Equilibrium Causal Digital Twins: Validation, Transport, and Identification Limits

Faraz Dadgostari; Neda Nazemi

REVIEW 2 major objections 5 minor 12 references

Equilibrium Causal Digital Twins: Validation, Transport, and Identification Limits

T0 review · 2 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read For feedback systems, experimental agreement alone cannot certify a digital twin's counterfactuals.

desk verdict The paper's central impossibility result has a proof gap involving the selection rule's dependence on the solution set, but the surrounding theory is careful and worth engaging. read the letter →

arxiv 2607.21667 v1 pith:3VII6AE6 submitted 2026-07-23 stat.ME cs.MAcs.SYeess.SYmath.OC

Faraz Dadgostari , Neda Nazemi This is my paper · ORCID

classification stat.MEcs.MAcs.SYeess.SYmath.OC MSC 62D2062F03

keywords equilibriumcausalgamedigitaltwincounterfactualvalidationtransportabilitycyclicstructuralmodelpartialidentificationinference

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks when a digital twin of a feedback-driven system can be trusted to answer intervention questions, and when that trust survives a change of domain. It claims that, for equilibrium causal games, population-level agreement with a sufficient set of staged interventional distributions does identify per-unit counterfactuals, provided the mechanisms are strictly monotone, the equilibrium selection rule is stable, and the design is query-sufficient (Theorem 5). Without those structural assumptions, a finite collection of experiments cannot certify a cross-world prediction: two systems can match every experimental law in the design and still disagree on the counterfactual (Theorem 31). The paper therefore draws the boundary between what can be validated from data alone and what requires an a priori class assumption, and it gives transport rules for reusing a source twin in a target domain, including a hybrid that re-estimates only the changed mechanisms.

What carries the argument

The central object is the equilibrium causal game (ECG): a cyclic structural equation model $V_i = f_i(V_{\mathrm{pa}(i)}, U_i)$ with independent noises and a declared measurable equilibrium selection rule Sel. The argument is carried by three devices: Lemma 4 (quantile abduction under feedback), which rewrites each monotone mechanism through its interventional quantile function so that factual observations fix noise ranks; the complete legal fibre, the set of all models matching the retained observational and staged laws, on which query identification is exactly constancy; and the cyclic selection diagram $\mathcal{D} = G \cup \{S_i \to i : i \in D\}$, which locates the changed mechanisms. Theorem 31's witness-admissible reshuffle, a parent-dependent, measure-preserving rank reflection at a node the design never probes, is the engine of the impossibility result.

What would settle it

Enumerate all two- and three-node linear equilibrium causal games with strictly monotone additive noises and a fixed selection rule; for a query-sufficient design, test whether any two models agree on every staged interventional distribution yet differ on a per-unit counterfactual, since Theorem 5 predicts no such pair exists, and any such pair found would refute the positive identification claim.

Watch

Extended reading notes

Core claim

The central discovery is that validation of equilibrium counterfactuals is query-specific and class-conditional. Under strict noise monotonicity (T1), a factual observation recovers each unit's private-noise rank; with a stable selection rule (T3) and a design whose experiments pin the query-relevant interventional kernels (T4), agreement on complete staged interventional distributions forces equality of full-factual per-unit counterfactuals almost everywhere (Theorem 5). Partial factual information requires a stronger singleton condition on the complete legal fibre (T2). Conversely, Theorem 31 constructs, for any finite design, a witness-admissible node whose mechanism can be reshuffled by a parent-dependent, measure-preserving transformation that is invisible to every law in the design yet changes the counterfactual by a positive gap; hence cross-world validation is possible only conditional on class membership. Transport results include cyclic selection diagrams, direct reuse under ancestral separation, and hybrid transport that replaces only the changed mechanism–noise pairs on the post-surgery ancestral closure of the query.

Load-bearing premise

The positive results assume the exogenous noises are mutually independent across nodes in both domains (diagonal covariance), so a change of domain can only alter per-node mechanisms and marginals; if the joint noise distribution can change while every node marginal stays fixed, the validation, transport, and identification criteria do not apply.

