{"id":"cbd2e5a8-9409-4cf6-8a8b-93cb16e903bd","arxiv_id":"1908.04002","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Bayesian sequential Monte Carlo inference is proposed for latent Andersson-Madigan-Perlman Gaussian chain graphs, illustrated on graduation-rate and tenofovir pharmacokinetics data.","lead":"This paper builds a Bayesian method for recovering hidden AMP chain graphs, which are networks that mix directed arrows and undirected connections. The authors use a particle-based sequential Monte Carlo algorithm and test it on simulated data plus two real datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The appendix omits the Metropolis-Hastings acceptance probability for the joint graph/parameter moves, leaving the central SMC algorithm's correctness unverifiable.","rationale":"I read the paper as proposing a full Bayesian framework for latent AMP chain graphs, with the SMC sampler as the computational engine. For the central claim to hold, the MCMC kernels used inside the SMC must be pi_t-invariant. The appendix describes the moves but omits the acceptance probabilities for dimension-changing graph/parameter updates, so the invariant distribution cannot be checked. This is an omission in the core algorithm, not merely a scalability limitation. The initialization issue the reader identified is real and acknowledged by the authors (Section 4), but it affects large-p applicability; the kernel issue, if real, invalidates the reported posterior even for small p. In good faith, the authors may well have included the appropriate ratios, but the paper does not show them, and the simulation evidence (one empty-graph dataset, no error bars) is too weak to detect a target misspecification. Thus my stress-test supports the CONDITIONAL verdict, adding an explicit condition: provide the acceptance-probability derivation or a validation against an exact posterior for a small p.","tokens_in":15228,"tokens_out":17174,"duration_ms":181351,"concrete_test":"Re-derive the Metropolis-Hastings acceptance probability for the 0->1 and 1->0 transitions in Appendix A from the reversible-jump equations; verify that the factor 1/phi(omega_ij; 0, sigma_2^2) appears in the acceptance probability. If it is absent, the kernel is not reversible with respect to pi_t. To confirm empirically, implement the algorithm on a p=3 or p=4 problem where the exact posterior can be obtained by enumerating all valid graphs and integrating over Omega and B numerically; compare the SMC estimates of P(a_ij=0|y) with the exact values. A deviation beyond Monte Carlo error indicates the SMC sampler is biased.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 and Appendix A describe the SMC sampler and the MCMC kernel K_t. The joint move that updates the graph A together with Omega and B changes the dimension of the parameter space: adding an undirected edge (0->1) introduces a new off-diagonal entry of Omega, and removing it (1->0) deletes that entry; similarly, directed-edge moves change non-zero entries of B. The proposal distributions are specified (e.g., new Omega entries drawn from N(0, sigma_2^2), new B entries from N(xi, kappa), removed entries set to zero deterministically), but the paper only states that 'the acceptance probability for this move is easily calculated' without giving the formula. Under reversible-jump MCMC, the acceptance probability must include the ratio of proposal densities between forward and reverse moves; for the 0->1 vs 1->0 pair this is 1/phi(omega_ij; 0, sigma_2^2), and for 0->2 vs 2->0 it is 1/phi(b_ji; xi, kappa). If this ratio is omitted, K_t is not pi_t-invariant, so the SMC weights are computed with a biased kernel and the final posterior estimates are incorrect. This is more load-bearing than the initialization difficulty, because even for the small p used in the paper the sampler would target the wrong distribution. The single simulated example (empty graph, p=10) and case studies would not reveal such a bias.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a Bayesian model for Gaussian AMP chain graphs with an unobserved graph structure. The graph is encoded by an adjacency matrix with four edge states; a Dirichlet-multinomial prior with a validity constraint is placed on the graph, a G-Wishart prior on the precision matrix Ω, and independent Gaussian priors on nonzero entries of the coefficient matrix B. Posterior inference is performed with an adaptive sequential Monte Carlo sampler (Algorithm 1) over a tempered sequence of targets. The method is illustrated on one simulated empty graph and two real case studies (university graduation rates and a tenofovir pharmacokinetics study), with comparisons to structural equation model fits.","tokens_in":15597,"tokens_out":8811,"duration_ms":90777,"significance":"The