{"id":"21894935-f860-438b-8a06-7df790007b9c","arxiv_id":"2601.20405","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Causal estimands are classified by whether they depend only on marginal potential outcome distributions (intervention layer), joint or nested distributions (counterfactual layer), or individual-level outcomes.","lead":"This paper maps Pearl's three-level causal hierarchy onto the probability distributions of potential outcomes, so that each causal estimand is classified by how much probabilistic detail it needs. A generalist might read it to understand why some causal questions are easy to answer while others require strong, often unverifiable assumptions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ATT is classified as second layer in Example 1, but Pearl's hierarchy treats E[Y(0)|A=1] as a layer-3 counterfactual, so the marginal/joint criterion as stated does not faithfully recast Pearl's hierarchy.","rationale":"The reader's weakest assumption was that the marginal/joint/individual classification is asserted rather than derived from Pearl's SCM hierarchy. My stress test finds a concrete instance where that equivalence actually fails: ATT, listed as a second-layer estimand in Example 1, is a layer-3 counterfactual under Pearl's definitions because it conditions on observed treatment and targets E[Y(0)|A=1]. This shows the Section 3.1 criterion cannot be taken as a faithful operationalization as stated. The paper is otherwise carefully organized, with a useful survey of identification strategies and many correct classifications, so the flaw appears fixable by revising the definition to exclude conditioning on the treatment variable and reclassifying ATT (and any analogous estimands) as third-layer counterfactual marginals. For that reason I would keep the reader's CONDITIONAL verdict rather than escalate to rejection, but the required revision is now a substantive correctness issue, not merely a novelty claim.","tokens_in":15301,"tokens_out":14076,"duration_ms":126287,"concrete_test":"For a binary treatment with a confounder U, set U~Bernoulli(0.5), A=U, and Y=U+epsilon with epsilon~N(0,1). Compute ATT = E[Y(1)-Y(0)|A=1] and ATE = E[Y(1)]-E[Y(0)]. Show that E[Y(0)|A=1] differs from E[Y(0)] (equivalently from P(Y|do(A=0))), so ATT requires the conditional distribution of U given A=1, which is not available from interventional marginals alone. Then apply the paper's Section 3.1 classification rule verbatim: because ATT's components are marginals conditional on A=1, the rule assigns layer 2; the SCM hierarchy assigns layer 3 because P(Y(0)|A=1) is a counterfactual conditional requiring abduction. This single counterexample settles whether the proposed criterion is a faithful operationalization of Pearl's hierarchy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mapping in Section 3.1 defines layer-2 estimands as functionals of marginal potential-outcome distributions, explicitly allowing conditioning on observed treatment, and Example 1 classifies ATT = E[Y(1)-Y(0)|A=1] as second layer. Under the SCM hierarchy the paper itself cites (Section 2.2, Bareinboim et al. 2022), a query of the form P(Y(0)|A=1) is a counterfactual conditional: it conditions on the endogenous treatment value A=1 and requires knowledge of the joint dependence between A and Y(0). Interventional marginals P(Y|do(A=0)) and P(Y|do(A=1)) do not determine this dependence in general. For example, in a simple SCM with U~Bernoulli(0.5), A=U, and Y=U+epsilon, P(Y(0)|A=1) differs from P(Y(0)) and requires the distribution of U given A=1 (abduction). Thus ATT is a layer-3 quantity under Pearl's definitions. The paper's criterion would classify it as layer 2 because it depends only on marginal distributions conditional on A=1, revealing that the 'marginal vs joint' distinction is not equivalent to Pearl's hierarchy. The parenthetical in the Section 3.1 definition introduces observed-treatment conditioning into 'marginal'; Figure 2 only shows P(Y(a)|X), not P(Y(a)|A=1), so Example 1 is inconsistent with the paper's own figure. This is not a mere formal quibble: layer 2 information (all interventional distributions) is insufficient for ATT in general, so the classification would mislead readers about which layer a query belongs to.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a potential-outcomes interpretation of Pearl's causal hierarchy. The central rule (Section 3.1) classifies an estimand as second-layer if it is a functional of marginal potential-outcome distributions, and third-layer if it requires the joint distribution of potential outcomes, nested/cross-world quantities, or individual-level counterfactuals. The authors apply this rule to a broad set of estimands (ATE, ATT, QTE, DRF, PN, TBR, persuasion rate, ITE distribution, principal causal effects, mediation, and fairness metrics) and review identification strategies such as monotonicity, association parameters, copulas, rank preservation, and conformal inference. The paper is written as a perspective/survey rather than as a new technical result.","tokens_in":15694,"tokens_out":10957,"duration_ms":100856,"significance":"If the proposed classification were valid, it would provide a practical and useful map from scientific questions to the probabilistic objects required for identification, organizing a large and fragmented literature. The paper's strengths are its broad coverage, its many explicit examples, and its systematic review of identification strategies for joint and individual-level estimands. However, the central equivalence between the proposed classification and Pearl's SCM-based hierarchy is not established, and there is