Pith. sign in

REVIEW 3 major objections 4 minor 20 references

The paper claims that posterior inference for interval-censored continuous-time multistate transition models can be amortized: after one offline training phase on simulated datasets, a neural density estimator supplies posterior means, cred

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

2026-07-31 23:49 UTC pith:NYAMKVIS

load-bearing objection Useful, honest method paper, but the consistency theorems do not cover the main simulations because the sufficiency lemma requires homogeneous covariate strata. the 3 major comments →

arxiv 2607.23294 v1 pith:NYAMKVIS submitted 2026-07-25 stat.ME stat.AP

AIM: Amortized Inference for Multistate Transition Models

Yuxi Zhu , Rui Zhang This is my paper
classification stat.ME stat.AP MSC 62F1562M0562-08
keywords multistate transition modelamortized inferenceinterval censoringcontinuous-time Markov chainneural posterior estimationmixture density networksummary statisticsposterior consistency
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.

AIM is an amortized Bayesian inference framework for interval-censored continuous-time multistate transition models. Instead of running a separate likelihood-based fit for every new dataset, AIM trains a neural density estimator once on simulated parameter–dataset pairs, then conditions on a fixed-dimensional summary of the observed panel to produce posterior inference instantly. The paper claims that this summary—interval-specific transition counts, state occupancy proportions, and covariate-stratum proportions—is sufficient for the panel likelihood, that the population summaries identify the model parameters, and that the induced posterior, its neural approximation, and the posterior mean are consistent. Simulations show point estimates close to truth, near-nominal interval coverage, and median online speedups of roughly 154- to 2308-fold over repeated likelihood-based fitting. A real heart-transplant panel application yields estimates concordant with a standard likelihood-based analysis, supporting calibration.

Core claim

AIM replaces dataset-specific likelihood optimization with a learned neural conditional density. For a prespecified model class—fixed transition graph, visit schedule, and covariate strata—it summarizes each panel by interval transition counts, state occupancies, and stratum proportions, then trains a mixture density network to approximate the posterior of the transition-intensity parameters given that summary. The paper proves that these transition counts are sufficient under a discretized-covariate likelihood, that the population summary map is injective under reachability and full-rank design conditions, and that the summary posterior, its neural approximation, and the resulting posterior

What carries the argument

The central object is the fixed-dimensional summary vector S_N, composed of interval-specific transition counts n_rs^{(k,g)}, state-occupancy proportions π_r^{(g)}(t_k), and stratum proportions ω_g. This summary is matched to the panel likelihood, so the observed likelihood factorizes through the counts (Lemma 1, sufficiency), the population version of the summary identifies the generator and regression parameters (Lemma 2, identifiability), and posterior concentration follows (Theorem 1). The neural component is a mixture density network q_φ(θ|x) trained by minimizing the negative conditional log-likelihood over prior-predictive simulations; Theorems 2–4 show that a sufficiently accurate MD

Load-bearing premise

The load-bearing premise is that within each covariate stratum all subjects share the representative covariate value, so interval-specific transition counts are sufficient for the panel likelihood; in the covariate simulations with continuous covariates split at the median, this premise is violated and the stated consistency guarantees do not strictly apply to the evaluated configuration.

What would settle it

Generate a test dataset from the covariate model with continuous covariates, compute AIM's posterior from its summary, and compute the exact panel-likelihood posterior by MCMC. If the two distributions differ systematically and the discrepancy does not shrink as the sample size grows, then the sufficiency-based consistency chain does not hold for the implemented procedure. Alternatively, if AIM's posterior mean bias persists at large N while the likelihood-based estimator's bias vanishes, the consistency theorem is not covering the evaluated setup.

