Pith. sign in

REVIEW 2 major objections 5 minor 30 references

Causal treatment effects on cure rates and survivor survival can be identified and estimated by writing the mixture cure model likelihood as a mixture over latent principal strata.

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 · deepseek-v4-flash

2026-08-01 13:34 UTC pith:Q7VMCFED

load-bearing objection Solid Bayesian principal stratification for cure models with a genuinely new estimand, but the causal claims rest on conditional monotonicity, which the paper's own simulations show is fragile—so read the robustness arguments carefully. the 2 major comments →

arxiv 2607.19050 v1 pith:Q7VMCFED submitted 2026-07-21 stat.ME

A Bayesian Approach to Causal Cure Models

Emma Torrini , Ma\"ilis Amico , Nicolas Molinari , Cl\'ement Berenfeld This is my paper
classification stat.ME MSC 62N0162F1562N02
keywords Bayesian inferencecausal inferencecure modelsprincipal stratificationsurvival analysismixture cure modelsidentifiabilitytime-to-event
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.

The paper argues that when a fraction of patients will never experience the event (are 'cured'), causal questions about treatment—does it prevent the event, and does it delay it among those who remain susceptible—can be answered by viewing the classical mixture cure likelihood as a mixture over four latent principal strata defined by potential cure status. This decomposition turns the mixture cure model's two ingredients, the uncured probability and the uncured survival function, into stratum-specific causal quantities. The contribution is a Bayesian estimation framework that targets effects on the cure fraction, on an always-uncured stratum, and on a new union-of-strata estimand for the non-always-cured population. Identification is shown under monotonicity plus either a substitution-relevance condition or a Cox-model genericity condition, the latter avoiding the need to locate a special covariate. A worked application shows the framework is usable in practice and can fold in auxiliary cure-status information.

Core claim

The paper's central claim is that the causal cure model likelihood—the product, over treatment arms, of mixture cure likelihoods—decomposes exactly into the four principal strata CC, CU, UC, UU. The uncured probability under each arm equals the sum of two stratum probabilities, and the uncured survival function equals a π-weighted average of two stratum-specific survival functions (Equations 5–8). This gives a model-based route to principal-stratification estimands: the Bayesian posterior is defined directly on stratum probabilities and stratum-specific survival curves. Under Assumptions 1–6 plus either Assumption 7.A or 7.B, the strata proportions and stratum-specific survivor functions are

What carries the argument

The central object is the principal-stratified likelihood, built by writing each treatment arm's mixture cure likelihood in terms of four stratum probabilities π_g(X) and stratum-specific survival functions S_g^(z)(t|X). The link is the decomposition in Equations (5)–(8): p^(1)=π_UC+π_UU, p^(0)=π_CU+π_UU, and each S_U^(z) is a weighted average of two stratum survivals. This object carries the argument because it turns a classical non-causal model into a causal one while preserving the observed-data likelihood. Identification is then delivered by Proposition 2 (monotonicity pins down π_UU as the minimum of the two uncured probabilities) and Proposition 3 (either a substitution variable or the

Load-bearing premise

The load-bearing premise is conditional monotonicity (Assumption 6): the always-uncured stratum probability equals the smaller of the two treatment-specific uncured probabilities, so that no patient can be 'harmed' in the sense of being uncured under the investigational treatment but cured under control within any covariate stratum; if this fails, the stratum proportions—and with them the causal cure and survival effects—are not identified.

What would settle it

Search numerically for two distinct Cox parameter sets (Λ_UU, β_UU, Λ_gz, β_gz) and stratum probabilities that produce identical observed mixture-cure survival S(z)(t|x) for all t and x; the proof of Proposition 3 asserts this is impossible under the stated genericity condition, so finding such a pair—especially with d=4 or less—would refute the identification claim.

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

If this is right

  • Estimates of the treatment effect on cure probability δ = π_CU − π_UC and on restricted mean survival time for the always-uncured stratum follow directly from the posterior of the stratum parameters.
  • A new estimand, the survival effect on the union of non-always-cured strata (¬CC), gives a clinically interpretable effect on the susceptible population, without dilution by the always-cured.
  • When a Cox survival model is adopted, the identification condition replaces the hard-to-verify substitution-relevance assumption, so the method can be applied without choosing a special covariate.
  • If censored patients can be clinically identified as cured (for example, discharged), that binary information enters the likelihood and narrows credible intervals.
  • Because the framework is Bayesian, the identifying assumptions (notably conditional monotonicity) can be relaxed or varied in sensitivity analysis within the same implementation.