Editorial extensions

If this is right

Validation protocols for feedback twins must test complete interventional kernels, not just means and covariances; the paper shows that moment agreement does not identify distributional queries, even in linear models (a Gaussian and a shifted exponential can share every additive moment yet differ in tail probabilities).
Direct reuse of a source twin in a target is valid when the post-intervention ancestral set of the query contains no changed mechanism and the retained mechanisms satisfy cross-domain invariance; otherwise a hybrid that re-identifies only the changed mechanism–noise pairs on that ancestral closure suffices.
A finite experiment design cannot certify cross-world counterfactuals without class assumptions; Theorem 31 gives explicit witness-admissibility conditions (an unprobed node, a non-descendant parent moved by the intervention, and a nonzero resolvent coefficient) under which indistinguishable twins differ.
In linear models, identifying changed mechanisms requires an intervention count that depends on the observation model and graph support: all discrepancy rows in the ambient unknown-support branch, $m-1$ anchors for a full rotational source block, and potentially fewer for query-specific designs when the query is constant on the remaining fibre.
When point identification fails, the sharp identified set is the image of the complete legal fibre, which under semialgebraic conditions is a finite union of points and intervals with explicit endpoints.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension: any practical validation pipeline for feedback twins should report the width of the counterfactual range implied by the complete legal fibre whenever the monotone boundary cannot be certified, rather than a single predicted value.
The impossibility result implies that certification standards relying on behavioral equivalence under a finite test suite are sound only if the structural class itself is enforced by design, for example when monotone mechanisms are physically guaranteed; this consequence for engineering practice is left implicit in the paper.
The hybrid transport rule suggests a model-agnostic recipe for domain adaptation under feedback: freeze invariant mechanisms estimated in the source, re-estimate the target rows, and re-solve the full equilibrium; the paper does not connect this to the broader domain-adaptation literature.
The separation between interventional-layer testability and cross-world non-testability could be probed empirically by fitting two models to a real feedback system that agree on all available experiments yet differ in the paper's witness construction, then measuring the counterfactual divergence under a hidden intervention.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's my take. The paper has a genuine centerpiece—Theorem 31's finite-design impossibility—and that centerpiece has a proof gap that matters. The rest is solid.

What's new: cyclic selection diagrams, a validation/transport theory for equilibrium causal games, and linear intervention counts. The paper is unusually careful: each positive result states its conditions, Theorem 8 shows the four conditions are necessary by construction, and the limitations section is honest. The distinction between validation, transport, and identification is well drawn.

The problem: Theorem 31's construction of S^- uses a parent-dependent rank-band reflection at r. The proof claims this preserves 'every conditional mechanism law and hence every regime in the finite design.' That inference fails when the selection rule Sel depends on both the rank vector and the solution set. After the reshuffle, the equations change, so the solution set changes for some rank draws. Sel(tau, B'(tau)) can then pick a different equilibrium than Sel(tau, B(tau)), changing the law in a design regime. The two-node cycle example in the stress-test note makes this concrete. So the impossibility result is not established for multi-equilibrium systems; it needs an added condition (unique equilibria, or Sel invariant under the reshuffle-induced map on solution sets). This is not a minor caveat—it's the load-bearing result. The theorem may be repairable, but as stated it's too strong.

Other soft spots: key support is deferred to two companion preprints (Section 2.2, Corollary 7), which slows verification. And a lot of the guarantees are existential—no numerical constants or code. That's acceptable for a theory paper, but worth knowing.

Who gets value: anyone working on counterfactuals in feedback systems, transportability, or digital twin validation. It deserves a serious referee, but the referee should press hard on Theorem 31. I'd send it to peer review with a request for major revision, and specifically require the authors to fix the selection-rule issue.

Referee Report

2 major / 5 minor

Summary. The paper develops a theory of validation, transport, and identification for counterfactual queries in feedback systems modeled as equilibrium causal games (ECGs). It separates four conditions -- monotone noise mechanisms, a complete legal fibre condition, stable equilibrium selection, and query-sufficient design -- and shows how, under these conditions, population-level agreement with staged interventional distributions identifies per-unit equilibrium counterfactuals. It introduces cyclic selection diagrams for transport, gives linear-model intervention requirements, characterizes partial identification when point identification fails, and provides statistical procedures for mean and covariance comparison. It also states an impossibility theorem asserting that no finite experimental design can validate cross-world counterfactual predictions without structural class assumptions.