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

If this is right

  • After one offline training run, any compatible new panel can be analyzed without likelihood optimization; online cost is summary construction plus one forward pass through the network.
  • Repeated analyses across centers, time periods, subgroups, or bootstrap resamples become cheap, which is the intended operating regime for amortized inference.
  • The sufficiency–identifiability–consistency chain gives theoretical backing to amortized posterior inference for this class of multistate models.
  • The real-data application shows that the amortized posterior can reproduce established likelihood-based estimates, supporting calibration on non-simulated data.
  • The speedups (median 154- to 2308-fold) quantify the practical gain when the same model class is applied to many datasets.

Where Pith is reading between the lines

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

  • The covariate simulations use continuous N(0,1) covariates split at medians, but Lemma 1 assumes all subjects in a stratum share the representative covariate value z̄_g; that premise is violated, so the sufficiency and consistency proofs do not formally cover the evaluated setup, and any bias from the non-sufficient summary is unquantified.
  • If covariate discretization is coarse relative to effect sizes, AIM's posterior may systematically miss covariate effects; a practical safeguard would be to compare AIM against exact likelihood-based inference on a small validation set before deployment.
  • The mild undercoverage seen in several simulation cells suggests that calibration adjustments—for example ensemble methods or normalizing flows—could tighten uncertainty quantification in edge-of-prior regimes.
  • The same amortization template could extend to irregular observation times by encoding visit gaps or to learned encoders that recover information lost by fixed interval counts.

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

3 major / 4 minor

Summary. The manuscript proposes AIM, an amortized Bayesian inference framework for interval-censored continuous-time multistate transition models. The method trains a mixture density network on simulated parameter-dataset pairs, conditioning on a fixed-dimensional summary consisting of interval-specific transition counts, state occupancies, and covariate-stratum proportions. The paper proves sufficiency of the count summaries under a shared-covariate-within-stratum assumption, establishes identifiability of population summaries, and derives consistency results for the summary posterior and for the neural approximation. Simulation studies compare AIM with maximum-likelihood fits from the msm package in three scenarios, and a real-data CAV analysis is used as a calibration check. The claimed contribution is reusable posterior inference whose online cost is one forward pass through a neural network.

Significance. The paper addresses a real computational bottleneck in applied multistate modeling: repeated likelihood-based fitting for each new dataset. The amortization idea, training once on simulations and then performing posterior inference by a network feed-forward, is well motivated and potentially useful for registry or multicenter applications. The authors provide a deliberately model-facing summary representation tied to the panel likelihood, which is a strength. The simulation studies are extensive, with multiple scenarios, sensitivity analyses, and a real-data concordance check against a gold-standard R package. The reported speedups are large and the point estimates are broadly accurate. However, the theoretical guarantees are currently conditional on assumptions that the main simulation evaluations do not satisfy, and the neural-consistency theorem is an inheritance result under an unproved KL-convergence premise. These issues undermine the abstract's stronger claims of a fully consistent neural posterior estimator and need to be resolved before the contribution is fully established.

