acceptodds
Withdrawn from ICLR 2027

PEC: Improving Posterior Edge Confidence for Constraint-Based Causal Discovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Causal discovery has attracted growing attention across scientific fields, as most causal downstream analyses rely on the recovered causal graph. In practice, however, finite samples, incomplete data, and violated assumptions frequently yield erroneous edges, and this problem is aggravated in constraint-based algorithms, where errors in conditional independence (CI) tests can propagate through the subsequent decisions. In response, several methods also provide edge-wise scores that indicate how reliable each edge is, which helps verify the recovered causal graph. Yet these scores typically quantify the strength of causal influence or provide conservative bounds for statistical error control over the set of discovered edges, rather than the posterior probability of edge existence. We introduce Posterior Edge Confidence (PEC), a framework that estimates posterior edge probabilities for constraint-based causal discovery. PEC models the distributions of the test statistic under both hypotheses and combines them by probability rules derived from the logical structure of the skeleton search, rather than taking a single representative statistic. In experiments on synthetic datasets, we show PEC ranks edges more accurately than an existing representative-statistic baseline and improves the baseline as well.

Primary area:
causal reasoning
Keywords:
causal discovery, causal graph, edge confidence

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_nDYQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers the PC algorithm using Fisher’s Z tests and models the absolute values of the test statistics collected during its skeleton search as i.i.d. observations from a two-component Gaussian mixture model. It estimates test-level priors and likelihoods and converts them into posterior probabilities of edge existence. The method is evaluated on synthetic data generated from linear Gaussian DAG models.

Critical strengths

None

Critical weaknesses
  1. The modeling assumptions are insufficiently justified, and the interpretation of the resulting scores as posterior probabilities remains unsupported.

    • The paper assumes independence among Fisher’s Z test statistics. However, many tests reuse the same samples and involve highly overlapping conditioning sets, making complete independence an overly simplifying assumption. The authors neither analyze the effects of ignoring these dependencies on posterior estimation and calibration nor provide corresponding empirical validation.

    • The paper assumes that the Fisher’s Z test statistics are identically distributed. The authors fit a single mixture model to statistics obtained from different variable pairs and conditioning set sizes. However, the alternative effect distributions may differ across these tests. The paper does not justify the shared model or examine whether probabilities are well calibrated within each category of tests.

    • The paper assumes a linear Gaussian SEM. While this is not inherently problematic as a simplifying modeling choice, the resulting joint distribution is multivariate Gaussian. Under this setting, the experiments should compare against Bayesian methods that directly model graph structure in terms of edge probabilities, calibration, and computational cost, to establish the practical value of the proposed approximation.

  2. The empirical benefits are limited. PROPeR is the only external baseline, and comparisons with other uncertainty estimation methods are missing. Moreover, posterior calibration is actually worse (Table 1, PROPeR (MLE)).

Questions for authors
  1. The Introduction discusses several uncertainty estimation methods. Could the authors include representative methods as additional experimental baselines?

  2. Could the authors include a simple bootstrap baseline? Although bootstrap frequencies are not posterior probabilities, their ranking performance, probabilistic prediction performance, and computational cost can still be compared.

  3. Why is qmax⁡=3q_{\max}=3 fixed? What happens when truly non-adjacent variables require a separating set larger than 3?

  4. How does the method perform in non-linear and non-Gaussian settings?

  5. Under the paper’s linear Gaussian setting, the variables follow a multivariate Gaussian distribution. What advantages does the proposed method offer over frequentist uncertainty assessment methods or Bayesian structure learning methods that directly exploit this distributional structure? Could the authors explain why their modeling approach is necessary?

Reviewer_RwMoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper focuses on quantifying the uncertainty of the edges from the output of a constraint-based causal discovery algorithm. It proposes a framework named Posterior Edge Confidence to model the posterior probability of edge existence based on test statistic from hypothesis testing.

Critical strengths

None

Critical weaknesses
  • I find it difficult to understand the motivation of this work. There is no real-world use case supported by experiments in this paper. The performance shown by the experiment also seems very marginal.
  • The estimate of edge uncertainty methods may vary based on the distribution of the data. In Line 259-260, the paper chose mixture of normal distributions. For linear Gaussian data, one can simply rely on the p-value for each beta derived to account for the uncertainty of each edge. The only difference I see is to just estimate the posterior of the beta in this case. Also, a simple score subtraction from score-based discovery can already give a score to quantify edge uncertainty. It is unclear what the proposed method is adding besides calling it a new framework on top of the PC algorithm.
Questions for authors
  • Constraint-based causal discovery generally recover up to Markov equivalence class. With causal sufficiency, we have a CPDAG. Why is it important to know the posterior of some undirected edges in this case? Also, why is it designed for only constraint-based discovery methods? Do the authors have a specific use case for this problem setup where other methods cannot do anything about it?
  • Why can’t score-based causal discovery methods such as GES handle the edge uncertainty with decomposable scores? Particularly, suppose I use GES to get a CPDAG GG, now, we can compare the score difference between G−{e}G - \{e\} vs GG to get a score for an edge ee.
  • Why do the authors set the maximum conditioning set size in the experiment (Line 303)?
Reviewer_JcTeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PEC, a framework for estimating edge-existence probabilities in constraint-based causal discovery. During the skeleton search of the PC algorithm, PEC collects conditional independence test statistics across variable pairs and conditioning sets. It fits a two-component mixture model to the pooled statistics to estimate hypothesis-specific distributions and prior probabilities. The framework then uses probability combination rules based on the logical relationship between conditional independence statements and edge existence, producing a posterior confidence score for each variable pair.

Critical strengths
  1. The paper addresses the problem of quantifying uncertainty in individual edges returned by constraint-based causal discovery. This provides a clear methodological direction beyond conventional edge-stability frequencies or significance-based bounds.
  2. The proposed framework has a coherent and relatively simple implementation. It gathers test statistics during PC skeleton discovery, estimates a mixture distribution, and combines test-level evidence through explicit union and intersection rules. Applying the combination rules to an existing method also helps separate their contribution from the specific distribution estimator used by PEC.
  3. The reported synthetic experiments show improvements in average edge-discrimination metrics when the combination rules are used.
Critical weaknesses
  1. Equation (10) characterizes edge existence using a union or intersection over conditioning sets, but Algorithm 1 combines only tests actually performed during an adaptive PC search. With a maximum conditioning-set size of three and early stopping after an independence decision, relevant sets may be untested.
  2. The absolute null statistic follows a half-normal asymptotic distribution, and the pooled mixture proportions are test-level quantities rather than obviously edge-level priors.
  3. The evaluation is restricted to linear Gaussian synthetic DAGs with 10–20 variables and a maximum conditioning-set size of three. They do not establish whether the confidence scores remain useful for larger graphs, nonlinear mechanisms, or real datasets. A downstream graph-screening evaluation would also strengthen the practical motivation.
Questions for authors

Please refer to the weaknesses listed above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.