acceptodds
Under review as a conference paper at ICLR 2027

OrthoGen: A Generative Orthogonal Learner for Time-Varying Treatments

Abstract

Estimating conditional distributional potential outcomes (CDPOs) over time is important in medicine (e.g., to estimate patient-specific risks under different treatment sequences). However, this task is challenging because of time-varying confounding, yet existing adjustment strategies for this task are limited. In this paper, we aim to learn CDPOs under time-varying treatments using flexible generative models. Our contributions are two-fold. (1) We introduce a tailored adjustment strategy for our setting, namely, generative recursive g-computation. Our adjustment strategy recursively propagates full conditional outcome distributions rather than conditional means, modeling the variables of interest directly rather than full trajectories. Building on our adjustment strategy, we formulate simple generative learners for CDPO estimation. However, these learners can be sensitive to nuisance estimation errors, which motivates an orthogonal learner. (2) We thus introduce OrthoGen, a Neyman-orthogonal and doubly robust generative learner. Importantly, we show that OrthoGen further achieves rate double robustness and quasi-oracle efficiency under suitable conditions. Our learners are flexible and can be instantiated with different generative backbones (e.g., normalizing flows and diffusion models). Across experiments with synthetic, semi-synthetic and real-world datasets, we find that OrthoGen is highly effective. To the best of our knowledge, we are the first to propose a generative orthogonal learner for estimating CDPOs under time-varying treatments.

Primary area:
causal reasoning
Keywords:
Generative models, orthogonal learners, time-varying treatments

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5GLLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes OrthoGen for estimating conditional outcome distributions under time-varying treatments. It fits continuation distributions through backward generative regression and trains a target generator using a longitudinal AIPW objective. It establishes orthogonality and a product-error bound under regularity assumptions. Experiments use normalizing flows and diffusion models on synthetic, semi-synthetic, and real data.

Critical strengths

The framework combines backward distribution learning with a model-agnostic orthogonal objective. Comparisons using common architectures help isolate the effect of the adjustment strategy. The semi-synthetic results show consistent improvements across the tested configurations.

Critical weaknesses
  1. The motivation for learning the whole distribution needs stronger support. The backward sequential fitting and AIPW correction build on established methods. The paper should give concrete examples of quantities that users need and demonstrate why one generative model is preferable to targeted sequential estimation. It does not quantify the computational savings from using one distribution model to answer multiple queries.

  2. My main concern is accuracy for the quantities ultimately reported. Learning a whole distribution may introduce approximation error for a particular quantity compared with estimating it directly. The bound controls distance to the best generator within the chosen model class, rather than error relative to the true target quantity. Therefore, it does not resolve whether distribution learning improves or worsens bias for means, threshold risks, or quantiles.

Questions for authors
  1. What concrete collection of outcome queries motivates learning the whole distribution? For those queries, how do accuracy and total computational cost compare with targeted sequential AIPW estimators?

  2. What guarantee does OrthoGen provide for a quantity calculated from its generated distribution, such as a threshold risk or quantile? Can the authors separate generator estimation error from model-class approximation bias and provide a direct empirical comparison with targeted estimation?

4 comments

We want to thank the reviewer for their careful review and constructive feedback, as well as for the positive evaluation of our work. We have uploaded a new revision, with several changes highlighted in blue, in response to the concerns of all the reviewers. Below, we address all your concerns.

  • Motivation for learning the full counterfactual distribution.

    Thank you for encouraging us to further clarify the practical motivation for learning CDPOs rather than predefined functionals. We have added a paragraph clarifying the motivation for learning whole distributions in Sec. 3, titled “Why learn the full distribution?” and a clinical example paragraph. In these paragraphs, we summarize the following points:

    First, the quantities relevant to treatment decisions may be patient-specific and need not be specified before model training. For example, the risk threshold considered clinically relevant may depend on the patient's health status, comorbidities, or preferences. A fitted CDPO allows clinicians to evaluate such individualized risks, as well as additional quantities of interest identified during clinical assessment, without retraining a separate estimator for each query.

    Second, the distribution itself can provide useful information beyond individual numerical summaries. Visualizing the predicted outcome distribution under alternative treatment sequences allows decision-makers to examine differences in dispersion, tail behavior, and multimodality, which, for some distributions, may be obscured by a small collection of predefined functionals.