major comments (3)
  1. [Section 2.2 (Lemma 1), Section 4.1, Theorem 1 (C3)] The sufficiency lemma and all downstream consistency results are proved only under the assumption that within each covariate stratum all subjects share the representative covariate value zbar_g, so that all subjects in a stratum have the same generator Q_g(theta) and the same interval transition matrices. In Scenarios B and C the covariates are continuous N(0,1) variables and the strata are median splits, so subjects within a stratum have heterogeneous transition matrices P_k(z_i;theta). The interval transition counts are then not sufficient for the actual panel likelihood; the factorization in Appendix A does not hold with the likelihood that generated the evaluated data. Consequently Theorem 1, and the neural consistency and posterior mean consistency that depend on it, do not apply to the main simulation settings. The Discussion acknowledges that the theory assumes discretized strata,
  2. [Section 3.3-3.4 (Theorems 2 and 3)] Theorem 3 assumes, rather than proves, that KL{Pi(.|S_N) || q_hat(.|S_N)} converges to zero in probability along the observed summary sequence, and then concludes concentration of q_hat. This is nearly tautological: it shows that the neural posterior inherits consistency from the summary posterior conditional on the very convergence that is the main object of a 'neural consistency' claim. The paper states in the Abstract that it proves 'consistency of the induced summary posterior and its neural approximation,' but Theorem 3 does not provide such a proof. Moreover, Theorem 2 establishes expected KL convergence under the prior predictive distribution p(S), not under the data-generating distribution P_{theta_0}(S_N) of the observed summary; the paper does not bridge this distributional gap or show that expected convergence implies pointwise convergence at S_N. The KL convergence in Theorem
  3. [Section 4.4, Tables 1-2] The abstract and Section 4.4 claim 'reliable uncertainty quantification' and coverage 'generally near the nominal 95% level,' but the reported AIM coverage values include 0.85 and 0.87 for several parameters in the covariate and four-state scenarios. Some deviation is expected with 100 replications, but a systematic pattern of undercoverage in the more complex scenarios is consistent with the concern that the summary is not sufficient and the theoretical calibration guarantees do not apply. The manuscript should explicitly discuss these coverage shortfalls as approximation error, report Monte Carlo standard errors for the coverage estimates, and ideally provide a calibration diagnostic as suggested in the Discussion.
minor comments (4)
  1. [Section 2.3 vs Section 3.3] Notation for the MDN parameters is inconsistent: phi is used in Section 2.3 and psi in Section 3.3. Using the same symbol throughout would improve readability.
  2. [Theorem 1 proof] The transition probability p(k,g)_rs(theta) is used in the proof but is not defined in the theorem statement. Define it as (P_k^{(g)}(theta))_{rs} for clarity.
  3. [Theorem 1, condition (C6)] The notation n(k,g)_{r.}/N is introduced without definition. Specify that it is the empirical proportion of subjects in stratum g and interval k whose origin state is r.
  4. [Discussion] The Discussion correctly lists the fixed-schedule and discretized-covariate assumptions as limitations, but this appears only after the formal theorems and the main simulations violate the latter assumption. The body of the paper would benefit from explicitly noting at the start of Section 3 that the theorems apply to discretized covariates, and that the continuous-covariate simulations are an extrapolation.

Circularity Check

0 steps flagged

No circular derivation: AIM fits a well-defined summary posterior from simulations and does not pass fitted inputs off as predictions; the main caveats are unproved premises and a covariate-discretization mismatch, not circularity.

full rationale

The paper's derivation chain is not circular. AIM trains an MDN against simulated (θ, S) pairs by minimizing a proper scoring rule, so the learned network targets a well-defined summary posterior π(θ | S_N) ∝ p(S_N | θ)π(θ); no fitted constant is later relabeled as a prediction. Lemma 1 proves sufficiency of interval transition counts through the explicit factorization L_g(θ; X) = h_g(X) g_g(T_g(X), θ), which is a standard factorization, not a self-fulfilling identity. Lemma 2 and Theorem 1 are ordinary KL-based consistency arguments from an identifiable product-multinomial likelihood. The main theoretical weakness is that Theorem 3 ('Neural posterior consistency') assumes 'KL{Π(· | S_N) ∥ q_hat(· | S_N)} → 0' and then derives posterior concentration; the stronger pointwise KL-convergence premise is not established by Theorem 2, which only gives expected KL convergence over the summary distribution. This is an unproved premise / gap, not the conclusion being used as the sole input. The simulation mismatch — continuous covariates in Scenarios B/C with median splits while Lemma 1 requires within-stratum shared z̄_g — means the consistency theorems do not cover the evaluated setting; the Discussion concedes 'the present theory and implementation assume a common observation schedule and discretized covariate strata.' This is a scope or correctness limitation, not circularity. Self-citations (Zhu et al. 2021, 2024, 2025) are background references and are not load-bearing. Hence no circular step is identified.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The method rests on standard CTMC and Bayesian machinery; no invented entities are introduced. The largest costs are the within-stratum constant-covariate assumption, which is violated in the continuous-covariate simulations, and unstated prior/representative-value choices that block exact reproduction.