Significance. If the main claims hold, the paper is a substantive contribution to causal inference in cyclic systems, where acyclic transport and validation methods are known to fail. Its strengths are the explicit separation of abduction, selection, and design support; the necessity constructions in Theorem 8; the query-specific design framework; and the sharp partial-identification bounds in Section 9. The main qualification is that the central impossibility result, Theorem 31, is not fully proved as stated for multi-equilibrium systems.

major comments (2)

[Section 8, Theorem 31; Appendix A.3] The assertion that the reshuffled system S^- agrees with S^+ on every law of every regime in the finite design is not established for multi-equilibrium systems. After the parent-dependent rank reset rho_pi, the selected outcome under S^- is Sel(tau, B(rho_pi(tau))), whereas under S^+ it is Sel(tau, B(tau)). The map rho_pi being measure-preserving for each parent value implies only that tau and rho_pi(tau) have the same marginal law; it does not imply that the laws of Sel(tau, B(rho_pi(tau))) and Sel(tau, B(tau)) coincide when Sel depends on the actual rank vector tau, which T3 explicitly allows. The proof's statement that preserving conditional mechanism laws preserves every regime law is therefore insufficient under feedback. The theorem needs an additional argument or an additional condition, such as uniqueness of equilibrium in every relevant regime, or invariance of the selection rule under the reshuffle-induced map on solution sets. The same gap affects Corollary 32, which applies Theorem 31 in the switch-augmented model.
[Section 2.2 and Corollary 7] Several claims are explicitly deferred to companion preprints: payoff and curvature interventions are covered only after the reward-layer conditions of the companion reward paper hold, and the nonlinear source-block positive result cited in Corollary 7 is attributed to a companion representation paper. Because these results are used in the main text, the manuscript should either include the needed arguments or clearly mark those statements as conditional on unpublished work, so that the journal version can be assessed as a standalone contribution.

minor comments (5)

[Section 3, Lemma 4] The notation T^+ is used before it is formally defined; please define the class T^+ (and T^+(D)) at its first occurrence.
[Section 4, Proposition 6 and Theorem 5] There are formatting artifacts such as inline line breaks in displayed formulas and the phrase "the y+x 2 counterexample"; the latter should presumably read "the y+x^2 counterexample." Please clean up these artifacts.
[Section 2.4] The standing exogenous-independence assumption excludes shared-marginal copula shifts, and the paper notes this exclusion. Because this materially limits the transport and identification statements, the limitation should be highlighted in the abstract or introduction, not only in Section 2.4.
[Section 8, Theorem 31] The statement should specify explicitly whether S^- inherits the same selection rule Sel as S^+; the proof currently leaves this implicit, which contributes to the gap described in the first major comment.
[Section 11, Proposition 50] The rank-degenerate covariance branch is clearly scoped, but the exposition is extremely dense; a short intuitive summary of when hard-clamp versus soft-probe rank behavior differs would help readers apply the result correctly.

Circularity Check

2 steps flagged · score 3.0 of 10

Mostly self-contained; fibre-identification statements are definitional restatements, but core validation/transport and impossibility results have independent content.

self definitional [Section 5, Definition 16 and Proposition 17]
"Proposition 17 (Pointwise transport identification). The query is pointwise transport-identified at (d,θ⋆) if and only if QI(Fd(θ⋆)) is a singleton. ... Proof. This is the definition of identification on the complete evidence fibre; no global measurable factorization through Ed is required."

Fd(θ⋆) is defined as the fibre of parameter values with identical evidence Ed, and QI is the query law. 'Pointwise transport identification' is then defined as constancy of QI on exactly that fibre. The proposition's iff is therefore a restatement of the definition, and the proof explicitly concedes this. It has no independent derived content; it is a definitional equivalence presented as a proposition.
self definitional [Section 10, Theorem 42]
"Put ∆χ(D) = sup θ′,θ′′∈KD ∥χ(θ′)−χ(θ ′′)∥Q, V qry min(χ) = min{|D|: ∆ χ(D) = 0}, ... Thus V qry min(χ) is the minimum query-identifying intervention-block cost, and population query identification is exactly complete-fibre constancy."