Where Pith is reading between the lines

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

  • The ¬CC union estimand could serve as a summary endpoint in clinical trials where the always-cured fraction is large, since it isolates the treatment effect on the patients who are potentially going to fail.
  • The Cox-genericity identification suggests that other parametric survival families with identifiable parameterizations would yield analogous results, so the approach may generalize beyond the piecewise-exponential implementation used here.
  • A testable extension would compare estimates from this method against direct long-term follow-up classifications of cure status (when available) to validate the monotonicity assumption at the individual level.
  • If the assumptions are only partially credible, the Bayesian posterior still concentrates under partial identification; the accompanying sensitivity analyses thus become an integral part of the causal conclusion, not a post hoc check.

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

2 major / 5 minor

Summary. The paper develops a Bayesian principal-stratification framework for causal inference in time-to-event settings with a cured fraction. It defines four principal strata from the joint potential cure statuses, introduces a new estimand for the union of non-always-cured strata, and shows that the classical mixture cure model likelihood can be rewritten as a decomposition of principal-stratum-specific likelihood components (Theorem 1, Equations (5)–(8)). Identification of strata proportions and stratum-specific survival curves is analyzed under conditional monotonicity (Assumption 6) and either a substitution-relevance assumption (7.A) or a Cox-model genericity condition (7.B) (Propositions 1–3). The framework is implemented in Stan, evaluated in four simulation scenarios against Wang et al. (2024), and applied to the NIV AS randomized trial. The paper also incorporates auxiliary information on cure status (Proposition 4).

Significance. If the results hold, the paper makes a useful contribution to causal cure-model methodology. The likelihood decomposition provides a transparent link between classical mixture cure models and principal-stratification estimands, and the Bayesian formulation avoids the need to preselect a substitution variable, which is a practical advantage over existing frequentist approaches. The new ¬CC estimand is clinically interpretable, and the inclusion of auxiliary cure information is a sensible extension. The identification results are nontrivial and the proofs in Appendix A are mostly careful. The simulation study and the NIV AS application demonstrate the method in realistic settings, and the Stan implementation supports reproducibility. The main weakness is that the headline causal estimands rely on the untestable Assumption 6, and the paper's claim that this assumption can be relaxed is not backed by simulation evidence under violation.

major comments (2)
  1. [§4.3, Table 3] In Scenario 4, where Assumption 6 is violated, the proposed method (with ρ fixed at 1) shows substantial bias: 0.083 in π_UU, 0.087 in π_CC, and -0.448 in τ_RMST_UU. The text states that different parameterizations allow this assumption to be relaxed, and Appendix C reports an application-level sensitivity analysis (Model 2), but no simulation demonstrates that an unconstrained or ρ-prior version recovers the truth when Assumption 6 is violated. Because π_UU, and hence δ and the ¬CC union estimand, are identified only through Assumption 6 (Propositions 2–3 and Equation (12)), this is a load-bearing gap. Please add a simulation in which ρ is estimated under a prior (or a multinomial model is used) under a DGP resembling Scenario 4, reporting bias and interval coverage for the headline estimands.
  2. [Appendix A.1, Proposition 3 (Assumption 7.A)] The proof claims: 'for a given value of U_i = u, we can always find two values v1 and v2 such that q(u,v1) ≠ q(u,v2)'. Assumption 7.A, as stated, only asserts V_i ⊥̸⊥ B_i(z) | U_i, B_i(1−z)=1, which ensures dependence somewhere in the distribution of U but not necessarily at every u. The theorem as stated is stronger than the assumption. Either strengthen Assumption 7.A to require that the conditional dependence holds for (almost) every u in the support, or provide a continuity/approximation argument to cover all u. As written, this is a logical gap in one of the two identification routes.
minor comments (5)
  1. [§4.2–4.3, Table 3] The simulation evaluation reports only point estimates and empirical SEs across 100 replications. Please add posterior interval coverage or a calibration measure, and consider reporting Monte Carlo standard errors for the bias estimates. One hundred replications is adequate for rough comparisons but not for high-precision claims.
  2. [§4.2, Assumption 7.B] The simulation DGP uses four covariates (including the intercept), so the d > 4 condition in Assumption 7.B is never satisfied. If Assumption 7.B is intended as a genuinely alternative identification route, a simulation with five or more covariates would strengthen the paper's empirical claims.
  3. [§5.2, Appendix C] The statement that 'The conclusions are quite robust to the monotonicity assumption' is stronger than the sensitivity analysis suggests. In Model 2, δ has 95% CI (−0.05, 0.12) and τ_RMST_¬CC has 95% CI (−1.71, 6.24), so the 95% intervals cross zero for δ and are much wider. Please temper this wording or emphasize the posterior-probability interpretation.
  4. [§4.1] The parameterization π_UU = ρ min(p1,p0) + (1−ρ) p1 p0 is not guaranteed to be a valid probability for arbitrary ρ ∈ (−1,1) and arbitrary p1,p0; for some values it can be negative. This does not affect the reported simulations, but the permissible range of ρ for given (p1,p0) should be stated or the parameterization should be constrained.
  5. [§3, after Proposition 4] There is a typo: 'uknown' should be 'unknown'. Also in the proof of Equations (9)–(10), 'nineth' should be 'ninth'.

