{"id":"255e5da2-d161-4189-8e3d-9fa8e630dee9","arxiv_id":"2611.17074","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"GraphInfer is a pure R package that provides a unified likelihood framework for decomposable graphical models, including structure inference, exact tests, missing data handling, and accumulation tree models.","lead":"The paper presents GraphInfer, an R package for statistical inference with decomposable graphical models, covering continuous, discrete, and mixed data. It provides a unified likelihood framework for structure inference, model fitting, exact tests, missing data handling, and models for irreversible event accumulation on trees.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Exactness of the conditional test for discrete–discrete edges relies on a uniform-sampling claim that is not fully justified for the clique-by-clique Patefield scheme described.","rationale":"The reader's verdict is ACCEPT with high confidence, and the weakest assumption identified is the MAR assumption for missing data.  I disagree: the most load-bearing concern is the unsupported claim of exactness for the conditional test for discrete–discrete edges.  The paper states that the p-value is exact conditionally on the margins, but the algorithm described (clique-by-clique Patefield) is not justified and likely does not produce a uniform draw from the correct conditional distribution.  This is a technical flaw in the central claim of exact finite-sample tests.  The MAR assumption is acknowledged and standard; the discrete test exactness is presented as a novel contribution (Proposition 7) and is not standard.  The paper's own validation (Table 11) only checks level for a specific design (three discrete variables, two continuous) and does not test the discrete–discrete edge test in isolation; the row 'Discrete, n=40' tests two discrete edges jointly, but the test is conditional on the margins and the paper reports exact level.  However, the simulation may have used the same algorithm, so it would not detect a flaw if the algorithm is wrong but the level holds by coincidence for that particular configuration.  A more thorough check would be needed.  The paper does not provide a proof that the clique-by-clique permutation yields a uniform draw from the conditional distribution given the clique margins; the argument in Appendix C is insufficient.  Therefore the claim of exactness is not adequately supported, and the paper should be CONDITIONAL on providing a correct justification or correction.","tokens_in":40735,"tokens_out":1482,"duration_ms":17106,"concrete_test":"Implement a brute-force enumeration for a small but non-trivial decomposable log-linear model (e.g., three binary variables with graph A—B—C, so clique margins are AB and BC).  Enumerate all 2x2x2 tables with given AB and BC margins (say, each margin has row sums (5,5) and column sums (5,5), so total n=10).  Compute the exact conditional distribution of the table under the model (uniform over tables with those margins?  Actually the model is log-linear, so the conditional distribution given the margins is uniform only if the model is the saturated one on the margins; for a decomposable model the conditional distribution given the clique margins is not uniform over tables—it is uniform over the set of tables consistent with those margins only under the null of no three-way interaction?","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper claims (Proposition 7, Section 5.3, Appendix C) that the conditional test for edges between discrete variables is exact because the discrete data are drawn uniformly from all sequences with the same clique margins of the smaller model, using Patefield's algorithm applied clique by clique.  The description says: 'the sampler keeps the first clique and, for each later clique C_j with separator S_j and R_j = C_j \\ S_j, assigns the observed values of R_j within each cell of S_j to the observations of that cell in random order.'  This is not obviously equivalent to drawing uniformly from the conditional distribution given the clique margins of G_0[Δ].  The clique margins of a decomposable log-linear model are not independent across cliques; the joint distribution of the table given those margins is not generally uniform over all tables with those margins, and a sequential conditional-uniform assignment does not in general produce a uniform draw from the correct conditional distribution.  The paper cites Patefield (1981) for the two-way case, but the extension to multi-way decomposable margins is non-trivial and is not argued or referenced.  The claim that 'every completion is equally likely' (Appendix C, proof of Proposition 7(b)) is asserted without proof and is not obviously true for the described algorithm.  If this claim fails, the p-values for discrete–discrete edges are not exact, which undermines a central advertised feature (exact finite-sample tests).  The paper acknowledges that the test does not account for selection (Section 10), but that is a separate issue; the concern here is about the claimed exactness of the conditional test itself.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents GraphInfer, an R package for statistical inference with decomposable graphical models (Gaussian, discrete, and conditional Gaussian). It provides structure inference (minimum forests, stepwise, global search), exact finite-sample likelihood-ratio tests for certain edges, missing data handling (EM, structural EM, multiple imputation), and models for irreversible event accumulation on trees. The package is validated against independent references and simulation studies.","tokens_in":41121,"tokens_out":1485,"duration_ms":15410,"significance":"If the claims hold, GraphInfer provides a unified and computationally efficient framework for decomposable graphical models, with exact tests and missing data capabilities, which could be valuable for practitioners. The package is a successor to the archived gRapHD and offers additional features like global search and accumulation tree models.","major_comments":[{"comment":"The claim that the conditional test for discrete-discrete edges is exact relies on the assertion that the clique-by-clique Patefield scheme draws uniformly from the conditional distribution given the clique margins of the smaller model. The description of the sampler (Appendix C, proof of Proposition 7(b)) is insufficient to justify this. The clique margins of a decomposable log-linear model are not independent across cliques, and a sequential conditional-uniform assignment does not in general produce a uniform draw from the correct conditional distribution. The paper cites Patefield (1981) for the two-way case, but the extension to multi-way decomposable margins is non-trivial and is not argued or referenced. This is a load-bearing issue because the exactness of the discrete-discrete edge test is a central advertised feature. The authors should provide a rigorous justification or a reference that establishes the uniformity of the described sampler for decomposable margins.","section":"Section 5.3, Appendix C, Proposition 7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The simulation study of the exact tests' level (Table 11) shows that the exact test holds its level, but the study is limited in scope. For the discrete-discrete edge test, the design involves only three discrete variables with small numbers of levels. The study does not examine scenarios with more complex discrete structures or larger numbers of variables, where the conditional test's exactness might be more challenging to maintain. The authors should