a concrete counterexample (ATT) where the paper's rule contradicts the SCM hierarchy it cites. Because this equivalence is the paper's main conceptual contribution, the current version cannot be accepted as a faithful recasting of Pearl's hierarchy.","major_comments":[{"comment":"The paper classifies ATT = E[Y(1)-Y(0)|A=1] as a second-layer estimand because it is written in terms of marginals conditioned on observed treatment. Under the SCM hierarchy cited in Section 2.2, however, E[Y(0)|A=1] is a counterfactual conditional: it conditions on the endogenous treatment value A, and the interventional distributions P(Y(0)) and P(Y(1)) do not determine it in general. For example, in the SCM with U~Bernoulli(0.5), A=U, and Y=A+U+epsilon, E[Y(0)] = 0.5+E[epsilon] but E[Y(0)|A=1] = 1+E[epsilon]; the two interventional marginals are the same across models that differ in this conditional, so layer-2 information is insufficient for ATT. Pearl's hierarchy therefore places ATT at layer 3, not layer 2. This is not a terminological quibble: ATT is the first worked example, and the parenthetical definition of 'marginal' is what drives the misclassification. The proposed criterion is not equivalent to Pearl's hierarchy as stated.","section":"Section 3.1, Example 1"},{"comment":"The second-layer definition permits conditioning on observed treatment and covariates, but Figure 2 displays only P(Y(1)|X) and P(Y(0)|X), not P(Y(a)|A=1). This inconsistency matters because Pearl's hierarchy distinguishes pre-treatment covariates X, which may be conditioned on at layer 2, from post-treatment variables such as A, conditioning on which is a counterfactual (abduction) operation. Collapsing X and A into a single notion of 'marginal' produces the ATT error and leaves the classification rule ambiguous for other conditional estimands. The paper should state formally which variables may be conditioned on at each layer; if conditioning on A is moved to layer 3, then ATT and similar conditional-on-treatment estimands must be reclassified, and the 'marginal vs joint' dichotomy needs to be replaced by a more precise criterion.","section":"Section 3.1, Figure 2"},{"comment":"The claimed correspondence with Pearl's hierarchy is asserted rather than derived. Section 3.1 and Figure 2 present the classification as the definition of the layers, but no proof or formal argument establishes that membership in Pearl's layers is equivalently characterized by whether an estimand depends on marginal, joint, or individual-level potential outcomes. The ATT counterexample shows that this is not merely a missing proof: the asserted equivalence is false as stated. The authors should either revise the classification rule and re-derive the layer assignments, or explicitly present their taxonomy as a distinct potential-outcomes hierarchy that is related to, but not identical with, Pearl's.","section":"Section 3.1 (general framework)"}],"minor_comments":[{"comment":"The comparison table uses symbols such as '/' and ',' in the 'Rank Preservation' and 'Conformal Inference' columns without explaining their meaning; the comparison would be clearer with explicit statements of whether each method has weaker conditions, point identification, and generalizability.","section":"Table 3"},{"comment":"The sentence 'a Gaussian copula assumes that (Y(1),Y(0)) follows a joint Gaussian distribution' is imprecise; a Gaussian copula with arbitrary marginal distributions does not imply joint normality of the potential outcomes.","section":"Section 3.3.2"},{"comment":"Counterfactual parity is stated as an equality constraint, P(S(1)=1)=P(S(0)=1), rather than as an estimand; the intended estimand or target of inference should be clarified.","section":"Example 10"},{"comment":"The notation P(Y(a)|A=a',Y(a')) in the layer-3 row of Table 1 is not defined in the text; it should be defined or cross-referenced to the SCM notation in Section 2.2.","section":"Table 1"},{"comment":"There are several encoding artifacts in the references and text (e.g., 'Hern´ an', '¤ects', missing spaces in citations); these should be cleaned before publication.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on a cluster of recent works by the same authors (Wu et al. 2024a, 2025a; Wang et al. 2025b; Wu and Mao 2025) for load-bearing identification claims. Independent verification or a clearer separation between literature review and the authors' own contributions would strengthen the manuscript. The fit with the journal is acceptable as a perspective piece, but the central equivalence to Pearl's hierarchy must be corrected before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clear, systematic survey that sorts a wide range of causal estimands by whether they need marginal, joint, or individual-level potential-outcome information. That's worth having. But there is a load-bearing classification error: Example 1 puts the ATT in the second layer, and under the SCM hierarchy the paper itself cites, that is wrong. P(Y(0)|A=1) is a counterfactual conditional; knowing all interventional distributions does not determine it in general. The paper's parenthetical 'conditional on observed treatment' in the layer-2 definition is an attempt to make room for ATT, but it breaks the correspondence with Pearl and is inconsistent with Figure 2, which only shows marginals conditional on X. This is not a minor labeling quibble – it misleads readers about what information ATT really requires.