{"id":"8f8ec12a-4416-4892-9f74-ad09e32a1dd3","arxiv_id":"2507.16107","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A constructive pattern-mixture identification method, PM-ID, with an extension for unsupported missingness patterns, yields a Gibbs imputation algorithm (MISPR) that beats MICE in MNAR simulations.","lead":"This paper develops a new way to identify the full data distribution in graphical missing data models, even when some missingness patterns never appear in the data. It uses this to build MISPR, a Gibbs-sampling imputation algorithm that handles both MAR and MNAR mechanisms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MISPR's consistency under MAR/MNAR is asserted but unproven: the algorithm's recursive use of imputed values as covariates in estimated Gibbs factors can introduce feedback bias, and the simulations provide no error bars or code to rule it out.","rationale":"The paper's core theoretical contribution is constructive identification via PM-ID/PM-ID+. That part is well supported: Theorem 2 and Corollaries 3–5 give explicit recursive constructions under stated positivity and graph assumptions, and the pattern-DAG conditions are carefully tied to the local Markov property. The central practical claim, however, is that MISPR is a consistent imputation method under MAR and MNAR, including with sparse pattern support (abstract, Section 1, Section 5). This claim is never given a formal statement, let alone a proof. The algorithm's recursive structure—fitting Gibbs factors on rows that have themselves been imputed from earlier fitted conditionals—introduces a feedback loop whose bias is exactly the reason MICE can fail under misspecification. The population identification equalities only say the true conditionals coincide across patterns; they say nothing about whether the sequential estimation scheme converges to them. Footnote 9 explicitly acknowledges the risk of incongenial specifications. The simulations are a single run per DGP, reporting point L2/L∞ distances without standard errors or code; they cannot distinguish a consistent estimator from one with slow-but-nonvanishing bias. This is a load-bearing gap because the paper's advertised advantage over MICE rests on the consistency claim. The fix is either a proof (e.g., showing the estimated Gibbs factors converge to the truth and the Gibbs sampler mixes to the target extrapolation density) or a revised claim that MISPR is a heuristic imputation method motivated by identification, with consistency still open. Under either route, the reader's conditional verdict remains appropriate.","tokens_in":36833,"tokens_out":23047,"duration_ms":254258,"concrete_test":"Specify and prove a consistency theorem for MISPR under explicit regularity conditions (consistent Gibbs-factor estimators, number of Gibbs iterations and imputed datasets growing with N), or exhibit a counterexample. A concrete computational check: on the strict-positivity MNAR DGP of Figure 7(b), run MISPR at N = 10^3, 10^4, 10^5, 10^6 with 20 independent replicates each, and test whether the distribution of imputed X values converges to the true extrapolation density (e.g., via a two-sample test against oracle draws from the true model). If the discrepancy does not shrink toward zero as N grows, the consistency claim in the abstract is false. Report standard errors and release code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 (Algorithm 3) and the abstract claim MISPR is 'consistent under both MAR and MNAR settings,' but no theorem or proof establishes this. In the algorithm, each Gibbs factor p(X_i | X_{-i}, O, R=r) is fit using rows from parent patterns er in the pattern DAG, and those rows' missing entries have been imputed in earlier iterations from estimated conditionals (lines 10–16). Thus later fits use imputed values as covariates. The pattern-DAG equalities (e.g., Eq. 8) guarantee only population-level equality of conditionals; they do not by themselves ensure that the recursive estimation procedure converges to the true extrapolation densities. This is the same feedback-bias mechanism that breaks MICE under misspecification, and footnote 9 concedes the risk of incongenial specifications. The numerical evidence (Table 4) is a single run per DGP without Monte Carlo error bars, repeated seeds, or shared code, so it cannot substantiate the asymptotic claim. The identification theorems (Corollaries 3 and 5) are sound under correct G and (P1,P2)/(P1,P3); the gap is specifically the unproven statistical consistency of the imputation algorithm built on them.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a constructive pattern-mixture approach to identification of the full data law in graphical models of missing data, for m-DAGs, m-ADMGs, and m-CGs. It introduces PM-ID for strictly positive laws and PM-ID+ for settings where some missingness patterns have zero support, and it packages the resulting identifying functionals into an imputation algorithm, MISPR, that imputes each supported pattern using Gibbs factors recursively borrowed from previously identified patterns. Theorems 2 and Corollaries 3-5 give the identification results; Algorithm 3 implements the imputation; Section 6 compares MISPR to MICE on synthetic DGPs. The paper claims in the abstract that MISPR is consistent under both MAR and MNAR settings and that it handles unsupported patterns without assumptions beyond the missing data model.","tokens_in":37088,"tokens_out":12622,"duration_ms":143979,"significance":"The identification contribution