acceptodds
Under review as a conference paper at ICLR 2027

(Almost) Automating Causal Inference With Tractable Probabilistic Circuits

24 views

Abstract

Approximate inference can lead to wrong conclusions when working with causal effects. For exact causal inference, the community has started using Probabilistic Circuits (PCs). However, these PCs have only been shown to work on discrete domains and only for backdoor adjustment, and as we show, existing previous methods fail to represent the dependencies present in the data under general vtree structures. We therefore develop a general-purpose causal inference engine for PCs that cover continuous and mixed domains, general vtree structures, and identifiable interventional queries. To this end, we pinpoint where existing circuits fall short: their conditioning operation is not exactly evaluable for a broad class of estimands, and their construction loses the capacity to model dependencies on general vtree structures. We develop the tools that overcome both: a diversity-preserving circuit construction, and disjoint continuous leaves that extend exact inference to mixed domains. We introduce T-ID, a tractability-aware extension of the ID algorithm, whose soundness and completeness we prove. T-ID takes as input a query, a causal graph, and a circuit, and outputs identifiability, tractability relative to the weak marginal determinisms available in the circuit, and a compositional query with a stated complexity bound. Reproducible experiments on synthetic continuous and mixed structural causal models corroborate our theoretical predictions.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Tractable Probabilistic Circuits, Causal Inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yy67OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper develops a general-purpose framework for exact causal inference using probabilistic circuits (PCs), addressing limitations of existing methods that primarily support discrete variables and backdoor adjustment. It identifies two key shortcomings in prior approaches of Wang & Kwiatkowska (2023), and introduces T-ID, a tractability-aware extension of the ID algorithm, and proves its soundness and completeness for determining causal query identifiability and tractability. The paper validates the theoretical results through experiments on synthetic continuous and mixed structural causal models, reporting results consistent with the theoretical predictions.

Critical strengths

The paper appears to make a significant technical advance over the existing work of Wang and Kwiatkowska (2023).

Critical weaknesses

-The paper is very difficult to follow because many concepts are introduced through abstract definitions without sufficient intuition or illustrative examples. The presentation appears to assume substantial familiarity with probabilistic circuits and the prior work of Wang and Kwiatkowska (2023). Concrete examples demonstrating the key concepts would significantly improve the paper. This difficulty also prevented me from confidently assessing the technical soundness of the work.

-The claim that causal effects are mostly computed approximately is not accurate. The paper does not acknowledge the extensive literature on estimating backdoor estimands, including inverse probability weighting (IPW), propensity score matching, double/debiased machine learning (DML), and many more. Recent work has also investigated the estimation of general causal expressions produced by the ID algorithm, including Graph-based Complexity for Causal Effect Evaluation by Plug-in (AISTATS 2025), Estimating Causal Effects from Learned Causal Networks (ECAI 2024), and Estimating Identifiable Causal Effects through Double Machine Learning (AAAI 2021). The paper needs to clarify the specific advantages of probabilistic circuits. In particular, I'd expect a comparison of computational complexity and scalability.

Questions for authors

The paper states a complexity bound of O(C^K) for T-ID, where C denotes circuit size and K is associated with the complexity of composition. How do these quantities relate to the size of the underlying structural causal model (SCM) or causal DAG or causal estimand? How do they scale with the graph structure or the complexity of the query, and under what conditions does the bound remain practically tractable as the problem size increases?

How does the proposed approach compare with well-known estimators such as IPW and DML, as well as recent methods for estimating general expressions produced by the ID algorithm? What are the main advantages and limitations of the proposed approach in terms of computational complexity, scalability, and the range of causal queries it can handle?

Reviewer_YfSbOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper extends probabilistic circuits, making them more practical for causal inference: it extends PCs by adding an explicit conditioning operation, allowing it to evaluate identified causal estimands, and the method works on discrete/continuous/mixed data. The method is compared to one other, across four datasets, outperforming it in three of them.

Critical strengths
  • This seems like "the right" way to do causal inference with deep generative models: it integrates with the ID algorithm, provides exact inference, and certifies complexity of evaluating a causal estimand.
Critical weaknesses
  • I feel it's not written for the right audience: it solves a problem in causal inference, but it doesn't introduce PCs gently and clearly enough and doesn't motivate them enough for a causal inference reader.
Questions for authors

Major questions:

  1. Can you elaborate on the limitation mentioned on line 413, "the returned familar must be anticipated when the vtree and its labels are designed"? What does this mean practically for someone doing causal inference? What happens if they fail to anticipate it? What about experimental results showing what happens under misspecification?
  2. Generally, can you explain more why someone doing causal inference should care about exact inference and use this method? What's the benefit compared to using ID and then setting up a neural network or other standard estimator? Can you provide experiments to corroborate this?

Minor questions:

  1. on line 120, doesn't the expression need to mention YY? Also, C\mathcal{C} hasn't been formally introduced.
  2. on line 235, isn't that vacuously true? The only other place I see MM mentioned in line 206, where it's specified as a mediator and thus X≠MX\neq M.
Reviewer_DKZQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extends Wang & Kwiatkowska's (2023) method for causal inference using probabilistic circuits. It first introduces a weak marginal determinism property that generalizes marginal determinism and supports efficient conditioning, then presents a circuit construction algorithm that produces a circuit satisfying this property. Next, the paper extends the framework to continuous and mixed domains and proposes an algorithm that combines causal identification with circuit tractability. The paper evaluates learned distributions against analytical ground truth in synthetic experiments and reports improvements over MD-Net on several Bayesian-network benchmarks.

Critical strengths
  1. The generalized weak marginal determinism is a useful new property; it supports conditioning and runs in linear time. This is a reusable result for the community.
  2. The proposed method connects the identified causal formula to the structured requirements on the circuit and the computational cost of evaluating it. This is a valuable direction beyond manually constructing circuits for individual adjustment formulas.
Critical weaknesses

The continuous extension approach discretizes the treatment space into intervals. Consequently, despite continuous variables, the learned interventional distribution remains piecewise over the treatment value. Although Section 8 mentions this, a clearer evaluation of its effect on causal estimation would be beneficial.

The writing is not easy to follow. For example, key construction steps and the complete T-ID algorithm are pushed to the appendix. Adding a running example back to the main text to illustrate a causal query, the resulting T-ID output, and the necessary circuit properties would significantly improve readability.

Questions for authors

For a continuous treatment X in P(Y|do(X)), is the learned distribution constant within each treatment interval? How sensitive are the causal estimates to the number of intervals? What advantage does this approach offer over standard discretization: for instance, simply binning X into N categories and using a discrete model?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.