\n\nThe paper does a lot well. The placement of PN, TBR, persuasion rate, ITE distribution, principal causal effects, and NDE/NIE in the third layer is accurate. The review of identification strategies – independence, monotonicity, copulas, data fusion, conformal inference – is solid and clearly organized. The sublayer distinction between cross-world and individual-level counterfactuals is useful. If the table and examples were cleaned up, this would be a handy reference for applied researchers.\n\nThe novelty claim is stronger than the evidence: 'no work has formally examined Pearl's causal hierarchy from a potential outcomes lens' is not true in substance. The distributional treatment-effects literature (Fan and Park 2009; Heckman et al. 1997) and the SWIG formulation already build on the marginal/joint distinction. The authors cite these works but don't engage them as direct predecessors. They should reposition the paper as a systematic classification rather than a first formalization.\n\nThe paper also leans on several of the authors' own preprints for key results. That's not a sin, but for a survey it would be better to flag which claims are independently verified and which are new.\n\nBottom line: send it to peer review. The framework is worth publishing after the ATT classification and the novelty framing are fixed. A serious referee would catch these issues, and they are fixable.","headline":"Useful systematic classification of causal estimands by potential-outcome information, but the ATT placement in layer 2 is a real error that contradicts the SCM-based hierarchy the paper itself cites.","tokens_in":16162,"tokens_out":6547,"would_cite":true,"duration_ms":57042,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper recasts Pearl's causal hierarchy in potential-outcomes terms, sorting causal estimands by whether they require marginal, joint, or individual counterfactual information.","keywords":["causal hierarchy","potential outcomes","counterfactuals","intervention layer","joint distribution of potential outcomes","individualized treatment effects","causal estimands","identification assumptions"],"falsifier":"Construct any causal estimand the paper assigns to Layer 3 and show, for some data-generating process, that it is point-identified from the marginal potential-outcome distributions alone under randomization — or symmetrically, an estimand assigned to Layer 2 whose identified set depends on the joint distribution. A concrete test: in a binary-treatment, binary-outcome randomized trial, compute the sharp identified set for the persuasion rate $P(Y(1)=1\\mid Y(0)=0)$ from the marginals $P(Y(1))$ and $P(Y(0))$ alone; if this set is ever a single point without any dependence assumption, the claimed boundary collapses.","tokens_in":15082,"feed_emoji":"🧩","tokens_out":9549,"duration_ms":76588,"temperature":0.7,"pith_summary":"Pearl's causal hierarchy sorts causal questions into association, intervention, and counterfactuals. This paper argues that the boundary between the intervention and counterfactual layers can be drawn precisely in potential-outcomes language: an estimand belongs to the intervention layer exactly when it depends only on the marginal distributions of potential outcomes, and to the counterfactual layer when it depends on their joint distribution, on nested outcomes across different interventions, or on individual-level counterfactual outcomes. On this view, randomization identifies the intervention layer, while counterfactual estimands need extra assumptions about dependence between potential outcomes. The paper applies this criterion across a broad range of estimands, including benefit and harm rates, persuasion, principal effects, fairness metrics, and mediation, and uses it to explain why individual-level decisions cannot rest on conditional average treatment effects alone. The practical payoff is a map from a scientific question to the estimand, the probabilistic object it requires, and the assumptions needed for identification.","feed_headline":"Marginal, joint, or individual: a new rule sorts causal estimands","feed_subtitle":"A potential-outcomes lens reveals what information and hidden assumptions each causal estimate requires.","key_machinery":"The load-bearing object is the pair of probabilistic distinctions the paper draws: the marginal distributions of potential outcomes versus their joint distribution, and population-level joint quantities versus individual realized counterfactual outcomes. The classification rule is that an estimand belongs to the intervention layer exactly when it can be written as a functional of marginal distributions under single interventions, and to the counterfactual layer when it requires the joint distribution, nested cross-world outcomes, or individual-level outcomes. Because each unit realizes only one potential outcome, the joint distribution is never directly observed, so every identification strategy for a counterfactual-layer estimand must add an assumption that pins down the dependence between potential outcomes; the paper catalogs these assumptions and shows how they achieve point identification or partial identification.","core_discovery":"The central claim is that Pearl's three-layer causal hierarchy can be operationalized at the level of estimands by a distributional criterion. The intervention layer consists of estimands that are functionals of the marginal potential-outcome distributions $P(Y(a)\\mid X)$; the counterfactual layer consists of estimands that are functionals of the joint distribution $P(Y(0),Y(1)\\mid X)$, involve nested potential outcomes such as $Y(1,M(0))$, or target individual realized counterfactual outcomes. The paper splits the counterfactual layer into cross-world queries and individual-level queries, and argues that this ordering tracks identifiability demands: randomization identifies