{"id":"4df9b9a3-4a10-4f35-add8-febf23fd6964","arxiv_id":"2509.03410","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Markov missing graphs plus the Principle of Available Information identify the full-data distribution and lead to a flexible imputation risk-minimization framework.","lead":"Missing data can bias analyses, and this paper offers a new imputation framework built on an undirected graph of the study variables. The method assumes each missing value depends only on nearby observed variables, giving researchers a principled way to impute under missing-not-at-random mechanisms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 is mathematically correct but inherits its force from PAI, an untestable cross-pattern equality; the identification claim is assumption-relative, not model-free.","rationale":"I read the paper's central claim as the identification theorem. The proof in Appendix D.1 is a correct two-line argument; I have no objection to the algebra. My concern is that the theorem's conclusion is only as strong as equation (2), and equation (2) is a genuine missing-not-at-random restriction. The constructed-law check makes precise the saturation property: because the MMG-PAI model does not constrain the observed-data distribution, no amount of data can distinguish PAI from alternatives. This matches the reader's weakest_assumption, so I agree with that diagnosis. The reader's conditions (tangent-space derivation, out-of-family simulations) remain additional reasons for the CONDITIONAL verdict; my check does not replace them. I therefore recommend no change to the reader's verdict. I also note the missing positivity condition in Theorem 1 as a small technical gap; it should be added as a regularity condition, but it does not overturn the framework when complete cases exist.","tokens_in":28405,"tokens_out":22658,"duration_ms":229003,"concrete_test":"Run the following analytical consistency check. Take any observed-data distribution with positive mass on {R_{\\bar N(s)}=1} for every connected pattern s, and construct the full-data law p(x,r)=p(x_r,R=r) * product over connected components k of p(x_{s_k}|x_{N(s_k)}, R_{\\bar N(s_k)}=1). Compute the R=r marginal of this constructed law for each pattern r and compare it with the input observed distribution. If the marginals match for every r, the model is saturated: PAI has zero testable implications, and the identification claim is strictly conditional on an untestable assumption. If any marginal fails to match, Theorem 1 is internally inconsistent and would need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1's Theorem 1 is algebraically sound: given equations (1)-(2), each p(x_s|x_N(s),R=r) is replaced by p(x_s|x_N(s),R_{\\bar N(s)}=1), which is identified from rows with the whole neighborhood observed. The load-bearing assumption is the equality in equation (2). It is not a consequence of MAR, of the graphical model, or of any observed-data property. Under MNAR, conditioning on R=r in equation (1) can induce dependence between missing components even if the complete data are faithful to G, so the factorization in equation (1) is itself an additional assumption, not a free by-product of choosing G. Moreover, Theorem 1 requires a positivity condition (positive mass on R_{\\bar N(s)}=1 for every connected pattern s) that is never stated. Because equations (1)-(2) impose no restrictions on the observed-data distribution, the assumption cannot be tested: any observed law with support on the conditioning events admits a full-data law satisfying MMG and PAI. Section 2.2's caveat that it is impossible to know the true model is the paper's own admission of this. The practical risk is not internal inconsistency but that all downstream claims, including identification, efficiency, and the NACC conclusions, are conditional on PAI being true; if PAI is false, MMG imputations are biased in a way no check on observed data can detect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Markov missing graph (MMG), an undirected-graph framework for imputation. The full-data law is decomposed pattern-mixture style, and the extrapolation density is factorized over connected components of the missingness pattern (eq. 1). The Principle of Available Information (PAI, eq. 2) identifies each factor with the corresponding conditional density among rows in which the relevant neighborhood is fully observed, yielding Theorem 1. The paper then frames estimation as empirical risk minimization (Section 3), with Gaussian, Ising, and mixture-of-product instantiations. Theoretical results connect MMG to CCMV (Theorems 3-4), prove recovery under MCAR and an efficiency gain for G-MMG (Theorem 5), and give an augmented