{"id":"033d8a96-970b-4bb4-8f63-f33345a08115","arxiv_id":"1908.07301","paper_version":8,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper shows that causality questions reduce to specifying sequential functional relationships among random variables, reconciling the major schools and giving elementary proofs of Pearl's intervention rules.","lead":"A statistics monograph argues that standard probability, built on random variables defined as functions of a sample point, is enough to model cause and effect, and that the approaches of Pearl, Robins, Rubin and others fit into one conventional framework. It rederives canonical results in causal inference, including parts of Pearl's intervention calculus, with elementary probability arguments.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's 'first time' proof claim for Pearl's Rules 1–2 is contradicted by the paper's own description of Appendix A as an unfinished attempt; this advertised novelty is not delivered.","rationale":"The paper's elementary derivations in Sections 4–6 are largely self-contained and compelling: the reconciliation of Rubin's potential outcomes with model (4.1) is explicit in remark (ii) of Section 4, the worked examples are instructive, and the broad thesis that standard probability can treat causal questions is supported independently of the Pearl calculus. I therefore do not treat the acyclic functional representation as a fatal flaw; the paper targets Pearl-type recursive models and explicitly acknowledges limitations in footnote 17. The most immediately decisive weakness is internal: the abstract's boldest claim, that Pearl's Rules 1 and 2 are proved for the first time, is undercut by the author's own statements that Appendix A is unfinished and unused. This is not a disagreement with consensus or a complaint about novelty level; it is an inconsistency between what the paper advertises and what it says it contains. The right remedy is the reader's conditional verdict: either complete the appendix with a genuine proof of Rules 1 and 2 under Pearl's conditions, or temper the abstract's claim. My agreement with the reader is partial because the reader's formal 'weakest_assumption' field points at the Section 4 modeling assumption, whereas I would locate the load-bearing concern in the appendix/abstract mismatch, although the reader's rationale also flags this issue.","tokens_in":63537,"tokens_out":5056,"duration_ms":57904,"concrete_test":"Read Appendix A (pp. 79–80) and check whether, for an arbitrary semi-Markovian causal graph, Rule 1 (insertion/deletion of observations) and Rule 2 (action/observation exchange) are derived from the defining equations without invoking the rule under proof or restricting to graphs where the conclusion is trivial. As a specific benchmark, instantiate Pearl's canonical graph for Rule 2 with a hidden common cause of T and Y and verify the algebra from Section 4 yields P(Y=y | do(T=t), Z=z) = P(Y=y | T=t, Z=z) only under the paper's stated conditions; if the derivation relies on imposing unconfoundedness that is stronger than Pearl's d-separation condition, the 'first time' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest advertised claim is that Pearl's calculus of intervention is 'examined anew, and its first two rules are formulated and proved by means of elementary probability for the first time.' But the manuscript itself repeatedly disclaims exactly this: Section 1, footnote 8 calls Appendix A 'an unfinished attempt', and Section 6 (p. 52) says 'Appendix A presents an attempt at understanding Pearl's calculus, but we shall make no use of it here.' An attempt that is explicitly unfinished cannot support a 'proved for the first time' claim. If the appendix does not contain a complete derivation of Rules 1 and 2 under Pearl's stated conditions, the abstract's central novelty must be withdrawn or downgraded to 'attempted.' This concern is independent of the rest of the paper's useful reconciliation, which may well stand; but it is the one advertised novel contribution singled out in the abstract, and the manuscript's own self-assessment says it is not delivered.