discuss the limitations of the simulation evidence and ideally provide additional simulations or theoretical justification for the test's validity in broader settings.","section":"Section 9.4, Table 11","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The calibration study of the accumulation test (Section 9.5) is acknowledged to be limited ('These figures concern few configurations and replicates: they show no substantial calibration problem when the fitted model is well determined, and the test should not be regarded as calibrated in general.'). This is an honest limitation, but it means that the test's reliability in practice is uncertain. The paper should provide clearer guidance on when the test can be trusted and perhaps suggest diagnostic checks for boundary estimates.","section":"Section 7.3, Table 4, Section 9.5","authors_acknowledge":"These figures concern few configurations and replicates: they show no substantial calibration problem when the fitted model is well determined, and the test should not be regarded as calibrated in general.","authors_acknowledge_p":0.95},{"comment":"The structural EM method uses a weighted pseudo-sample based on the unscented transform to reproduce expected sufficient statistics. The paper claims that this pseudo-sample is exact for the expected criterion. However, the unscented transform is designed for capturing the mean and covariance of a distribution, not necessarily all higher-order moments that might be relevant for discrete data or for the likelihood of a graphical model. The authors should justify why the 2q-point symmetric set is sufficient for the exact reproduction of the expected sufficient statistics for all vertex sets, especially in the mixed data case with discrete variables.","section":"Section 6.2, structural EM","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof of independence of step statistics in Proposition 6 relies on Basu's theorem and the completeness of the sufficient statistic for the conditional Gaussian family given the discrete data. The paper states that the family is a regular exponential family of full rank, but this requires that the parameter space is open and the sufficient statistic is complete. In the heterogeneous model, the covariance matrices can vary across cells, and the parameter space may have boundaries (e.g., positive definiteness). The authors should verify that the completeness argument holds under the stated conditions, especially when some cell probabilities are estimated at zero (boundary of the parameter space).","section":"Section 5.3, Proposition 6","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The paper is well-written and comprehensive. The package appears to be a significant contribution to the R ecosystem for graphical models.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The definition of the dimension for the heterogeneous model (dhet) includes a term for the covariance matrix per cell. It might be helpful to clarify that the number of parameters for each covariance matrix is gA(gA+1)/2, and that this is summed over occupied cells.","section":"Section 3.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The description of the exact test for discrete-discrete edges could be clearer. The paper mentions that the test conditions on the margins of the cliques of G0[Δ], but it is not immediately obvious how this conditioning is implemented in practice. A brief explanation of the algorithm would be helpful.","section":"Section 5.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table 5 compares forward and backward search results. It would be useful to also include the results of the global search from multiple starts to show the variability.","section":"Section 8.2, Table 5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The comparison of the fast and classic stepwise engines is convincing. The paper should mention that the fast engine is the default and that the classic engine is available for verification.","section":"Section 9.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The simulation study on missing data is thorough. The paper could discuss the computational cost of the two-stage structural EM more explicitly, as it is mentioned only in passing.","section":"Section 9.6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof of Proposition 7(b) is sketchy. The statement 'every completion is equally likely' is asserted without proof. The authors should provide a more detailed argument or a reference that establishes the uniformity of the sampler for decomposable margins.","section":"Appendix C","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This is a careful, well-tested R package that fills the gap left by the archived gRapHD and adds genuinely useful machinery: the weighted pseudo-sample for structural EM in conditional Gaussian models, the obstruction result for tree moves, and the full accumulation model with noise. The validation is exemplary—closed-form checks, independent references, full enumeration for small cases, and a test suite with 13,724 expectations. The simulation studies are honest, including the calibration limits of the bootstrap tests. Credit where due: this is not a toy; the author has done the engineering and the statistics seriously.\n\nThe paper is upfront about most of its soft spots: the MAR assumption is standard and acknowledged, the global search gives no optimality certificate, and the bootstrap tests are not calibrated beyond the studied configurations. Those are fine.\n\nThe one place I would push back is the exactness claim for discrete–discrete edges. The stress-test note lands: Proposition 7(b) says the conditional distribution of the discrete data given the clique margins is uniform on all sequences with those margins, and the clique-by-clique Patefield sampling draws from it. That uniform claim is not obviously true. The clique margins of a decomposable log-linear model are not variation-independent across cliques, and a sequential conditional-uniform assignment does not in general produce a uniform draw from the correct conditional distribution. The proof jumps from 'every completion is equally likely' to the uniform claim without demonstrating that the sequential scheme preserves uniformity. For a two-way table, Patefield's algorithm works, but the multi-way decomposable extension is nontrivial and the paper does not justify it. If that step fails, the claimed exactness of the discrete-edge tests would be compromised, and the level of those tests in Section 9.4 would be empirical rather than exact.\n\nThat said, the empirical level results look fine, and the continuous-endpoint tests are on solid ground (the proofs in Appendix C are careful and complete). The discrete-edge issue is a genuine concern but still testable—the author could simulate from the conditional distribution with an independent exact sampler and compare. I would not desk-reject over it, but I would want a referee to check it thoroughly.\n\nThis is a paper for statistical software users and developers: methodologically conservative, honest about limitations, and with a real contribution to the field. It deserves serious peer review.","headline":"A serious software paper with real novel bits and an honest limitations section, but the claimed exactness of the discrete-edge conditional test rests on a sampling justification that needs closer scrutiny.","tokens_in":41750,"tokens_out":608,"duration_ms":8790},"pith_extraction":null,"created_at":"2026-11-13T10:18:42.832743+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":null,"supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}