    Third, this flexibility is particularly valuable for multivariate outcomes (our fully synthetic dataset, for instance, has three-dimensional outcome), where clinically relevant decisions may depend on dependencies between outcome components. For example, consider three clinical outcomes with potentially patient-specific thresholds. A treatment decision may depend on the probability that at least two of these outcomes fall below their respective thresholds. This probability cannot generally be recovered from marginal outcome distributions, because it depends on their joint dependence structure. A fitted joint CDPO permits such probabilities to be evaluated for arbitrary combinations of thresholds and outcomes.

    Finally, we would like to highlight that the problem of distributional potential outcomes estimation has attracted considerable interest in recent years, with many papers addressing that problem in the static setting (see our extended related work; [1, 2, 3, 4, 5, 6]). Our novelty is to bring doubly robust distributional potential outcomes to the time-varying setting, which is non-trivial.

    Action: We have added the paragraphs titled “Why learn the full distribution?” and “clinical example” in Sec. 3, and we have further developed the interest of distributional estimation with a motivation example in the new Appendix I “Functional Accuracy and Computational Cost”.

  • Comparison with targeted models.

    Thank you for this suggestion. We have extended our evaluation with our new Appendix I (Functional Accuracy and Computational Cost) and our new Tables 9 and 10, where we compare the conditional means and quantiles implied by our generative models against estimators directly targeting these quantities. We also include a summary in Sec. 7.

    For conditional means, we consider four dedicated learners based on PI, RA, IPTW, and doubly robust (DR) estimation. For conditional quantiles, we implement a direct IPTW estimator based on weighted pinball loss. We use IPTW because extending recursive g-computation and doubly robust estimation to conditional quantiles requires additional quantile-indexed nuisance functions and a more involved sequential construction that is underdeveloped in the existing literature and lies beyond the scope of our paper, as discussed in our new appendix.

    On the semi-synthetic benchmark (τ\tau=3), OrthoGen achieves mean and 0.8-quantile RMSEs of 0.0639 ± 0.0143 and 0.0700 ± 0.0132, respectively, compared with 0.0556 ± 0.0098 for the dedicated DR mean learner and 0.0587 ± 0.0135 for the dedicated IPTW quantile learner.

    => While targeted estimators can achieve higher accuracy for the specific quantities they are trained to target, OrthoGen remains competitive and, importantly, it learns the full conditional distribution rather than optimizing either functional directly. As a consequence, the same fitted model provides both quantities without retraining.

    Action: We extended our empirical evaluation with comparisons against dedicated conditional-mean and quantile estimators (new Tables 9–10, Appendix I.1–I.2), and summarized the findings in Section 7.

  • Guarantees of quantities estimated from the distribution.

    We thank the reviewer for raising this important point. Our approach aims to learn the optimal generator within a predefined generative model class, which, for sufficiently flexible generative models, can closely approximate the true counterfactual outcome distribution. Since quantities such as conditional means, threshold probabilities, and quantiles are functionals of this distribution, they can be directly obtained from the learned generator.

    To further clarify the guarantees for such quantities, we have added a discussion in our new Appendix I.4 (Eq. 150) that decomposes the error of any functional into two components: (i) the estimation error relative to the optimal generator within the chosen model class, and (ii) the model-class approximation bias relative to the true counterfactual distribution.

    Theorem 2 controls the former through a model-specific generator discrepancy, while the latter depends on the expressiveness of the generative class. Under suitable continuity conditions, distributional approximation guarantees can be translated into guarantees for individual functionals. For example, the absolute error in the conditional mean is bounded by the Wasserstein distance between the estimated and true distributions: [∣EP^[Y]−EP[Y]∣≤W1(P^,P)≤W2(P^,P)][|\mathbb E_{\hat P}[Y]-\mathbb E_P[Y]|\leq W_1(\hat P,P)\leq W_2(\hat P,P)]. For conditional quantiles, analogous guarantees can be obtained under regularity conditions.

    Importantly, the generative objectives employed by our backbones are designed to approximate the true outcome distribution. Maximum likelihood for NFs minimizes the forward KL divergence, while denoising score matching for DMs targets the scores of the underlying perturbed distributions. Under suitable regularity conditions and sufficient model capacity, these objectives can recover the true distribution at their population optima, making the model-class approximation bias small.

    => Thus, flexible generative models provide a principled approach to estimating multiple distributional functionals, while the decomposition above makes explicit the role of finite-capacity approximation error.