IPW-type estimator with a claimed efficient influence function (Theorems 6-7) and graphical properties (Propositions 8-10). Simulations and a NACC application compare MMG with CC, MICE, and missForest, with a sensitivity analysis over the graphical lasso threshold.","tokens_in":28801,"tokens_out":12171,"duration_ms":105177,"significance":"The identification result, though assumption-relative, is clean and correct under the stated PAI condition, and it provides a new and simple way to build MNAR imputation models with local structure. The MCAR recovery theorem and the efficiency comparison to complete-case analysis are valuable, as is the demonstrated practicality on a large Alzheimer's data set with an R package. If the efficiency claims are appropriately revised or completed, the framework could be a useful addition to the missing-data toolbox. The main weakness is that the efficiency theory is not yet fully established.","major_comments":[{"comment":"The theorem needs an explicit positivity condition. For every connected pattern s that can occur, PAI conditions on the event R_{\\bar N_G(s)}=1; if this event has zero probability, the right-hand side of (2) is undefined and the imputation submodel is not identifiable. The statement that MMG and PAI nonparametrically identify the full-data distribution is therefore too strong without adding the support assumption.","section":"§2.1, Theorem 1 and equation (2)"},{"comment":"The pathwise-derivative calculation is performed in the unrestricted model for p0(x1,xNG(sj),r), without characterizing the tangent space of the MMG-PAI model. Since equations (1)-(2) impose equality constraints across response patterns, the model's tangent space is smaller than the one used in the derivation; consequently, the paper has not shown that the displayed expression is the efficient influence function in the MMG model. The estimator may still be a reasonable augmented IPW estimator, but the efficiency claim in the abstract and Section 4.3 is not supported.","section":"§4.3.3, Theorem 6"},{"comment":"The multiple-robustness proof only treats the case where the odds model is correct and the regression model is misspecified; the symmetric case is asserted without proof. More importantly, the proof uses the expansion from Theorem 6, which per the previous comment is not established in the MMG model, so the consistency argument for the oracle estimator does not go through as written.","section":"§4.3.3, Theorem 7"},{"comment":"The identification formulas rely on the MMG factorization (1) applied to the extrapolation density for the given pattern r. Under MNAR, conditioning on R=r can create dependencies among missing components even if the complete-data law is faithful to G, so equation (1) is a substantive assumption about the missingness mechanism, not a consequence of choosing G. The paper should state this caveat explicitly in the statements of the identification results, not only in the informal discussion of Section 2.2.","section":"§4.3.1-4.3.2, equations (11)-(13)"}],"minor_comments":[{"comment":"The text 'X001 = (X1, X2)^T' appears to be a typo; it should likely read 'X110 = (X1, X2)^T'.","section":"§1.1"},{"comment":"The shorthand 'R≥ pattern' (for example, R≥11110) is used without a formal definition; define it once near the notation section.","section":"§2.1"},{"comment":"The statement that G-MMG is asymptotically more efficient than complete-case analysis presumes the graph G used by MMG is the true graph; if G is estimated from data, the result requires an additional consistency assumption on graph estimation, which should be stated explicitly.","section":"§4.2, Theorem 5"},{"comment":"The phrase 'codes 96' or 996'' appears to be a typo; clarify whether the intended test codes are 96 and 996.","section":"§6.1"},{"comment":"The assumption that X1, Osj, and msj are uniformly bounded is restrictive for continuous X1; consider replacing it with a bounded-influence-function condition or suitable moment assumptions, since the NACC application includes continuous test scores.","section":"§4.3.3, assumption (A2)"},{"comment":"The simulation uses the true graph for G-MMG, which gives the method an advantage over competitors; the practical setting where the graph is estimated is only addressed in the real-data sensitivity analysis, so this limitation should be acknowledged in the simulation section.","section":"§5.1"}],"recommendation":"major_revision","confidential_remarks":"The central identification theorem is correct and the paper is readable; the main blocker is the efficiency theory. I would encourage the authors to either derive the MMG tangent space and verify the efficient influence function within it, or substantially soften the efficiency claims (for example, presenting the estimator as doubly robust rather than efficient). I also think the label 'Principle of Available Information' risks being read as a general principle rather than a specific, untestable MNAR assumption; the main text should state this plainly at the point of Theorem 1."