V_qry_min is defined as the minimum number of design blocks for which the query diameter on the complete fibre KD is zero. The theorem then concludes that the minimum query-identifying cost is exactly complete-fibre constancy. Since 'query-identifying' was defined by that same constancy, the identification criterion is true by construction. The surrounding inequalities involving V_FO_min and K_ID_min are substantive, but the identification statement itself is a tautological unpacking of the definition.

full rationale

The paper's derivation chain is mostly self-contained. Lemma 4 derives per-unit counterfactual equality from shared interventional kernels, rank abduction via strict monotonicity, and a rank-measurable selection rule; Theorems 5, 10, 12, and 15 apply that lemma under explicit structural and design hypotheses. Theorem 8 and Theorem 31 are constructive necessity/impossibility witnesses backed by explicit calculations, not circular reductions. No fitted parameter is renamed as a prediction, and no external benchmark is claimed. The identification statements at the fibre level, however, are definitional: Proposition 17 is admittedly just the definition of identification on the complete evidence fibre, and Theorem 42 defines the query-identifying cost as the minimal fibre-width-zero design before asserting that query identification is exactly fibre constancy. These are transparent definitional restatements and are not what carries the substantive validation/transport theorems, which rely on Lemma 4 and the explicit witnesses. The companion-paper citations (Dadgostari 2026; Dadgostari and Nazemi 2026) appear, but the cited passages are scope/backbone comments rather than the proof of the main theorems; they are not load-bearing enough to constitute circularity. The skeptical concern about Theorem 31's selection rule is a mathematical correctness question about whether the reshuffled twin is observationally equivalent under T3, not a circularity, so it does not affect this score. Overall score 3: minor definitional restatements and self-citations, while the central claims have independent content.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The central contribution is a set of conditional identification results. All load-bearing premises are explicit assumptions about the model class, solvability, selection, and definability. No parameters are fitted to data, and no physical entities are introduced; the only new objects (cyclic selection diagrams, legal fibres) are mathematical constructs defined within the paper.

assumptions (7)

domain assumption T1 strict noise monotonicity: each mechanism f_i(v,.) is strictly increasing in its private noise.
Used in Lemma 4 for rank abduction; Theorem 8 and Theorem 31 show this condition is necessary for the validation conclusion.
ad hoc to paper Exogenous independence (standing): U_i are mutually independent with diagonal covariance Omega in both source and target domains.
Section 2.4 explicitly excludes shared-marginal copula shifts that could move the equilibrium law without changing node marginals.
domain assumption Shared graph, shared intervention semantics, and shared SCC-local selection rule Sel across domains (T3).
Theorems 10, 12, and 15 require the same selection rule; the T3 necessity construction in Theorem 8 shows that different selection rules can change counterfactuals by more than 17/5.
ad hoc to paper Complete legal fibre and design sufficiency (T4, Definition 2): the query kernel is constant on the fibre of models matching the design responses.
This is the definition of query identification; Theorem 42 shows it is exactly the minimal experiment-design criterion.
domain assumption SCC solvability and well-posedness of the augmented model, including the Forre-Mooij hypothesis for mixed S-profiles.
Carried verbatim in Rule-D (Theorem 14); the linear route avoids it via Lemma 9.
standard math Definability and compactness conditions for the Lojasiewicz modulus: the quotient is compact, closed, Hausdorff, and definable in a common polynomially bounded o-minimal expansion.
Needed for Proposition 6 and Corollary 7; imported from van den Dries-Miller and Miller's dichotomy.
domain assumption Target hybrid lies in T+ (monotone noise, complete-fibre pushforward-kernel singleton, shared Sel) for unit-level counterfactual transport.
Theorem 15 requires these boundary-class conditions; below the boundary only bounded transfer holds.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Equilibrium Causal Digital Twins: Validation, Transport, and Identification Limits." pith.science (2026). https://pith.science/paper/3VII6AE6