    Action: We have clarified these theoretical connections in Appendix I.4 and complemented them with direct empirical comparisons against dedicated mean and quantile estimators in Appendix I.1–I.2.

We thank the reviewer again for their constructive feedback, which has helped us strengthen both the motivation and empirical evaluation of our work. We hope that the additional comparisons and theoretical discussion clarify the benefits and limitations of learning full counterfactual distributions relative to directly targeting individual functionals. We would be happy to provide any further clarifications or address additional questions or suggestions.

References:

[1] Kennedy, Edward H., Sivaraman Balakrishnan, and Larry A. Wasserman. "Semiparametric counterfactual density estimation." Biometrika 110.4 (2023): 875-896.

[2] Melnychuk, V., & Feuerriegel, S. (2026, April). GDR-learners: Orthogonal learning of generative models for potential outcomes. In International Conference on Learning Representations (Vol. 2026, pp. 105267-105301).

[3] Yoon, J., Jordon, J., & Van Der Schaar, M. (2018, February). GANITE: Estimation of individualized treatment effects using generative adversarial nets. In International conference on learning representations.

[4] Melnychuk, V., Frauen, D., & Feuerriegel, S. (2023, July). Normalizing flows for interventional density estimation. In International Conference on Machine Learning (pp. 24361-24397). PMLR.

[5] Ma, Y., Melnychuk, V., Schweisthal, J., & Feuerriegel, S. (2024). DiffPO: A causal diffusion model for learning distributions of potential outcomes. Advances in Neural Information Processing Systems, 37, 43663-43692.

[6] Spiegelhalter, David. "Risk and uncertainty communication." Annual Review of Statistics and Its Application 4 (2017): 31-60.

Thank you for the careful response and the new comparison with dedicated mean and quantile estimators, which I appreciate and which makes the paper clearer. My remaining concern is the practical implication. In the new results the targeted estimators are somewhat more accurate for the quantities they target, so the main advantage of learning the full distribution appears to be avoiding retraining, and I am not sure retraining is a real bottleneck in causal inference practice. I think the work is careful and I remain slightly positive.

Thank you for your thoughtful follow-up and for your continued positive assessment of our work.

We would like to emphasize that, although dedicated estimators may be preferable when targeting a single predefined functional, the value of learning full counterfactual distributions extends beyond avoiding retraining. In particular, full distributions provide a unified representation of counterfactual uncertainty, enabling the visualization and comparison of entire outcome distributions under alternative treatments, including differences in tail behavior, dispersion, and multimodality that may be obscured by individual summary statistics. This flexibility can be particularly valuable for patient-specific decision-making, where the quantities of interest may vary across patients or evolve with clinical objectives.

We believe that enabling this richer form of counterfactual inference under time-varying treatments, while maintaining competitive accuracy for individual functionals and providing orthogonality and double robustness, constitutes an important contribution.

We sincerely appreciate the reviewer's constructive feedback, which has helped us further strengthen and clarify our work.

Reviewer_EuvqOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
3(+1)
Read review
Paper summary

This paper studies estimation of conditional distributional potential outcomes (CDPOs) under time-varying treatments. The authors start from the observation that naive history adjustment is biased due to time-varying confounding: intermediate covariates affected by earlier treatments influence both subsequent treatment assignment and the terminal outcome, so conditioning on the observed future treatment sequence selects a non-representative subgroup. To address this, they propose a recursive distributional g-formula: rather than conditioning on observed future treatments, it propagates the outcome distribution backward from the final treatment step, at each step averaging over the naturally observed one-step-ahead history. This avoids both the selection bias of naive conditioning and the need to simulate full covariate trajectories. Building on this, they develop a two-stage Neyman-orthogonal learner, OrthoGen, which combines the recursive g-computation with IPTW so that the target model is first-order insensitive to nuisance estimation errors. Experiments on synthetic, semi-synthetic, and real-world datasets demonstrate the advantages of the proposed method, especially at longer treatment horizons.

Critical strengths
  1. Estimating CDPOs under time-varying treatments is practically important and relevant to real-world decision-making.

  2. The proposed method is well-motivated. Starting from the time-varying confounding problem, the proposed recursive distributional g-formula cleanly addresses both the selection bias of naive conditioning on future treatments and the burden of simulating full future covariate trajectories.

  3. The method is evaluated on synthetic, semi-synthetic, and real-world datasets with two different generative backbones, and the results consistently demonstrate the advantages of OrthoGen, especially at longer treatment horizons.