free parameters (4)
  • MDN architecture hyperparameters (H, hidden units, learning rate) = H=12, 2×256 units, lr=1e-3
    Hand-selected; the approximation quality of q_phi depends on them; no architecture ablation is reported.
  • Offline training sample size M = 50,000 (sensitivity 10k–100k)
    Finite M controls neural approximation error; consistency requires M→∞.
  • Covariate stratum representative value z̄_g = not stated (median split implied)
    The summary's sufficiency relies on within-stratum constant z̄_g; the actual choice affects the target posterior and is not specified.
  • Prior distribution π(θ)
    A prior is an input to the Bayesian method, but the paper never states the prior used in simulations or the CAV application; posterior results depend on it.
axioms (6)
  • domain assumption The process is a time-homogeneous continuous-time Markov chain with proportional-intensity covariate effects.
    Defines the model class; AIM is trained for this class only (Section 2.1).
  • domain assumption Within each stratum, all subjects share the representative covariate value z̄_g.
    Needed for Lemma 1 sufficiency; false in the continuous-covariate simulations (Section 2.2 vs Section 4.1).
  • domain assumption Initial state distribution is free of θ and treated as nuisance.
    Used in Lemma 1 and Theorem 1.
  • domain assumption Generator identifiability from interval transition matrices (I2) holds for the transition graphs considered.
    Assumed for acyclic progressive and competing-pathway structures; not proved for general graphs (Section 3.1).
  • standard math Parameter space is compact and the prior is continuous and positive near θ0.
    Conditions C1–C2 for summary posterior consistency.
  • ad hoc to paper KL{Π(·|S_N) || q_hat(·|S_N)} → 0 in probability along the observed summary sequence.
    This is the premise of Theorem 3; it is assumed rather than established by Theorem 2.

pith-pipeline@v1.3.0-alltime-deepseek · 14047 in / 17022 out tokens · 155200 ms · 2026-07-31T23:49:23.713413+00:00 · methodology

0 comments
read the original abstract

Interval censored continuous time multistate transition models (MSTMs) are widely used to characterize disease progression and other dynamic processes. Existing inference procedures are predominantly likelihood-based and require a separate model fit for each dataset, involving repeated evaluation of transition probability matrices. This repeated computation can become burdensome when the same model is applied across centers, time periods, or newly collected cohorts. We propose AIM, an amortized Bayesian inference framework for interval censored continuous time MSTMs. AIM learns the relationship between informative summary statistics and model parameters from simulated datasets generated during an offline training stage. Once trained for a prespecified model class, AIM provides posterior inference for new datasets without repeated likelihood optimization. Under fixed observation schedules and discretized covariate strata, we show that interval-specific transition counts are sufficient for the observed panel likelihood and establish identifiability of the corresponding population summaries. We further prove consistency of the induced summary posterior and its neural approximation, yielding a consistent posterior mean estimator. Simulation studies under progressive and competing-pathway MSTMs demonstrate accurate point estimation and reliable uncertainty quantification. AIM completed online inference in milliseconds, with median speedups ranging from approximately 154-fold to 2308-fold relative to repeated likelihood-based estimation. These results establish AIM as a scalable and theoretically grounded framework for reusable Bayesian inference in interval censored MSTMs.

Figures

Figures reproduced from arXiv: 2607.23294 by Rui Zhang, Yuxi Zhu.