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified, conventional-probability foundation for statistical causality. Its core model (4.1) represents each observation as functions of standard uniform random variables, with treatment generated from situation and external factors and the response generated from situation, treatment and further external factors; unconfoundedness is identified with conditional independence of those external factors. In this framework, Section 4 derives the standard identification of potential-outcome laws from observed conditional laws (Eq. 4.6) and the adjustment/integration formulas (4.7)–(4.8), with worked examples including Simpson's paradox and Lord's paradox. Section 5 develops second-level causal models and a reduction to first-level causality, including a version of Pearl's back-door criterion, and derives a Robins-type formula for joint treatment effects in a two-stage model (Eq. 5.19). Section 6 solves several benchmark examples from the causal-inference literature using elementary probability, including smoking and genotype, eelworms and fumigants, direct effects in a two-stage plan, sex discrimination, and case-control studies. An appendix is said to treat Pearl's calculus of intervention, but the text repeatedly describes it as an unfinished attempt that is not used in the paper.","tokens_in":63686,"tokens_out":5647,"duration_ms":62205,"significance":"If the central derivations are accepted, the paper makes a valuable contribution by showing that many results from the Pearl, Robins and Rubin schools can be derived in a single probability-space framework using elementary random-variable constructions, without invoking a separate causal calculus. The worked examples in Sections 4 and 6 are mostly self-contained and genuinely illuminating, and the treatment of unconfoundedness as conditional independence of the driving uniforms is conceptually clear. The paper is also honest about the observational assumptions that underlie causal conclusions. However, the strongest advertised novelty — a first elementary proof of Pearl's Rules 1 and 2 — is not delivered by the manuscript as it stands, since the relevant appendix is described by the author themselves as unfinished. In addition, the proof of Pearl's criterion in Section 5.1 is heuristic and the author explicitly states dissatisfaction with existing proofs. These issues affect the abstract's central claim and the completeness of Section 5.1, though they do not, in my view, undermine the value of the remaining derivations.","major_comments":[{"comment":"The abstract states that Pearl's first two rules of intervention are 'formulated and proved by means of elementary probability for the first time,' but the manuscript itself says in Section 1, footnote 8 that Appendix A is 'an unfinished attempt at understanding Pearl's calculus of intervention,' and Section 6 says 'Appendix A presents an attempt at understanding Pearl's calculus, but we shall make no use of it here.' Since Appendix A is the only place where Rules 1 and 2 could be proved, the advertised novelty is not supported by the manuscript as it stands. This is a load-bearing discrepancy: either the appendix must be completed to contain a rigorous derivation of the two rules, or the abstract and Section 1 must be revised to state that the calculus is only examined or attempted, not proved.","section":"Abstract; §1, footnote 8; §6, opening paragraph (p. 52)"},{"comment":"The proof of Pearl's criterion is presented as a sequence of informal 'constraint' arguments and concludes that 'eventually' the conditioning equations reduce to forms such as (5.13)–(5.14). This does not constitute a proof of the graph-theoretic criterion for arbitrary second-level models: the argument does not rigorously treat collider configurations, descendants of colliders, or non-atomic distributions, and footnote 50 admits 'We have not been satisfied with any proof of this result.' As written, Section 5.1 provides a heuristic justification rather than a theorem. Since the criterion is used repeatedly in the paper, the section should either be replaced by a complete proof or be explicitly presented as a motivation with the criterion attributed to Pearl and with the central claims of the paper not relying on it as a proved result.","section":"§5.1, Eqs. (5.10)–(5.14) and footnote 50"}],"minor_comments":[{"comment":"The derivation multiplies and divides by assignment probabilities such as P(τ(Ũ, x) = t) and P(τ'(Ũ', x', t, r) = t') without stating positivity conditions; the paper should state the support assumptions needed for these displayed identities, especially because the paper elsewhere works with discrete variables and zero-probability conditioning events.","section":"§5.2, derivation of Eq. (5.19)"},{"comment":"The identification formulas are written for discrete situations and treatments, with continuous cases addressed only in a footnote; since these formulas are central, a brief formal statement with the appropriate density or kernel notation for the continuous case would improve clarity.","section":"§4, Eqs. (4.3)–(4.6)"},{"comment":"There are several typographical errors, including 'it it is undeniable' in Section 1 and 'footone 66' in Section 6.4; these should be corrected in a final pass.","section":"Throughout"},{"comment":"The worked examples are valuable but often rely on support/positivity assumptions implicitly (e.g., dividing by P(Y = y, Z = z) in Section 6.1); adding