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the core identification result (Theorem 1) is mathematically correct, but it inherits all its force from PAI, a cross-pattern equality that no observed-data check can validate. The paper is honest about this, but the honesty does not reduce the assumption load.\n\nWhat is new: the local Markov decomposition of the imputation model over an undirected graph is a genuine departure from the DAG/ADMG/chain-graph missing-data literature. Using PAI to borrow conditional distributions from rows where a graph neighborhood is fully observed is a clean, useful idea. The reduction to CCMV (Theorems 3-4), the MCAR recovery result (Theorem 5), and the IPW/regression/multiply-robust estimators are all correct and well proved. The IRM framing is a nice unification, and the R package makes the method usable.\n\nSoft spots, in proportion: the identification claim is assumption-relative, not model-free. PAI is a substantive MNAR assumption, and the paper's Section 2.2 freely admits that the true model is unknowable without additional information. The stress-test is right that a positivity condition is never stated: Theorem 1 needs positive mass on R_{\\bar N(s)}=1 for every connected pattern s; this is a minor fix but should be made. The efficiency theory is incomplete: Theorem 6 computes a pathwise derivative in an unrestricted model without characterizing the tangent space under the MMG/PAI restrictions, so calling it the efficient influence function is not fully justified. That is a moderate gap, not fatal. The simulations are in-family — data generated from Gaussian graphical models for G-MMG and mixtures of Gaussian graphical models for MP-MMG — so the empirical results show internal consistency rather than independent validation. The NACC application is a nice illustration, and the sensitivity analysis over glasso thresholds is a good start, but it does not close the out-of-family gap.\n\nWho this is for: missing-data methodologists and applied statisticians who want a principled undirected-graph alternative to pattern-mixture models. It deserves serious peer review. The referee should require a tangent-space derivation and out-of-family simulations, but the central identification result holds up.","headline":"MMG is a genuinely new undirected-graph framework for MNAR imputation, but the identification rests entirely on PAI, an untestable assumption the paper itself admits.","tokens_in":29203,"tokens_out":3624,"would_cite":true,"duration_ms":29733,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D10","62G05","62H12"],"pacs":[],"model":"deepseek-v4-flash","headline":"An undirected graph and a local information principle identify the full data distribution without assuming missing at random.","keywords":["missing data imputation","undirected graphical model","missing not at random","nonparametric identification","principle of available information","imputation risk minimization","pattern-mixture model","semiparametric efficiency"],"falsifier":"Take a fully observed dataset, mask values with a mechanism in which the probability of each response pattern still depends on the missing block's values after conditioning on the observed neighborhood, and compare the nonparametric MMG imputed distribution of the masked block with its true distribution; a systematic discrepancy at growing sample size would show that equation (2) is doing the identification work and cannot be inferred from the observed data. Alternatively, with a validation subset, estimate $p(x_s|x_{N_G(s)},R=r)$ and $p(x_s|x_{N_G(s)},R_{\\bar N_G(s)}=1)$ separately and test their equality for patterns sharing the same neighborhood.","tokens_in":28194,"feed_emoji":"📊","tokens_out":15572,"duration_ms":125930,"temperature":0.7,"pith_summary":"The paper claims that missing-data imputation can be organized around an undirected graph of the study variables. Given such a graph, the imputation model for any missingness pattern splits into independent pieces, one for each connected component of missing variables, and each piece is identified by borrowing information from rows in which that component's graph neighborhood is fully observed. The paper names this the principle of available information and shows it turns imputation into a local empirical-risk-minimization problem that accommodates Gaussian, Ising, and mixture-of-product