{"id":"ea369043-1a08-409b-ab74-191acae01991","arxiv_id":"2507.15842","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper gives sound and complete methods for identifying conditional causal effects when the causal graph is known only up to an MPDAG.","lead":"This paper develops ways to estimate conditional causal effects, such as how a treatment works within a subgroup, when the causal graph is known only up to a partially oriented graph called an MPDAG. It contributes an identification formula, a do calculus, and a complete algorithm, which matter for analyzing observational data when experts can add background knowledge.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Completeness proof of Theorem 7 invokes Lemma 57 without first establishing the lemma's two standing assumptions in Case 1.","rationale":"The reader's weakest_assumption is causal sufficiency, which is an explicitly acknowledged scope limitation rather than a flaw in the central argument within its stated setting. My stress-test instead identified a concrete gap in the proof of the paper's central completeness claim (Theorem 7, Appendix G): Lemma 57 is used before its hypotheses are shown to hold, and the missing WLOG reasoning is nontrivial enough that the proof as written does not rigorously establish completeness. The gap is likely repairable, so the appropriate verdict is CONDITIONAL acceptance: the result should be accepted once the authors supply the missing justification or amend the proof. This is not a rejection; the examples and the overall strategy are credible, and the causal sufficiency limitation is honestly disclosed.","tokens_in":61619,"tokens_out":51325,"duration_ms":520154,"concrete_test":"Independently re-derive the proof of Theorem 7 in Appendix G, focusing on the first invocation of Lemma 57 in Case 1. Check whether, before invoking Lemma 57, the proof establishes: (i) if a possibly causal path from X to Y starts undirected and avoids Z, then Proposition 41 finishes the case; and (ii) a possibly causal path from X to Z that starts undirected and avoids Y, obtained as a shortest prefix of the theorem's given path that ends at a Z-node. If these steps are absent, the written proof is incomplete; if they can be inserted without changing the construction of D1' and D2', then completeness is sound but requires amendment.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Appendix G, the proof of Theorem 7 splits into cases based on whether S3 is empty. In Case 1 (S3 nonempty), the proof selects a d-connecting path p from X to Y, defines the collider set C, and chooses shortest causal paths q_i from each collider to Z. It then invokes Lemma 57 to assert that the q_i share no nodes with p or with each other. However, Lemma 57 is stated only under two assumptions: (i) there is no possibly causal path from X to Y that starts undirected and avoids X\\{X} ∪ Z, and (ii) there is a possibly causal path from X to Z that starts undirected and avoids X\\{X} ∪ Y. The proof of Theorem 7 never verifies these assumptions before using Lemma 57. The theorem's hypothesis provides a path from X to Y∪Z, not necessarily to Z; if the path ends at Y and avoids Z, Proposition 41 already gives non-identifiability, and if it contains Z, a shortest prefix ending at a Z-node gives (ii), while (i) follows because any such path to Y would also finish the proof via Proposition 41. These WLOG steps are missing from the written proof. Since Theorem 7 is the entire completeness argument for Algorithm 1, this is a load-bearing gap, albeit one that appears repairable.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript addresses identification of conditional causal effects f(y|do(x),z) when the causal structure is known only up to an MPDAG and all variables are observed. It presents three results: a closed-form identification formula (Theorem 3) under the assumption that the conditioning set contains no possible descendants of the treatment; a three-rule do calculus for MPDAGs (Theorem 6) that is claimed to reduce to Pearl's calculus in the DAG case; and an algorithm (CIDM, Algorithm 1) that iteratively applies Rule 2 and then Theorem 3, with a completeness theorem (Theorem 7) stating that failure implies non-identifiability. The appendices contain long proofs, including explicit linear SEM counterexamples with numerical coefficients. The paper clearly states in Section 6 that latent confounding is outside the scope of the setting.","tokens_in":61895,"tokens_out":14306,"duration_ms":159024,"significance":"If the results are correct, the paper fills a genuine gap: prior MPDAG work either focused on unconditional effects or on conditional adjustment sets, which need not exist for identifiable conditional effects. The proposed do calculus and the completeness of Algorithm 1 are natural and nontrivial extensions of the existing theory, and the completeness theorem is the main technical contribution. The paper's strengths include explicit SEM-based non-identifiability constructions with concrete coefficients, detailed appendix proofs, and a clear statement of the causal-sufficiency limitation. The main caveat is that the proofs are long and not machine-checked, and one load-bearing step in the completeness proof is not fully justified as written.","major_comments":[{"comment":"The