Circularity Check

0 steps flagged

No significant circularity; the likelihood derivation and identification results are self-contained conditional on explicit assumptions.

full rationale

The paper's derivation chain is not circular. Theorem 1 derives the observed-data likelihood directly from the joint distribution of (Y, Delta, Z, X) under Assumptions 1, 2, and 4, and Proposition 4 extends it to incorporate auxiliary cure-status information. Equations (5)-(8) are algebraic identities linking the mixture cure model components p^(z) and S_U^(z) to principal-stratum quantities, and the causal cure model likelihood L_n is obtained by substituting these identities into the Theorem 1 likelihood; this is a transparent reparameterization, not a fit of the target estimands. The identification results in Propositions 1-3 are conditional on explicit assumptions: Assumption 6 supplies the identifying restriction pi_UU = min(p^(1), p^(0)), and Proposition 3 adds either substitution relevance (7.A, attributed to external work of Wang et al. 2024) or a Cox/genericity condition (7.B) with a proof provided in Appendix A. These are assumptions, not conclusions smuggled from the estimands; the paper even shows in Simulation Scenario 4 that when Assumption 6 is violated the estimator is biased, which confirms the results are not forced by construction. The Bayesian estimands delta, tau_UU, and tau_¬CC are posterior functions of the model parameters computed via Equations (9)-(10), rather than fitted constants renamed as predictions. Benchmarks use external Wang et al. (2024) code and simulation truths, and the only self-citations are literature reviews (Amico and Van Keilegom 2018; Voinot et al. 2025) that are not load-bearing for the central derivation. No circular step is present.

Axiom & Free-Parameter Ledger

0 free parameters · 9 axioms · 0 invented entities

No ad hoc fitted constants are introduced; all model parameters (logistic/multinomial coefficients, piecewise-exponential hazards) are estimated via posterior. The hand-chosen priors and number of time intervals are standard modeling choices. The central assumptions are listed in axioms.

axioms (9)
  • domain assumption SUTVA (Assumption 1): observed outcome equals potential outcome under assigned treatment; no interference
    Standard in causal inference; needed to link observed and potential outcomes in likelihood.
  • domain assumption Ignorability (Assumption 2): treatment assignment independent of potential outcomes given covariates
    Justified by randomization in the application; needed for identifying causal survival curves.
  • domain assumption Positivity of treatment (Assumption 3): 0 < P(Z=1|X) < 1
    Ensures both treatment groups are possible for every covariate profile.
  • domain assumption Conditionally independent censoring (Assumption 4): C_i(z) ⟂ T_i(z) | X, Z
    Allows factoring out censoring from the likelihood.
  • domain assumption Positivity of censoring (Assumption 5): P(C>t|X)>0 up to t*
    Ensures tail of survival is observed; needed for cure fraction identifiability.
  • domain assumption Conditional monotonicity (Assumption 6): π_UU = min(p^(1), p^(0))
    Identifies strata proportions; violated in Scenario 4 where both methods are biased.
  • domain assumption Substitution relevance (Assumption 7.A): exists V_i ⊥ T_i(z) | Z, G, U and V_i ⊥̸ B_i(z) | U, B_i(1-z)=1
    One of two routes to identify stratum-specific survival; hard to verify.
  • ad hoc to paper Cox model + genericity (Assumption 7.B): survival follows Cox; either supports of π_g and π_UU differ or their ratio has full-rank Hessian (d>4)
    Introduced specifically to provide an alternative identification route; the Hessian genericity condition is tailored to the proof and simulations have d=4, so it is not tested.
  • standard math Bayes' theorem and law of total probability used in proofs
    Standard probabilistic tools used to derive likelihood and identifiability equations.

