acceptodds
Under review as a conference paper at ICLR 2027

Generalization Bounds and Representation Learning for Distribution-Valued Potential Outcomes

Abstract

Heterogeneous treatment effect estimation is central to understanding how causal effects vary across units with different observed characteristics. In many applications, however, each unit is observed through repeated post-treatment measurements that are more naturally represented by a probability distribution. Treatment can then change dispersion, tails, or shape even when the conditional mean is unchanged. In this paper, we study treatment-specific conditional Wasserstein barycenters for distribution-valued potential outcomes and define their conditional contrast through the difference of barycenter quantile functions. We develop a shared-representation estimator for learning these treatment-specific barycenters from observational data. We characterize the approximation induced by the learned representation, establish counterfactual generalization and finite-sample guarantees for target-population barycenter and contrast risks, and clarify how finite repeated measurements affect identification and estimation of the latent distribution-valued targets. Experimental results show that (i) shared fitting achieves lower mean contrast risk than separate fitting in all reported settings and lower treatment-specific barycenter risk in each synthetic family, and (ii) controlled experiments support the theoretically predicted effects of repeated measurements, quantile discretization, and representation learning.

Primary area:
causal reasoning
Keywords:
Causal Inference, Treatment Effect Estimation, Distribution-Valued Potential Outcome

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

The paper presents a framework for learning distribution-valued potential outcomes via treatment-conditioned Wasserstein-barycenter estimation. For this, they present (to the best of my knowledge) a novel estimation target, and provide a shared-representation learning method similar to Tarnet for the point-valued case. The paper i) derives a counterfactual risk bound that splits the error into factual prediction, treatment-group discrepancy, and reconstruction under additional lipschitz assumptions, ii) a finite sample error bound of the fitted estimator, and iii) a consistency result. The work is validated on syntehtic and semi-synthetic datasets.

Critical strengths
  • The paper appears to be a (almost) complete piece of work (some assumptions are missing, see below), is mostly self-contained, and provides all proofs in the appendix (which I checked briefly)

  • The paper presents a, to my knowledge, novel estimand and a new way to estimate it from finite data. I would partially see the novel estimand as a contribution since it presents a new way to think about potential outcomes, but I also have concerns here (see Weaknesses below).

  • The empirical results on (semi-)synthetic data look strong.

Critical weaknesses
  • While I generally think the introduction of Wasserstein-barycenter estimands is somewhat interesting, one of my main concerns is the impact of this work, and especially, if the Wasserstein-barycenter framing makes things more complicated than necessary. The authors essentially assume that the potential outcome itself is a distribution (i.e., Y(a)∈PY(a) \in \mathcal P), and then have the (artificial?) problem of creating an estimand that involves a complicated new machinary to estimate it from data. A large body of work, which is completely unreferenced here, is concerned with estimating the distribution of potential outcomes (e.g., Yoon et al., 2018; Ma et al., 2024; Näf et al., 2026, Melnychuk et al. 2023; 2026 and the cited literature therein). I do not see the necessity why one would move from "potential outcome is a scalar, and we estimate the distribution" with various methods and, for example, semi-parametrically efficient learners, to "we change the whole framework and let the potential outocme be a distribution itself". This is an interesting idea theoretically, but given that the proposed estimand involves Wasserstein distances, quantile functions, etc., the new problem becomes much more difficult than it needs to be.

  • My other main concern is on identifiability: The proposed estimand depends on latent Y(a)Y(a), but the estimator only ever sees finitely many ZijZ_{ij}. Then, for fixed MiM_i, even if we let j→∞j\to \infty, different latent distributions can generate the same distribution of the observed samples. Therefore, the conditional barycenter is not identifiable without extra assumptions (I double-checked this with help of an LLM). We would therefore need MiM_i going to infinity, which is a heavy additional layer for identifiability. For example, GDR learners by Melnychuk et al. (2026) do not require extra layer for identifiability, since they rely on the standard setup (scalar potential outcomes --> then estimate the distribution instead of defining the distribution as potential outcomes).

  • The paper is a hard read. The motivation is a bit poor (see Weakness 1), and notation is very dense due to the complicated estimation problem. To be fair, I do not know how this could be resolved if one insists on this particular setup, but this again confirms my intuition that the problem is made more complicated than necessary.


Jinsung Yoon, James Jordon, and Mihaela van der Schaar. GANITE: Estimation of individualized treatment effects using generative adversarial nets. In ICLR, 2018.

Yuchen Ma, Valentyn Melnychuk, Jonas Schweisthal, and Stefan Feuerriegel. DiffPO: A causal diffusion model for learning distributions of potential outcomes. In NeurIPS, 2024

Jeffrey Näf, Junhyung Park, and Herbert Susmann. Causal-DRF: Conditional kernel treatment effect estimation using distributional random forest. In AISTATS, 2026.

Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. Normalizing flows for interventional density estimation. In ICML, 2023.

Valentyn Melnychuk, and Stefan Feuerriegel. GDR-LEARNERS: Orthogonal learning of generative models for potential outcomes. In ICLR, 2026.

Questions for authors
  • Could the authors motivate what advantages the new estimand provides over the (not referenced) related work on distribution of (scalar-valued) potential outcomes? This is related to weakness 1: I wonder why one would choose the hard route of defining the potential outcome itself as a distribution.