models. If the central theorem is right, the full joint distribution of the variables and the missingness indicators is nonparametrically identifiable, so missing-not-at-random settings can be handled without a parametric model for the missingness mechanism. The framework is then shown to reduce to the classical complete-case missing-value assumption in special cases, to recover the true model under missing completely at random, and to yield multiply robust efficient estimators.","feed_headline":"Missing-data imputation becomes identifiable from graph neighborhoods","feed_subtitle":"The key is borrowing each missing block's distribution from rows where its graph neighborhood is fully observed.","key_machinery":"The load-bearing object is the Markov missing graph (MMG), an undirected graph $G=(V,E)$ whose neighborhoods specify which variables may be used to impute a missing variable. The argument runs on two equations: equation (1) decomposes the imputation model $p(x_{\\bar r}|x_r,R=r)$ into a product over connected components $s_k$ of the missing set, and equation (2), the principle of available information, identifies each factor with the same conditional distribution computed from rows where the component and its neighbors are all observed. Around these, the paper builds an imputation risk minimization (IRM) framework: for each connected pattern $s$, the parameter $\\theta_s$ minimizes an empirical loss restricted to observations with $R_{\\bar N_G(s)}=1$, so Gaussian, Ising, and mixture-of-product likelihoods become plug-in choices. For estimating a mean functional, the efficient influence function combines an inverse probability weight $O_s$ with a regression adjustment $m_s$, giving a multiply robust estimator.","core_discovery":"The paper's central discovery is a nonparametric identification theorem. Starting from the pattern-mixture decomposition $p(x,r)=p(x_{\\bar r}|x_r,R=r)p(x_r,R=r)$, it defines a Markov missing graph (MMG): an undirected graph $G$ whose connected components of the missing set $s_1,\\ldots,s_K$ factor the extrapolation density as $p(x_{\\bar r}|x_r,R=r)=\\prod_{k=1}^K p(x_{s_k}|x_{N_G(s_k)},R=r)$. The principle of available information (PAI) identifies each factor with $p(x_{s_k}|x_{N_G(s_k)},R_{\\bar N_G(s_k)}=1)$, the same conditional law computed from rows whose component and neighbors are all observed. Theorem 1 states that under equations (1) and (2) the full-data distribution $p(x,r)$ is nonparametrically identified: every extrapolation density is a product of identifiable conditional distributions. The paper further shows that a fully connected graph reduces PAI exactly to the complete-case missing value assumption, that under monotone missingness a chain-faithful complete-case law makes the two equivalent, and that under MCAR with a faithful graph NP-MMG recovers the true model while the Gaussian variant is asymptotically more efficient than complete-case analysis.","pith_inferences":["A testable consequence the paper leaves implicit: if PAI holds, then any two missingness patterns that agree on a missing component and its observed neighborhood must share the same conditional law for that missing block; this equality can be checked directly in a validation study with fully observed holdout rows.","Because PAI trains each local submodel only on rows where the relevant neighborhood is complete, the effective sample size is the size of that local stratum; users with highly fragmented missingness patterns may need to verify that these strata are large enough before relying on nonparametric versions.","The graph itself is an assumption carrier: estimating the graph from complete cases is justified primarily under MCAR and faithfulness, so outside that setting a misspecified graph can bias imputations; the paper's sensitivity analysis suggests stability until the graph becomes very sparse, which could be tested routinely in applications.","The paper's locality propositions imply that edits to a graph far from a missing component do not change that component's imputation model, so large or noisy graphs can be simplified locally; this is a direct corollary, though the paper stops short of proposing an automatic local simplification procedure."],"forward_implications":["Imputation under missing-not-at-random becomes a local modeling problem: each missing block is drawn from a conditional distribution learned from rows whose graph neighborhood is fully observed, so no parametric model for the full data or the missingness mechanism is required.","When the graph is fully connected, PAI coincides exactly with the classical complete-case missing value assumption, so MMG nests a standard complete-case restriction as a boundary