Pith. sign in

REVIEW 2 major objections 3 minor 33 references

Identifying Conditional Causal Effects in MPDAGs

T0 review · 2 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read 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.

desk verdict Solid completeness result for MPDAG conditional effects; the completeness proof has a repairable gap that a referee should ask to be fixed. read the letter →

arxiv 2507.15842 v1 pith:WXMM4B6U submitted 2025-07-21 cs.AI stat.MEstat.ML

Sara LaPlante , Emilija Perković This is my paper · ORCID
classification cs.AIstat.MEstat.ML MSC 62D2005C20
keywords conditionalcausaleffectsMPDAGidentificationdocalculusequivalenceclasssufficiencyidentifiabilitypartialordering
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

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

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

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

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.

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 (2)
  1. [Appendix G, proof of Theorem 7, Case 1] 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.
  2. [Section 5.4, completeness argument for Algorithm 1] 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.
minor comments (3)
  1. [Theorem 3, Equation (2)] 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.
  2. [Algorithm 2, lines 12-13] 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.
  3. [References and typography] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: claimed identifications are proven from graph lemmas and Pearl's do calculus; self-citations are not load-bearing.

full rationale

I walked the derivation chain. Theorem 3's expression (2) is derived in Appendix D from Lemma 38, whose proof applies Pearl's do calculus to an arbitrary DAG in [G]; the buckets and PCO output are graph constructs, and the final expression is not presupposed. Theorem 6 is obtained from Pearl's rules via Lemmas 49, 51, and 54, which are proven from MPDAG path properties; no rule of the MPDAG calculus is assumed to prove another rule. Theorem 7 (completeness) is a non-identifiability proof: it constructs two SEM families with f1(v) = f2(v) but different interventional means, using Wright's rule and multivariate normal computations; the conclusion is not an input to the construction. The self-citations (Perković 2020 for PCO and Lemma D.1; LaPlante and Perković 2024 for conditional adjustment; Venkateswaran and Perković 2024 for future work) are not load-bearing in a circular sense: PCO is restated in Appendix B, Lemma D.1 is used only as an analogy while Lemma 38 is proved independently, and the 2024 papers are context and examples. I do flag one manuscript-level issue: in Appendix G, the proof of Theorem 7 invokes Lemma 57 (and Lemma 58 in Case 2) without explicitly verifying Lemma 57's two standing assumptions in Case 1; this appears repairable by the without-loss-of-generality arguments (if a possibly causal undirected-starting path to Y exists, Proposition 41 finishes; otherwise use a prefix to Z). That is a proof-completeness gap, not a circular definition or renamed input. Section 6's stated limitation (no latent confounding) is an assumption, not circularity. No fitted parameter is renamed as a prediction, and no target quantity is used to define its own identifying expression.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

No free parameters are fitted; the paper is purely theoretical. The central derivations rest on standard graph-theoretic facts, Pearl's do calculus, and prior MPDAG results (Perkovic 2020, Meek 1995). The main domain assumptions are causal sufficiency and positivity, both stated explicitly.

assumptions (6)
  • domain assumption Markov compatibility of observational density with an MPDAG via factorization over a DAG in [G]
    Used throughout (Section 2) to relate graph separations to conditional independences.
  • domain assumption Positivity: f(v)>0 for all values
    Assumed in Section 2; needed for conditional densities and for the fractional form in Algorithm 1 (denominator f(z_D|do(x'),z_N)).
  • domain assumption Causal sufficiency: all variables observed, no latent confounding
    Stated in the abstract and used in the consistency definition; limits the scope of the results.
  • standard math Rules of Pearl's do calculus are valid for densities consistent with a DAG
    Invoked in proofs of Theorem 3, Theorem 6, and Proposition 5 via Theorem 10 (Appendix B).
  • standard math Recorded MPDAG properties from Meek (1995) and Perkovic (2020), including correctness of the PCO algorithm
    Borrowed as prior external results; the central proof structure relies on them (Appendices B-D).
  • standard math All DAGs represented by an MPDAG have the same adjacencies, unshielded colliders, and d-separations
    Stated in Section 2; foundational for defining densities consistent with an MPDAG.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Identifying Conditional Causal Effects in MPDAGs." pith.science (2026). https://pith.science/paper/WXMM4B6U

@misc{pith2026250715842,
  author       = {Pith},
  title        = {Pith review of: Identifying Conditional Causal Effects in MPDAGs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXMM4B6U}},
  note         = {Machine review of arXiv:2507.15842}
}
read the original abstract

We consider identifying a conditional causal effect when a graph is known up to a maximally oriented partially directed acyclic graph (MPDAG). An MPDAG represents an equivalence class of graphs that is restricted by background knowledge and where all variables in the causal model are observed. We provide three results that address identification in this setting: an identification formula when the conditioning set is unaffected by treatment, a generalization of the well-known do calculus to the MPDAG setting, and an algorithm that is complete for identifying these conditional effects.