a sentence about such assumptions in each example would make the derivations fully rigorous.","section":"§6.1–6.5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a long survey with a strong advertised novelty claim. Given the author's own description of Appendix A as an unfinished attempt, the abstract's 'first time' proof claim should be corrected or realized before publication. The book announcement in the note may explain the provisional character of the appendix, but the arXiv version under review should still meet the journal's standards for the claims it makes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll tell you up front what to take from this: it's a teaching survey, not a proof-of-Pearl's-calculus paper. The expository idea—define all variables recursively on one sample space, treat observed data as evaluations at a single draw ω—is genuinely useful for statisticians who want Pearl, Rubin, and Robins in one language. The worked examples (Simpson, Lord, mediation, case-control, IV) are clear and mostly self-contained. But the abstract's 'first time' claim about proving Pearl's rules 1 and 2 is contradicted by the manuscript itself: Appendix A is called 'an unfinished attempt,' and Section 6 says the calculus is not used in the paper. You can't advertise a first proof that the paper disowns.\n\nWhat's actually new? Not much mathematics. Model (4.1) is a standard structural-equation model with uniform noise. The main body re-derives adjustment formulas by elementary probability; that's valuable for pedagogy, not as a novel result. The author is honest about the limits: no feedback, no interventions that change the functions, and unconfoundedness is a modeling assumption, not something data can verify. The examples are worked correctly as far as I can see, and the author credits Freedman, which is right.\n\nThe real soft spot is the gap between claim and delivery. If the two rules are proved in the appendix, the author should say so plainly. If they are not, the abstract has to be toned down. History makes it worse: Freedman is cited for elementary derivations, so 'for the first time since they were stated' is dubious no matter what the appendix contains. Minor issue: 84 pages for a survey with no new theorems is long; the Section 1 polemic could be trimmed.\n\nDoes it deserve peer review? Yes. The survey is coherent and the examples are worth checking. A serious editor should not desk-reject it. But the referee should demand a one-page statement of what is proved, what is conjectured, and what the actual contributions are beyond exposition. After that revision, it would be a solid foundations-of-statistics piece. As is, I wouldn't cite the novelty claim, only the examples.","headline":"Good teaching survey, but the advertised first proof of Pearl's rules is contradicted by the paper's own appendix disclaimer.","tokens_in":64199,"tokens_out":3625,"would_cite":false,"duration_ms":38543,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62A99","62F99","62E99","60J99","60K99","62P10","62P20","62P25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Standard probability, not a separate causal calculus, is sufficient for the core results of statistical causality.","keywords":["causality","Markovian models","identifiability","confounding","potential outcomes","unconfoundedness","intervention calculus","random variables"],"falsifier":"Simulate or observe a system with feedback, for example one where the current treatment and current response influence each other, apply the paper's unconfounded-adjustment estimator $\\sum_x E(R\\,|\\,T=t,X=x)P(X=x)$ to the observational data, and compare it with the mean response under a randomized intervention that sets $T=t$; if the two disagree while the data satisfy the paper's conditional-independence checks, the representation (4.1) is not the true mechanism.","tokens_in":63318,"feed_emoji":"🎲","tokens_out":9600,"duration_ms":97608,"temperature":0.7,"pith_summary":"This paper claims that questions of cause and effect in statistics can be posed and answered within standard probability, using nothing more than random variables, conditional expectations, and carefully chosen functional equations. Its proposal is a single representational scheme for causal modelling in which each variable is defined as a function of earlier variables and of unobserved uniform \"noise\" variables, evaluated at one random draw. Within that scheme, a conditional-independence condition called unconfoundedness makes the law of the potential response to a fixed treatment equal to the observed conditional law of the response given treatment and situation, and that identity is shown to drive the main results of the intervention calculus and of potential-outcome analysis. The paper also gives elementary probability