acceptodds
Withdrawn from ICLR 2027

Probabilistic Counterfactual Inference for Discrete Outcomes in Gaussian-Process Causal Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Counterfactual inference in Gaussian-process structural causal models (GP-SCMs) has been developed primarily for continuous endogenous variables, limiting applicability to causal graphs that contain discrete child nodes with continuous parents. We introduce a unified probabilistic framework for counterfactual inference with heterogeneous variable types by pairing GP predictors with explicit exogenous noise mechanisms. For discrete outcomes, we derive exact conditional noise-abduction procedures using a uniform threshold for binary variables, a Gumbel-max race for nominal categories, and a latent Gaussian cut-point model for ordinal ones. In each case, we propagate abducted noise through interventions while accounting for posterior uncertainty in the GP latent functions, and prove that the resulting mechanisms reproduce the fitted model’s observational and interventional distributions. On synthetic SCMs with known ground-truth counterfactuals, we evaluate estimation accuracy, consistency, and robustness to coupling misspecification. A key finding is that applying a categorical coupling to ordinal data inflates counterfactual error roughly threefold even when observational fit remains comparable, and that this error does not diminish with more data: as the training set grows the fitted structural equation converges to the truth while the counterfactual error flattens onto a floor. In the reverse direction, forcing a false order onto nominal data instead degrades the fitted equation itself. The choice of coupling must therefore be justified on structural grounds rather than read off the fit.

Primary area:
causal reasoning
Keywords:
counterfactual inference, structural causal models, Gaussian processes, noise abduction, heterogeneous variable types

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

The paper studies counterfactual inference for discrete variables under Gaussian process SCM with a known DAG. The paper proposes a unifying framework for binary, categorical, and ordinal outcomes. The authors proposes a method for estimating the conditional noise distributions given the observation. The paper provides synthetic experiments to investigating the effects of misspecifying coupling between the counterfactual and the observations. This demonstrates that accurate conditional probability estimation, i.e., predictive fit, does not automatically lead to accurate counterfactual estimations.

Critical strengths
  • carefully designed ablation experiments that demonstrate the effects of wrong couplings
Critical weaknesses
  • The contributions seem limited and incremental. The ablation study is largely empirical. The computational method seems lacking advancement.
  • The justification for the selected counterfactual couplings seems not strong enough. Ordered categories seem to me do not automatically imply preservation of an individual’s latent rank across interventions.
Questions for authors
  • What assumptions beyond outcome type justify the chosen coupling, and how sensitive are the results to alternative couplings?

  • Can it be demonstrated that posterior averaging improves counterfactual uncertainty calibration compared with point estimates, since Table 3 evaluates only point-estimation error?

Reviewer_FR9KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extends Gaussian-process-based counterfactual inference to binary, categorical, and ordinal outcomes. It combines GP classifiers with different noise-abduction mechanisms and evaluates their counterfactual accuracy on synthetic causal models.

Critical strengths
  1. The paper addresses counterfactual inference with heterogeneous variable types, a relevant problem in causal modeling.
  2. The experiments distinguish counterfactual accuracy from observational and interventional fit.
Critical weaknesses
  1. The proposed framework mainly combines existing GP classification, Gumbel-max coupling, and ordinal threshold models. The theoretical results largely establish properties that follow from these known constructions.
  2. The paper argues that variable type should determine the coupling mechanism. However, knowing whether a variable is categorical or ordinal does not uniquely determine its cross-world dependence structure.
  3. The synthetic data are generated using the same coupling mechanisms assumed by the proposed estimators. Showing that correctly specified models outperform deliberately misspecified ones provides limited evidence of general effectiveness.
  4. Experiments cover only four relatively simple synthetic SCMs, without real-world evaluation or broader comparisons against alternative counterfactual methods. The generalizability of the findings remains unclear.
Questions for authors
  1. Why should variable type alone be sufficient to select an appropriate counterfactual coupling?
  2. How does the method compare with alternative coupling strategies under the same fitted distributions? and how demonstrate effectiveness on more complex or real-world causal models?
Reviewer_zLxBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies counterfactual inference in Gaussian-process structural causal models (GP-SCMs). Existing methods are primarily designed for continuous endogenous variables, which limits their application in causal systems that have discrete child nodes with continuous parents. To address this, the authors propose a unified probabilistic framework for heterogeneous variable types including binary, categorical, and ordinal by combining Gaussian process predictors with explicit noise mechanisms. Experiments on synthetic SCMs demonstrate that using an incorrect coupling leads to noticeable counterfactual errors, showing the importance of choosing correct mechanism.

Critical strengths
  1. Broad Scope
    Prior counterfactual inference in GP-SCMs focused almost exclusively on continuous variables, whereas real-world systems often involve discrete choices and decisions. This paper extends the framework to binary, nominal categorical, and ordinal variables in a unified manner, expanding the practical utility of GP-SCMs.

  2. Complete Methodological Formulation
    The authors provide concrete noise-abduction and interventional prediction derivations for each discrete variable type while accounting for Gaussian process posterior uncertainty. The theoretical derivation is clear.

  3. Empirical Guidance The experiments show that even if a model achieves a strong fit, misspecifying the coupling mechanism still incurs substantial error, demonstrating that the choice of mechanism should be careful.

Critical weaknesses
  1. Limited Experimental Scope
    The experiments are conducted on small synthetic causal graphs with at most three nodes and small label alphabets (C≤3C \le 3). The paper lacks experiments on more complex causal systems and lacks semi-synthetic or real-world examinations.

  2. Need Strict Prior Categorization
    The approach assumes that we can classify discrete variables as strictly nominal or strictly ordinal. However, in reality, it can not be always satisfied. Take a partially ordered setting as an example: a,b > c,d, but no order within a,b or c,d.

  3. Lack of Uncertainty Quantification
    Infinitely many valid joint distributions (couplings) can share the same observational and interventional marginals. The paper adopts specific mechanisms (Gumbel-max and comonotone couplings) to produce point estimates, but does not explore uncertainty bounds. which may yield overly confident counterfactual predictions if the true mechanism deviates from the assumed family.

Questions for authors
  1. Experiments on more complex settings Could the authors conduct experiments on deeper causal DAGs or scenarios with larger category sets (C>3C > 3)? It will be better if further incorporate semi-synthetic benchmarks or real-world settings?.
  2. Enlarge the range of variable type Could the author further consider settings that are not strictly nominal or strictly ordinal (e.g., partially ordered sets)? It will make the framework more practicable.
  3. Uncertainty Quantification
    Given the fundamental unidentifiability of counterfactual couplings, have the authors considered integrating partial identification techniques to provide bounding intervals on counterfactual probabilities alongside point estimates?

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

Related papers

Discussion (0)

Sign in to comment.