@misc{pith2026260721667,
  author       = {Pith},
  title        = {Pith review of: Equilibrium Causal Digital Twins: Validation, Transport, and Identification Limits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3VII6AE6}},
  note         = {Machine review of arXiv:2607.21667}
}

read the original abstract

Digital twins are often used to predict how a system would respond to an intervention. In systems with feedback, a twin must reproduce an equilibrium counterfactual, and a twin developed in one domain may fail after mechanisms change. We study when these predictions can be validated and transported. For equilibrium causal games, we give conditions on the mechanisms, equilibrium selection, and intervention design under which agreement with experimental distributions identifies the counterfactual of interest. We show why agreement of means and covariances is insufficient for distributional queries. We then introduce cyclic selection diagrams and derive criteria for direct reuse and for hybrid models that combine invariant source mechanisms with target information. An impossibility result constructs systems that agree under every experiment in a finite design but disagree on the target counterfactual, showing that validation requires structural assumptions. For linear models, we derive intervention requirements that depend on the mechanisms that changed, the observation model, and graph support. When point identification fails, we characterize the remaining range of query values. We also provide statistical tests for reconstructed means and covariances and illustrate the theory in synthetic feedback systems.

Figures

Figures reproduced from arXiv: 2607.21667 by the authors.

**Figure 1.** Interventions reduce rotational ambiguity in the displayed source block. Each additional query-relevant row removes rotational degrees of freedom; the residual dimension reaches zero after |D| − 1 rows are pinned. This calculation concerns the stated rotational orbit and does not give a universal intervention count [PITH_FULL_IMAGE:figures/full_fig_p028_1.png] view at source ↗

**Figure 2.** Different model discrepancies require different diagnostics. The cells summarize the displayed constructions. A staged intervention distinguishes the moment-matched latent rotation, whereas structural-class analysis distinguishes the parentdependent reshuffle and the selection-branch change. The loop-gain row is marked “not established” because no matching construction is claimed here. 28 [PITH_FULL_IMAGE:figures/… view at source ↗

**Figure 3.** Hybrid reconstruction versus direct source reuse. In the displayed synthetic systems, the hybrid model of Theorem 12 reproduces the target response to numerical precision, whereas direct reuse of the source model retains the domain discrepancy [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Design relativity and the acyclification hazard. Panel A: in these examples, the target-intervention budget scales with the discrepancy set rather than the full system; no universal count is implied. Panel B: an acyclic reading misses the nonzero equilibrium effect in…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 6 canonical work pages

[5]

Stephan Bongers, Patrick Forré, Jonas Peters, and Joris M

doi: 10.1007/s41060-016-0028-8. Stephan Bongers, Patrick Forré, Jonas Peters, and Joris M. Mooij. Foundations of structural causal models with cycles and latent variables.The Annals of Statistics, 49(5):2885–2915,

work page doi:10.1007/s41060-016-0028-8
[1994]

Arash Nasr-Esfahany, Mohammad Alizadeh, and Devavrat Shah

doi: 10.2307/2160869. Arash Nasr-Esfahany, Mohammad Alizadeh, and Devavrat Shah. Counterfactual identifiability of bijective causal models. InInternational Conference on Machine Learning (ICML), PMLR 202, pages 25733–25754,

work page doi:10.2307/2160869
[1996]

doi: 10.1215/S0012-7094-96-08416-1. 40

work page doi:10.1215/s0012-7094-96-08416-1
[2008]

The digital twin counterfactual framework: A validation architecture for simulated potential outcomes.arXiv preprint arXiv:2604.01325,

Olav Laudy. The digital twin counterfactual framework: A validation architecture for simulated potential outcomes.arXiv preprint arXiv:2604.01325,

work page arXiv
[2012]

Elias Bareinboim and Judea Pearl

doi: 10.1609/aaai.v26i1.8232. Elias Bareinboim and Judea Pearl. Causal transportability with limited experiments. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 27, pages 95–101,

doi:10.1609/aaai.v26i1.8232
[2013]

Elias Bareinboim and Judea Pearl

doi: 10.1609/aaai.v27i1.8692. Elias Bareinboim and Judea Pearl. Causal inference and the data-fusion problem.Proceedings of the National Academy of Sciences, 113(27):7345–7352,

work page doi:10.1609/aaai.v27i1.8692
[2014]