proof invokes Lemma 57 without first verifying the lemma's two standing assumptions. After defining the collider set C and the shortest causal paths q_i from each collider to Z, the proof states that 'by Lemma 57, Ci is the only node on both p' and q'_i, and there is no shared node between the causal paths q'_i and q'_j.' However, Lemma 57 is stated under the assumptions that (i) there is no possibly causal path from X to Y that starts undirected and avoids X\\{X} ∪ Z, and (ii) there is a possibly causal path from X to Z that starts undirected and avoids X\\{X} ∪ Y. Theorem 7's hypothesis supplies a path from X to Y∪Z, not necessarily to Z. If that path ends at Y and avoids Z, Proposition 41 already gives non-identifiability; if it contains a node of Z, a shortest prefix ending at a Z-node gives assumption (ii), and any violation of assumption (i) would also allow the proof to conclude via Proposition 41. These without-loss-of-generality reductions are not stated, so the written proof of Theorem 7 has a load-bearing gap, although the gap appears readily repairable.","section":"Appendix G, proof of Theorem 7, Case 1"},{"comment":"The completeness argument does not explicitly connect a FAIL at a later while-loop state (X', Z') to the original query (X, Z). Algorithm 1 can reach its FAIL branch only after some variables have been moved from the do-set to the conditioning set using Rule 2 of Theorem 6. Theorem 7, as stated, applies to arbitrary disjoint sets and would show non-identifiability of f(y|do(x'), z'), but one must additionally use the soundness of the preceding Rule-2 transformations to conclude non-identifiability of the original f(y|do(x), z). This equivalence is not spelled out in the proof of completeness; it should either be stated as an invariant of the algorithm or argued explicitly before Theorem 7 is applied.","section":"Section 5.4, completeness argument for Algorithm 1"}],"minor_comments":[{"comment":"The integration notation is ambiguous: the theorem defines B = An(Y, G_{V\\X}) \\ (Z ∪ Y), while the bucket decomposition is computed on An(Y, G_{V\\X}) \\ Z, which includes the bucket containing Y. As written, the integral in Equation (2) appears to integrate over all bucket variables, including Y. Example 1 makes clear that the intended integral is over all buckets except Y, but the notation should be made precise.","section":"Theorem 3, Equation (2)"},{"comment":"The pseudocode uses two sequential 'Return' statements instead of combining the results into a single multiset. The intended behavior is presumably 'return the multiset union of CIDME(X', Z', Y, G1) and CIDME(X', Z', Y, G2)', and the pseudocode should say so explicitly.","section":"Algorithm 2, lines 12-13"},{"comment":"The reference list contains repeated typos such as 'Uncertianty of Artifical Intelligence', and Lemma 57 contains a typo in the phrase 'which leads us to a contradiction with the choice fo p'. A careful proofread is needed.","section":"References and typography"}],"recommendation":"major_revision","confidential_remarks":"I do not see a reason to doubt the main claims, and the missing steps in the completeness proof appear repairable. If the authors add the missing WLOG argument for Lemma 57 and explicitly justify the reduction from fail-state sets to the original query, the completeness theorem would be convincing. The paper is long but within the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real advance for MPDAG identification. The do calculus for MPDAGs (Theorem 6), the conditional identification formula (Theorem 3), and the CIDM algorithm (Algorithm 1) with its completeness theorem (Theorem 7) fill a gap left by LaPlante/Perkovic 2024 and Perkovic 2020. The paper's own examples are concrete, and the non-identifiability examples use explicit SEMs with coefficient values, which is more than most papers in this area do. The proofs are long but organized, with supporting lemmas in appendices. I did not find internal contradictions.\n\nThe main soft spot is in the completeness proof of Theorem 7. In Case 1 (S3 nonempty), the proof selects a shortest d-connecting path p, defines collider set C, and then invokes Lemma 57 to claim the causal paths from colliders to Z do not overlap with p or each other. But Lemma 57 is stated under two standing assumptions: no possibly causal path from X to Y that starts undirected and avoids X\\{X}∪Z, and a possibly causal path from X to Z with analogous avoidance. The proof of Theorem 7 never verifies these assumptions before applying the lemma. The stress-test note correctly identifies this as a gap in the written proof. That said, the gap looks repairable: the theorem's hypothesis gives a path from X to Y∪Z; if it ends at Y, Proposition 41 already gives non-identifiability, and if it contains Z, a shortest prefix to a Z-node gives the second assumption, while the first follows because any such path to Y would finish via Proposition 41. A referee should ask for these WLOG steps spelled out rather than reject the paper.\n\nA smaller caveat: the whole enterprise assumes causal sufficiency (no latent confounding), which the authors acknowledge in Section 6. That restricts scope but does not undermine the result.