  4. OrthoGen leverages Neyman-orthogonalization, which makes the target model first-order insensitive to nuisance estimation errors.

Critical weaknesses
  1. Presentation clarity is a major concern.
  • The core challenge is not sufficiently explained. In particular, the paper states that naive history adjustment is biased (Eq. 2) but does not clearly spell out what the naive procedure does or why it is biased. A causal diagram illustrating time-varying confounding would help readers understand why conditioning on the observed future treatment sequence induces selection bias.
  • The paper motivates avoiding full covariate-trajectory simulation, but does not adequately discuss why directly modeling/simulating the future history trajectory is undesirable.
  • A clearer statement of the two concrete challenges being solved, (i) selection bias from conditioning on future treatments, and (ii) the burden of simulating future covariate trajectories, would strengthen the motivation.
  • An end-to-end data-flow diagram clarifying the inputs, outputs, and loss functions of each stage would make the method much easier to follow.
  1. The experimental comparison relies on a single baseline (IPTW). In addition to IPTW, it would be informative to compare against point-estimate methods for time-varying treatments. Although these methods target the conditional mean rather than the full distribution, the authors could compare the mean implied by OrthoGen's predicted distribution against their point estimates.
Questions for authors

Please see Weaknesses.

5 comments

We want to thank the reviewer for the time and effort dedicated to reviewing our paper and the constructive feedback. We have uploaded a new revision of our work, with several changes highlighted in blue, in response to the concerns of all the reviewers. Below, we address all your concerns.

Weaknesses:

  • 1a. Naive history adjustment bias and time-varying confounding.

    Thank you for highlighting the need for a clearer explanation. To clarify this issue, we have revised Sec. 3 as follows: (i) we explicitly describe the naive generative learning procedure, (ii) we explain why observational conditioning differs from intervention, and (iii) we introduce a new causal graph (Fig. 2) illustrating treatment–confounder feedback and the resulting time-varying confounding.

    A naive approach would consist of directly training a generative model for the terminal outcome (Yt+τY_{t+\tau}), conditioned on the observed history (Hˉt\bar H_t) and the subsequent treatment sequence. Such a model would approximate the observational distribution on the right-hand side of Eq. (2), rather than the interventional distribution on its left-hand side, which is our target.

    These distributions generally differ because of time-varying confounding [1, 2, 3, 4]. In particular, covariates observed after the prediction origin may be simultaneously influencing subsequent treatment assignments and the terminal outcome. Consequently, conditioning on the observed future treatment sequence selects patients according to these intermediate covariates, whose distribution may differ from that induced by the intervention. This introduces selection bias relative to the target interventional distribution.

    Action: We expanded Section 3 to explain the bias introduced by naive observational conditioning, explicitly distinguish observational and interventional distributions (Eq. 2), and introduce a new causal graph illustrating time-varying confounding (Fig. 2).

  • 1b. Full covariate-trajectory simulation.

    Thanks for bringing up this important question. To address it, we have expanded the opening discussion of Sec. 4 to clarify the limitations of full covariate-trajectory simulation. The main limitation of full covariate-trajectory simulation is the additional modeling complexity it introduces. Such approaches require learning the joint longitudinal data-generating process, rather than directly targeting the distribution of the outcome of interest. This can be particularly demanding when the covariates are high-dimensional relative to the target outcome, as the generative model must accurately represent many additional variables that are not themselves the estimand.

    Furthermore, jointly modeling all variables may allocate substantial modeling capacity to the covariate distributions rather than the target outcome distribution, potentially compromising accuracy for the latter. At inference time, these approaches also require sequentially simulating future covariate trajectories, introducing additional computational costs and opportunities for error propagation.

    In contrast, our generative recursive g-computation directly models the conditional distribution of the target outcome and propagates this distribution backward across treatment steps.

    => This avoids explicit modeling and simulation of complete future covariate trajectories, while still accounting for the distribution of intermediate histories through the recursive adjustment.

    Action: We expanded the discussion in Section 4 to clarify the modeling complexity and potential error propagation associated with full covariate-trajectory simulation, highlighting how our recursive approach avoids these limitations.

  • 1c. Statement of the challenges:

    Thank you for this suggestion. Explicitly distinguishing the two challenges strengthens the motivation for our approach. We have therefore revised the Introduction (Sec. 1) to highlight: (i) the selection bias introduced by naively conditioning on observed future treatment sequences and (ii) the additional modeling complexity associated with full covariate-trajectory simulation.

