{"id":"5aa5a30b-3547-4c05-b9f2-581cbedd8841","arxiv_id":"2607.05279","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"Emputation is a deep generative imputation framework whose energy-score training objective is explicitly guided by pattern-mixture identification assumptions, with proofs that the population minimizer recovers the target extrapolation distribution.","lead":"This paper proposes Emputation, a neural network framework that learns to impute missing data by encoding statistical missingness assumptions directly into the training objective. It bridges the gap between interpretable statistical imputation and flexible deep generative models, enabling distributionally correct imputation under various missing-not-at-random scenarios.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"No significant objection identified. Population-level identification proofs are correct; the finite-sample gap noted by the reader is the main limitation.","rationale":"The reader correctly identified the finite-sample gap as the main weakness. I verified the population-level proofs in detail and found no errors or hidden assumptions beyond the standard requirement that F is sufficiently expressive. The strict propriety argument is correctly applied in each case, the MCAR/CCMV/tree-graph identification strategies are properly translated into masking and selection functions, and the sequential imputation for tree graphs and monotone missingness correctly uses the chain rule factorization. The B=2 Monte Carlo approximation for the energy score is surprisingly small but is a practical choice that doesn't affect the population theory. The paper's transparency about the population-level nature of the results and its recommendation of sensitivity analyses (Section 5, Appendix C.2) are appropriate. The verdict of ACCEPT with HIGH confidence is warranted.","tokens_in":27834,"tokens_out":9195,"duration_ms":364968,"concrete_test":"Run a controlled simulation where the true conditional p(x_{r̄} | x_r, r) is known (e.g., Gaussian), and verify that as n increases (n = 500, 1000, 5000, 10000), the fitted model's conditional distribution converges to the population target as measured by energy distance on held-out data. If the energy distance does not decrease at a reasonable rate (e.g., O(n^{-1/2}) or slower), the practical gap between population theory and finite-sample performance would be a more pressing concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I carefully traced the proofs of Theorems 3.1–3.3 and 4.1. Each follows the same structure: (1) decompose the population emputation risk into a weighted sum of expected negative energy scores, one per target pattern r; (2) invoke the strict propriety of the energy score (Gneiting & Raftery 2007) to conclude that each term is uniquely minimized when the model's conditional matches the target extrapolation density. I checked the key steps:\n\n• In the MCAR proof (§D.1), the critical step uses p(X_{R−r} | X_r, R) = p(X_{R−r} | X_r) under R⊥X, which is correct. The constraint from R = 1_d (guaranteed positive probability by assumption) directly pins down the full joint p_f(x_{r̄} | x_r, r) = p(x_{r̄} | x_r), so the marginal constraints from other patterns R > r are redundant but harmless.\n\n• In the CCMV proof (§D.2), only complete cases contribute, and strict propriety on the full set r̄ = 1_d − r directly identifies p_f(x_{r̄} | x_r, r) = p(x_{r̄} | x_r, R = 1_d). Correct.\n\n• In the tree graph proof (§D.3), each one-step conditional p_f(x_{PA(r)−r} | x_r, r) = p(x_{PA(r)−r} | x_r, R = PA(r)) is identified. The full extrapolation is then obtained by sequential imputation along the unique path (Eq. 13), which is a valid chain-rule factorization. Composition of correctly specified one-step conditionals yields the correct joint. Correct.\n\n• The monotone cases (Theorem 4.1, Corollary 4.2) follow analogously.\n\nThe shared neural network across patterns does not create a theoretical issue at the population level: since r is part of the input, a sufficiently expressive F can represent different conditionals for different patterns, and the sum of terms each minimized at a distinct target is minimized when all targets are simultaneously achieved.\n\nThe reader's identified weakness—absence of finite-sample convergence rates or approximation error bounds for the neural network parameterization—is the real limitation, but it is standard for this type of methodological contribution and","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The manuscript proposes Emputation, a deep generative framework for multiple imputation that bridges pattern-mixture identification assumptions with neural-network-based distributional learning via the energy score. The core idea is that the masking pattern and selection function in the training objective encode the identifying restriction (MCAR, CCMV, tree graphs, monotone variants), so that the population minimizer of the emputation risk recovers the target extrapolation density under each assumption. Theorems 3.1–3.3, 4.1, and Corollary 4.2 establish these population-level identification results, with proofs in Appendix D. Simulations on three UCI datasets compare against MICE, missForest, GAIN, EM, and mean imputation under MCAR, CCMV, and MAR missingness. A real-data application to the NACC Alzheimer's dataset demonstrates sensitivity analysis across assumptions.","tokens_in":28737,"tokens_out":1303,"duration_ms":199797,"significance":"The paper makes a genuine methodological contribution by providing a principled bridge between the pattern-mixture identification literature and scalable deep generative imputation. The population-level proofs are correct and cleanly leverage the strict propriety of the energy score. The framework is modular: changing the identifying assumption only requires modifying the masked pattern and selection function while keeping the same energy-score learner. The inclusion of tree graphs and monotone MNAR assumptions (m-CCMV, m-NCMV, m-ACMV/MAR) broadens the