Figures

Figures reproduced from arXiv: 2507.15842 by the authors.

Figure 1
Figure 1. Illustrating an MPDAG. Let the DAG in (a) represent an unknown causal model. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. MPDAGs used in Examples 1-3 thus, Equation (2) reduces to the exact form of the unconditional identification formula shown in Equation (4). For the second setting, suppose, as in Theorem 3, that Z ∩ PossDe(X, G) = ∅. But instead of excluding a subset of possibly causal paths from X to Y, consider a stronger assumption that Y ∩ PossDe(X, G) = ∅. In this setting, identification of the conditional effect reduces to an … view at source ↗
Figure 3
Figure 3. MPDAG used in Example 5 does this with the help of well-known probability rules, and it does this for an effect that our identification formula was able to find. We provide Examples 5 and 6 to highlight that our do calculus is able to find conditional effects that our identification formula cannot. Example 4 (Multiple Routes) Reconsider Example 1, where we found the form of a conditional effect using our identificat… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: MPDAGs used in Examples 7-9 Step (1) happens inside the while loop of lines 2-8. Here the algorithm successively transfers nodes from do(X) into Z using Rule 2 of the do calculus for MPDAGs (Theorem 6). The while loop ends once there is no longer a proper possibly caus…
Figure 5
Figure 5. Figure 5: DAGs used in Example 13 5.4 Soundness and Completeness Algorithm 1 is sound for identifying conditional causal effects given an MPDAG. That is, any expression the algorithm outputs will be an identifying form of f(y | do(x), z). This follows directly from Rules 2-3 of …
Figure 6
Figure 6. Figure 6: MPDAG used in Example 14 To calculate the same expectation under f1, recall that f1(v | do(x)) := f∗(v), where f∗(v) is the multivariate normal density generated by the following SEM with independent errors: V1 ← εv1 ; εv1 ∼ N (0, 1); (15) X ← x; Z ← 1 2 V1 + 1 2 x + ε…
Figure 7
Figure 7. Figure 7: An MPDAG used in Remark 27. Proof of Lemma 28. ⇐ Holds by Lemma 18. ⇒ Let p be a possibly causal path from X to Y in G = (V, E) and let p ∗ = ⟨X = V0, . . . , Vk = Y ⟩, k ≥ 1, be a shortest subsequence of p in G. By Lemma 24, p ∗ forms an unshielded, possibly causal pa…
Figure 8
Figure 8. Figure 8: Proof structure of Theorem 6. (i) All colliders on p are unshielded. (ii) The path in G corresponding to p is of definite status. Proof of Lemma 44. Let p = ⟨Z = V1, . . . , Vk = Y ⟩, Z ∈ Z, Y ∈ Y, k > 1. (i) Let Vi , i ∈ {2, . . . , k − 1}, be an arbitrary collider on…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 32 canonical work pages

  1. [1]

    Bareinboim, E., Brito, C., and Pearl, J. (2012). Local characterizations of causal B ayesian networks. In Graph Structures for Knowledge Representation and Reasoning: Second International Workshop , pages 1--17

    work page 2012

  2. [2]

    Chickering, D. M. (2002). Optimal structure identification with greedy search. Journal of Machine Learning Research , 3:507--554

    work page 2002

  3. [3]

    and Bareinboim, E

    Correa, J. and Bareinboim, E. (2020). A calculus for stochastic interventions: Causal effect identification and surrogate experiments. In The AAAI Conference on Artificial Intelligence , volume 34, pages 10093--10100

    work page 2020

  4. [4]

    and He, Y

    Fang, Z. and He, Y. (2020). IDA with background knowledge. In Uncertianty of Artifical Intelligence , pages 270--279

    work page 2020

  5. [5]

    Guo, F. R. and Perkovi\'c, E. (2021). Minimal enumeration of all possible total effects in a M arkov equivalence class. In International Conference on Artificial Intelligence and Statistics

    work page 2021

  6. [6]

    Henckel, L., Perkovi \'c , E., and Maathuis, M. H. (2022). Graphical criteria for efficient total effect estimation via adjustment in causal linear models. Journal of the Royal Statistical Society: Series B , pages 579--599

    work page 2022

  7. [7]

    Jaber, A., Ribeiro, A., Zhang, J., and Bareinboim, E. (2022). Causal identification under M arkov equivalence: C alculus, algorithm, and completeness. In Advances in Neural Information Processing Systems , volume 35, pages 3679--3690

    work page 2022

  8. [8]

    Jaber, A., Zhang, J., and Bareinboim, E. (2018). A graphical criterion for effect identification in equivalence classes of causal diagrams. In International Joint Conference on Artificial Intelligence

    work page 2018

Show all 33 references
  1. [9]

    Kivva, Y., Etesami, J., and Kiyavash, N. (2023). On identifiability of conditional causal effects. In Uncertainty in Artificial Intelligence

