\theta f}$, $H_{ff}$ and regularization $\zeta_n$) enters Proposition 2's score $\phi=\partial_\theta\ell-H_{\theta f}(H_{ff}+\zeta_n)^{-1}\partial_f\ell$, while Proposition 3's score uses a kernel-logistic nuisance $g_k$ that plays the role of an inverse treatment-probability weight marginalized over time. These scores are Neyman-orthogonal, meaning the Gateaux derivative with respect to the nuisance parameters vanishes at the truth, so slow kernel learning rates for $f$ and $g_k$ do not bias $\theta$.

What would settle it

Simulate data where treatment assignment depends on a covariate transformation that is not representable in the chosen RKHS (so the density ratio $d\nu_k/d\nu_0$ lies outside the kernel space), then run the debiased estimator for increasing $n$; if the $t$-statistics drift away from $N(0,1)$, the range-inclusion half of Assumption 9 is load-bearing and false in that setting.

Watch

Extended reading notes

Core claim

The central discovery is that the maximum-likelihood solution of a baseline-hazard-free exponential model $h(t|A_t,X_t)=\exp(\theta'A_t+f(X_t))$, with $f$ in a reproducing-kernel Hilbert space, is a causally interpretable quantity: under Assumptions 1–8, $\theta^*_k$ equals the log ratio of counterfactual hazards at time $t$ for two treatment schedules that agree before $t$ and differ only on a small interval, conditional on the observed covariate history. This removes the multiplicity of contradictory interpretations that the standard semiparametric proportional-hazards model permits, because the time evolution of the risk set is modeled explicitly by $f$ (and optionally latent variables) instead of being absorbed by an unestimated baseline hazard. The paper then constructs Neyman-(near-)orthogonal scores—one based on the likelihood Hessian operator (Proposition 2) and one based on inverse-probability-weighted integral balance (Proposition 3)—and shows that, when Assumption 9 holds, the debiased ML estimator is $\sqrt{n}$-consistent and asymptotically normal.

Load-bearing premise

The entire debiasing argument rests on Assumption 9—that nuisance estimators converge at a polynomial $n^{-\beta}$ rate and that the Hessian block $H^*_{f\theta}$ lies in the range of $H^*_{ff}$—which the authors themselves leave as a conjecture in Remark 4-1 and Appendix IX; if either piece fails, the claimed asymptotic unbiasedness of the debiased estimator is unsupported.

Editorial extensions

If this is right

Hazard ratios estimated in this framework are causal contrasts between counterfactual treatment schedules, so they can be reported with the same interpretation as randomized hazard contrasts.
The debiased estimator is asymptotically normal with a plug-in standard error, allowing significance tests even though the kernel nuisance $f$ converges slower than $\sqrt{n}$.
Model misspecification that leaves temporal risk changes unmodeled shows up as a violation of the time-homogeneity assumption, detectable by comparing Bayesian model evidence with and without time-elapsed covariates.
The method extends to latent-variable models, so unobserved risk factors that select the risk set can be modeled explicitly while the hazard ratio stays estimable with minimal bias.
The approach provides a machine-learning-friendly alternative to marginal structural Cox models for observational data with dynamic treatment and many time-dependent covariates.

Reading between the lines

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

If the range-inclusion condition $H^*_{f\theta}=H^*_{ff}\rho$ fails for a data-generating process with a propensity outside the chosen kernel space, the Proposition 2 score may lose its orthogonality; the Proposition 3 density-ratio score may then be the safer default.
The time-homogeneity assumption is a strong biological claim; a testable extension would allow a smooth baseline trend in $f$ and check whether $\theta$ estimates remain stable across model choices.
Comparing Bayesian model evidence with and without time-elapsed covariates can be read as a diagnostic for unobserved risk factors—an implicit screen that could be formalized as a model-selection rule.
For continuous or nonlinear treatment effects, Proposition 3 does not apply but Proposition 2 does, suggesting a path to dose-response analyses beyond the binary treatment setting.

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.

Desk Editor's Note

read the letter

The paper does something real: it drops the baseline hazard from the hazard model to dodge the Cox model's selection-induced interpretability problem, fits an exponential hazard with an RKHS nuisance f, and builds two Neyman-orthogonal score constructions for debiased ML estimation of the treatment log-hazard ratio. Proposition 1 gives a clean causal interpretation of the target under explicit assumptions, and Propositions 2 and 3 are genuinely new score constructions; the density-ratio form in Proposition 3 is elegant and practically appealing. The simulations show the debiasing removes most of the naive ML bias, which is a useful sanity check.