contribution is potentially useful: the AMP chain graph family is less studied than DAGs or undirected graphical models, and the paper addresses the harder inferential task of jointly learning graph and parameters. The SMC framework is a reasonable route for posterior sampling on a constrained graph space, and the incorporation of G-Wishart priors and covariate information is sensible. However, the manuscript as submitted does not yet establish the correctness of the central algorithm: the Metropolis-Hastings kernel is not fully specified, the simulation evidence is a single run, and the main case study uses an informative prior derived from the very graph used for comparison. If these points are addressed, the paper would provide a useful methodological contribution. The paper does not include code or reproducible numerical scripts, and no machine-checked proofs are provided.","major_comments":[{"comment":"Appendix A: The joint (A,Ω,B) Metropolis-Hastings move is specified as a list of proposals, but the acceptance probability is never given; the text only states that it is 'easily calculated.' This is a load-bearing omission because the move changes the dimension of the parameter space: 0→1 introduces two new off-diagonal entries of Ω, 0→2/0→3 introduce a new entry of B, and the reverse moves delete entries deterministically. Correct reversible-jump acceptance requires a common reference measure, the ratio of target densities, and the forward/reverse proposal-density ratio, e.g., for 0→1 vs 1→0 a factor involving 1/φ(ω_ij;0,σ_2^2) enters. Without the formula, the claim that K_t admits π_t as invariant distribution is not verifiable, and the SMC weights are not justified because they rely on a correct mutation kernel. The authors should provide the full acceptance probability or a proof of invariance for all twelve move types.","section":"Appendix A"},{"comment":"Section 3.2: The informative prior α=(0.39,0.25,0.36,0.05) is constructed by matching the empirical edge-type frequencies in Figure 2 of [13], and the posterior graph obtained under this prior is then compared with Figure 2. As the authors acknowledge, this prior 'forces the posterior distribution of the edges to be similar to Figure 2.' The agreement with Figure 2 is therefore not independent evidence of the method's graph-recovery performance. The comparison is also used to argue that the algorithm can 'take advantage of available prior information and perform better'; this can only be evaluated with a prior that is not derived from the target graph. Please re-run the graduation case with a prior constructed independently of Figure 2, or explicitly frame the informative-prior result as a prior-predictive check rather than as validation.","section":"Section 3.2"},{"comment":"Section 3.1: The only simulation experiment is a single dataset generated from an empty graph with p=10, m=100. No repeated replicates, random seeds, standard errors, or alternative methods are reported, and the estimated posterior probabilities in Table 1 vary between 0.738 and 0.952 with no uncertainty quantification. The statement that the algorithm 'is able to recover the structure of the graph' is based on one run and, as reported, could be a favorable draw. The paper should report repeated simulation summaries (mean/range of posterior probabilities, graph recovery rates, Hamming distances) and, ideally, a nonempty graph with a known chain structure, before claiming that performance is 'stable and robust.'","section":"Section 3.1"},{"comment":"Section 2.3: The initialization step requires exact i.i.d. sampling from π0, but the only description is 'one simulates exactly from π0 using rejection sampling.' No proposal distribution, acceptance-rate calculation, or algorithm is given. The graph prior is proportional to a product over edges times an indicator of the AMP chain-graph constraint; for p=10 with α=(3,1,1,1) this may be feasible, but the authors themselves state in Section 4 that the sampler 'is not suitable when the number of nodes p is large, as the initializing step is difficult.' Because every particle is initialized this way, the method's feasibility and reproducibility depend on details that are currently missing. Please provide the rejection-sampling algorithm and report its acceptance rate in the experiments.","section":"Section 2.3"},{"comment":"Sections 3.2 and 3.3: The graph estimate is formed by taking the elementwise posterior mode over each unordered pair, a_ij = arg max_k P(a_ij=k|y,α). Because the posterior support is the set of AMP chain graphs, an elementwise mode can in principle fall outside that support, for example by producing a semi-directed cycle. The paper does not verify that the reported