acceptodds
Under review as a conference paper at ICLR 2027

Beyond Additivity: Causal Discovery in Location-Scale Noise Models with Hidden Variables

Abstract

We study causal discovery with hidden variables when causes modulate both the mean and the spread of their effects, a location-scale noise model; existing functional-model methods for hidden variables assume additive noise. We introduce LSNM-UV, a two-level location-scale model in which the hidden parents of a variable act inside the scale set by its observed parents, so that observed-parent effects can be divided out: every hidden effect collapses into an effective noise whose dependence structure encodes unobserved backdoor and causal paths. The resulting residual-independence patterns identify directed edges, non-edges and hidden-variable pairs, and identify bow-free acyclic directed mixed graphs. Under an explicit search-coverage condition, an oracle answering residual-independence queries at the population level makes the LSNM-UV search sound and complete. A thresholded HSIC statistic on fitted residuals answers every query on the oracle path correctly with probability tending to one under four explicit conditions on admissible population targets, targeting of independence witnesses, consistency of the empirical dependence statistic and a vanishing threshold, which gives graph consistency for a fixed number of variables; a fixed test level need not have this property. Controlled simulations diagnose the statistical conditions and their failure modes, an oracle-driven search shows that correct answers yield the correct graph, and a ten-node benchmark and the NetSim fMRI benchmark with hidden regions show the method at finite samples.

Primary area:
causal reasoning
Keywords:
causal discovery, hidden variables, acyclic directed mixed graphs, location-scale noise models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HsGYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses the problem of causal discovery by assuming that a functional causal model follows the form of a location-scale noise model. Here, the authors extend the existing theory by allowing hidden variables, including confounders and hidden intermediates, whose effects enter through the location and scale of an effective noise term. For this setting, conditional identifiability results are given. The method has been evaluated on synthetic datasets.

Critical strengths

The paper aims to nicely extend the location-scale noise model to exploit asymmetries between cause and effect for causal discovery with hidden variables. This is a promising extension and contribution, although there are some issues (see below).

Critical weaknesses

There are two core issues:

  • The paper’s correctness argument assumes that relevant candidate parents have already been collected, without showing that the algorithm has collected them. Restricting the setting to just two observed variables could simplify this, but correctness of the resulting bivariate procedure would still need a separate justification.
  • The experimental evaluation is not sufficiently broad for this kind of work. I would expect comparisons across a wider range of datasets and, in particular, real-world datasets. Further, the paper builds on prior work on location-scale noise models but does not compare with those methods in the experiments.
Questions for authors

The extension of location-scale noise models to hidden-variable settings is a promising direction. However, there are currently too many unresolved issues with this paper:

  • The correctness argument requires that the algorithm has collected the relevant candidate parents in non-bivariate causal discovery settings. Where would this information come from? The paper might be better positioned if it focused on the bivariate case, where this particular issue is simpler and which is also the setting of the prior work.
  • Even for bivariate cases, you establish (important) identifiability conditionally on Assumptions 2 and 3 (Residual irreducibility), but it remains unclear which concrete bivariate model families would satisfy these assumptions. An explicit characterization or sufficient conditions would strengthen the result here.
  • The experimental evaluation is not sufficiently broad for this work. In particular, since you extend location-scale noise models, why is there no direct comparison with the related work in the experiments (or did I miss it)? While they might not be applicable to graph discovery, you can at least compare with them in a bivariate setting. In this regard, you can check the Tübingen cause-effect pairs dataset for the bivariate case, which would help to clarify how it compares with causally sufficient alternatives. Or is there a particular reason for not including it?
Reviewer_7TUaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new causal discovery model, called LSNM-UV, for settings with latent variables. Starting from a two-level location-scale noise model, the authors derive residual-independence criteria that characterize invisible pairs, visible directed edges, and visible non-edges (Proposition 1). These characterizations in turn establish identifiability of the bow-free ADMG (Theorem 3).

Based on this identification logic, the paper proposes the LSNM-UV algorithm. Its first stage builds and prunes parent-candidate sets and classifies observed pairs if possible, using residual-independence tests; its second stage revisits unresolved or apparently invisible pairs using larger candidate regression sets. Under an additional search-coverage assumption, the algorithm correctly recovers the target bow-free ADMG (Theorem 4).

Finally, under high-level conditions on regression targeting, consistency of the empirical dependence statistic, and the threshold sequence, the empirical residual-independence decisions converge to their population-oracle counterparts (Theorem 5). This eventually yields consistency of the estimated target graph (Corollary 6).

Critical strengths

The paper extends location-scale causal modeling to a latent-variable. Existing LSNM methods established identifiability for bivariate cause-effect relations or heteroscedastic DAGs under causal sufficiency (Immer et al., 2023; Sun and Schulte, 2023; Yin et al., 2024; Lin et al., 2025), with only limited robustness analyses involving latent confounding. By introducing a two-layer LSNM in which observed and hidden parents enter through a common location-scale structure, the paper provides a tractable model for studying hidden-variable causal discovery under heteroscedastic noise. This is important because it moves beyond the additive-noise setting while retaining a residual representation that can still be used for causal identification.