pith-pipeline@v1.3.0-alltime-deepseek · 25965 in / 15599 out tokens · 153262 ms · 2026-08-01T13:34:30.902705+00:00 · methodology

0 comments
read the original abstract

Time-to-event data often include individuals who will never experience the failure event, and are therefore considered cured. In such settings, frequently encountered in clinical research, a substantial proportion of patients may remain event-free throughout the observation period, leading to the appearance of a survival plateau that is commonly interpreted as evidence of a cured fraction. Analysing such data requires inference on the cure fraction and on the survival function for the uncured subpopulation, tasks which are traditionally achieved with mixture cure models. However, assessing the causal effect of a treatment on these quantities is non-trivial. We consider principal stratification causal estimands, which have been proposed to evaluate effects on the cure fraction and on the survival for an always-uncured stratum. We additionally introduce a novel estimand, which considers the causal effect on the survival for a non-always-cured union of strata. We frame the problem from a Bayesian model-based perspective, which provides a flexible and unified estimation strategy while maintaining a direct link with classical mixture cure model quantities. The reliability of the proposed approach is validated through simulations, demonstrating competitive and robust performance relative to existing methods. Finally, we illustrate its practical usefulness through an application to a randomized trial comparing non-invasive ventilation with standard oxygen therapy in patients with hypoxemic respiratory failure following abdominal surgery.

Figures

Figures reproduced from arXiv: 2607.19050 by Cl\'ement Berenfeld, Emma Torrini, Ma\"ilis Amico, Nicolas Molinari.

Figure 1
Figure 1. Figure 1: Simulation results for the four scenarios considered. Boxplots and violin plots represent [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Kaplan-Meier curves by treatment group in the NIVAS study. [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Posterior summaries of the relevant quantities. Panel (a): posterior distributions of [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Model 1. Posterior summaries of the marginal survival curves. Panel (a): posterior [PITH_FULL_IMAGE:figures/full_fig_p030_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Model 2. Posterior summaries of the relevant quantities. Panel (a): posterior distribu [PITH_FULL_IMAGE:figures/full_fig_p031_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Model 3. Posterior summaries of the relevant quantities. Panel (a): posterior distribu [PITH_FULL_IMAGE:figures/full_fig_p032_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Model 4. Posterior summaries of the relevant quantities. Panel (a): posterior distribu [PITH_FULL_IMAGE:figures/full_fig_p033_7.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

30 extracted references · 1 linked inside Pith

  1. [1]

    , author=

    Estimating causal effects of treatments in randomized and nonrandomized studies. , author=. Journal of educational Psychology , volume=. 1974 , publisher=

    1974

  2. [2]

    Biometrics , volume=

    Principal stratification in causal inference , author=. Biometrics , volume=. 2002 , publisher=

    2002

  3. [3]

    Lindley, Dennis Victor , year=

  4. [4]

    Biometrical Journal , volume=

    Ballerini, Veronica and Bornkamp, Bj. Biometrical Journal , volume=. 2025 , publisher=

    2025

  5. [5]

    Bayesian analysis , year=

    Assessing causal effects in the presence of treatment switching through principal stratification , author=. Bayesian analysis , year=

  6. [6]

    Biometrics , volume=

    Principal stratification analysis of noncompliance with time-to-event outcomes , author=. Biometrics , volume=. 2024 , publisher=

    2024

  7. [7]

    Carpenter, Bob and Gelman, Andrew and Hoffman, Matthew D and Lee, Daniel and Goodrich, Ben and Betancourt, Michael and Brubaker, Marcus and Guo, Jiqiang and Li, Peter and Riddell, Allen , journal=

  8. [8]