point estimates in Figures 4 and 5 are valid chain graphs. The authors should either prove that elementwise modes preserve the chain-graph property, or compute a jointly coherent posterior mode under the validity constraint.","section":"Sections 3.2-3.3"}],"minor_comments":[{"comment":"There is a typo in Section 2.1: 'AMP Markox property' should be 'AMP Markov property.'","section":"Section 2.1"},{"comment":"Table cross-references in Section 3.3 are inconsistent: the SEM summary is called Table 7 but appears as Table 6, and the modification indices are called Table 5 but appear as Table 7. Please renumber the tables and update all references.","section":"Section 3.3"},{"comment":"The adaptive tempering schedule is not described; the text only cites [37] for the choice of {φ_t} and [15] for the kernel parameters. A reader cannot reproduce the algorithm without consulting these references. Please state the schedule or give the adaptive rule explicitly.","section":"Algorithm 1"},{"comment":"No code, seeds, or running-time details are provided, so the numerical results in Sections 3.1-3.3 are not directly reproducible as reported.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an early-stage methodological contribution with a clear research question, but the omitted reversible-jump acceptance probability currently prevents verification of the central algorithm, the informative-prior case study is circular as evidence, and the simulation study is too thin to support the stability and robustness claims. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a plausible new combination—Bayesian inference for latent AMP chain graphs via an adaptive SMC sampler—but the paper as written does not give me enough to trust the algorithm's correctness, and the empirical evidence is thinner than the claims.\n\nWhat's new and good: AMP chain graphs are genuinely awkward because of the semi-directed cycle constraint, and the authors correctly identify that existing MCMC work on mixed graphs ([28], [32]) doesn't carry over directly. The SMC framework with tempering is a sensible choice, and the model setup (G-Wishart prior on Omega, Gaussian priors on B entries, Dirichlet prior on edge indicators) is clean. They are also honest about the main limitation in the conclusion: the initialization does not scale to large p.\n\nThe problems, in rough order of importance. First, the MCMC kernel in Appendix A does not give the acceptance probability for the joint graph/parameter moves. Those moves change the dimension of the parameter space—adding or removing an edge introduces or deletes a nonzero entry of Omega or B. A reversible-jump acceptance ratio needs the proposal density ratio between forward and reverse moves; the paper only says 'the acceptance probability is easily calculated.' Without that formula, the pi_t-invariance of K_t is unverifiable, and the SMC weights inherit any bias. This is a load-bearing omission, not a cosmetic one. It may be that the authors did include the ratio and just didn't write it down, but the paper needs to show it.\n\nSecond, the informative prior in Section 3.2 is built from the very graph the posterior is then compared against. The authors concede this ('forces the posterior distribution of the edges to be similar to Figure 2'). That comparison proves little. The uniform-prior result is the more honest benchmark, and it recovers less of the known structure.\n\nThird, the simulation evidence is a single empty-graph example with no repeated runs and no error bars. The claim of 'stable and robust' performance is not supported by one run. The real-data comparisons rely on SEM p-values and AIC/BIC, which are a weak substitute for a structure-learning baseline.\n\nNone of these strike me as fatal in principle. The core idea is a reasonable extension, the SMC machinery is standard, and the missing pieces are fixable. But the algorithm's correctness has to be verifiable, and the empirical support has to be more than one run.\n\nRecommendation: send it to a serious referee, with the expectation of major revision. The method is worth engaging with; the current presentation is not ready to be trusted.","headline":"A genuine but under-verified SMC approach to Bayesian AMP chain graph inference: the missing MH acceptance ratio and the circular prior are load-bearing gaps that need fixing before the results can be trusted.","tokens_in":16090,"tokens_out":2751,"would_cite":false,"duration_ms":29010,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62H12","60J22"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that an adaptive sequential Monte Carlo sampler can recover the structure of a latent AMP chain graph while jointly inferring the graph's parameters from data.","keywords":["AMP