acceptodds
Under review as a conference paper at ICLR 2027

Identifying a Canonical Latent Causal Model from Multi-Environment Mixed Observations

Abstract

Scientific variables are often observed through heterogeneous measurements rather than directly: some coordinates are continuously recorded after unknown monotone distortions, whereas others are ordinal or binary coarsenings. We study a linear Gaussian latent DAG shared across environments, with invariant measurement mechanisms and environment-varying independent structural-noise variances. We show that measurement invariance aligns latent marginal scales across environments and identifies a common-reference latent covariance family. When the structural noises have pairwise distinct cross-environment variance profiles, this family identifies a canonical latent structural causal model and the full latent DAG. The identified canonical model also determines hard-intervention laws in observed environments; for strictly increasing measurements that may be discontinuous, we further characterize the sharp set of physical measured responses and the exact condition for point identification. We develop a mixed-data covariance front end and a regularized coefficient estimator with fixed-dimensional finite-sample guarantees. Across eight graph topologies, three sample sizes, and strong and weak heterogeneity regimes, recovery improves with sample size and variance-profile separation; under weak separation, our estimator attains higher mean area under the precision–recall curve than same-front direct diagonalization and BACKSHIFT. A BRFSS 2015 application illustrates real-data behavior through held-out covariance fit and bootstrap edge-selection stability.

Primary area:
causal reasoning
Keywords:
Causal discovery, latent variable models, multi-environment learning, mixed data, identifiability, structural equation 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_eWF1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposed a latent causal discovery method to infer causal relationships among latent variables who have one-to-one mapping to observed variables. The observed variables can be either continuous or ordinal (binary included). They assume both the causal mechanism and measurement mechanism are invariant across environments. Then they are able to identify the covariance of the latents and subsequently the causal DAG.

Critical strengths

The measurement model for continuous variables appears to be fully nonparametric, although the ordinal measurement seems to be parametric (probit) and causal relationship is linear Gaussian. They provide an identifiability result as well as estimation theory. The paper appears to be overall sound.

Critical weaknesses

My main concern is the literature doesn't seem to be well reviewed. I am not directly working in this area, but a quick Google search finds the following two papers that appear to be very much related to this paper's topic (I'm sure there are more since I didn't search for it very extensively):

  • Causal discovery with linear non-Gaussian models under measurement error: Structural identifiability results
  • Independence testing-based approach to causal discovery under measurement error and linear non-Gaussian models. This is based on my understanding of the paper: the main problem the authors tried to address is the measurement problem in causal discovery. Especially for continuous variables, that's really a measurement error problem with a non-parametric error structure, right? Therefore, I think the authors should position their paper against the causal discovery methods that allow measurement errors.
Questions for authors
  1. What do you lose by using rank-based correlation estimator?
  2. Do you need to fix tau in practice? how? it looks like it is fixed at 0.2 in real data and in simulations it seems not chosen because AUC doesn't require one.
  3. What's the convergence rate for G?
  4. Are there existing mixed-data causal discovery methods or multiple-environment causal representation learning methods you can compare with (treat ordinal as continuous if necessary; for a large number of categories, ordinal data are similar to continuous data in practice)?
  5. Are the multi-environment assumptions reasonable in the real application and is it possible to check them?
Reviewer_VHF4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies causal identification from mixed continuous, ordinal, and binary measurements across environments. Its central contribution is an identification chain that aligns latent scales through measurement invariance, then exploits changing structural-noise variances to recover a canonical coefficient matrix and the latent DAG, and it distinguishes identified canonical interventions from potentially set-identified measured responses.

Critical strengths
  1. Recovering a separate latent correlation matrix per environment is insufficient for variance-based causal discovery; explicitly identifying relative marginal scales and assembling a common-reference covariance family is a useful contribution.
  2. Well-motivated canonical target. Distinguishing raw coefficients from coefficients in reference-environment units makes the residual scale ambiguity explicit while preserving identifiable graph structure and signs.
  3. The distinction between canonical latent interventions and measured responses at discontinuities, with a sharp interval characterization, clarifies why observational equivalence need not imply identical responses at targeted latent points.
Critical weaknesses
  1. Empirical comparisons lack repeated-sampling uncertainty and use unequal run subsets. Results average over eight fixed graph instances rather than repeated datasets, and C1 failures are excluded, leaving 6/8 and 7/8 graphs in the strong and weak n=500 cells while competitors use all eight. The means are therefore not fully paired and may favor methods that fail on hard instances. Please report paired comparisons on common successes, failure rates as a separate outcome, repeated-sampling variability, and ideally a prespecified fallback so the complete procedure can be evaluated.
  2. The guarantees do not characterize the implemented algorithm. Theorem 4 bounds a global minimizer of a nonconvex objective, whereas the implementation checks stationarity after optimization and polishing, and CV-selected regularization need not satisfy the theorem’s conditions (Appendix F.3).
Questions for authors