Ilya Shpitser and Judea Pearl

doi: 10.1214/14-STS486. Ilya Shpitser and Judea Pearl. What counterfactuals can be tested. InProceedings of the 23rd Conference on Uncertainty in Artificial Intelligence (UAI), pages 352–359,

doi:10.1214/14-sts486
[2016]

39 Elias Bareinboim, Juan D

doi: 10.1073/pnas.1510507113. 39 Elias Bareinboim, Juan D. Correa, Duligur Ibeling, and Thomas Icard. On Pearl’s hierarchy and the foundations of causal inference. InProbabilistic and Causal Inference: The Works of Judea Pearl, pages 507–556. ACM,

doi:10.1073/pnas.1510507113

Show all 12 references

[2021]

Vincent Chin, John P

doi: 10.1214/21-AOS2064. Vincent Chin, John P. A. Ioannidis, Martin A. Tanner, and Sally Cripps. Effect estimates of covid- 19 non-pharmaceutical interventions are non-robust and highly model-dependent.Journal of Clinical Epidemiology, 136:96–132,

doi
[2022]

Gilles Blondel, Marta Arias, and Ricard Gavaldà

doi: 10.1145/3501714.3501743. Gilles Blondel, Marta Arias, and Ricard Gavaldà. Identifiability and transportability in dynamic causal networks.International Journal of Data Science and Analytics, 3(2):131–147,
[2023]

Judea Pearl and Elias Bareinboim

arXiv:2302.02228. Judea Pearl and Elias Bareinboim. External validity: From do-calculus to transportability across populations.Statistical Science, 29(4):579–595,

arXiv
[2026]

Version Number:

URLhttps://arxiv.org/abs/2607.19531. Version Number:

arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [5]

Stephan Bongers, Patrick Forré, Jonas Peters, and Joris M

doi: 10.1007/s41060-016-0028-8. Stephan Bongers, Patrick Forré, Jonas Peters, and Joris M. Mooij. Foundations of structural causal models with cycles and latent variables.The Annals of Statistics, 49(5):2885–2915,

work page doi:10.1007/s41060-016-0028-8

[2] [1994]

Arash Nasr-Esfahany, Mohammad Alizadeh, and Devavrat Shah

doi: 10.2307/2160869. Arash Nasr-Esfahany, Mohammad Alizadeh, and Devavrat Shah. Counterfactual identifiability of bijective causal models. InInternational Conference on Machine Learning (ICML), PMLR 202, pages 25733–25754,

work page doi:10.2307/2160869

[3] [1996]

doi: 10.1215/S0012-7094-96-08416-1. 40

work page doi:10.1215/s0012-7094-96-08416-1

[4] [2008]

The digital twin counterfactual framework: A validation architecture for simulated potential outcomes.arXiv preprint arXiv:2604.01325,

Olav Laudy. The digital twin counterfactual framework: A validation architecture for simulated potential outcomes.arXiv preprint arXiv:2604.01325,

work page arXiv

[5] [2012]

Elias Bareinboim and Judea Pearl

doi: 10.1609/aaai.v26i1.8232. Elias Bareinboim and Judea Pearl. Causal transportability with limited experiments. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 27, pages 95–101,

doi:10.1609/aaai.v26i1.8232

[6] [2013]

Elias Bareinboim and Judea Pearl

doi: 10.1609/aaai.v27i1.8692. Elias Bareinboim and Judea Pearl. Causal inference and the data-fusion problem.Proceedings of the National Academy of Sciences, 113(27):7345–7352,

work page doi:10.1609/aaai.v27i1.8692

[7] [2014]

Ilya Shpitser and Judea Pearl

doi: 10.1214/14-STS486. Ilya Shpitser and Judea Pearl. What counterfactuals can be tested. InProceedings of the 23rd Conference on Uncertainty in Artificial Intelligence (UAI), pages 352–359,

doi:10.1214/14-sts486

[8] [2016]

39 Elias Bareinboim, Juan D

doi: 10.1073/pnas.1510507113. 39 Elias Bareinboim, Juan D. Correa, Duligur Ibeling, and Thomas Icard. On Pearl’s hierarchy and the foundations of causal inference. InProbabilistic and Causal Inference: The Works of Judea Pearl, pages 507–556. ACM,

doi:10.1073/pnas.1510507113