chain graphs","Bayesian inference","Sequential Monte Carlo","Graph structure learning","Gaussian graphical models","G-Wishart prior","Latent graph"],"falsifier":"Generate data from a known AMP chain graph with, say, $p=20$ nodes and $m=200$ observations under the paper's own likelihood, then run the sampler with the stated hyperparameters; if the rejection-sampling initialization cannot produce the required number of particles in feasible time, or if the estimated posterior probabilities of the true edges do not exceed 0.5, the paper's central claim would be undercut.","tokens_in":15042,"feed_emoji":"📈","tokens_out":5490,"duration_ms":59073,"temperature":0.7,"pith_summary":"The paper develops a Bayesian method for learning the structure of a partially observed AMP chain graph, a graphical model that mixes directed and undirected edges and forbids semi-directed cycles. The graph itself, the regression matrix $\\mathbf{B}$, and the precision matrix $\\Omega$ are all treated as unknown and inferred jointly from observed data using an adaptive sequential Monte Carlo sampler. If the method works as claimed, practitioners obtain a full posterior distribution over chain graphs and parameters, including quantified uncertainty about which edges exist and in which direction, rather than a single graph estimate. The approach is demonstrated on simulated data, university graduation rates, and a tenofovir pharmacokinetics study.","feed_headline":"SMC sampler recovers hidden chain graphs from data","feed_subtitle":"A full posterior over graph, B, and Omega replaces single estimates, tested on graduation and drug data.","key_machinery":"The load-bearing object is the AMP chain graph adjacency matrix $A\\in\\{0,1,2,3\\}^{p\\times p}$ paired with the covariance factorization $\\Sigma=(\\mathbf{I}-\\mathbf{B})^{-1}\\Omega^{-1}(\\mathbf{I}-\\mathbf{B}')^{-1}$, where the zero pattern of $\\Omega$ encodes undirected edges and the zero pattern of $\\mathbf{B}$ encodes directed edges. The sampler is an adaptive sequential Monte Carlo scheme: it builds a sequence of tempered posteriors from the prior at $\\phi_0=0$ to the full posterior at $\\phi_T=1$, initializes by rejection sampling from the prior, resamples particles according to effective sample size, and applies Metropolis-Hastings kernels that propose single-edge changes and elementwise random walks in $\\Omega$ and $\\mathbf{B}$. The AMP constraint, no semi-directed cycles, is enforced by rejecting any graph move that produces such a cycle, and $\\Omega$ is given a G-Wishart prior conditional on the graph.","core_discovery":"The authors claim that their adaptive SMC sampler enables full Bayesian inference for latent AMP chain graphs. The posterior target factorizes over chain components, each component following a Gaussian conditional on its parents, with the covariance constraint $\\Sigma = (\\mathbf{I}-\\mathbf{B})^{-1}\\Omega^{-1}(\\mathbf{I}-\\mathbf{B}')^{-1}$ enforcing the AMP Markov property. The sampler uses a tempering sequence from prior to posterior, exact rejection sampling from the initial distribution, and Metropolis-Hastings kernels that propose single-edge changes among the four adjacency states (no edge, undirected edge, and two directed orientations) while keeping the graph in the valid chain graph space. In simulation with an independent graph, the estimated posterior probabilities of no edge are mostly above 0.9, which the authors read as recovery of the generating structure. In the graduation-rates application, the chain graph selected under an informative Dirichlet prior has lower AIC and BIC than the comparison graphs, and in the tenofovir study the highest-posterior graph under a connection-favoring prior reproduces several clinically expected associations.","pith_inferences":["Because the graph prior is a product over edges truncated to valid AMP graphs, the effective prior mass per graph is not uniform; posterior edge probabilities should be read relative to this product prior, not as if the prior were uniform over chain graphs.","A natural testable extension would be to replace the rejection-sampling initialization with a sequential construction of valid AMP graphs, for example by building chain components one at a time, which could push the approach to much larger $p$ while keeping the SMC tempering framework.","In the pharmacokinetics case, the modification-index analysis points to adding an edge between TFVrectal and TFVplasma; refitting the model with that edge explicitly allowed would provide a direct posterior check of the clinical pathway discussed in the paper."],"forward_implications":["Posterior probabilities for each edge type (absent, undirected, or directed in either direction) can be reported directly, so