case.","Under monotone missingness with a chain-faithful complete-case law, PAI and the complete-case missing value assumption give the same imputation model.","Under MCAR with a faithful graph, using the same graph in MMG recovers the true imputation model, and the Gaussian variant is asymptotically more efficient than complete-case analysis.","The IRM framework produces regression-adjustment, inverse-probability-weighted, and multiply robust estimators for means, and the NACC application replaces 351 raw missingness patterns with 84 representative model patterns while yielding narrower confidence intervals than complete-case analysis."],"supporting_citations":[{"why":"It supplies the pattern-mixture decomposition and the complete-case missing value assumption that MMG generalizes.","marker":"Little (1993)"},{"why":"It frames nonparametric identifiability of saturated missing-data models, the target of Theorem 1.","marker":"Robins (1997)"},{"why":"It establishes the directed-graph missing-data paradigm whose focus MMG shifts to undirected graphs.","marker":"Mohan et al. (2013)"},{"why":"It reviews graphical missing-data models and testability, motivating the undirected gap MMG fills.","marker":"Mohan and Pearl (2021)"},{"why":"It introduces pattern graphs as identifying restrictions, a close relative of PAI used as a comparison point.","marker":"Chen (2022)"},{"why":"It provides the Donsker and uniform-entropy results used in the multiple-robustness proof.","marker":"Van der Vaart (2000)"},{"why":"It supplies missForest as a benchmark imputation method in the simulations and NACC analysis.","marker":"Stekhoven and Buhlmann (2012)"},{"why":"It supplies MICE as a benchmark imputation method in the simulations and NACC analysis.","marker":"Van Buuren (2018)"}],"fun_headline_variants":["Graph-based imputation: borrowing neighborhoods to fill gaps","Identifiable missing data imputation via Markov missing graphs","Graph neighborhoods unlock nonparametric imputation identifiability","Impute missing data via graph-local distributions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the principle of available information: for each missing block, the conditional distribution given its observed graph neighborhood is assumed to be the same whether the row belongs to the target missingness pattern or to any pattern where that neighborhood is fully observed; this is an untestable missing-not-at-random assumption, and if it fails, the identification proof collapses.","fun_headline_variants_meta":{"raw":{"variants":["Graph-based imputation: borrowing neighborhoods to fill gaps","Identifiable missing data imputation via Markov missing graphs","Graph neighborhoods unlock nonparametric imputation identifiability","Impute missing data via graph-local distributions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000845,"raw_usage":{"total_tokens":3683,"prompt_tokens":952,"completion_tokens":2731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":2670}},"tokens_in":568,"tokens_out":2731,"duration_ms":15132,"temperature":1.0,"reasoning_tokens":2670,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:30:23.824838+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fully observed dataset, mask values with a mechanism in which the probability of each response pattern still depends on the missing block's values after conditioning on the observed neighborhood, and compare the nonparametric MMG imputed distribution of the masked block with its true distribution; a systematic discrepancy at growing sample size would show that equation (2) is doing the identification work and cannot be inferred from the observed data. Alternatively, with a validation subset, estimate $p(x_s|x_{N_G(s)},R=r)$ and $p(x_s|x_{N_G(s)},R_{\\bar N_G(s)}=1)$ separately and test their equality for patterns sharing the same neighborhood.","supporting_citations":[{"cited_title":"Graphical models for processing missing data","cited_arxiv_id":null,"evidence_quote":"It reviews graphical missing-data models and testability, motivating the undirected gap MMG fills."},{"cited_title":"Graphical models for inference with missing data","cited_arxiv_id":null,"evidence_quote":"It establishes the directed-graph missing-data paradigm whose focus MMG shifts to undirected graphs."},{"cited_title":"Pattern graphs: a graphical approach to nonmonotone missing data","cited_arxiv_id":null,"evidence_quote":"It introduces pattern graphs as identifying restrictions, a close relative of PAI used as a comparison point."},{"cited_title":"Flexible Imputation of Missing Data","cited_arxiv_id":null,"evidence_quote":"It supplies MICE as a benchmark imputation method in the simulations and NACC analysis."}],"review_version":2}