acceptodds
Under review as a conference paper at ICLR 2027

Learning Balanced Causal Representations via Differentiable Propensity-Score Subclassification

Abstract

Neural causal representation learning is increasingly used for average treatment effect (ATE) estimation. Many existing methods estimate propensity scores using binary cross-entropy loss, treating treatment-prediction accuracy as a proxy for adjustment quality. However, accurate propensity prediction from a learned representation does not in general ensure covariate balance or conditional independence between treatment assignment and covariates given the estimated score. We propose a plug-and-play subclassification-based balancing module that promotes covariate balance within propensity-score subclasses. The method incorporates subclassification into end-to-end neural training by jointly using a differentiable relaxation, dynamic programming to select subclass intervals, and iterative optimization. On the ACIC and IHDP benchmarks, our method reduces absolute ATE bias by approximately 0.035 and 0.062, respectively, compared with the strongest baseline. These results demonstrate that explicitly enforcing covariate balance can provide benefits beyond propensity-score prediction accuracy alone.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation Learning, Causal Inference, Subclassification, Covariate Balancing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_XKS6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper starts from an important observation: good propensity-score prediction does not necessarily imply good covariate balance in the original covariate space after representation learning. In particular, if Z = \psi(X) is a learned representation, the usual balancing property applies to Z, T \perp Z | e(Z), but this does not automatically imply T \perp X | e(Z). The authors therefore introduce a differentiable propensity-score subclassification module that learns a representation and propensity score; sorts samples by the estimated propensity score; partitions them into contiguous propensity-score subclasses; alternates between neural-network optimization and subclassification optimization. The paper then plugs the module into Dragonnet and BCAUSS, and evaluates it on IHDP and ACIC. The reported results show substantial ATE-bias improvements, especially for Dragonnet + subclassification and BCAUSS + subclassification. The paper also provides an asymptotic equivalence theorem between the hard and smoothed subclassification losses.

Critical strengths
  1. The motivation is fundamentally correct. The paper identifies an important problem that is sometimes overlooked in neural causal representation learning. A highly predictive propensity model is not necessarily a good balancing representation. In fact, a representation could preserve information useful for predicting treatment while discarding information that is less useful for treatment classification but still important for adjustment.
  2. Local balance is more informative than only global balance. This is another strong aspect. The paper argues that a global balance metric can hide substantial imbalance in particular propensity-score regions. This is a legitimate concern. The dynamic-programming formulation is also elegant because propensity-score subclasses are constrained to be contiguous after sorting by propensity score.
  3. The method is genuinely plug-and-play. I like the fact that the authors do not introduce an entirely new causal architecture. The proposed module can be attached to existing representation-learning models equipped with propensity branches.
Critical weaknesses
  1. The method optimizes only low-order moments, but the paper discusses conditional independence. This is the most important theoretical issue. The optimization objective only measures differences in means and coordinate-wise variances.The balance loss is based on an aggregated SMD of these quantities.
  2. The representation-learning interpretation is not fully justified. The authors claim that optimizing balance in raw X-space encourages the representation to preserve confounding-relevant information. Therefore the model may be encouraged simply to produce propensity scores that partition observations in a way that makes raw covariates balanced. That is not necessarily equivalent to learning a better low-dimensional representation.
  3. Dynamic programming is elegant but computationally expensive. This can become expensive for modern causal representation-learning datasets. The authors themselves identify computational cost as a limitation and propose replacing DP with heuristic search in future work. It is difficult to judge whether the proposed method is practically scalable.
Questions for authors
  1. What exactly is the theoretical target of the method?
  2. Why should zero SMD imply useful causal balance?
  3. Can the learned propensity score become intentionally distorted?
  4. Why are all subclasses equally weighted?
Reviewer_5MD5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a plug-and-play propensity model for representation-based causal estimation approaches that encourages covariate balance with propensity-score subclassification. This is done by optimizing for balance in the original covariate space. Experiments on ACIC and IHDP demonstrate the capabilities of the proposed method.

Critical strengths

The general of concept of propensity-score sub-classification could improve identification of causal effects as argued by Rosenbaum and Rubin (1984).

Critical weaknesses

The authors describe the proposed method, however, very little is provided in terms of justification, motivation or reasoning behind the model and the choices made to structure it. For example, the balance losses in Section 3.1 are introduced without justification, assumptions or references for the use of the standardized mean-difference discrepancy. In Section 3.2, propensity scores are ordered, but it is not clear how such ordering is restricted to sub-classification.

The main theoretical result is not surprising (or informative) giving Assumption 3.3.

The results are underwhelming once the error bars are considered in Table 1 and Table 2(b) and the performance of BCAUSS in Table 3.

Questions for authors

Why is the optimal value of the objective in Section 3.2 D[n,K]? Also, what is a legal gap in the ordered propensity scores?

The paper could benefit from proofreading to improve readability. For example, the notation is unnecessary bloated (why the need of the n subindex), the only numbered equation in the main body of the paper is not used and some wording seems incomplete (title of Section 3).

Why are the means in Figures 4 and 5 not covered by the boxplots?

Reviewer_g4VpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes adding a within-propensity-subclass covariate balance penalty to neural ATE estimators. Subclass boundaries are then chosen by dynamic programming, and smooth weights make subclass membership differentiable.

Critical strengths

The paper is well-motivated. Accurate treatment prediction from a learned representation need not balance raw covariates.

Critical weaknesses
  1. The methodological novelty appears limited. The idea of targeting balance rather than prediction has been well studied with CBPS and balancing weights. Counterfactual regression also uses balance objectives in representation learning. Here, the main addition seems to be the adaptive partitioning. It would be nice to see whether dynamic programming actually helps compared with simply updating equal-frequency subclasses during training, using the same smooth penalty.
  2. The theory does not say much about ATE estimation. Theorem 3.1 shows that the hard and smooth losses become close under an assumption that their membership weights become close. The algorithm does not guarantee this assumption. There are no results connecting the penalty to ATE error or showing that the alternating procedure converges.
  3. It is unclear how the balance objective relates to the reported ATE estimator. Section 4.1 appears to average predicted outcome differences, so the propensity score may affect the estimate through shared training rather than enter the final formula directly. This is plausible, but within-subclass mean balance alone does not establish lower ATE error. Please specify the final estimator and provide evidence for this connection. Results with AIPW would also be informative.
  4. The tuning procedure needs clarification. Choosing K, delta_n, and the smoothing function by true ATE error is not possible in a real observational study. Were these choices made without looking at the test splits, and were baselines tuned the same way? The different Dragonnet numbers in Tables 5 and 6 also need an explanation.
  5. The experiments need more support. With only ten realizations and no uncertainty estimates for paired differences, it is hard to judge whether the reported gains are reliable. More runs and confidence intervals for paired differences would help. The permutation test is also not described. A larger p-value does not show better balance; reporting within-subclass SMDs on held-out data would be more useful.
Questions for authors
  1. Which ATE estimator is used, and do gains persist with AIPW?
  2. Are representation learning, partition selection, and tuning restricted to the training data?
  3. Does dynamic programming outperform simpler updated partitions under matched budgets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.