uncertainty about edge presence and orientation is quantified instead of hidden in a single fitted graph.","The sampler can incorporate informative prior knowledge about edge frequencies through the Dirichlet parameters $\\alpha$, and the graduation-rate case suggests that doing so yields lower AIC and BIC than the baseline graphs.","Because SMC is naturally parallelizable, the algorithm can exploit multiple cores to reduce runtime in settings where the initialization step is feasible.","The paper's own scaling limits imply that the method as presented is suited to small and moderate numbers of nodes; large-network applications require a cheaper initialization or a different first proposal."],"supporting_citations":[{"why":"Defines the AMP Markov property and the chain graph class, which supplies the graph constraint the sampler enforces.","marker":"[2]"},{"why":"Gives the likelihood framework for Gaussian AMP chain graph models on which the paper's likelihood is based.","marker":"[13]"},{"why":"Introduces sequential Monte Carlo samplers, the algorithmic foundation used here.","marker":"[10]"},{"why":"Provides the adaptive SMC variant and the choices for tempering and kernel parameters.","marker":"[15]"},{"why":"Gives the G-Wishart prior and its computational treatment for precision matrices conditional on a graph.","marker":"[21]"},{"why":"Defines the adaptive sequence for the tempering exponents $\\phi_t$ that the implementation follows.","marker":"[37]"},{"why":"Supplies the university graduation data and the causal background used in the first real-data comparison.","marker":"[14]"},{"why":"Supplies the tenofovir pharmacokinetics study data used in the second real-data application.","marker":"[3]"}],"fun_headline_variants":["SMC sampler learns full posterior over chain graphs","Tempering SMC for Bayesian latent chain graphs","Exact rejection and MH moves for chain graph inference","Recover hidden chain graphs with SMC and tempering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The sampler starts by drawing exact samples from the initial distribution $\\pi_0$ via rejection sampling, but the paper gives no algorithm for this step and notes that it becomes difficult as the number of nodes grows; if those initial draws effectively never occur for larger graphs, the entire method cannot get started.","fun_headline_variants_meta":{"raw":{"variants":["SMC sampler learns full posterior over chain graphs","Tempering SMC for Bayesian latent chain graphs","Exact rejection and MH moves for chain graph inference","Recover hidden chain graphs with SMC and tempering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1669,"prompt_tokens":839,"completion_tokens":830,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":779}},"tokens_in":455,"tokens_out":830,"duration_ms":7775,"temperature":1.0,"reasoning_tokens":779,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:55:29.875708+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate data from a known AMP chain graph with, say, $p=20$ nodes and $m=200$ observations under the paper's own likelihood, then run the sampler with the stated hyperparameters; if the rejection-sampling initialization cannot produce the required number of particles in feasible time, or if the estimated posterior probabilities of the true edges do not exceed 0.5, the paper's central claim would be undercut.","supporting_citations":[{"cited_title":"A.,Madigan, D","cited_arxiv_id":null,"evidence_quote":"Defines the AMP Markov property and the chain graph class, which supplies the graph constraint the sampler enforces."},{"cited_title":"&Eichler, M","cited_arxiv_id":null,"evidence_quote":"Gives the likelihood framework for Gaussian AMP chain graph models on which the paper's likelihood is based."},{"cited_title":"&Jasra, A","cited_arxiv_id":null,"evidence_quote":"Introduces sequential Monte Carlo samplers, the algorithmic foundation used here."},{"cited_title":"A., Doucet, A","cited_arxiv_id":null,"evidence_quote":"Provides the adaptive SMC variant and the choices for tempering and kernel parameters."},{"cited_title":"&Dobra, A","cited_arxiv_id":null,"evidence_quote":"Gives the G-Wishart prior and its computational treatment for precision matrices conditional on a graph."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the adaptive sequence for the tempering exponents $\\phi_t$ that the implementation follows."},{"cited_title":"J &Glymour, C","cited_arxiv_id":null,"evidence_quote":"Supplies the university graduation data and the causal background used in the first real-data comparison."},{"cited_title":"A.,Cranston, R","cited_arxiv_id":null,"evidence_quote":"Supplies the tenofovir pharmacokinetics study data used in the second real-data application."}],"review_version":1}