    2023

  2. [10]

    and Perkovi \'c , E

    LaPlante, S. and Perkovi \'c , E. (2024). Conditional adjustment in a M arkov equivalence class. In Artifical Intelligence and Statistics

    2024

  3. [11]

    L., Dawid, A

    Lauritzen, S. L., Dawid, A. P., Larsen, B. N., and Leimer, H.-G. (1990). Independence properties of directed M arkov fields. Networks , 20(5):491--505

    1990

  4. [12]

    Maathuis, M. H. and Colombo, D. (2015). A generalized back-door criterion. Annals of Statistics

    2015

  5. [13]

    H., Colombo, D., Kalisch, M., and B \"u hlmann, P

    Maathuis, M. H., Colombo, D., Kalisch, M., and B \"u hlmann, P. (2010). Predicting causal effects in large-scale systems from observational data. Nature methods , 7(4):247--248

    2010

  6. [14]

    H., Kalisch, M., and B \"u hlmann, P

    Maathuis, M. H., Kalisch, M., and B \"u hlmann, P. (2009). Estimating high-dimensional intervention effects from observational data. Annals of Statistics , 37:3133--3164

    2009

  7. [15]

    V., Kent, J

    Mardia, K. V., Kent, J. T., and Bibby, J. M. (1980). Multivariate Analysis (Probability and Mathematical Statistics) . Academic Press London

    1980

  8. [16]

    Meek, C. (1995). Causal inference and causal explanation with background knowledge. In Uncertianty of Artifical Intelligence , pages 403--410

    1995

  9. [17]

    Moffett, L., Weissman, A., McCormick, M., Weiland, C., Hsueh, J., Snow, C., and Sachs, J. (2023). Enrollment in pre-k and children's social-emotional and executive functioning skills: To what extent are associations sustained across time? Journal of Educational Psychology , 115(3):460

    2023

  10. [18]

    H., and Richardson, T

    Nandy, P., Maathuis, M. H., and Richardson, T. S. (2017). Estimating the effect of joint interventions from observational data in sparse high-dimensional settings. Annals of Statistics , 45(2):647--674

    2017

  11. [19]

    Pearl, J. (1995). Causal diagrams for empirical research. Biometrika , 82:669--688

    1995

  12. [20]

    Pearl, J. (2009). Causality: Models, Reasoning, and Inference . Cambridge University Press

    2009

  13. [21]

    Perkovi \'c , E. (2020). Identifying causal effects in maximally oriented partially directed acyclic graphs. In Uncertianty of Artifical Intelligence , pages 530--539

    2020

  14. [22]

    Perkovi \'c , E., Kalisch, M., and Maathuis, M. H. (2017). Interpreting and using CPDAG s with background knowledge. In Uncertianty of Artifical Intelligence

    2017

  15. [23]

    H., et al

    Perkovi \'c , E., Textor, J., Kalisch, M., Maathuis, M. H., et al. (2018). Complete graphical characterization and construction of adjustment sets in markov equivalence classes of ancestral graphs. Journal of Machine Learning Research , 18(220):1--62

    2018

  16. [24]

    ausler, D., Ernest, J., and B\

    Rothenh\"ausler, D., Ernest, J., and B\"uhlmann, P. (2018). Causal inference in partially linear structural equation models: identifiability and estimation. Annals of Statistics , 46:2904--2938

    2018

  17. [25]

    and Pearl, J

    Shpitser, I. and Pearl, J. (2006). Identification of conditional interventional distributions. In Uncertainty in Artificial Intelligence , pages 437--444

    2006

  18. [26]

    Shpitser, I., VanderWeele, T., and Robins, J. M. (2010). On the validity of covariate adjustment for estimating causal effects. In Uncertainty in Artificial Intelligence

    2010

  19. [27]

    Spirtes, P., Glymour, C., and Scheines, R. (2000). Causation, Prediction, and Search . MIT Press, second edition

    2000

  20. [28]

    Van der Zander, B., Liskiewicz, M., and Textor, J. (2014). Constructing separators and adjustment sets in ancestral graphs. In Uncertianty of Artifical Intelligence , pages 11--24

    2014

  21. [29]

    and Perkovi \'c , E

    Venkateswaran, A. and Perkovi \'c , E. (2024). Towards complete causal explanation with expert knowledge. arXiv preprint arXiv:2407.07338

    2024 arXiv

  22. [30]

    Wright, S. (1921). Correlation and causation. Journal of Agricultural Research , 20(7):557--585

    1921

  23. [31]

    Wright, S. (1934). The method of path coefficients. The Annals of Mathematical Statistics , 5(3):161--215

    1934

  24. [32]

    Zhang, J. (2006). Causal Inference and Reasoning in Causally Insufficient Systems . PhD thesis, Carnegie Mellon University

    2006

  25. [33]

    Zhang, J. (2008). Causal reasoning with ancestral graphs. Journal of Machine Learning Research , 9(7)

    2008

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.