Figure 1
Figure 1. Figure 1: Simulation performance of AIM and the likelihood-based [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Online inference time for AIM and the likelihood-based [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity analysis with respect to the observed sample size and the number of simulated [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

20 extracted references · 1 linked inside Pith

  1. [1]

    , title =

    Wood, Simon N. , title =. Nature , volume =. 2010 , pages =

    2010

  2. [2]

    Inferring coalescence times from

    Tavar\'. Inferring coalescence times from. Genetics , volume =. 1997 , number =

    1997

  3. [3]

    and Seielstad, Mark T

    Pritchard, Jonathan K. and Seielstad, Mark T. and Perez-Lezaun, Anna and Feldman, Marcus W. , journal=. Population growth of human

  4. [4]

    and Zhang, Wenyang and Balding, David J

    Beaumont, Mark A. and Zhang, Wenyang and Balding, David J. , journal=. Approximate

  5. [5]

    Journal of Machine Learning for Modeling and Computing , issn =

    Rui Zhang and Oksana Chkrebtii and Dongbin Xiu , title =. Journal of Machine Learning for Modeling and Computing , issn =. 2025 , volume =

    2025

  6. [6]

    2025 , journal=

    Likelihood-free Posterior Density Learning for Uncertainty Quantification in Inference Problems , author=. 2025 , journal=

    2025

  7. [7]

    CPT: Pharmacometrics & Systems Pharmacology , volume=

    A multistate transition model for statin-induced myopathy and statin discontinuation , author=. CPT: Pharmacometrics & Systems Pharmacology , volume=. 2021 , publisher=

    2021

  8. [8]

    BMC Medical Research Methodology , volume=

    Joint modeling of multistate survival processes with informative examination scheme: application to progressions in diabetes , author=. BMC Medical Research Methodology , volume=. 2025 , publisher=

    2025

  9. [9]

    Journal of the american statistical association , volume=

    The analysis of panel data under a Markov assumption , author=. Journal of the american statistical association , volume=. 1985 , publisher=

    1985

  10. [10]

    Statistical Methods in Medical Research , volume=

    Uniformization and bounded Taylor series in Newton--Raphson method improves computational performance for a multistate transition model estimation and inference , author=. Statistical Methods in Medical Research , volume=. 2024 , publisher=

    2024

  11. [11]

    Biometrika , volume=

    Maximum likelihood estimation for semiparametric regression models with interval-censored multistate data , author=. Biometrika , volume=. 2024 , publisher=

    2024

  12. [12]

    Journal of statistical software , volume=

    Multi-state models for panel data: the msm package for R , author=. Journal of statistical software , volume=

  13. [13]

    Journal of the American Statistical Association , volume=

    A Bayesian approach to multistate hidden Markov models: application to dementia progression , author=. Journal of the American Statistical Association , volume=. 2020 , publisher=

    2020

  14. [14]

    Biometrics , volume=

    Approximate Bayesian inference for discretely observed continuous-time multi-state models , author=. Biometrics , volume=. 2019 , publisher=

    2019

  15. [15]

    Transplantation , volume=

    Diagnostic accuracy of coronary angiography and risk factors for post--heart-transplant cardiac allograft vasculopathy , author=. Transplantation , volume=. 2003 , publisher=

    2003

  16. [16]

    Proceedings of the National Academy of Sciences , volume=

    The frontier of simulation-based inference , author=. Proceedings of the National Academy of Sciences , volume=. 2020 , publisher=

    2020

  17. [17]

    Advances in neural information processing systems , volume=

    Fast -free inference of simulation models with bayesian conditional density estimation , author=. Advances in neural information processing systems , volume=

  18. [18]

    The 22nd international conference on artificial intelligence and statistics , pages=

    Sequential neural likelihood: Fast likelihood-free inference with autoregressive flows , author=. The 22nd international conference on artificial intelligence and statistics , pages=. 2019 , organization=

    2019

  19. [19]

    International conference on machine learning , pages=

    Automatic posterior transformation for likelihood-free inference , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  20. [20]

    arXiv preprint arXiv:2306.16015 , year=

    BayesFlow: Amortized Bayesian workflows with neural networks , author=. arXiv preprint arXiv:2306.16015 , year=

    Pith/arXiv arXiv