\n\nThe citation pattern is fine; self-citations are to prior work they are genuinely extending. The paper deserves a serious referee. I would send it to review and ask for a fix to the Theorem 7 proof gap.","headline":"Solid completeness result for MPDAG conditional effects; the completeness proof has a repairable gap that a referee should ask to be fixed.","tokens_in":62383,"tokens_out":1459,"would_cite":true,"duration_ms":15912,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","05C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves a complete algorithm, CIDM, that decides exactly when a conditional causal effect is identifiable from observational data when the graph is known up to an MPDAG, and derives the identifying expression whenever one exists.","keywords":["conditional causal effects","MPDAG","causal identification","do calculus","equivalence class","causal sufficiency","identifiability","partial causal ordering"],"falsifier":"Enumerate all small MPDAGs and, for each triple (X,Y,Z), compare CIDM's verdict against an exhaustive do-calculus search over the DAGs in [G]; any case where CIDM returns FAIL yet an identifying expression exists, or returns an expression that differs from the true interventional distribution under a linear Gaussian density consistent with G, would refute the soundness and completeness claims.","tokens_in":61458,"feed_emoji":"🎯","tokens_out":9564,"duration_ms":92893,"temperature":0.7,"pith_summary":"Conditional causal effects ask how an intervention changes outcomes within a subpopulation defined by observed covariates. This paper asks when those effects can be computed exactly from observational data when the causal graph is known only up to a maximally oriented partially directed acyclic graph (MPDAG), a partially oriented graph that encodes a class of causal DAGs consistent with both observational data and any expert background knowledge. The paper's central claim is a complete solution: an algorithm, CIDM, that outputs an identifying expression whenever one exists and reports FAIL exactly when no expression exists, together with an identification formula for the common case where the conditioning covariates are unaffected by treatment and a three-rule do calculus for MPDAGs that generalizes the classical do calculus. The results are proven under causal sufficiency, meaning all variables are observed, so the practical payoff is a decision procedure for subgroup effect identification in fully measured systems with uncertain edge directions.","feed_headline":"An algorithm decides exactly when subgroup effects are identifiable","feed_subtitle":"Given a partially oriented causal graph, the method returns an identifying formula or proves none exists.","key_machinery":"The central objects are maximally oriented partially directed acyclic graphs (MPDAGs), which represent the set [G] of DAGs sharing the same adjacencies and directed edges, the same unshielded colliders, and a partial causal ordering imposed by background knowledge. The argument is carried by three mechanisms: the PCO algorithm, which partitions node sets into buckets ordered by a partial causal ordering and yields the factorization in Theorem 3; the possibly causal path, a path with no backward edges whose presence or absence determines identifiability—in particular, a proper possibly causal path from X to Y that starts with an undirected edge is the obstruction that makes conditional effects unidentifiable; and mutilated MPDAGs GX, GXZ, and GX,Z'(W), obtained by deleting edges into X or out of Z, whose d-separations drive the three rules of the generalized do calculus. Algorithm 1 works by repeatedly moving treatment variables into the conditioning set via Rule 2 until no such undirected-starting possibly causal path remains, then applying Rule 3 or the ratio of two applications of Theorem 3.","core_discovery":"The paper establishes that, for any causal MPDAG G and pairwise disjoint node sets X, Y, Z, the conditional interventional density f(y|do(x),z) is identifiable from the observational density of any density f consistent with G if and only if Algorithm 1 (CIDM) returns an expression; when it returns FAIL, the authors construct two densities consistent with G that are observationally identical yet disagree on f(y|do(x),z), so no identifying expression exists. This completeness result is Theorem 7, and it is obtained by combining three results: Theorem 3, an identification formula written in terms of bucket decompositions from the PCO algorithm that applies when Z is unaffected by treatment and no proper possibly causal path from X to Y starts with an undirected edge; Theorem 6, a do calculus for MPDAGs whose three rules mirror the classical do calculus but use d-separations in mutilated MPDAGs such as GX and GXZ; and the fractional decomposition f(y|do(x'),z') = f(y,zD|do(x'),zN)/f(zD|do(x'),zN) used by the algorithm to handle conditioning variables that are possible descendants of treatment. The calculus is shown to reduce exactly to the classical do calculus when the MPDAG is a DAG, and the identification formula reduces to the earlier unconditional MPDAG formula when Z is empty.","pith_inferences":["The completeness boundary suggests that the hardness of conditional identification in causally sufficient systems lies entirely in the interaction between undirected-edge uncertainty and conditioning on