    Action: We revised the Introduction (Section 1) to explicitly distinguish the two main challenges motivating our approach: (i) selection bias from naive conditioning and (ii) the additional modeling complexity of full trajectory simulation.

  • 2. Mean estimators comparison:

    Thanks for this suggestion. We have extended our experimental evaluation to include four dedicated conditional-mean estimators based on plug-in (PI), regression adjustment (RA), inverse propensity weighting (IPTW), and doubly robust (DR) learning (see the new Appendix I and a brief summary in Sec. 7).

    We compare their predictions against the conditional means obtained by Monte Carlo sampling from the corresponding generative models. On the semi-synthetic MIMIC-III benchmark (=3) (Table 9), OrthoGen achieves a conditional-mean RMSE of 0.0639 ± 0.0143, compared with 0.0556 ± 0.0098 for the dedicated DR mean estimator, which is highly competitive. Interestingly, the generative PI and RA learners outperform their dedicated mean-based counterparts, while the mean-based IPTW and DR learners achieve lower errors.

    => These results show that our generative learners remain competitive for conditional-mean estimation, despite being trained to recover the full conditional outcome distribution rather than only its expectation. Importantly, the learned distributions also allow other quantities, such as quantiles, to be estimated without retraining.

    Action: We extended our empirical evaluation with four dedicated conditional-mean estimators (PI, RA, IPTW, and DR), reporting their comparison against our generative learners in the new Appendix I (Table 9) and summarizing the findings in Section 7.

We thank the reviewer again for their constructive feedback, which has helped us substantially improve the clarity and presentation of our work. We hope that the revisions adequately address the concerns raised, particularly regarding the motivation and explanation of our approach. We would be happy to provide any further clarifications or address additional questions or suggestions.

[1] Robins, J. M., Hernan, M. A., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550-560.

[3] Melnychuk, V., Frauen, D., & Feuerriegel, S. (2022, June). Causal transformer for estimating counterfactual outcomes. In International conference on machine learning (pp. 15293-15329). PMLR.

[3] Hess, Konstantin, et al. "Igc-net for conditional average potential outcome estimation over time." International Conference on Learning Representations. Vol. 2026. 2026.

[4] Frauen, D., Hess, K., & Feuerriegel, S. (2025, May). Model-agnostic meta-learners for estimating heterogeneous treatment effects over time. In International Conference on Learning Representations (Vol. 2025, pp. 85750-85779).

Thank you for the response.

The presentation concerns have been fully addressed, and the additional experiments are more convincing.

However, the compared methods remain limited to the original three baselines, and there is still a lack of comparison with a broader range of causal inference methods for time-varying treatments, including but not limited to [1-3].

Nevertheless, I will raise my score, as the current presentation now sufficiently conveys the paper's contributions and novelty.

Reference

[1] Shirakawa, T., Li, Y., Wu, Y., Qiu, S., Li, Y., Zhao, M., ... & Van Der Laan, M. (2024). Longitudinal targeted minimum loss-based estimation with temporal-difference heterogeneous transformer. Proceedings of machine learning research, 235, 45097.

[2] Wang, X., Lyu, S., Yang, L., Zhan, Y., & Chen, H. (2024, July). A dual-module framework for counterfactual estimation over time. In Forty-first International Conference on Machine Learning.

[3] Boughdiri, A., Josse, J., & Scornet, E. (2025, July). Quantifying treatment effects: Estimating risk ratios via observational studies. In Forty-second International Conference on Machine Learning.

Reviewer_2HBEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes OrthoGen, a generative orthogonal learning method for estimating potential outcome distributions under time-varying treatments. It combines recursive g-computation with doubly robust estimation and evaluates the method using normalizing flows and diffusion models.

Critical strengths

The paper makes a substantial connection between causal inference and generative modeling. The theoretical analysis is interesting, and the evaluation covers multiple datasets and generative backbones.

Critical weaknesses

The experiments do not directly validate the claimed double robustness and Neyman orthogonality. Controlled experiments with misspecified propensity and outcome models would help establish whether the proposed method is less sensitive to nuisance estimation errors.

The empirical improvements are not consistent across generative backbones. In Table 4, the diffusion-based OrthoGen model does not outperform the plug-in baseline at treatment horizons 1 and 3, and its improvement at horizon 5 is relatively small.

