acceptodds
Withdrawn from ICLR 2027

Q-Loss: A Training Objective for Direct CATE Optimization Across Meta-Learner Families

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Conditional average treatment effect (CATE) estimation increasingly relies on model-agnostic meta-learners that decompose the problem into standard supervised-learning subproblems. These learners, however, are trained on surrogate losses only loosely aligned with treatment-effect quality, and a recent large-scale benchmark finds that 62% of fitted CATE estimators are degenerate—no better than predicting no effect at all. The Q-statistic of Yu & Sun (2025) provides an unbiased estimate of CATE risk that requires no counterfactuals and has been used to evaluate and rank estimators. We develop a differentiable, end-to-end algorithm for training meta-learners directly on this statistic—with learned feature gates that select treatment-effect modifiers—and instantiate it across the S-learner, X-learner, and DML-Learner families. On real and synthetic randomized-trial benchmarks, we then characterize when this helps: the apparent failure of the raw inverse-propensity target is largely a matter of propensity misspecification, which estimating the propensity and adopting a doubly-robust target resolves; and the resulting gains reflect genuine improvements in treatment-effect quality rather than the training statistic itself. Together, these analyses clarify when direct optimization of the Q-statistic is preferable to standard baselines.

Primary area:
causal reasoning
Keywords:
Causal Inference, Metalearners, Conditional Average Treatment Effect, Non-parametric Loss

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_8LPNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Q-loss, a learning objective that can be used to directly train meta learners with the Q-statistic without access to counterfactuals. Motivated by a study on real world data in which no heterogeneity was found, the authors assert that standard meta learners are insufficient for CATE estimation as they rely on surrogates, and instead propose an end-to-end learning framework. The proposed learner is validated on real and synthetic data.

Critical strengths
  • The paper is mostly self-contained.
  • The idea of directyl learning CATE estimators without second-stage regression sounds interesting.
  • Empirical results look somewhat promising.
Critical weaknesses
  • The authors propose Q-loss. However, completign the squares in their loss exactly yields standard pseudo outcome regression. There is nothing novel in their estimand (see for example Curth et al., 2021 and Kennedy, 2023). To be fair with the authors, they acknowledge this equivalence, but as a result, they strongly overclaim their contribution.

  • Proposition 1, as stated, is incorrect. The minimizer is only CATE on population level, not for an empirical finite sample estimator.

  • Proposition 2 is also, as stated, incorrect. Consistency requires assumptions on the neural network class / complexity class of the estimator.

  • Proposition 3 is overclaiming what the feature gates are doing. If a feature is irrelevant for CATE estimation, this only means that optimal prediction does not use this feature. Without regularization, there is no reason to assume that the feature gate will go to 0.

  • Experiments are insufficient: The most important comparison would be against neural DR learners, which are missing in the experiments.

  • Lastly, but this should not be the main reason for rejection: The motivation is questionable. I checked the referenced paper by Yu & Sun (2025), which serves as the main motivation. Here, a number of CATE learners was evaluated on RCTs, and a large number of them revealed no HTE. The authors call this degeneracy of the learners. I have severe doubts that this can be asserted so easily. Instead, what is more likely is either: i) there was no treatment effect (hence, τ≈0\tau\approx 0 is correct since τ≈0\tau \approx 0), or ii) there was a treatment effect, but no heterogeneity (and hence τ(x)≈0\tau(x)\approx 0 is correct as well.

The only novelty I see in this work is the integration of a Gumbel softmax function, but this is not enough for an ICLR contribution.


Alicia Curth & Mihaela van der Schaar. Nonparametric Estimation of Heterogeneous Treatment Effects: From Theory to Learning Algorithms. In AISTATS, 2021.

Edward H. Kennedy. Towards Optimal Doubly Robust Estimation of Heterogeneous Causal Effects. arXiv preprint, arXiv:2004.14497.

Questions for authors
  • Could the authors explain where exactly the novelty in their estimator lies, especially compared to standard learners?

  • Could the authors correct Propositions 1-3?

  • Could the authors explain why neural DR estimators are missing as baselines in their experiments?