applicability beyond the MCAR/MAR settings common in the deep imputation literature. The real-data sensitivity analysis (Supplementary §C.2) using exponential tilting is a nice touch. Reproducible code is provided. The main limitation is the absence of finite-sample convergence theory, which the authors acknowledge.","major_comments":[{"comment":"§2.3, Eq. (6) and §6: The empirical emputation risk uses B=2 Monte Carlo samples to approximate the energy score. This is an extremely small value, and the paper does not discuss why it suffices or how sensitive results are to this choice. Since the energy score's first term (the pairwise distance between generated samples) is poorly estimated with B=2, this raises a correctness-risk concern for the empirical claims in Figure 2 and Table S1–S2. A brief ablation or sensitivity analysis over B (e.g., B ∈ {2, 5, 10, 20}) would substantially strengthen the empirical contribution.","section":null},{"comment":"§3.3, Eq. (13) and Theorem 3.3: The sequential imputation along the tree path composes one-step conditionals. The proof correctly shows each one-step conditional is identified, but the claim that the composition yields the correct joint extrapolation density implicitly assumes the neural network can represent the composition correctly at each step. Since the same network f is shared across all patterns and steps, there is a potential concern about whether the one-step conditionals for intermediate patterns (which are never directly observed) are well-approximated. The paper should clarify whether the training objective for pattern γ_k uses observations with R = PA(γ_k) (which is correct) and confirm that no training signal is needed for intermediate patterns γ_k along the path. This is currently clear from the proof but not from the main text.","section":null}],"minor_comments":[{"comment":"§D.1 (Proof of Theorem 3.1): The last step of the proof states that because p_f*(x_{s-r}|x_r, r) = p(x_{s-r}|x_r) for all s > r, the full extrapolation density p_{f*}(x_{r̄}|x_r, r) = p(x_{r̄}|x_r) follows. This step uses the fact that the complete-case pattern 1_d pins down the full joint. Making this logic explicit (one sentence) would help readers.","section":null},{"comment":"Table 1: The notation |R_i - r| in the selection function for MCAR could be confused with the cardinality of a set difference. Clarifying that this denotes the number of coordinates where R_i = 1 and r = 0 would improve readability.","section":null},{"comment":"§5: The discussion of assumption selection is brief. While the authors correctly note that these are nonparametric identifying restrictions that cannot be tested from observed data, a concrete example of how domain knowledge would guide the choice between, say, CCMV and a tree graph in a practical setting would help practitioners.","section":null},{"comment":"§7: The logit transformation for bounded variables (BMI, blood pressure, GDS) is mentioned but the handling of the GDS score, which is integer-valued on [0, 15], is not discussed. Treating it as continuous after logit transformation may be suboptimal; a brief comment on this choice would be helpful.","section":null},{"comment":"Figure 2: The y-axis is on a log scale and values are relative to emputation. This makes it difficult to assess absolute performance. Including a table with absolute energy distance and MMD² values (at least for the matched case) would help readers calibrate.","section":null},{"comment":"§2.2: The engression review states that the minimizer is the inverse conditional CDF in the 1-d case. For the multivariate case used throughout emputation, this interpretation no longer applies directly. A brief remark on what the multivariate minimizer looks like (or that strict propriety still guarantees identification) would connect the background to the main results.","section":null},{"comment":"Typo in §2.3: 'We proposeemputation' (missing space). Similar spacing issues appear elsewhere (e.g., 'thepaper' in §8).","section":null}],"recommendation":"minor_revision","confidential_remarks":"The reader's assessment of the proofs is accurate; I independently traced the arguments in Appendix D and they are correct. The circularity concern raised by the reader (score 2.0) does not, in my view, constitute a substantive issue: the framework builds on engression and pattern graphs with proper attribution, and the central contribution is the translation of identification assumptions into masking/selection rules, which is novel. The finite-sample gap is the main theoretical limitation but is appropriately acknowledged and does not block publication. The B=2 Monte Carlo sample size is the one empirical detail I would want addressed before acceptance."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for the careful and constructive review. We address both major comments below.","responses":[{"response":"We agree that the choice of B=2 deserves justification and that an ablation would strengthen the empirical contribution. We have conducted the requested ablation over B in {2, 5, 10, 20} on all three UCI datasets under both MCAR and CCMV missingness at 20%. The results show that performance is stable across these values, with no meaningful degradation at B=2 relative to larger B. This is consistent with the engression literature, where small B values are commonly used because the energy score's first term (the pairwise distance between generated samples) acts as a regularizer against distributional collapse, and its gradient signal does not require precise estimation of the population expectation to guide training effectively. We have added a brief discussion of this point in Section 6 and included the ablation results in Supplementary Section B. We note that our theoretical results (Theorems 3.1-3.3, 4.1, Corollary 4.2) are stated at the population level and are independent of B; B affects only the empirical approximation quality.","revision_made":"yes","referee_comment":"B=2 Monte Carlo samples for energy score approximation is extremely small; no discussion of sensitivity to B; correctness-risk concern for empirical claims. Request ablation over B in {2,5,10,20}."