Now the soft spots, in proportion. The load-bearing issue is Assumption 9. The paper itself labels it a conjecture in Remark 4-1 and again in Appendix IX, and the central debiasing guarantee depends on it. The range condition H*_fθ = H*_ffρ is what turns the Gateaux derivative of the score from O(n^-β) into O(n^-α-β); without it, near-orthogonality at the required o(n^-1/2) rate fails. Appendix XI then derives conditions like β > 63/154, but that entire chain inherits the conjecture. So the asymptotic normality and unbiasedness claims are conditional on an unproved regularity assumption. That is not a fatal flaw if the conjecture is true, but it does mean the paper's central theorem is not fully proven. The plausibility argument in Appendix IX (density ratio belonging to the RKHS) is suggestive, not a proof.

Other concerns are smaller. The simulations do not benchmark against Cox partial likelihood or marginal structural models, which would help calibrate the efficiency tradeoff they discuss. The latent-variable simulation initializes the EM near the true values, so it does not stress the optimization. And the code is only promised as supplementary on acceptance; for this kind of method, a public repository with the code would be the norm.

Is the paper worth serious peer review? Yes. The idea is novel, the causal framing is careful, and the authors are explicit about what they have not proven. A good referee could either extract a proof of Assumption 9 under reasonable data-generating conditions or force the authors to state openly that the debiasing guarantee is conditional. I would send it out rather than desk reject. I would not cite it myself until the conjecture is resolved, but I would put it on a reading group agenda to think about whether the range condition is generic.

Referee Report

3 major / 5 minor

Summary. The paper proposes abandoning the unspecified baseline hazard in favor of an exponential hazard model h(t)=exp(θ'A_t+f(X_t)) with f in a reproducing-kernel Hilbert space, and develops debiased machine-learning estimators of the log-hazard-ratio θ* based on Neyman-orthogonal scores. It gives a causal interpretation of θ* under explicit assumptions (Proposition 1), constructs two debiasing scores (Propositions 2 and 3), presents a cross-fitting algorithm with model selection by Bayesian model evidence, and reports simulation studies for observed confounders and for unobserved heterogeneity via a latent variable. The claimed asymptotic guarantees for the debiased estimator rely on Assumption 9, which the paper itself labels a conjecture in Remark 4-1 and Appendix IX.

Significance. If the statistical claims were fully established, this would be a substantive contribution: it directly targets the well-known uninterpretability of Cox hazard ratios in dynamic observational settings and extends double/debiased machine learning to survival models with kernel-based nuisance functions. The paper is transparent about its assumptions, provides detailed derivations in the appendices, self-identifies Assumption 9 as a conjecture, and includes reproducible simulation code as supplementary material. The core mathematical construction (Propositions 2 and 3) is plausible and non-trivial. However, because the key convergence rates and the range condition for the Hessian are not proved, the paper currently delivers a conditional framework rather than the unconditional asymptotic-unbiasedness result announced in the abstract.

major comments (3)

[Section II.C, Assumption 9 and Appendix VII, Eqs. (58)-(59)] Assumption 9 is introduced in Section II.C and explicitly left as a conjecture in Remark 4-1 and Appendix IX, yet the near-orthogonality of Proposition 2 depends on it. The bound in Eq. (59) is O(n^{-α-β}) only after substituting H*_{fθ_k}=H*_{ff}ρ_k; absent that range condition, the Gateaux derivative in Eq. (58) is O(n^{-β}), which is not o(n^{-1/2}) because β<1/2. Thus Neyman near-orthogonality condition (16) can fail, and the asymptotic unbiasedness of the debiased estimator is unsupported unless Assumption 9 is proved. Appendix IX provides only a density-ratio plausibility argument for ρ_k∈H_k and explicitly leaves the stochastic-process convergence of the averaged Hessians open.
[Appendix XI, Proposition 4] The verification of the Chernozhukov et al. conditions (Assumptions 3.3 and 3.4) in Proposition 4 derives the requirement β>63/154 using the same unproved nuisance convergence rates and the range condition from Assumption 9. Consequently, the statement in the abstract and Section II.C that the paper proves 'necessary convergence results' overstates what is established. The main theorem should be stated explicitly as conditional on Assumption 9, or Assumption 9 should be replaced by a verifiable condition with a proof for a reasonable class of data-generating processes.
[Section III, simulations] The numerical experiments do not probe the load-bearing parts of Assumption 9. Both simulation DGPs use smooth Gaussian-kernel components and logistic density-ratio structures, precisely the regime in which the range condition H*_{fθ}=H*_{ff}ρ is most plausible, and the nuisance functions are estimated by correctly specified kernel models. The simulations therefore cannot distinguish the proposed debiased estimator from one that would fail when the range condition is violated or when nuisance convergence is slower than n^{-β}. A stress test with a misspecified or non-smooth nuisance, or with a deliberately violated range condition, would materially strengthen the evidence.

minor comments (5)