possible descendants of treatment; a similar boundary for systems with latent variables would need structure beyond the chordal components that make ancestral-graph algorithms work.","The fractional output A/B may make plug-in estimation practical: both numerator and denominator are ordinary observational quantities, so nonparametric estimators of each component could be combined, although the paper does not analyze finite-sample behavior.","A natural test of the method is to apply CIDM to an MPDAG learned from data with known background knowledge and compare its identified expression to estimates from a randomized trial on the same variables; agreement would empirically validate the completeness claim."],"forward_implications":["Any conditional effect that is identifiable given an MPDAG can be written down by CIDM; no identification algorithm in this setting needs to be more powerful.","For pre-treatment covariates (Z unaffected by X), identifiability is fully characterized by the absence of a proper possibly causal path from X to Y that starts undirected, giving a fast graphical check.","The do calculus for MPDAGs reproduces the classical do calculus exactly when the known graph is a DAG, so the new rules are a strict generalization rather than a competing framework.","When the effect is not identifiable, Algorithm 2 (CIDME) enumerates candidate expressions by orienting each undirected edge of a shortest obstruction, producing a multiset that can be used for bounding the effect.","Unconditional identification is a special case (Z = ∅) of the conditional formula, unifying the new result with earlier MPDAG identification work."],"supporting_citations":[{"why":"Provides the PCO algorithm and unconditional identification formula in MPDAGs that Theorem 3 generalizes.","marker":"Perković (2020)"},{"why":"Supplies the original do calculus, truncated factorization, and the notion of identifiability that Theorems 3 and 6 build on.","marker":"Pearl (2009)"},{"why":"Defines MPDAGs and the orientation rules R1–R4 used throughout, including the claim that [G] contains DAGs with both orientations of each undirected edge.","marker":"Meek (1995)"},{"why":"Defines conditional adjustment sets in MPDAGs and provides examples where no such set exists, motivating the need for CIDM.","marker":"LaPlante and Perković (2024)"},{"why":"Gives the CIDP algorithm for PAGs; Appendix H shows a naive translation is incomplete for MPDAGs, motivating the new completeness proof.","marker":"Jaber et al. (2022)"},{"why":"Extends the do calculus to ancestral graphs, the main comparator for Theorem 6.","marker":"Zhang (2008)"},{"why":"Establishes the IDC algorithm for conditional identification in DAGs, the setting CIDM generalizes.","marker":"Shpitser and Pearl (2006)"}],"fun_headline_variants":["Complete algorithm for identifying conditional causal effects","When are conditional causal effects identifiable in MPDAGs?","MPDAGs: do calculus and complete identification algorithm","Deciding identifiability of subgroup effects in MPDAGs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume that every variable influencing two or more measured variables is itself measured, so no hidden common causes exist; if hidden common causes are present, the identification formula, the do-calculus rules, and the completeness guarantee can all fail.","fun_headline_variants_meta":{"raw":{"variants":["Complete algorithm for identifying conditional causal effects","When are conditional causal effects identifiable in MPDAGs?","MPDAGs: do calculus and complete identification algorithm","Deciding identifiability of subgroup effects in MPDAGs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1542,"prompt_tokens":897,"completion_tokens":645,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":579}},"tokens_in":513,"tokens_out":645,"duration_ms":7183,"temperature":1.0,"reasoning_tokens":579,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:22:03.438915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate all small MPDAGs and, for each triple (X,Y,Z), compare CIDM's verdict against an exhaustive do-calculus search over the DAGs in [G]; any case where CIDM returns FAIL yet an identifying expression exists, or returns an expression that differs from the true interventional distribution under a linear Gaussian density consistent with G, would refute the soundness and completeness claims.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines MPDAGs and the orientation rules R1–R4 used throughout, including the claim that [G] contains DAGs with both orientations of each undirected edge."},{"cited_title":"and Perkovi \\'c , E","cited_arxiv_id":null,"evidence_quote":"Defines conditional adjustment sets in MPDAGs and provides examples where no such set exists, motivating the need for CIDM."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the CIDP algorithm for PAGs; Appendix H shows a naive translation is incomplete for MPDAGs, motivating the new completeness proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends the do calculus to ancestral graphs, the main comparator for Theorem 6."},{"cited_title":"and Pearl, J","cited_arxiv_id":null,"evidence_quote":"Establishes the IDC algorithm for conditional identification in DAGs, the setting CIDM generalizes."}],"review_version":1}