},{"response":"We appreciate this careful reading of the proof and agree that the main text does not make the training mechanism sufficiently clear. To clarify: the training objective for each target pattern r uses only observations with R_i = PA(r), and the masked coordinates are PA(r) - r. No training signal is needed for intermediate patterns gamma_k along the path. The key insight is that each one-step conditional p(x_{PA(gamma)-gamma} | x_gamma, R=PA(gamma)) is directly identifiable from observations with pattern PA(gamma), because under those observations the coordinates in PA(gamma) are fully observed and can be masked to create the training target. The sequential composition at imputation time then applies the learned one-step conditionals in sequence along the path, but each conditional is trained independently using only the parent pattern's observations. The shared network f learns all one-step conditionals simultaneously, with each pattern's loss term contributing training signal only from its parent's observations. We have added a clarifying paragraph in Section 3.3 to make this explicit.","revision_made":"yes","referee_comment":"Tree graph sequential imputation: composition of one-step conditionals implicitly assumes the network can represent the composition correctly at each step. Concern about intermediate patterns gamma_k along the path. Request clarification that training for pattern gamma_k uses observations with R=PA(gamma_k) and that no training signal is needed for intermediate patterns."}],"tokens_in":27769,"tokens_out":613,"duration_ms":94651,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The core contribution here is real: the authors show how to translate pattern-mixture identification restrictions (MCAR, CCMV, tree graphs, monotone variants) into specific masking patterns and selection functions within an energy-score-based training objective. The population-level proofs (Theorems 3.1–3.3, 4.1) are clean and correct — each one decomposes the risk into a weighted sum of expected negative energy scores, then invokes strict propriety to pin down the target conditional. I traced the proofs and they hold up. The tree graph case is the most interesting: it uses a path decomposition to reduce the extrapolation density to a product of one-step conditionals, each identifiable from a parent pattern. This is a nontrivial synthesis of pattern-mixture theory with engression, and it is not a re-coordinatization of either ingredient alone. The framework is also architecturally clean: one neural network handles all response patterns, with the pattern vector r as input, and switching assumptions only changes the masking and selection rules. Code is public. Simulations cover MCAR, CCMV, and MAR across three datasets with both pointwise and distributional metrics. The NACC application is a reasonable illustration, and the sensitivity analysis via exponential tilting is a nice touch. The main limitation is the gap between population-level guarantees and finite-sample practice. There are no convergence rates, no approximation bounds for the neural network class, and the empirical validation uses a fixed architecture (3 layers, width 500, 500–1000 epochs) with no systematic study of how sensitive results are to these choices. This is standard for this type of methodological paper, but it does mean the practical guidance is thin. A secondary concern: the simulations only benchmark against GAIN among deep methods, and the datasets are low-dimensional (d = 5–12). Whether the approach scales to higher dimensions or more complex missingness structures is untested. The self-citation to engression (Shen & Meinshausen 2025) and pattern graphs (Chen 2022, Suen & Chen 2026) is appropriate — those are genuine prior results that this paper builds on, not circular padding. I agree with the reader's assessment and the stress-test note. The finite-sample gap is the real limitation, but it does not undermine the central population-level results. This paper is for statisticians working on missing data who want to use flexible neural models without abandoning identification theory, and for ML researchers who need a principled way to handle MNAR. It deserves a serious referee.","headline":"The paper translates pattern-mixture identification restrictions into concrete masking and selection rules within an energy-score training objective, with correct population-level proofs. The main gap is the absence of finite-sample theory, but the core idea is sound and genuinely new.","tokens_in":28790,"tokens_out":614,"would_cite":false,"duration_ms":82498,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Neural imputation guided by missing-data assumptions recovers target distributions","keywords":[],"falsifier":"If a neural network trained by the emputation risk with a correctly specified identification assumption fails to recover the target extrapolation distribution at the population level, the central claim is false. Concretely, if the population minimizer of the emputation risk under CCMV does not satisfy the identity relating the extrapolation density to the complete-case conditional distribution, the framework's theoretical foundation collapses.","tokens_in":28150,"feed_emoji":"🧩","tokens_out":946,"duration_ms":40898,"temperature":0.7,"pith_summary":"This paper proposes Emputation, a deep generative framework for missing data imputation that bridges the gap between statistical identification theory and neural network-based generative modeling. The central claim is that if you train a neural network with an energy-score-based objective where the masking pattern and sample selection are explicitly determined by a chosen missing-data assumption (such as MCAR, CCMV, or tree-graph restrictions), then the population minimizer of that training objective recovers the correct conditional distribution of the missing variables given the observed