[Table I] The table heading contains a typo: 'Mathmatical notations' should read 'Mathematical notations'.
[Section II.E] The sentence 'On can observe whether incorporation or removal of a component function...' contains a typo; 'On' should be 'One'.
[Section III.A] The value of P2 used for the reported t-statistic histograms in Fig. 1(G,H) is not stated; the text says different values are used without specifying which one produced the displayed results.
[Section II.F] The cross-fitting split in Step 2 (training D\(D(m)∪D(m+1 mod M)), validation D(m+1 mod M), held-out D(m)) is nonstandard and should be clarified: the validation set is used only for hyperparameter selection and is never used for scoring, so the text should state explicitly that this preserves the required independence.
[Data availability] The source codes are described as available as supplementary materials but no permanent repository or DOI is given; consider providing an archival link.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: causal identification is proven from the likelihood model, and the debiasing results are conditional on explicitly stated (though partly conjectural) assumptions; nothing is fitted and then relabeled as a prediction.

full rationale

The paper's causal claim (Proposition 1) is not circular: theta* is defined as the maximizer of the expected log-likelihood of the exponential hazard model, and the equality to the counterfactual hazard contrast in Eq. (12) is derived from Assumptions 1-8 in Appendix VI. The likelihood equations identify theta* as the conditional hazard ratio; the causal interpretation then follows from consistency, ignorability, and regularity assumptions in a standard identification argument, not by restating the definition of theta*. The debiasing construction is also not circular: the score in Eq. (22) is the Neyman-orthogonalized likelihood score, and Proposition 2 proves near-orthogonality conditional on Assumption 9. The paper itself flags that Assumption 9 is left as a conjecture (Remark 4-1 and Appendix IX); in particular the range condition H*_fθ = H*_ffρ and the n^{-beta} convergence rates are not fully proved. That is a genuine gap in support for the asymptotic unbiasedness claim, but an unproved assumption is a limitation, not a circular reduction of the conclusion to the input. There are no self-citations, no fitted parameters relabeled as predictions, and no ansatz smuggled in via citation: the DML framework of Chernozhukov et al. [19] is external and applied transparently, and the simulations serve as independent numerical checks rather than inputs to the estimator.

Assumptions & free parameters 4 free parameters · 8 assumptions · 1 invented entities

The central claim rests on conventional causal assumptions (1-4), strong homogeneity and time-homogeneity assumptions (5-7), correct model specification (8), and an unproven convergence-rate assumption (9). The method also introduces tuning hyperparameters that are fit to data. No new physical entity is introduced; the latent variable is a statistical modeling component.

free parameters (4)

Kernel bandwidth hyperparameters σ_d = e.g., 1, 10, 100 (grid search)
Chosen via grid search and BME in Step 1; affect the RKHS functions f̂ and ĝ.
Regularization strengths λ_k = e.g., 10, 100, 1000 for Gaussian kernels; λ_k=0 for linear kernels
Selected by BME; control the complexity of the kernel ML nuisance estimators.
Ridge hyperparameter ζ_n for Hessian/g_k = e.g., ζ_{n,1}=1 or 70 in simulation
Used in the orthogonal score to stabilize inversion; tuned via CVErr_H or CVErr_g.
Convergence-rate exponents α, β = not specified (α > 1/2 - β, β > 63/154 etc.)
Theoretical rates assumed in Assumption 9; not derived, only constrained by inequalities in Appendix XI.

assumptions (8)

domain assumption Assumptions 1-4 (consistency, sequential ignorability, positivity, hazard regularity)
Standard causal inference assumptions, Section II B.
domain assumption Assumption 5 (hazard independent of past treatment/covariates)
Essentially a Markov/conditional independence condition; Section II B.
domain assumption Assumption 6 (homogeneous treatment effect)
Hazard contrast constant across covariate values; Section II B.
domain assumption Assumption 7 (time-homogeneous hazard)
Hazard independent of observation time given covariates; Section II B. Strong and central to the causal interpretation.
domain assumption Assumption 8 (correctly specified ML model)
There exists θ*,f* maximizing expected log-likelihood over all Borel functions; Section II B.
ad hoc to paper Assumption 9 (nuisance convergence rates and H*_{fθ}=H*_{ff}ρ)
Stated in Section II C and explicitly left as a conjecture in Remark 4-1 and Appendix IX. Required for DML convergence proofs in Appendix XI.
standard math RKHS background (bounded kernels, representer theorem)
Used to represent gradients and Hessians; cited from [22,23].
standard math Chernozhukov et al. DML framework
General theorem for asymptotic normality of Neyman-orthogonal scores; Lemma 1 cited from [19].

invented entities (1)

Latent binary variable Z (or W) for unobserved risk factor
purpose: Model risk-set selection caused by unobserved heterogeneity in the latent-variable extension (Section II D)
A statistical latent variable introduced to capture unobserved factors; no falsifiable handle outside the model's identifiability assumptions, and the authors note identifiability is a challenge.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Debiased maximum-likelihood estimators for hazard ratios under kernel-based machine-learning adjustment." pith.science (2026). https://pith.science/paper/FE2DO23F

