acceptodds
Under review as a conference paper at ICLR 2027

From Representation Identifiability to Counterfactual Identifiability in Latent Structural Causal Models

Abstract

This paper investigates whether latent structural causal models (SCMs) learned from factual data can be trusted on counterfactuals, that is, whether latent SCMs that fit the same factual data agree on every counterfactual. To answer this question, we define counterfactual identifiability on latent SCMs, introduce the latent SCM isomorphism that implies it, and establish sufficient assumptions for reaching the isomorphism: an auxiliary variable that modulates the exogenous variables and one perfect intervention on each non-root endogenous variable. We further show that the isomorphism supports causal effect analysis at the counterfactual level, identifying natural direct and indirect effects up to a common scaling. Our theoretical results extend causal representation learning by turning representation identification, in the form of the isomorphism, into a guarantee for counterfactual identification. Finally, we develop FlowSCM, an estimator built on our assumptions and theory, and show on synthetic and image data that it recovers the latent SCM and returns its counterfactuals and causal effects.

Primary area:
causal reasoning
Keywords:
causal representation learning, counterfactual identification, causal effects, identifiability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_s5kMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This study investigates the conditions under which a latent SCM learned from factual data can be generalized to counterfactual data. It shows that the latent SCM is identifiable up to isomorphism under assumptions involving regime-dependent distributions of exogenous variables and a single, unknown target of a perfect intervention in each regime. The study then shows that SCM isomorphism implies counterfactual identifiability (Chen and Du, ICML 2025), allowing certain causal effects to be identified up to a common scaling factor.

Critical strengths
  1. The study shows when a latent SCM learned from factual data can be generalized to counterfactual data, based on the concept of SCM isomorphism (Chen and Du, ICML 2025).

  2. The study provides a practical estimation framework for the proposed model.

Critical weaknesses
  1. The assumptions seem rather strong and restrictive. The framework requires variation in the distributions of exogenous variables across regimes, as well as perfect interventions on individual latent variables. The intervention assumption seems particularly restrictive and difficult to justify in practice, although some existing studies use a similar assumption (Kügelgen et al., NeurIPS 2023).

  2. The estimation framework has no consistency guarantee and seems difficult to optimize on high-dimensional data. In particular, computing the log-determinant of the Jacobian may be unstable. The numerical experiments consider only a small number of latent variables (n=5), leaving scalability and numerical stability at higher dimensions unclear.

Questions for authors

See Weaknesses.

Reviewer_mNtaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper established a connection between latent SCM isomorphism and counterfactual identifiability, i.e., the former implies the latter. Under certain assumptions, SCM isomorphism is proved to be guaranteed under observationally equivalence. A corresponding learning algorithm is also proposed, showing ideal identification in terms of latent variables, counterfactual prediction, and causal effect estimation on synthetic datasets.

Critical strengths
  1. This paper discusses counterfactual identifiability, which is an important but seldomly studied problem in previous works.
  2. Though the relation between SCM isomorphism and counterfactual identifiability seems to be obvious in my view, I appreciate the efforts that precisely formalizing all related concepts.
Critical weaknesses
  1. My main concern lies on the impact of the theoretical results. The assumptions for successful counterfactual identification are way too restrictive. The authors include (1) sufficient variability that requires multiple observational distributions, and (2) known intervention over each non-root node that requires multiple interventional distributions, for identifiability of an SCM. Such conditions can hardly be met in real applications, making this work mainly a theoretical contribution. As for the theoretical part, the connection between SCM isomorphism and counterfactual identifiability is obvious for me and does not bring much new understandings. Intuitively, such relation can further generalize by removing the affine class restriction, if the latent SCM does not insist a form of additive noise (Correct me if there are misunderstandings). Overall, my preliminary opinion leans toward rejection.
Questions for authors
  1. Please discuss the potential scenarios that FlowSCM can be applied. I mean practical scenarios that meet the requirements of training data, rather than a set of assumptions that restrict the SCM.
  2. In Definition 3, why “do-intervention” uses a separated symbol τ\tau instead of reusing ρ\rho? Is this necessary?
Reviewer_4A5fOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Counterfactual identifiability was introduced by Chen and Du in their 2025 paper Exogenous Isomorphism for Counterfactual Identifiability. It is defined as a property of an structural causal model (SCM) wrt a set of assumptions, such that any other SCM satisfying those assumptions generates the same counterfactual distributions, which are the joint distributions produced by a single sample of exogenous variables under any finite set of do-interventions. Note that counterfactual identifiability is weaker than identifying the exact model, which is the usual goal in causal representation learning (CRL).

The paper works with the standard latent SCM from CRL, like that of Brehmer et al in their 2022 paper Weakly supervised causal representation learning, restricted to additive noise mechanisms. For this particular setting, a modified definition of counterfactual identifiability is given, which also generalises to perfect and soft interventions. It is shown that to achieve counterfactual identifiability we need to be able to observe the following auxiliary variables

  1. Variation in the exogenous noise settings, assuming gaussian noise with the standard sufficient variability conditions from nonlinear ICA literature (Hyvärinen et al., 2019, Khemakhem et al., 2020)

  2. Variation in the interventional regime settings, assuming availability of all regimes with perfect single node interventions, taken from multi-environment CRL literature (von Kügelgen et al., 2023)

Provided we have these ingredients, the paper shows that the latent SCM is identifiable up to affine isomorphisms, which then implies counterfactual identifiability.

Finally, the paper proposes FlowSCM, a maximum likelihood estimator of the latent SCM evaluated on synthetic data and rendered images. It learns a separate noise distribution for each exogenous noise setting, as in iVAE (Khemakhem et al., 2020), and learns the target node for each intervention regime, building on DCDI (Brouillard et al., 2020) and ILCM (Brehmer et al., 2022).

Critical strengths

The paper brings the idea of counterfactual identifiability to CRL, which is likely to be far more useful than the usual claims of identifying the causal graph and latent variables up to diffeomorphisms. It's particularly interesting to see that element-wise diffeomorphisms between the latent variables does not imply counterfactual identifiability, as shown in lemma 2.

Critical weaknesses

The paper's headline claim that "latent SCMs that fit the same factual data agree on every counterfactual" for its given set of assumptions is misleading, because of the modified definition of counterfactual consistency for latent SCMs. So what "two models agreeing on a counterfactual" actually means here is that there exists a single intervention correspondence ω:ι↦ι^\omega: \iota \mapsto \hat{\iota} such that whenever an intervention ι\iota acts on the first model, the generated counterfactual distribution agrees with the second model when it's acted on by ι^=ω(ι)\hat{\iota} = \omega (\iota). The problem is that ω\omega is unknown, so we are still left unable to answer a specific real-world counterfactual, even if we had a fitted model that satisfied the assumptions.

Additionally, the additive-noise mechanism and the variability assumptions on available data are all very restrictive. No practical example is given anywhere in the paper.

Questions for authors
  1. Given a fitted FlowSCM and no ground truth, how can I recover the intervention correspondence, i.e. which latent variable is which and its scale and offset, in order to pose a specific real-world counterfactual?

  2. Is there a practical setting that provides both the auxiliary labels uu and vv, that also satisfies your assumptions?

  3. Do you have any insight into which of the current assumptions could be loosened, for future work?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.