is substantial. Unlike the earlier odds-ratio identification approach, PM-ID provides explicit identifying functionals for each pattern, which is a genuine practical advance and is the key enabler of the imputation algorithm. The extension to sparse pattern support via PM-ID+ is novel, and the pattern-DAG formulation unifies DAG, ADMG, and chain-graph missing data models in a single recursive construction. The paper also gives an explicit completeness statement for PM-ID in m-DAGs and m-ADMGs. These contributions are likely to be useful beyond the specific imputation algorithm. However, the paper's headline statistical claim that the imputation algorithm MISPR is consistent is not backed by a theorem, and the numerical evidence is too thin to establish it. The identification results are sound; the missing consistency theory is the main barrier.","major_comments":[{"comment":"The abstract and Section 5 describe MISPR as 'consistent under both MAR and MNAR settings', but no theorem or proof of statistical consistency is provided anywhere in the paper. In Algorithm 3, line 11, each Gibbs factor is fit using rows from parent patterns pa(r, i); by the induction built into lines 13-16, those rows have had their missing values filled in from estimated conditionals in earlier iterations. The pattern-DAG equalities in Eq. (5), (6), and (8) are population-level statements about the true law; they do not, by themselves, imply that the recursive estimator computed from imputed rows converges to the true extrapolation density. This is exactly the feedback mechanism that can bias chained-equation imputation when the imputation conditionals are misspecified, and footnote 9 concedes the risk of incongenial specifications. The numerical evidence in Section 6 and Table 4 consists of a single dataset per DGP with no Monte Carlo standard errors and no shared code, so it cannot substitute for a proof. Please either state and prove a consistency theorem for MISPR under explicit conditions on the graphical model, the conditional estimators, and Gibbs convergence, or revise the abstract and Section 5 to claim identification of the target law and algorithmic validity, rather than consistency.","section":"Abstract; Section 5, Algorithm 3"},{"comment":"The condition defining Z(r, i) in Eq. (8) is asserted to be sufficient for the Gibbs-factor equality p(X_i | X_{-i}, O, R=r) = p(X_i | X_{-i}, O, R=er), but no proof is given for any of the three graph classes, and for ADMGs and CGs the claim is not immediate from the text. The equality appears to follow from the local Markov property with the Markov blankets defined in Tables 2 and 3, because the conditioning set in Eq. (5) contains mbG(X_i). Since Eq. (8) is the engine of PM-ID+ and underlies Corollary 5, please add a short lemma with proof establishing that if R_j notin mbG(X_i) for all j in diff(r, er), then the equality holds for every distribution satisfying the relevant local Markov property. This will also clarify whether the ordered Markov blanket used for ADMGs needs any additional qualification.","section":"Section 4.2, Eq. (8) and Algorithm 2"}],"minor_comments":[{"comment":"The acronym MISPR is expanded inconsistently: the abstract says 'Multivariate Imputation via Supported Pattern Recursion', while Section 1 says 'Pattern mixture Imputaton with Gibbs Sampling (MISPR)'; the latter also contains a typo ('Imputaton').","section":"Section 1; Section 5"},{"comment":"The notation R12 in the display for the odds-ratio example is not defined; write OR({R1,R2}, R3 | ...) or define R12 explicitly.","section":"Section 3.2.1"},{"comment":"The text says that the full data law for Figure 7(e) has 24 out of 25 possible patterns and then repeats 'the one for Figure 7(e)' for the 476-out-of-2^20 case; the second reference should be to Figure 7(f).","section":"Section 6, Experiment 3"},{"comment":"Rows without a '+' report both MISPR with PM-ID+ and MISPR with PM-ID, but Algorithm 3 as written selects PM-ID whenever all patterns have positive support; please clarify how the PM-ID+ column was produced in those rows.","section":"Section 6, Table 4"},{"comment":"The numerical comparison reports no Monte Carlo repetitions, standard errors, hardware details, or code, which makes it difficult to assess whether the large timing and error differences in Table 4 are stable.","section":"Section 6"},{"comment":"The no-complete-case extension PM-ID* is presented only as an example; if this is intended as a contribution, a formal statement and proof analogous to Corollary 5 would be needed.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The identification theory is the strongest part of the paper and is likely publishable. The main risk is the unproven consistency claim for MISPR. If the authors cannot supply a consistency theorem, I would advise them to reframe the contribution as constructive identification plus an algorithm with empirical illustrations, and to remove 'consistent' from the abstract. The paper's scope fits well; the simulations should also be made reproducible before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the PM-ID/PM-ID+ identification framework is the real contribution here, and it looks sound. The MISPR algorithm built on it is reasonable, but the paper's claim that MISPR is 'consistent under both MAR and MNAR settings' is asserted, not proven. The simulations don't have error bars or code to back that up. If you treat the identification results as the main result and MISPR consistency as a conjecture needing proof, this is a strong paper.