The real-world evaluation uses factual CRPS rather than counterfactual distribution accuracy. While this is understandable given the lack of counterfactual ground truth, it limits the evidence for the method's causal estimation performance on real data.

Questions for authors

Can the authors provide controlled experiments where the propensity model or continuation outcome model is deliberately misspecified?

Why does the diffusion-based OrthoGen model show weaker improvements than the normalizing-flow version on MIMIC-III?

How sensitive is OrthoGen to weak treatment overlap and extreme inverse-propensity weights at longer treatment horizons?

2 comments

We would like to thank the reviewer for their time in reviewing our paper and the constructive feedback. We have uploaded a new revision of our work, with all key changes highlighted in blue color. Below, we address all your comments.

  • Empirical validation of double robustness (W1; Q1).

    Thank you for raising this important point and giving us the opportunity to provide further evidence of the double robustness of our approach.

    We have added a new nuisance misspecification experiment in our new Table 8. We provide the full experimental design described in our new Appendix H (Controlled Nuisance Misspecification Experiment). Specifically, we use the known structural oracle to obtain the true propensity scores and outcome-continuation distributions, and introduce controlled perturbations to either nuisance component separately or jointly.

    The results support the expected robustness behavior. With oracle outcome nuisances, OrthoGen remains stable under increasing propensity misspecification (Wasserstein costs between 0.584 and 0.591). Under increasing outcome misspecification with oracle propensities, OrthoGen deteriorates only slightly (0.591 to 0.636), whereas RA deteriorates substantially (0.564 to 1.056). OrthoGen also achieves the lowest error under joint misspecification.

    => These findings provide direct empirical evidence of robustness to controlled nuisance perturbations, consistent with our theoretical results.

    Action: We added new experiments to empirically validate the double robustness property of our method (see new Table 8 & new Appendix H).

  • Consistency of empirical improvements (W2; Q2).

    Thanks for highlighting the differences in empirical performance across generative backbones. We would like to clarify two important points.

    First, we would like to highlight that, in the specific real-world setting + diffusion models that you mention, OrthoGen is very close to the best model for all the time steps. In fact, when considering the standard deviation, we cannot say that either one is directly better. Further, in the same real-world setting but with normalizing flows or in the synthetic and semi-synthetic experiments (where we have ground truth and then counterfactual estimation is directly evaluated), OrthoGen is most of the times the best model.

    On the other hand, we would like to clarify that the theoretical properties of OrthoGen is the robustness to nuisance estimation errors, rather than a guarantee of uniformly superior finite-sample performance across generative backbones. In particular, the error bound in Theorem 2 includes both the target-model learning error and a higher-order nuisance estimation term. Even though Neyman-orthogonality reduces sensitivity to nuisance estimation errors, the overall performance also depends on how accurately the target generative model is learned. Consequently, differences in optimization and approximation accuracy between NFs and DMs can affect the magnitude of the empirical gains.

    => Even if performance may depend on each particular instantiation due to differences in optimization and approximation accuracy, OrthoGen is highly competitive in all the evaluated settings, being the best model in most of them.

    Action: We expanded the discussion in Section 7 to clarify the backbone-dependent differences in empirical performance and their relationship to our theoretical guarantees.

  • Real-world evaluation (W3).

    Thanks for raising this point. In our revision, we add a clarification on this in the Metrics paragraph (Sec. 7, Evaluation). While the main sources of counterfactual evaluation are the synthetic and semi-synthetic experiments, our factual CRPS metric in the real-world experiment is an important check that demonstrates the applicability to real-world patient data. Hence, we mainly interpret this as a demonstration of the practical value for medical stakeholders, rather than as direct evidence of counterfactual estimation accuracy. Unfortunately, directly evaluating counterfactual estimation on real world data is not possible due to the lack of ground truths ([1, 2, 3, 4]), but we hope to show how practitioners could, in principle, apply our method in practice.

    Action: We clarified the scope and limitations of the real-world evaluation in Section 7, by emphasizing that the main purpose is to demonstrate applicability of our method to real-world patient data, rather than for benchmarking.

  • Sensitivity to weak overlap (Question 3).

    Thank you for raising this important question. We have added two new controlled experiments (new Appendix K, Tables 12–13) investigating sensitivity to weak treatment overlap at the longest treatment horizon ((τ=5\tau=5)). We progressively increase the strength of history-dependent treatment assignment and consider both alternating and all-treated intervention sequences.