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

The authors propose to estimate the CATE using a neural network trained by minimizing Q-loss, which is equivalent to squared-error regression on an IPW pseudo-outcome. They combine this with Gumbel-Softmax feature selection to jointly learn the CATE estimator and feature gates in an end-to-end algorithm. They also consider an AIPW pseudo-outcome and evaluate the approach across several CATE meta-learner architectures on real and synthetic datasets.

Critical strengths

None

Critical weaknesses

The methodological novelty is extremely limited. The proposed Q-loss is algebraically equivalent to squared-error regression on an IPW pseudo-outcome, which is precisely the existing PW-learner (Curth and van der Schaar, 2021). Replacing the IPW pseudo-outcome with its AIPW counterpart recovers the standard DR-learner (Kennedy, 2023). The authors acknowledge these equivalences. The claimed contribution therefore appears to be training a neural network with this existing loss and adding Gumbel-Softmax feature selection. The former is not a new contribution, since PW- and DR-learners already accommodate arbitrary regression methods, including random forests, neural networks, and boosting. The latter is at most a minor algorithmic contribution, as Gumbel-Softmax feature selection is itself well established (Jang et al., 2017; Strypsteen and Bertrand, 2021). It is unclear what is methodologically new beyond combining existing techniques. Moreover, Propositions 1 and 2 follow directly from standard squared-error regression theory.

The emphasis on the IPW pseudo-outcome is also poorly motivated. The AIPW pseudo-outcome used in the DR-learner is doubly robust, has more favorable nuisance-error dependence, and is statistically more efficient when the outcome regressions are consistently estimated. Even when the propensity score is known, the AIPW pseudo-outcome with correctly specified outcome regressions has no greater conditional variance than the IPW pseudo-outcome. It is therefore unclear why the authors emphasize the raw IPW learner rather than the established DR-learner. Finally, if the novelty is an end-to-end neural network architecture, why not develop it for general pseudo-outcomes/metalearners?

The empirical results do not convincingly establish an advantage over existing methods. The main experiments compare the proposed neural learners primarily against an XGBoost S-learner, rather than well-tuned PW- and DR-learners using comparable neural architectures. Although the authors include some ablations, these do not provide a systematic evaluation of the incremental benefit of the proposed feature-selection mechanism over standard pseudo-outcome regression. Moreover, the reported improvements from replacing IPW with AIPW largely reflect established variance-reduction and double-robustness properties. Given that the proposed Q-loss is equivalent to existing PW/DR regression objectives, stronger comparisons with these methods are essential to establish the practical contribution.

References:

  • Curth and van der Schaar (2021), Nonparametric Estimation of Heterogeneous Treatment Effects: From Theory to Learning Algorithms, AISTATS.
  • Kennedy (2023), Towards Optimal Doubly Robust Estimation of Heterogeneous Causal Effects, Electronic Journal of Statistics.
  • Jang, Gu, and Poole (2017), Categorical Reparameterization with Gumbel-Softmax, ICLR.
  • Strypsteen and Bertrand (2021), End-to-End Learnable EEG Channel Selection for Deep Neural Networks with Gumbel-Softmax, Journal of Neural Engineering.
Questions for authors

None

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

This paper proposes Q-Learner, a framework for CATE estimation using the Q-statistic as a training loss. The method combines IPW or AIPW pseudo-outcome regression with Gumbel-Softmax feature selection. The authors apply the framework to several meta-learner families and evaluate it on real and synthetic trial datasets.

Critical strengths
  • S1. The paper addresses an important problem of degeneracy in CATE estimation.
  • S2. The experiments cover several trial datasets and generate confounded datasets from them, and meta-learners, including ablation studies.
Critical weaknesses
  • W1. The experimental setup is not sufficiently clear.
  • W2. For semi-synthetic datasets where the true nuisance functions and CATE are available, the authors should report the Q-statistic and the PEHE.
Questions for authors
  • Q1. How are the nuisance functions e(X)e(X), μ0(X)\mu_0(X), and μ1(X)\mu_1(X) estimated? What models and training procedures are used?
  • Q2. Can the authors report the estimated Q-statistic, the oracle Q-statistic using the true τ(X)\tau(X) , and PEHE for various meta-learnears including the proposed one with SynthUplift and Megafon datasets?

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

Related papers

Discussion (0)

Sign in to comment.