@misc{pith2026250717686,
  author       = {Pith},
  title        = {Pith review of: Debiased maximum-likelihood estimators for hazard ratios under kernel-based machine-learning adjustment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FE2DO23F}},
  note         = {Machine review of arXiv:2507.17686}
}

read the original abstract

Previous studies have shown that hazard ratios between treatment groups estimated with the Cox model are uninterpretable because the unspecified baseline hazard of the model fails to identify temporal change in the risk set composition due to treatment assignment and unobserved factors among multiple, contradictory scenarios. To alleviate this problem, especially in studies based on observational data with uncontrolled dynamic treatment and real-time measurement of many covariates, we propose abandoning the baseline hazard and using kernel-based machine learning to explicitly model the change in the risk set with or without latent variables. For this framework, we clarify the context in which hazard ratios can be causally interpreted, and then develop a method based on Neyman orthogonality to compute debiased maximum-likelihood estimators of hazard ratios, proving necessary convergence results. Numerical simulations confirm that the proposed method identifies the true hazard ratios with minimal bias. These results lay the foundation for developing a useful, alternative method for causal inference with uncontrolled, observational data in modern epidemiology.

Figures

Figures reproduced from arXiv: 2507.17686 by the authors.

**Figure 2.** FIG. 2. Numerical results for a clinically plausible data-g [PITH_FULL_IMAGE:figures/full_fig_p046_2.png] view at source ↗

**Figure 3.** FIG. 3. Two causal graphs describing the causal relationshi [PITH_FULL_IMAGE:figures/full_fig_p047_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 53 canonical work pages

[1]

Alternative analysis methods for time to event endpoi nts under nonproportional haz- ards: A comparative analysis,

Ray S. Lin, Ji Lin, Satrajit Roychoudhury, Keaven M. Ande rson, Tianle Hu, Bo Huang, Larry F Leon, Jason J.Z. Liao, Rong Liu, Xiaodong Luo, Pralay Mukhopadhyay, Rui Qin, Kay Tatsuoka, Xuejing Wang, Yang Wang, Jian Zhu, Tai-Tsang C hen, and Renee Iacona and, “Alternative analysis methods for time to event endpoi nts under nonproportional haz- ards: A compa...

arXiv 2020
[2]

The optimization was stopped, if the ℓ2-norm of the gradient gets smaller than ǫstop = 1.0 × 10−2

We solve the above optimization problem, by applying limited- memory BFGS algorithm [56] with backtracking and line search stoppe d by the Armijo condition. The optimization was stopped, if the ℓ2-norm of the gradient gets smaller than ǫstop = 1.0 × 10−2. For ML estimation, one can approximate the BME by using the Laplace approximation. Assigning the Hess...

work page
[3]

Here we have used the fact ‖(H ∗ f f + ζn)−1H ∗ f f‖ ≤ 1. 31 VIII. GRADIENT FUNCTIONAL AND HESSIAN OPERA TOR In this section, we provide the gradient functional and the Hessian operator of the sam- plewise negative log-likelihood: ℓ(D;θ,f ) = −I[0,C](T )(θ′AT +f (XT )) + ∫ T ∧C 0 exp (θ′At +f (Xt))dt. (60) Calculating the Gateaux derivative 1 ǫ (ℓ(D;θ,f +...

work page
[4]

Does cox analysis of a randomized survival study yield a causal treatment eﬀect?

Odd O. Aalen, Richard J. Cook, and Kjetil Røysland, “Does cox analysis of a randomized survival study yield a causal treatment eﬀect?” Lifetime Dat a Analysis 21, 579–593 (2015)

work page 2015
[5]

The hazards of per iod speciﬁc and weighted hazard ratios,

Jonathan W. Bartlett, Tim P. Morris, Mats J. Stensrud, Rh ian M. Daniel, Stijn K. Vansteelandt, and Carl-Fredrik Burman, “The hazards of per iod speciﬁc and weighted hazard ratios,” Statistics in Biopharmaceutical Research 12, 518–519 (2020), https://doi.org/10.1080/19466315.2020.1755722

arXiv 2020
[6]

The hazards of hazard ratios,

Miguel A. Hern´ an, “The hazards of hazard ratios,” Epide miology 21 (2010)

work page 2010
[7]

Intention-to-trea t comparisons in randomized trials,

Ross L Prentice and Aaron K Aragaki, “Intention-to-trea t comparisons in randomized trials,” Statistical Science 37, 380–393 (2022)

work page 2022
[8]

Subtleties in the inter- pretation of hazard contrasts,

Torben Martinussen, Stijn Vansteelandt, and Per Kragh A ndersen, “Subtleties in the inter- pretation of hazard contrasts,” Lifetime Data Analysis 26, 833–855 (2020)

work page 2020

Show all 62 references

[9]