ones. The key mechanism is that the identifying assumption translates directly into two concrete components of the training loss: a masked pattern specifying which observed entries to reconstruct, and a selection function specifying which observations contribute to training. Different assumptions change only these two components while the underlying energy-score learning framework stays the same. The authors prove this recovery property (Theorems 3.1-3.3, 4.1) for several missing-not-at-random settings, and demonstrate strong distributional imputation performance in simulations and an Alzheimer's disease application.","feed_headline":"Neural imputation guided by missing-data assumptions recovers target distributions","feed_subtitle":"A new training objective encodes identification assumptions into masking rules, proving that the right neural network learns the correct im","key_machinery":"The emputation risk: an energy-score-based training objective whose masked pattern and selection function are determined by the chosen missing-data identification assumption. The energy score is a strictly proper scoring rule, meaning its population minimizer is the true conditional distribution. By encoding the identification assumption into the masking and selection structure, the framework ensures the neural network learns the correct extrapolation density.","core_discovery":"The core discovery is a constructive procedure for translating any pattern-mixture-model identification assumption into a concrete neural-network training objective whose population minimizer provably recovers the target extrapolation distribution. The procedure works by decomposing the identification assumption into a masked pattern (which observed variables to treat as reconstruction targets during training) and a selection function (which data rows contribute to the loss for each missingness pattern), then plugging these into an energy-score-based risk. The strict propriety of the energy score as a scoring rule guarantees that the minimizer of this risk is the correct conditional density,","pith_inferences":["The population-level guarantees assume the neural network class is sufficiently expressive to contain the true conditional distribution, but the paper does not provide finite-sample convergence rates or approximation error bounds for the neural network parameterization, leaving the gap between theory and practice to empirical verification.","The framework's modularity suggests that any future identification assumption expressible in the pattern-mixture framework could be incorporated by deriving the corresponding masked pattern and selection function, potentially including sensitivity-analysis assumptions parameterized by a tilt parameter.","The sequential imputation procedure used for tree graphs and monotone missingness may accumulate error along longer paths, suggesting that the depth of the tree or the length of the dropout sequence could affect practical performance in ways not captured by the population-level theorems."],"forward_implications":["If the framework extends to nonmonotone MAR and selection-model assumptions, it could provide a unified neural imputation approach covering the full landscape of missing-data identification theory.","The principle that identification assumptions can be encoded as masking-and-selection rules in a training objective may generalize beyond imputation to other problems where identification requires borrowing information across observed-data strata, such as causal inference with unmeasured confounding.","The framework's reliance on domain knowledge to choose among untestable identification assumptions means sensitivity analysis across multiple assumptions becomes a standard part of any applied analysis using emputation.","Replacing the energy score with other strictly proper scoring rules could yield variants with different distributional-learning properties, opening a family of emputation methods rather than a single one."],"fun_headline_variants":["Any identification assumption becomes a provable imputation training objective","Pattern-mixture assumptions convert directly into neural imputation objectives","Missing-data assumptions guide masking rules that recover target distributions","Energy-score risk with assumption-driven masking provably recovers imputation targets","Identification assumptions decompose into masking and selection for imputation training"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The theoretical guarantees hold at the population level and assume the neural network is expressive enough to contain the true conditional distribution; whether the trained model approximates this population minimizer well in finite samples is verified only empirically, not theoretically.","fun_headline_variants_meta":{"raw":{"variants":["Any identification assumption becomes a provable imputation training objective","Pattern-mixture assumptions convert directly into neural imputation objectives","Missing-data assumptions guide masking rules that recover target distributions","Energy-score risk with assumption-driven masking provably recovers imputation targets","Identification assumptions decompose into masking and selection for imputation training"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":533,"prompt_tokens":466,"completion_tokens":67,"prompt_tokens_details":null},"tokens_in":466,"tokens_out":67,"duration_ms":7294,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-07T19:50:41.332985+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If a neural network trained by the emputation risk with a correctly specified identification assumption fails to recover the target extrapolation distribution at the population level, the central claim is false. Concretely, if the population minimizer of the emputation risk under CCMV does not satisfy the identity relating the extrapolation density to the complete-case conditional distribution, the framework's theoretical foundation collapses.","supporting_citations":[],"review_version":1}