marginals; monotonicity, association parameters, copula restrictions, rank preservation, or data fusion are needed for the joint; and individual outcomes require a deterministic view of counterfactuals plus an SCM-based abduction-action-prediction step or rank preservation, with conformal inference providing prediction intervals. This correspondence with Pearl's hierarchy is illustrated in Figure 2 and applied systematically in Examples 1–11.","pith_inferences":["A corollary the paper does not fully spell out is an 'information hierarchy' for causal functionals: each move up a layer must add exactly one new piece of dependence information, so the framework could be used to audit causal claims for hidden assumptions.","The same marginal-versus-joint lens could be extended to dynamic treatment regimes, where nested potential outcomes under sequences of interventions may form additional sublayers beyond the two the paper distinguishes.","The asserted equivalence between the potential-outcome layers and Pearl's SCM hierarchy is testable: if a formal SCM layer-3 query could not be translated into potential-outcome notation, the two hierarchies would not be isomorphic.","A practical extension would be an automated layer checker: given a symbolic expression for a causal estimand, detect whether it contains cross-world or individual-level terms and assign the corresponding layer."],"forward_implications":["Researchers can use the marginal/joint/individual criterion to check whether a chosen estimand matches the scientific question, rather than settling for a convenient proxy.","Because randomization identifies only marginal potential-outcome distributions, any counterfactual-layer conclusion drawn from a randomized trial is valid only under an additional, often implicit assumption about the dependence between potential outcomes.","Individual-level treatment recommendations cannot be justified by conditional average treatment effects alone; harm rates, benefit rates, or prediction intervals for the individual treatment effect are the relevant third-layer quantities.","Mediation estimands such as natural direct and indirect effects are inherently cross-world, so mediation analyses require assumptions beyond those identifying average treatment effects.","The framework clarifies when proposed identifiability assumptions are sufficient or overly restrictive for a given estimand, as summarized in the paper's Table 2."],"supporting_citations":[{"why":"Defines the structural causal model and the three-layer causal hierarchy that the paper recasts in potential-outcomes terms.","marker":"Pearl 2009"},{"why":"Provides the formal complete identification results for the causal hierarchy that set the layer distinctions being operationalized.","marker":"Shpitser and Pearl 2008"},{"why":"Serves as the authoritative reference for the scope and foundations of Pearl's hierarchy.","marker":"Bareinboim et al. 2022"},{"why":"Introduces the potential-outcomes notation the paper adopts throughout.","marker":"Rubin 1974"},{"why":"Establishes the potential-outcome model for randomized experiments that underlies the marginal-identification claims.","marker":"Neyman 1990"},{"why":"Supplies the textbook treatment of potential outcomes and average treatment effects used as the layer-2 baseline.","marker":"Imbens and Rubin 2015"},{"why":"Defines natural direct and indirect effects with nested potential outcomes, a central layer-3 example.","marker":"Pearl 2001"},{"why":"Introduces principal stratification, the basis for principal causal effects classified as layer 3.","marker":"Frangakis and Rubin 2002"}],"fun_headline_variants":["A distributional key for Pearl's causal hierarchy","New rule maps causal questions to hidden assumptions","Beyond marginals: classifying causal estimands by counterfactual depth","What your causal estimate secretly requires","From marginal to individual: a map for causal estimands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The classification assumes that Pearl's layer distinctions coincide exactly with whether an estimand depends on marginal, joint, or individual-level potential-outcome information; the paper illustrates this alignment with examples and Figure 2 but does not prove exhaustiveness from the formal SCM definitions.","fun_headline_variants_meta":{"raw":{"variants":["A distributional key for Pearl's causal hierarchy","New rule maps causal questions to hidden assumptions","Beyond marginals: classifying causal estimands by counterfactual depth","What your causal estimate secretly requires","From marginal to individual: a map for causal estimands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000587,"raw_usage":{"total_tokens":2753,"prompt_tokens":940,"completion_tokens":1813,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1740}},"tokens_in":556,"tokens_out":1813,"duration_ms":10601,"temperature":1.0,"reasoning_tokens":1740,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:39:19.573785+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct any causal estimand the paper assigns to Layer 3 and show, for some data-generating process, that it is point-identified from the marginal potential-outcome distributions alone under randomization — or symmetrically, an estimand assigned to Layer 2 whose identified set depends on the joint distribution. A concrete test: in a binary-treatment, binary-outcome randomized trial, compute the sharp identified set for the persuasion rate $P(Y(1)=1\\mid Y(0)=0)$ from the marginals $P(Y(1))$ and $P(Y(0))$ alone; if this set is ever a single point without any dependence assumption, the claimed boundary collapses.","supporting_citations":[],"review_version":2}