Causality and the cox regression m odel,

Torben Martinussen, “Causality and the cox regression m odel,” Annual Review of Statistics 39 and Its Application 9, 249–259 (2022)

2022
[10]

Treatment eﬀect quantiﬁcation for tim e-to-event endpoints–estimands, analysis strategies, and beyond,

Kaspar Ruﬁbach, “Treatment eﬀect quantiﬁcation for tim e-to-event endpoints–estimands, analysis strategies, and beyond,” Pharmaceutical Statist ics 18, 145–165 (2019), https://onlinelibrary.wiley.com/doi/pdf/10.1002/pst.1917

2019 doi
[11]

On defense of the hazard rati o,

Andrew Ying and Ronghui Xu, “On defense of the hazard rati o,” (2023), arXiv:2307.11971 [math.ST]

2023 arXiv
[12]

Causal interpretation of the ha zard ratio in ran- domized clinical trials,

Michael P Fay and Fan Li, “Causal interpretation of the ha zard ratio in ran- domized clinical trials,” Clinical Trials 21, 623–635 (2024), pMID: 38679930, https://doi.org/10.1177/17407745241243308

2024 doi
[13]

Estimating heterogeneous treatment eﬀects with right-censored data vi a causal survival forests,

Yifan Cui, Michael R Kosorok, Erik Sverdrup, Stefan Wag er, and Ruoqing Zhu, “Estimating heterogeneous treatment eﬀects with right-censored data vi a causal survival forests,” Jour- nal of the Royal Statistical Society Series B: Statistical M ethodology 85, 179–211 (2023), htt...

2023
[14]

Uses and limitations of the restricted mean survival time: Illustra tive examples from cardiovascular outcomes and mortality trials in type 2 diabetes,

David E. Kloecker, Melanie J. Davies, Kamlesh. Khunti, and Francesco Zaccardi, “Uses and limitations of the restricted mean survival time: Illustra tive examples from cardiovascular outcomes and mortality trials in type 2 diabetes,” Annals of Internal Medicine 172, 541–552 (20...

2020 doi
[15]

Treatment eﬀe ct measures un- der nonproportional hazards,

Steven Snapinn, Qi Jiang, and Chunlei Ke, “Treatment eﬀe ct measures un- der nonproportional hazards,” Pharmaceutical Statistics 22, 181–193 (2023), https://onlinelibrary.wiley.com/doi/pdf/10.1002/pst.2267

2023 doi
[16]

Design, implem entation, and inferential issues asso- ciated with clinical trials that rely on data in electronic m edical records: a narrative review,

Alan Leviton and Tobias Loddenkemper, “Design, implem entation, and inferential issues asso- ciated with clinical trials that rely on data in electronic m edical records: a narrative review,” BMC Medical Research Methodology 23, 271 (2023). 40

2023
[17]

Estimating heterogeneous treatment eﬀects on survi val outcomes using counterfactual censoring unbiased transformations,

Shenbo Xu, Raluca Cobzaru, Stan N. Finkelstein, Roy E. W elsch, Kenney Ng, and Zach Shahn, “Estimating heterogeneous treatment eﬀects on survi val outcomes using counterfactual censoring unbiased transformations,” (2024), arXiv:2401 .11263 [stat.ME]

2024
[18]

Orthogonal sur- vival learners for estimating heterogeneous treatment eﬀec ts from time-to-event data,

Dennis Frauen, Maresa Schr¨ oder, Konstantin Hess, and Stefan Feuerriegel, “Orthogonal sur- vival learners for estimating heterogeneous treatment eﬀec ts from time-to-event data,” (2025), arXiv:2505.13072 [cs.LG]

2025 arXiv
[19]

Double/debiased machine learning for treat- ment and structural parameters,

Victor Chernozhukov, Denis Chetverikov, Mert Demirer , Esther Duﬂo, Christian Hansen, Whitney Newey, and James Robins, “Double/debiased machine learning for treat- ment and structural parameters,” The Econometrics Journal 21, C1–C68 (2018), https://onlinelibrary.wiley.com/doi/...