see above

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

The paper studies a linear gaussian latent DAG shared across multiple environments--only structural noise variances change among environments. Each latent variable (LV) is observed through 1 measurement, the same in any environment. This is either an unknown strictly increasing (possibly discontinuous) mapping or an ordinal/binary thresholding. Within 1 environment, copula methods recover latent correlations on a standardized scale -- so the cross environment changes in LV variances (which variance based causal identification requires) are lost. The paper first recovers a common reference covariance family {Ω(e)}\{\Omega^{(e)}\}, that is, every environment's latent covariance in units of the ref environment's latent stds (Prop 1). It does so by applying the ref environment normal score transform to all envs, comparing invariant ordinal thresholds and inverting bivariate gaussian bridges for the correlations. Under H2; the noise-variance ratio profiles (relative to ref env) are pairwise distinct, then whitening & joint diagonalization recover the mixing directions up to permutation and scale -- then acyclicity removes the permutation, and normalization gives the canonical coefficient matrix B∘B^\circ and the full DAG (Thm 2). Hard interventions on the canonical LVs are then identified. The physical measured response at a jump of a continuous measurement map is identified up to the jump interval (Thm 3). The estimator comes with fixed dimensional rates and asymptotic support recovery (Thm 4). Simulations (p=10p = 10, 5 environments, 8 graphs, strong/weak variance separation) show it outperforms same front joint diagonalization and BACKSHIFT under weak separation.

Critical strengths

The paper identifies a gap between (1) mixed data copula methods which standardize scales within each environment and (2) variance-based multiple environment methods, which need exogenous variances on a common scale. Observing that "invariant measurement maps make cross environment scale-ratios identifiable" is worth investigating to link the two.

Critical weaknesses
  • W1: My biggest concern is limited novelty. The identification chain combines known results. Thm2 recovers mixing directions from changing source variances; recall nonstationary BSS (Pham and Cardoso, 2001)--the estimator in (8) is their log-determinant criterion. The permutation removal by acyclicity is from LiNGAM (Shimizu et al., 2006). Linear SEMs identification from cross-domain noise variance changes is also known (Ghassami et al., 2018). Montagna (2026) identifies the graph from source variance changes across environments with acyclicity--in a more general class than linear. The new addition is the measurement layer (Prop1), which combines the standard Gaussian-copula normal score Φ−1∘F\Phi^{-1} \circ F with threshold invariance across groups--standard in ordinal factor analysis. Neither of these refs except Montagna is cited. The proofs as far as I checked use standard arguments.

  • W2: The key assumptions are strong and not well examined. Examples: H2 asks every variable pair to have different variance profiles across environments; it fails when two noise variances stay unchanged (e.g., sparse or targeted shifts). When does H2 generally hold? I think simulations satisfy it because node variance changes in every environment. Acyclicity is also fundamental, without it neither B∘B^\circ nor the DAG are identified; yet BACKSHIFT (the main baseline) allows cycles.

  • W3: The paper is hard to read, not well written. Examples include (1) Assumptions are stated inline and referred to by labels (H1, O3, binary scale visibility) that are defined only in Appendix A; (2) The main text has no Thm1, and terms like BACKSHIFT and BRFSS and the interval [L(z),U(z)][L(z), U(z)] are all used way before they are defined; (3) Several undefined terms, to name a couple, Table 9's frozen T8 runtime summary and C1 estimator.

  • W4: Empirical evaluation is narrow. The simulations use p=10p = 10, 5 environments, 8 graphs, and 1 dataset per graph-setting. The estimator's cross-validation failures are excluded; not counted. Only one external multi-environment baseline, which is BACKSHIFT. The BRFSS application does not have ground truth and violates the model (rounded BMI, tied day counts); I don't see how it supports the causal claims?

Questions for authors

See weakness section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.