- 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., ), 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 , but the estimator only ever sees finitely many . Then, for fixed , even if we let , 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 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 are necessary, how does this solve the distributional PO any better than existing works on the distribution of potential outcomes?
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.