\n\nWhat's new: PM-ID gives explicit Gibbs-factor functionals for the extrapolation densities, unlike the non-constructive OR-ID of Nabi et al. and Malinsky et al. That is a genuine practical advance. PM-ID+ extends the construction to positivity violations by using the full Markov blanket constraints, not just the (S1) independences, and handles zero-support patterns in a principled way. The pattern DAG encoding is clean, and the proofs of Theorem 2 and Corollaries 3 and 5 are coherent. I also appreciate the honest discussion of PM-ID+ not being complete, and the appendix on the no-complete-case setting is a nice bonus.\n\nSoft spots: the consistency claim is the load-bearing overstatement. Algorithm 3 fits Gibbs factors using rows whose missing entries were imputed from earlier-estimated conditionals; that is textbook feedback bias territory, and the pattern-DAG equalities only hold at the population level, not for the estimated conditionals. Footnote 9 concedes the incongeniality risk. The simulations are single runs per DGP, no Monte Carlo error bars, and no shared code. Table 4's row for 7f+ shows an impressively small L2 distance, but there is no way to assess variability or reproducibility. This does not invalidate the identification results, but it means the consistency claim should be either proven under explicit conditions or tempered to 'consistent when the Gibbs factors are correctly specified.' The PM-ID+ condition in Eq. (8) for ADMGs and CGs is also terse; I would want a bit more justification there, but that is minor compared to the consistency gap. The assumption that the analyst supplies a correct m-graph is standard for this literature and not a flaw in the derivation.\n\nBottom line: the identification theory deserves a serious referee, and the paper merits a conditional accept with major revision on the consistency claim and simulation reporting. I would cite it for the PM-ID construction, not for the MISPR consistency assertion.","headline":"Solid constructive identification theory for MNAR with sparse pattern support; the MISPR consistency claim is asserted, not proven, and the simulations are too thin to back it.","tokens_in":37578,"tokens_out":2501,"would_cite":true,"duration_ms":27596,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the full data law in graphical missing data models can be identified by a constructive pattern-mixture recursion, and that the resulting imputer, MISPR, is consistent under both MAR and MNAR mechanisms even when…","keywords":["missing not at random","graphical models of missing data","multiple imputation","pattern mixture factorization","Gibbs sampling","positivity violations","m-graphs","full law identification"],"falsifier":"Simulate large samples from a known MNAR m-graph under (P1, P3), impute with MISPR using the true graph, and compare the estimated target law to the true one; separately, estimate a Gibbs factor directly from rows of a supported pattern and compare it with the factor MISPR copies from a parent pattern, since any systematic mismatch would refute the pattern-DAG equality.","tokens_in":36644,"feed_emoji":"🧩","tokens_out":9586,"duration_ms":99919,"temperature":0.7,"pith_summary":"This paper tries to establish that the full data distribution of a graphical missing data model can be identified constructively, not just shown to be identifiable, and that this constructive identification directly yields a consistent imputation procedure. The method's payoff is practical: it imputes data that are missing not at random without invoking MAR or parametric approximations, and it continues to work when many missingness patterns never occur in the observed sample, a regime that defeats standard imputation packages. The authors claim that their identification step, PM-ID/PM-ID+, is sound for m-graphs satisfying a structural condition, and that their Gibbs-sampling imputer MISPR built on it matches MICE when data are MAR and achieves lower bias when data are MNAR.","feed_headline":"Imputation that stays unbiased when data are missing not at random","feed_subtitle":"A recursive graph-based method fills missing values without MAR assumptions, even when many missingness patterns never appear.","key_machinery":"The central object is the pattern DAG, a directed acyclic graph whose vertices are missingness patterns and whose edge $\\tilde{r} \\to_i r$ records the equality $p(X^{(1)}_i \\mid X^{(1)}_{-i}, O, R=r) = p(X^{(1)}_i \\mid X^{(1)}_{-i}, O, R=\\tilde{r})$, allowing a Gibbs factor of one pattern to be borrowed from another. This edge set encodes the recursive identification order: the complete-case pattern is the root, and every missing variable in every supported pattern must have a path from the root ending in an edge labeled with that variable. The construction also uses the pattern mixture factorization (interpolation vs. extrapolation densities) and the local Markov property of m-DAGs, m-ADMGs, and m-CGs to decide when such equalities hold; PM-ID+ selects candidate parents by the condition that no indicator differing between patterns lies in the variable's Markov blanket.","core_discovery":"The central claim is that, under positivity assumptions that allow some missingness patterns to have zero support, the full law $p(X^{(1)}, O, R)$ is identified for any m-graph satisfying the structural condition (S1): no missingness indicator lies in the Markov blanket of the variable it indicates. The proof is constructive. Pattern mixture factorization splits each pattern's