The extension of residual-independence based identifiability beyond additive causal models is also noteworthy. CAM-UV has shown that residual-independence patterns can be used to distinguish parent-child relations from latent-variable induced paths in additive models (Maeda and Shimizu, 2021), but this logic does not directly extend to heteroscedastic settings because the noise scale itself may depend on the parents. In other words, LSNM-UV extended such causal separation in a substantially broader location-scale setting where additive methods are no longer applicable.

Critical weaknesses
  1. Insufficient characterization of the model class and reliance on high-level assumptions

The scope of the model class is not sufficiently characterized. The paper does not fully provide primitive conditions under which its key structural assumptions—particularly the common location-scale frame and the residual assumptions—hold, nor does it clarify how broad the resulting model class is. If it, as a variant of the existing models, tries to suggest a new identifiability under new assumptions, it would be important to state more explicitly how the resulting identifiable class differs from or extends previously studied model classes.

Assumption 4(ii) requires the candidate pool QijQ_{ij} of every unresolved pair to contain all relevant true parents, yet QijQ_{ij} is itself an intermediate output of the search rather than a primitive graph property. Thus, while Section 3 establishes population identifiability for bow-free ADMGs, it remains unclear which graphs are recoverable by the proposed search.

Finally, the identifiability theory only requires the existence of admissible residuals that achieve the required independence pattern, whereas the implemented procedure relies on a specific regression-induced target residual h∗h^*. Assumption 5 largely assumes that this target realizes the independence guaranteed existentially by the population theory, rather than deriving this connection from the regression procedure.

  1. The practicality of the structure of the model is not fully motivated

The paper clearly explains why the factorization is mathematically useful—namely, it allows the observed-parent location and scale to be divided out so that the hidden effects collapse into an effective noise. However, it provides little motivation for why real structural mechanisms should take the specific form in which both the hidden-parent effect and the latent noise are modulated by the same observed-parent scale gi1(Ki)g_i^1(K_i).

The empirical results also provide limited evidence for the practical benefit of imposing this additional structure. The clearest advantage of LSNM-UV is observed in simulations generated from the proposed two-layer model itself, whereas its advantage is substantially less consistent on the out-of-model NetSim benchmark. This raises the question of whether the additional structural restriction provides sufficient practical benefit beyond enabling the identifiability analysis.

Questions for authors
  1. Are there any ways to reduce assumption 4(ii) to more graph-specific conditions? Explicit graph-theoretic sufficient conditions, for example in terms of maximum indegree, sparsity, search depth, or hidden-path structure, would clarify the scope of Theorem 4 and the resulting consistency guarantee.

  2. Do the authors consider Simulation 8 sufficient empirical justification for Assumption 4? The reported 2–3% oracle-search failures are attributed to in-degree exceeding the search depth, but Assumption 4(ii) is a more specific candidate-pool coverage condition. Could the authors clarify this connection or directly report how often the required parent-coverage condition holds in the synthetic experiments?

  3. Can the theoretical connection between the regression-induced population residual h∗h^* and the admissible residuals hh used in the identifiability theory be established? Beyond the sufficient case in Proposition 2, can the authors characterize when the regression target itself realizes the independence witness whose existence is guaranteed by the population theory?

  4. Assumption 5 C3 appears to absorb both residual-estimation error and dependence-estimation error into a single high-level condition. Can C3 be derived from more primitive assumptions on the estimators of f∗f^* and g∗g^* and on the HSIC statistics?

  5. What primitive causal mechanisms, specified without assuming the proposed factorization from the outset, imply that the hidden-parent effect and latent noise share the same observed-parent scale gi1(Ki)g_i^1(K_i)? Can the authors give concrete examples and characterize how broad this class of mechanisms is, particularly in light of the less consistent results on NetSim?

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

The paper studies causal discovery with latent variables under a location-scale noise model. It introduces LSNM-UV as an extension of CAN-UV. It is a two-level location-scale model where the hidden parents of a variable act inside the scale set by its observed parents, so that observed-parent effects can be divided out. The residual independence patterns identify directed edges, non-edges and hidden-variable pairs, and identify bow-free acyclic directed mixed graphs.

Critical strengths
  1. Extending latent-variable causal discovery to location-scale mechanisms is interesting and important.
  2. The identifiability theory looks sound.
Critical weaknesses
  1. The model's factorization seems restrictive. It depends on a special form of two-level LSNM, which is generally incompatible with a global LSNM for all nodes. And it depends on which variables were measured.
  2. Under the restrictive special "common location scale frame", the novelty seems moderate as the argument follows trivially from CAM-UV.
  3. Detecting hidden variables is the main selling point but suffers from low power: In simulations with the paper's ordinary nonlinear functions, the detection is correct in only 7–37% of runs when n = 6400.
  4. Experiments are limited: the considered graphs are small with p <= 10, which raises concerns about its scalability.
Questions for authors
  1. Can you assess the robustness of the method when the two-level LSNM structure does not hold?
  2. Can you add experiments for a larger graph?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.