proofs of the first two rules of the calculus of intervention, and works out classical examples as routine calculations. If this is right, statistical causality is not a separate formalism but a chapter of applied probability, and the main schools of causal inference differ mostly in notation and emphasis rather than in substance.","feed_headline":"Causal inference is just applied probability","feed_subtitle":"A single random-variable model unifies intervention, potential outcomes, and confounding in conventional probability terms.","key_machinery":"The load-bearing object is the random-variable model (4.1): $R_n=\\rho(V_n,X_n,T_n)$ and $T_n=\\tau(U_n,X_n)$, with $U_n,V_n$ standard uniform conditionally on $X_n$ and all variables defined as functions on one probability space, so observed data are one draw $\\omega$. This model turns causal questions into questions about laws of random variables formed by freezing the treatment argument at a fixed value $t$, i.e. the potential outcomes; the key identity is (4.6), which equates $\\mathcal{L}(R_n\\,|\\,X_n=x,T_n=t)$ with $\\mathcal{L}(\\rho(V_n,x,t))$ under unconfoundedness. The intervention model (4.2), obtained by replacing $T_n$ by a fixed sequence, represents what would happen if treatments were forced, and the paper shows that under unconfoundedness observational data estimate exactly that intervention model. The temporal precedence and acyclicity built into the functions $\\rho$ and $\\tau$ are what make the probability space a causal model rather than merely a joint distribution.","core_discovery":"The central discovery is that a causal model is a set of random variables defined recursively as functions of primitive random variables, with the random draw acting as the single source of randomness; distributional assumptions alone do not encode causal order. Concretely, the model $R_n = \\rho(V_n,X_n,T_n)$, $T_n = \\tau(U_n,X_n)$, with $U_n$ and $V_n$ standard uniform conditionally on $X_n$, says that each observation consists of one realized value of a system whose situation precedes treatment and whose treatment precedes response. Under the extra condition that $U_n$ and $V_n$ are conditionally independent given $X_n$ -- unconfoundedness -- the causal effect of treatment is fully captured by the laws $\\mathcal{L}(\\rho(V_n,x,t)\\,|\\,X_n=x)$, and these laws coincide with the observed conditional laws $\\mathcal{L}(R_n\\,|\\,X_n=x,T_n=t)$. From this single identity follow the adjustment formula, the propensity-score reduction, the reduction of a richer causal graph to the basic model whenever a suitable set of variables is conditioned on, and a formula for the joint effect of two treatments applied over time. The same framework gives elementary proofs of the first two rules of the calculus of intervention, stated directly in probability language.","pith_inferences":["The paper does not say this, but if its reconciliation is right, the disputes among the main schools of statistical causality are largely notational: their identification formulas are the same identities expressed in different languages, so a common exposition is possible.","One testable extension is to carry the same elementary-probability derivations through the third rule of the intervention calculus and through continuous-time or dynamic settings, which the paper leaves largely unworked.","Because the paper insists that observed data alone cannot distinguish a causal order from a reversed order with the same joint distribution, it implies that no amount of purely observational data can identify causal direction without subject-matter assumptions.","A practical check suggested by the construction is to simulate from a known structural system, estimate the paper's adjustment formula from the simulated observational data, and compare it with the true randomized-intervention distribution; disagreement flags violations of unconfoundedness or of the assumed functional order."],"forward_implications":["If model (4.1) holds with unconfoundedness, causal effects can be estimated by stratifying or matching on the situation variables, recovering the classical adjustment formula and the propensity-score method.","The first two rules of the intervention calculus follow from elementary conditional-probability manipulations, so those rules can be taught and verified without specialized graph-algebra machinery.","A detailed second-level causal graph can be reduced to the basic model whenever a suitable conditioning set is chosen; the paper gives a criterion for admissible sets and shows that conditioning on the wrong variables can create confounding.","For a two-stage treatment plan, the joint effect of the two treatments on the final response can be written as a sum of observed conditional probabilities, a special case of the general