joint law into an interpolation density, which is observed and therefore identified, and an extrapolation density, whose Gibbs factors are matched one-by-one to Gibbs factors of already-identified patterns using conditional independences read off the graph. PM-ID+ repairs the recursion when a pattern has no support by re-routing each Gibbs factor through any earlier-identified pattern that the local Markov property makes equivalent, and the paper shows this succeeds on its examples. Equipped with the explicit identifying functionals, MISPR imputes each row by Gibbs sampling from the fitted extrapolation density of its pattern, with rows chosen for fitting according to the same pattern DAG.","pith_inferences":["A testable diagnostic follows that the authors do not develop: because each Gibbs factor may have several candidate parent patterns, discrepancies among the fitted copies of that factor across parents would signal misspecification of the m-graph before any downstream analysis.","The recursion's dependence on a user-supplied graph means the practical bottleneck is model specification, not estimation; automating pattern-compatible graph search would extend MISPR to settings where analysts cannot draw the graph.","If the constructive identification is as general as claimed, the same pattern-DAG recursion could be adapted to other missingness structures, such as longitudinal attrition, by replacing the pattern lattice with time-indexed support sets.","The paper leaves open whether PM-ID+ is complete; a characterization of when 'fail' coincides with genuine non-identifiability would settle the boundary of the method."],"forward_implications":["If the full law is identified and MISPR is consistent, analysts can impute MNAR data and then run any downstream procedure as though missingness never happened, without MAR-based corrections.","For m-DAGs and m-ADMGs satisfying (S1), PM-ID matches the identification power of prior odds-ratio methods while also outputting explicit identifying functionals, which is what makes imputation feasible.","When some patterns have no support but the complete case is present and (P3) holds, PM-ID+ still identifies the full law, converting a common high-dimensional data complication from a barrier into a routine input.","In the paper's simulations, MISPR has lower bias than MICE under MNAR and in MAR settings with positivity violations, and it scales to a 20-variable problem with 476 supported patterns where MICE could not finish.","PM-ID+ is sound but the paper makes no completeness claim, so a returned 'fail' does not by itself prove non-identifiability."],"supporting_citations":[{"why":"Prior odds-ratio identification whose soundness and completeness for (S1) PM-ID reproduces constructively.","marker":"Nabi et al. (2020)"},{"why":"Semiparametric estimators for no self-censoring models that motivate the need for explicit identifying functionals.","marker":"Malinsky et al. (2021)"},{"why":"Establishes identification results and graphical restrictions for missing data DAGs that the paper builds on.","marker":"Bhattacharya et al. (2019)"},{"why":"Introduces the pattern mixture factorization that PM-ID uses to decompose each pattern's law.","marker":"Little (1993)"},{"why":"Earlier constructive identification for the no self-censoring model, generalized by PM-ID to any m-graph and to sparse support.","marker":"Sadinle and Reiter (2017)"},{"why":"Defines pattern graphs for missing data assumptions, contrasted with the pattern DAGs used here.","marker":"Chen (2022)"},{"why":"MICE is the MAR-based imputation baseline that MISPR is compared against in all experiments.","marker":"Van Buuren and Groothuis-Oudshoorn (2011)"},{"why":"Gibbs sampling is the estimation engine for the extrapolation densities in MISPR.","marker":"Geman and Geman (1984)"}],"fun_headline_variants":["Imputation for MNAR data even when some patterns never appear","Recursive imputation that handles missing not at random","Imputation without MAR assumptions and without full pattern support","New graph-based method imputes MNAR data with sparse patterns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the analyst's missingness graph is correct and that the observed support of missingness patterns is compatible with that graph; if an edge in the true mechanism is omitted or spurious, the Gibbs-factor equalities that drive identification and imputation fail.","fun_headline_variants_meta":{"raw":{"variants":["Imputation for MNAR data even when some patterns never appear","Recursive imputation that handles missing not at random","Imputation without MAR assumptions and without full pattern support","New graph-based method imputes MNAR data with sparse patterns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001035,"raw_usage":{"total_tokens":4428,"prompt_tokens":1084,"completion_tokens":3344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":3276}},"tokens_in":700,"tokens_out":3344,"duration_ms":25440,"temperature":1.0,"reasoning_tokens":3276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:19:42.558415+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate large samples from a known MNAR m-graph under (P1, P3), impute with MISPR using the true graph, and compare the estimated target law to the true one; separately, estimate a Gibbs factor directly from rows of a supported pattern and compare it with the factor MISPR copies from a parent pattern, since any systematic mismatch would refute the pattern-DAG equality.","supporting_citations":[],"review_version":1}