    Biometrics , volume=

    Causal inference for time-to-event data with a cured subpopulation , author=. Biometrics , volume=. 2024 , publisher=

    2024

  9. [9]

    2023 , publisher=

    Modelling survival data in medical research , author=. 2023 , publisher=

    2023

  10. [10]

    arXiv preprint arXiv:2501.05836 , year=

    Voinot, Charlotte and Berenfeld, Cl. arXiv preprint arXiv:2501.05836 , year=

    arXiv

  11. [11]

    Journal of the Royal Statistical Society

    Maximum likelihood estimates of the proportion of patients cured by cancer therapy , author=. Journal of the Royal Statistical Society. Series B (Methodological) , volume=. 1949 , publisher=

    1949

  12. [12]

    Journal of the American Statistical Association , volume=

    Survival curve for cancer patients following treatment , author=. Journal of the American Statistical Association , volume=. 1952 , publisher=

    1952

  13. [13]

    1996 , publisher=

    Stochastic models of tumor latency and their biostatistical applications , author=. 1996 , publisher=

    1996

  14. [14]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume=

    Regression models and life-tables , author=. Journal of the Royal Statistical Society: Series B (Methodological) , volume=. 1972 , publisher=

    1972

  15. [15]

    Annual Review of Statistics and Its Application , volume=

    Cure models in survival analysis , author=. Annual Review of Statistics and Its Application , volume=. 2018 , publisher=

    2018

  16. [16]

    Communications in Statistics-Theory and Methods , volume=

    Estimating the causal effects in randomized trials for survival data with a cure fraction and non compliance , author=. Communications in Statistics-Theory and Methods , volume=. 2017 , publisher=

    2017

  17. [17]

    Biometrical Journal , volume=

    Estimating causal effects in observational studies for survival data with a cure fraction using propensity score adjustment , author=. Biometrical Journal , volume=. 2023 , publisher=

    2023

  18. [18]

    Biometrical Journal , volume=

    Survivor average causal effects for continuous time: a principal stratification approach to causal inference with semicompeting risks , author=. Biometrical Journal , volume=. 2025 , publisher=

    2025

  19. [19]

    Biostatistics , volume=

    Xu, Yanxun and Scharfstein, Daniel and M. Biostatistics , volume=. 2022 , publisher=

    2022

  20. [20]

    Biostatistics , volume=

    Causal inference for semi-competing risks data , author=. Biostatistics , volume=. 2022 , publisher=

    2022

  21. [21]

    The Annals of Applied Statistics , volume=

    Multiply robust estimation for causal survival analysis with treatment noncompliance , author=. The Annals of Applied Statistics , volume=. 2026 , publisher=

    2026

  22. [22]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Principal stratification analysis using principal scores , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2017 , publisher=

    2017

  23. [23]

    Journal Of the Amercian Medical Association , volume=

    Effect of noninvasive ventilation on tracheal reintubation among patients with hypoxemic respiratory failure following abdominal surgery: a randomized clinical trial , author=. Journal Of the Amercian Medical Association , volume=. 2016 , publisher=

    2016

  24. [24]

    Biometrics , volume=

    The use of mixture models for the analysis of survival data with long-term survivors , author=. Biometrics , volume=. 1982 , publisher=

    1982

  25. [25]

    BMC medical research methodology , volume=

    Restricted mean survival time: an alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome , author=. BMC medical research methodology , volume=. 2013 , publisher=

    2013

  26. [26]

    The annals of statistics , pages=

    Bayesian inference for causal effects in randomized experiments with noncompliance , author=. The annals of statistics , pages=. 1997 , publisher=

    1997

  27. [27]

    arXiv preprint arXiv:2606.11405 , year=

    Bayesian Causal Machine Learning for Cure Models , author=. arXiv preprint arXiv:2606.11405 , year=

    Pith/arXiv arXiv

  28. [28]

    Journal of the American statistical Association , volume=

    Identification of causal effects using instrumental variables , author=. Journal of the American statistical Association , volume=. 1996 , publisher=

    1996

  29. [29]

    Journal of Statistical Software , year =

    Stef. Journal of Statistical Software , year =

  30. [30]

    2010 , publisher=

    Zhou, Xiang and Reiter, Jerome P , journal=. 2010 , publisher=

    2010