2018 doi
[20]

Marginal structural models to estimate the joint causal eﬀect of nonrandomized treatmen ts,

Miguel A Hern´ an, Babette Brumback, and James M Robins, “Marginal structural models to estimate the joint causal eﬀect of nonrandomized treatmen ts,” Journal of the American Statistical Association 96, 440–448 (2001), https://doi.org/10.1198/016214501753 168154

2001 doi
[21]

Mark J Van der Laan and Sherri Rose, Targeted learning in data science (Springer, 2018)

2018
[22]

Alain Berlinet and Christine Thomas-Agnan, Reproducing kernel Hilbert spaces in probability and statistics (Springer Science & Business Media, 2011)

2011
[23]

An introduction to double/debiased ma chine learning,

Achim Ahrens, Victor Chernozhukov, Christian Hansen, Damian Kozbur, Mark Schaﬀer, and Thomas Wiemann, “An introduction to double/debiased ma chine learning,” (2025), arXiv:2504.08324 [econ.EM]

2025
[24]

Full likelihood inference s in the cox model: an empirical likelihood approach,

Jian-Jian Ren and Mai Zhou, “Full likelihood inference s in the cox model: an empirical likelihood approach,” Annals of the Institute of Statistic al Mathematics 63, 1005–1018 (2011)

2011
[25]

Marginal structuralmodels and causal inference in epidemiology,

James M. Robins, Miguel ´Angel Hern´ an, and Babette Brumback, “Marginal structuralmodels and causal inference in epidemiology,” Epidemiology 11 (2000)

2000
[26]

Kernel ba yes’ rule: Bayesian inference with positive deﬁnite kernels,

Kenji Fukumizu, Le Song, and Arthur Gretton, “Kernel ba yes’ rule: Bayesian inference with positive deﬁnite kernels,” Journal of Machine Learning Res earch 14, 3753–3783 (2013)

2013
[27]

Applica- tion of marginal structural models in pharmacoepidemiolog ic studies: a sys- tematic review,

Shibing Yang, Charles B. Eaton, Juan Lu, and Kate L. Lapa ne, “Applica- tion of marginal structural models in pharmacoepidemiolog ic studies: a sys- tematic review,” Pharmacoepidemiology and Drug Safety 23, 560–571 (2014), https://onlinelibrary.wiley.com/doi/pdf/10.1002/pds.3569

2014 doi
[28]

Functional analysis and semi-groups, 3rd printing of rev. ed. of 1957,

Einar Hille and Ralph S Phillips, “Functional analysis and semi-groups, 3rd printing of rev. ed. of 1957,” in Colloq. Publ , Vol. 31 (1974)

1974
[29]

4 (Springer, 2006)

Christopher M Bishop and Nasser M Nasrabadi, Pattern recognition and machine learning , Vol. 4 (Springer, 2006)

2006
[30]

History-adjusted marginal structural models and statically-optimal dynamic treatme nt regimens,

Mark J van der Laan, Maya L Petersen, and Marshall M Joﬀe, “ History-adjusted marginal structural models and statically-optimal dynamic treatme nt regimens,” The International Journal of Biostatistics 1 (2005)

2005
[31]

Theory of reproducing kernels,

Nachman Aronszajn, “Theory of reproducing kernels,” T ransactions of the American mathe- matical society 68, 337–404 (1950)

1950
[32]

Learning the kernel matrix with semideﬁnite progr amming,

Gert RG Lanckriet, Nello Cristianini, Peter Bartlett, Laurent El Ghaoui, and Michael I 41 Jordan, “Learning the kernel matrix with semideﬁnite progr amming,” Journal of Machine Learning Research 5, 27–72 (2004)

2004
[33]

Fast learning rate of multiple kernel learning: Trade-oﬀ between sparsity and smoothness,

Taiji Suzuki and Masashi Sugiyama, “Fast learning rate of multiple kernel learning: Trade-oﬀ between sparsity and smoothness,” The Annals of Statistics 41, 1381 – 1405 (2013)

2013
[34]

Sparsity in mult iple kernel learning,

Vladimir Koltchinskii and Ming Yuan, “Sparsity in mult iple kernel learning,” The Annals of Statistics 38, 3660–3695 (2010)

2010
[35]

Consistency of the group lasso and mult iple kernel learning

Francis R Bach, “Consistency of the group lasso and mult iple kernel learning.” Journal of Machine Learning Research 9 (2008)

2008
[36]

Hig h-dimensional additive modeling,

Lukas Meier, Sara Van de Geer, and Peter B¨ uhlmann, “Hig h-dimensional additive modeling,” The Annals of Statistics 37, 3779–3821 (2009)

2009
[37]

The asymptotic information in censored s urvival data,

David Oakes, “The asymptotic information in censored s urvival data,” Biometrika 64, 441–448 (1977)

1977
[38]

Regression models and life-tables,

D. R. Cox, “Regression models and life-tables,” Journa l of the Royal Statistical Society: Series B (Methodological) 34, 187–202 (1972), https://rss.onlinelibrary.wiley.com/doi/pdf/10.1111/j.2517-6161.1972.tb00899.x

1972
[39]

The eﬃciency of cox’s likelihood funct ion for censored data,

Bradley Efron, “The eﬃciency of cox’s likelihood funct ion for censored data,” Journal of the American Statistical Association 72, 557–565 (1977)

1977
[40]

Bayesian inference for cox proportional hazard models with partial likelihoods, nonl inear covariate eﬀects and correlated observations,

Ziang Zhang, Alex Stringer, Patrick Brown, and Jamie St aﬀord, “Bayesian inference for cox proportional hazard models with partial likelihoods, nonl inear covariate eﬀects and correlated observations,” Statistical Methods in Medical Research 32, 165–180 (2023), pMID: 36317395,...