    The results provide strong empirical evidence of OrthoGen's favorable behavior under challenging overlap conditions. Under the alternating regime, OrthoGen consistently achieves the lowest Wasserstein cost across all treatment-selection strengths. Under the all-treated regime, treatment overlap deteriorates substantially, with no complete-regime matches observed at the two highest selection strengths. Even under these extreme conditions, OrthoGen substantially outperforms IPTW and achieves the lowest mean error among all methods. At the highest selection strength, OrthoGen obtains a squared Wasserstein cost of 0.287, compared with 5.342 for IPTW, 0.313 for PI, and 0.308 for RA.

    => These findings provide further empirical evidence of OrthoGen's robustness to weak treatment overlap, demonstrating its ability to maintain competitive estimation accuracy even under severe limitations in observational support.

    Action: We added two new controlled overlap-sensitivity experiments, including detailed propensity-score and effective-sample-size diagnostics, in new Appendix K (Tables 12–13).

Thank you again for your constructive feedback, which we believe will help strengthen our work. We hope that our rebuttal and revision address your concerns, and are happy to provide any further clarifications or address additional questions or suggestions.

[1] Hernán, M. A., & Robins, J. M. (2010). Causal inference. Boca Raton, FL: CRC.

[2] Pearl, J. (2010). Causal inference. Causality: objectives and assessment, 39-58.

[3] Melnychuk, V., Frauen, D., & Feuerriegel, S. (2022, June). Causal transformer for estimating counterfactual outcomes. In International conference on machine learning (pp. 15293-15329). PMLR.

[4] Hess, Konstantin, et al. "Igc-net for conditional average potential outcome estimation over time." International Conference on Learning Representations. Vol. 2026. 2026.

Comments on the paper

1 comment

Dear reviewers and AC,

Thank you for your time in handling our submission and for the constructive reviews.

We have carefully addressed every concern raised by the reviewers, providing detailed replies and uploading a revised manuscript. Below, we summarize the main improvements:

  • Empirical validation of double robustness (new Appendix H, new Table 8). We added a new experiment using oracle nuisance components with deliberately misspecified propensity scores and outcome-continuation distributions. => The results show that OrthoGen is robust to nuisance misspecification, thus confirming our theoretical guarantees.
  • Comparison with dedicated estimators (new Appendix I, new Tables 9–11). We extended our evaluation to compare the conditional means and quantiles obtained from OrthoGen against estimators directly targeting these quantities. We additionally report computational costs. => OrthoGen is highly competitive but, at the same time, much more flexible.
  • Sensitivity to weak treatment overlap (new Appendix K, new Tables 12–13). We introduced two controlled experiments examining sensitivity to increasingly extreme treatment-selection mechanisms under different intervention sequences. The results demonstrate that OrthoGen maintains competitive counterfactual distribution estimation accuracy even under severe overlap deterioration, substantially outperforming IPTW and achieving the lowest mean error under the most challenging conditions. => These findings provide further empirical evidence of OrthoGen's robustness in settings where limited treatment overlap makes causal estimation particularly challenging.
  • Stronger motivation and clarification of functional guarantees (Sec. 3, new Appendix I). We expanded the motivation for learning full counterfactual distributions, by emphasizing patient-specific queries, multivariate outcomes, and the ability to evaluate multiple quantities without retraining. We also clarified how functional estimation error decomposes into generator estimation error and model-class approximation bias. => These additions strengthen the practical motivation for distributional learning and clarify the scope of our theoretical guarantees.
  • Improved presentation and methodological clarity. We expanded the explanation of time-varying confounding and the limitations of naive observational conditioning. We also introduced two new figures: a causal diagram illustrating treatment–confounder feedback (Fig. 2) and an end-to-end diagram of the OrthoGen training pipeline (Fig. 4). => These changes are aimed at making the underlying causal challenges and our proposed methodology easier to understand.

Taken together, these revisions provide new experimental evidence and a clearer understanding of OrthoGen's capabilities, theoretical properties, and motivation. We believe they substantially strengthen the paper and its contribution to treatment effect estimation and decision-making under uncertainty.

We are confident that, following the reviewers' constructive feedback and the resulting improvements, our work makes a valuable contribution to the community. We would like to thank the reviewers once again for their time and careful evaluation of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.