longitudinal formula.","Classical paradoxes such as Simpson's and Lord's become transparent illustrations of confounding and of the distinction between what happened and what may happen, once the underlying random-variable model is fixed."],"supporting_citations":[{"why":"Presents the intervention calculus and the criterion for admissible conditioning sets; this paper reformulates and proves the first two rules in elementary probability terms.","marker":"[47]"},{"why":"Supplies the do-operator, adjustment formula, and worked examples such as smoking and sex discrimination, which the paper re-derives with its random-variable model.","marker":"[54]"},{"why":"Provides an earlier standard-probability treatment of the smoking/genotype example that the paper follows in notation and extends to the potential-outcome framework.","marker":"[25]"},{"why":"Introduces the basic model (4.1) under unconfoundedness and gives direct verifications of the conditioning identities used in this paper.","marker":"[22]"},{"why":"Gives the potential-outcome formulation and the unconfoundedness condition shown in the paper to be a special case of model (4.1) with conditionally independent noises.","marker":"[31]"},{"why":"Introduces the propensity score, which the paper derives as a dimension-reduction tool within its model.","marker":"[59]"},{"why":"Contains the general longitudinal causal-effect formula; the paper's result for two sequential treatments is the special case worked out by elementary sums.","marker":"[57]"}],"fun_headline_variants":["Causality without new math","Unify Pearl, Robins, Rubin with one model","Standard probability proves causal calculus","One random variable model for all causality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach assumes that a real causal system can be captured by equations in which the situation is settled first, the treatment is then drawn as a function of the situation plus noise, and the response is then drawn as a function of the situation, treatment, and further noise, all observed at a single random draw; if the true system contains feedback, simultaneous determination, or interventions that change these functions, the identification formulas no longer apply.","fun_headline_variants_meta":{"raw":{"variants":["Causality without new math","Unify Pearl, Robins, Rubin with one model","Standard probability proves causal calculus","One random variable model for all causality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001751,"raw_usage":{"total_tokens":6935,"prompt_tokens":988,"completion_tokens":5947,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":5895}},"tokens_in":604,"tokens_out":5947,"duration_ms":36867,"temperature":1.0,"reasoning_tokens":5895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:19:47.812687+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or observe a system with feedback, for example one where the current treatment and current response influence each other, apply the paper's unconfounded-adjustment estimator $\\sum_x E(R\\,|\\,T=t,X=x)P(X=x)$ to the observational data, and compare it with the mean response under a randomized intervention that sets $T=t$; if the two disagree while the data satisfy the paper's conditional-independence checks, the representation (4.1) is not the true mechanism.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents the intervention calculus and the criterion for admissible conditioning sets; this paper reformulates and proves the first two rules in elementary probability terms."},{"cited_title":"and Jewel, N.P","cited_arxiv_id":null,"evidence_quote":"Supplies the do-operator, adjustment formula, and worked examples such as smoking and sex discrimination, which the paper re-derives with its random-variable model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides an earlier standard-probability treatment of the smoking/genotype example that the paper follows in notation and extends to the potential-outcome framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the basic model (4.1) under unconfoundedness and gives direct verifications of the conditioning identities used in this paper."},{"cited_title":"& Rubin, D.B","cited_arxiv_id":null,"evidence_quote":"Gives the potential-outcome formulation and the unconfoundedness condition shown in the paper to be a special case of model (4.1) with conditionally independent noises."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the propensity score, which the paper derives as a dimension-reduction tool within its model."},{"cited_title":"Causal diagrams for empirical re- search","cited_arxiv_id":null,"evidence_quote":"Contains the general longitudinal causal-effect formula; the paper's result for two sequential treatments is the special case worked out by elementary sums."}],"review_version":1}