2023 doi
[41]

On maximum likelihood estima tion of the semi-parametric cox model with time-varying covariates,

Mark Thackham and Jun Ma, “On maximum likelihood estima tion of the semi-parametric cox model with time-varying covariates,” Journal of Applie d Statistics 47, 1511–1528 (2020)

2020
[42]

D oubly robust estimation un- der a possibly misspeciﬁed marginal structural cox model,

Jiyu Luo, Denise Rava, Jelena Bradic, and Ronghui Xu, “D oubly robust estimation un- der a possibly misspeciﬁed marginal structural cox model,” Biometrika 112, asae065 (2024), https://academic.oup.com/biomet/article-pdf/112/1/asae065/60791527/asae065.pdf

2024
[43]

Identiﬁability of parameters in latent structure models with many observed variables,

Elizabeth S Allman, Catherine Matias, and John A Rhodes , “Identiﬁability of parameters in latent structure models with many observed variables,” (20 09)
[44]

Machine learn- ing approaches to evaluate heterogeneous treatment eﬀects i n randomized controlled trials: a 42 scoping review,

Kosuke Inoue, Motohiko Adomi, Orestis Efthimiou, Tosh iaki Komura, Kenji Omae, Akira On- ishi, Yusuke Tsutsumi, Tomoko Fujii, Naoki Kondo, and Toshi A. Furukawa, “Machine learn- ing approaches to evaluate heterogeneous treatment eﬀects i n randomized controlled trials: a 42 sc...

2024
[45]

On the ma ximum penalized likelihood ap- proach for proportional hazard models with right censored s urvival data,

Jun Ma, Stephane Heritier, and Serigne N. Lˆ o, “On the ma ximum penalized likelihood ap- proach for proportional hazard models with right censored s urvival data,” Computational Statistics and Data Analysis 74, 142–156 (2014)

2014
[46]

Nonparametri c ﬁnite translation hidden markov models and extensions,

Elisabeth Gassiat and Judith Rousseau, “Nonparametri c ﬁnite translation hidden markov models and extensions,” Bernoulli 22, 193–212 (2016)

2016
[47]

Parameter identiﬁability of discrete bayesian networks with hidden v ariables,

Elizabeth S Allman, John A Rhodes, Elena Stanghellini, and Marco Valtorta, “Parameter identiﬁability of discrete bayesian networks with hidden v ariables,” Journal of Causal Inference 3, 189–205 (2015)

2015
[48]

Inference in ﬁnit e state space non parametric hidden markov models and applications,

E. Gassiat, A. Cleynen, and S. Robin, “Inference in ﬁnit e state space non parametric hidden markov models and applications,” Statistics and Computing 26, 61–71 (2016)

2016
[49]

Estimating bayes fa ctors via thermodynamic integration and population mcmc,

Ben Calderhead and Mark Girolami, “Estimating bayes fa ctors via thermodynamic integration and population mcmc,” Computational Statistics and Data An alysis 53, 4028–4045 (2009)

2009
[50]

On structural and practical identiﬁability,

Franz-Georg Wieland, Adrian L. Hauber, Marcus Rosenbl att, Christian T¨ onsing, and Jens Timmer, “On structural and practical identiﬁability,” Cur rent Opinion in Systems Biology 25, 60–69 (2021)

2021
[51]

25 (Cambridge uni- versity press, 2009)

Sumio Watanabe, Algebraic geometry and statistical learning theory , Vol. 25 (Cambridge uni- versity press, 2009)

2009
[52]

Pierre Moral, Feynman-Kac formulae: genealogical and interacting parti cle systems with ap- plications (Springer, 2004)

2004
[53]

A widely applicable bayesian informa tion criterion,

Sumio Watanabe, “A widely applicable bayesian informa tion criterion,” Journal of Machine Learning Research 14, 867–897 (2013)

2013
[54]

A bayesian informat ion criterion for singular models,

Mathias Drton and Martyn Plummer, “A bayesian informat ion criterion for singular models,” Journal of the Royal Statistical Society Series B: Statisti cal Methodology 79, 323–380 (2017)

2017
[55]

Kernel independent compone nt analysis,

F.R. Bach and M.I. Jordan, “Kernel independent compone nt analysis,” Journal of Machine 43 Learning Research (2003)

2003
[56]

, An introduction to sequential Monte Carlo , Vol

Nicolas Chopin, Omiros Papaspiliopoulos, et al. , An introduction to sequential Monte Carlo , Vol. 4 (Springer, 2020)

2020
[57]