  • Could the authors clarify the identifiability issue, and state additional necessary assumptions that would make estimation identifiable? If infinite MiM_i are necessary, how does this solve the distributional PO any better than existing works on the distribution of potential outcomes?

Reviewer_cEu2OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper focuses on heterogeneous treatment effect estimation when the outcome is a distribution. For the causal target, which the difference between the conditional quantile functions for the treatment and control, the paper develops a counterfactual risk bound, an estimator, and its finite-sample theory. The numerical performance of the proposed estimator is demonstrated by simulation experiments.

Critical strengths

None

Critical weaknesses
  1. The novelty of this paper beyond existing works seems weak. First, this paper is an extension of Johansson et al. (2022) from a scalar-valued outcome to a distribution-valued outcome. Due to such change of the outcome type, it is a natural choice to use the Wasserstein metric and thus study conditional quantile functions. Moreover, comments about the connections between developed theoretical results and existing works are missing. It is unclear what methodological and/or theoretical obstacles this paper has overcome or why existing methods fail here.

  2. The practical benefit of the proposed method is not convincing. First, the paper does not include any real data applications so the performance of the proposed method in real-world scenarios is unknown. Second, the scope of the simulation study is limited. For example, Katta et al. (2024) is not included in the comparison at all. Panel B only compares the proposed methods. Sensitivity analysis is also not provided.

References:

Johansson, F. D., Shalit, U., Kallus, N., & Sontag, D. (2022). Generalization bounds and representation learning for estimation of potential outcomes and causal effects. Journal of Machine Learning Research, 23(166), 1-50.

Katta, S., Parikh, H., Rudin, C., & Volfovsky, A. (2024). Interpretable causal inference for analyzing wearable, sensor, and distributional data. In International Conference on Artificial Intelligence and Statistics (pp. 3340-3348). PMLR.

Questions for authors
  1. What does ZijZ_{ij} in Eq. (1) refer to? Given what variables are ZijZ_{ij} assumed conditionally independent with law YiY_i? Is the conditional independence assumption reasonable in practice?

  2. The definitions of "weak overlap" and "fixed overlap" are missing.

  3. Code for the proposed estimator and simulation studies should be submitted and also make publicly assessable.

Reviewer_U3RxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers the estimation of conditional Wasserstein barycenter differences for units with repeated measurements. The proposed estimator uses an encoder–decoder architecture with three types of loss: Wasserstein prediction loss, balancing, and covariate reconstruction. The authors also provide a theoretical analysis accounting for these three parts. Synthetic and semi-synthetic experiments demonstrate benefits from the shared-learning structure, although the additional regularizers do not consistently improve performance.

Critical strengths

Not apparently.

Critical weaknesses

The estimation target has been studied before, as mentioned by the authors. I therefore assess the paper mainly through its method and theory.

  1. Method. The reconstruction penalty encourages recovery of the full covariate vector from the learned representation, which is stronger than preserving the information needed to predict the conditional barycenters. In particular, covariates irrelevant to both barycenter quantile functions need not be recovered. Requiring their reconstruction may hurt prediction. As for the balancing penalty, reducing this discrepancy may remove outcome predictive information, especially when the groups differ substantially along predictive covariates. Balancing is inappropriate in the absence of overlap. Although these points are acknowledged by the authors, I think they may limit the method’s practical usefulness, so I raise them here.

  2. Theory. The theoretical results use dense notation and are hard to digest. For example, Theorem 5.1 says EΔ(θ^)≤GK+2inf⁡θ∈ΘKJλ,K(θ)+4Γλ,K,n(δ)+4Ξ‾M(δM)+2λDbOT+2ϵopt+2(CD−λD)+Dθ^+2(CR−λR)+Rθ^.\mathcal{E}_{\Delta}(\hat{\theta}) \leq G_K + 2\inf_{\theta\in\Theta_K}J_{\lambda,K}(\theta) + 4\Gamma_{\lambda,K,n}(\delta) + 4\overline{\Xi}_M(\delta_M) + 2\lambda_D b_{\mathrm{OT}} + 2\epsilon_{\mathrm{opt}} + 2(C_D-\lambda_D)_{+}D_{\hat{\theta}} + 2(C_R-\lambda_R)_{+}R_{\hat{\theta}}. In addition, the overall statistical sharpness of the bound is not established, for example through a matching lower bound. This makes the theoretical results harder to evaluate.

Questions for authors

My questions concern the weaknesses discussed above.

  1. Could the authors comment on how to choose the additional penalties to avoid harming the prediction performance?

  2. Could the authors comment on the tightness of their theoretical results?

2 comments

I thank the authors for the detailed reply.

For the comment on the theoretical results, I appreciate the rates provided for different cases. However, the rates themselves are not sufficient to show the results can not be further improved/tightness. (The argument can be strengthened by providing (approximately) matching lower bounds.)

Overall, I am still not fully convinced by the necessity of recovering the full vector from the representations (may contain non-relevant info.). and enforcing the balance between the treatment and control group (may discard the info. that is helpful for a single arm prediction but not shared). The writing can be improved as well. Thus, after reading other reviewers' comments, I have decided to maintain my score.

Dear Reviewer U3Rx,

Thank you so much for your followup. We want to express our deep appreciation for your constructive comments again, and will try our best to revise the paper according to your reviews. These suggestions are valuable to us for future submission! Wish you all the best!

Best regards,

Authors of paper 16739

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.