Probability asymptotics: notes on not ation,

Svante Janson, “Probability asymptotics: notes on not ation,” (2011), arXiv:1108.3924 [math.PR]

2011 arXiv
[58]

S tatistical consistency of kernel canonical correlation analysis,

Kenji Fukumizu, Francis R. Bach, and Arthur Gretton, “S tatistical consistency of kernel canonical correlation analysis,” Journal of Machine Learn ing Research 8, 361–383 (2007)

2007
[59]

for multiple ways of density-ratio estimation based on the kerne l method). ForH ∗ θf and H ∗ f f of the latent-variable model, we cannot represent ρk as a density ratio, but it is still plausible enough that H ∗ θkf and the functions in the range of H ∗ f f are the integral o...

2022
[60]

On the limited memory bfgs method for large scale opti- mization,

Dong C Liu and Jorge Nocedal, “On the limited memory bfgs method for large scale opti- mization,” Mathematical Programming 45, 503–528 (1989)

1989
[61]

2 (MIT press Cambridge, MA, 2006)

Christopher KI Williams and Carl Edward Rasmussen, Gaussian processes for machine learn- ing, Vol. 2 (MIT press Cambridge, MA, 2006)

2006
[62]

Theoretical analysis of density ratio estimation,

Takafumi Kanamori, Taiji Suzuki, and Masashi Sugiyama , “Theoretical analysis of density ratio estimation,” IEICE transactions on fundamentals of e lectronics, communications and computer sciences 93, 787–798 (2010). 44 0 100 200 300 400 500 600 700 800 900 1000 10 100 1000 -4...

2010

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Alternative analysis methods for time to event endpoi nts under nonproportional haz- ards: A comparative analysis,

Ray S. Lin, Ji Lin, Satrajit Roychoudhury, Keaven M. Ande rson, Tianle Hu, Bo Huang, Larry F Leon, Jason J.Z. Liao, Rong Liu, Xiaodong Luo, Pralay Mukhopadhyay, Rui Qin, Kay Tatsuoka, Xuejing Wang, Yang Wang, Jian Zhu, Tai-Tsang C hen, and Renee Iacona and, “Alternative analysis methods for time to event endpoi nts under nonproportional haz- ards: A compa...

arXiv 2020

[2] [2]

The optimization was stopped, if the ℓ2-norm of the gradient gets smaller than ǫstop = 1.0 × 10−2

We solve the above optimization problem, by applying limited- memory BFGS algorithm [56] with backtracking and line search stoppe d by the Armijo condition. The optimization was stopped, if the ℓ2-norm of the gradient gets smaller than ǫstop = 1.0 × 10−2. For ML estimation, one can approximate the BME by using the Laplace approximation. Assigning the Hess...

work page

[3] [3]

Here we have used the fact ‖(H ∗ f f + ζn)−1H ∗ f f‖ ≤ 1. 31 VIII. GRADIENT FUNCTIONAL AND HESSIAN OPERA TOR In this section, we provide the gradient functional and the Hessian operator of the sam- plewise negative log-likelihood: ℓ(D;θ,f ) = −I[0,C](T )(θ′AT +f (XT )) + ∫ T ∧C 0 exp (θ′At +f (Xt))dt. (60) Calculating the Gateaux derivative 1 ǫ (ℓ(D;θ,f +...

work page

[4] [4]

Does cox analysis of a randomized survival study yield a causal treatment eﬀect?

Odd O. Aalen, Richard J. Cook, and Kjetil Røysland, “Does cox analysis of a randomized survival study yield a causal treatment eﬀect?” Lifetime Dat a Analysis 21, 579–593 (2015)

work page 2015

[5] [5]

The hazards of per iod speciﬁc and weighted hazard ratios,

Jonathan W. Bartlett, Tim P. Morris, Mats J. Stensrud, Rh ian M. Daniel, Stijn K. Vansteelandt, and Carl-Fredrik Burman, “The hazards of per iod speciﬁc and weighted hazard ratios,” Statistics in Biopharmaceutical Research 12, 518–519 (2020), https://doi.org/10.1080/19466315.2020.1755722

arXiv 2020

[6] [6]

The hazards of hazard ratios,

Miguel A. Hern´ an, “The hazards of hazard ratios,” Epide miology 21 (2010)

work page 2010

[7] [7]

Intention-to-trea t comparisons in randomized trials,

Ross L Prentice and Aaron K Aragaki, “Intention-to-trea t comparisons in randomized trials,” Statistical Science 37, 380–393 (2022)

work page 2022

[8] [8]

Subtleties in the inter- pretation of hazard contrasts,

Torben Martinussen, Stijn Vansteelandt, and Per Kragh A ndersen, “Subtleties in the inter- pretation of hazard contrasts,” Lifetime Data Analysis 26, 833–855 (2020)

work page 2020