acceptodds
Under review as a conference paper at ICLR 2027

High-dimensional counterfactual analysis with latent neural differential equations

Abstract

Counterfactual estimation is important in longitudinal data analysis, particularly in healthcare, where the goal is to infer how an observed patient trajectory would have evolved under an alternative intervention. This problem is especially challenging in real-world settings with longitudinal data, where measurements are typically high-dimensional and exhibit complex temporal structure. Although causal inference has been studied extensively, existing approaches are often tailored to low-dimensional data and do not naturally accommodate continuous-time dynamics. We address these limitations by proposing CLL-NODE, a Counterfactual analysis of high-dimensional Longitudinal data using Latent Neural ODE model. Our model parameterizes latent continuous-time dynamics as a function of treatment assignment and instance-level auxiliary covariates, enabling treatment and covariate dependent evolution while supporting flexible interpolation and extrapolation in time. To facilitate reliable counterfactual estimation, we introduce an adversarial regularization term that encourages treatment-invariant representations of the initial latent state. We evaluate CLL-NODE on multiple datasets and demonstrate competitive performance relative to strong deep-learning baselines for high-dimensional counterfactual modeling.

Primary area:
learning on time series and dynamical systems
Keywords:
causal inference, counterfactual prediction, neural ode, variational inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_yXBYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors have presented a continuous-time causal latent variable model designed for counterfactual analysis of high-dimensional longitudinal data. This approach, CLL-NODE, parameterizes latent dynamics using Neural ODEs conditioned on treatment assignments and auxiliary covariates, and incorporates an adversarial regularization term to encourage treatment-invariant initial latent states. The authors evaluated the model on Causal MNIST, the Physionet ICU dataset, and a semi-synthetic variant where treatment effects follow a predefined ODE. Performance was assessed using a proxy metric based on the classification error difference of an oracle Transformer applied to factual versus counterfactual trajectories, alongside standard reconstruction and prediction metrics. The authors claim state-of-the-art counterfactual quality and improved causal identifiability.

Critical strengths
  • The authors have approached a critical gap in the literature. While a consistent suite of efforts such as G-Transformer has been presented for the continuous domain, continuous time-setting has been an oft ignored setup. This extension of deep causal latent variable modeling to continuous-time settings via Neural ODEs is well-motivated and addresses a meaningful gap in longitudinal causal inference.
  • The paper was well motivated and easy to read and comprehend. The paper reports two strong empirical results on established benchmarks that demonstrate competitive counterfactual generation and temporal interpolation/extrapolation capabilities.
  • Finally, the integration of adversarial regularization to decouple initial latent states from treatment assignment is a thoughtful design choice that aligns with standard causal identifiability assumptions.
Critical weaknesses

There are a few aspects in the paper that might need some deliberation

  • The proxy evaluation metric lacks theoretical grounding and may not reliably correlate with true counterfactual fidelity. This metric seems to essentially measures how well the model fools a downstream predictor rather than causal correctness.
  • The semi-synthetic evaluation was appreciated but could have had a stronger impact if instead of a hand-crafted ODE for treatment effects it was constructed/presented using ground-truth causal graphs. This would have improved the ability to validate whether the model actually recovers the underlying causal structure or merely fits the imposed dynamics.
  • The authors didn't report key practical concerns such as computational complexity and training stability, especially considering Neural ODE-based architectures with adversarial components potentially applied to high-dimensional, irregularly sampled clinical data.
  • Finally, the theoretical claims regarding improved identifiability from the adversarial component are asserted without formal proofs, and the experimental design lacks rigorous ablation studies to isolate the contribution of the adversarial term versus the Neural ODE dynamics.
Questions for authors
  • How does the adversarial regularization interact with the reconstruction and prior terms in the ELBO? Were there empirical or theoretical mechanisms are in place to prevent mode collapse or posterior collapse during training? Have you observed degenerate latent solutions or training instability across different random seeds?

  • The proxy counterfactual quality metric relies on a Transformer classifier’s error difference. Can you justify why this metric correlates with true counterfactual accuracy, or provide alternative evaluations such as intervention consistency checks, sensitivity analyses, or ground-truth counterfactuals on fully synthetic causal graphs?

2 comments

We thank the reviewer for their careful evaluation of our manuscript and for the constructive comments and suggestions. We address each point below.

  1. Counterfactual evaluation metric.

    We agree that there is no formal mathematical guarantee that the proposed evaluation metric correlates with the true counterfactual prediction error. The metric we employ is related to counterfactual effectiveness, extensively discussed by Monteiro et al. (2023) [1]. Our evaluation procedure, which employs an oracle predictor, follows De Sousa Ribeiro et al. (2025) [2].

    Counterfactual effectiveness assesses whether generated counterfactual outcomes reflect the intended intervention. Specifically, if the generated counterfactual trajectories adequately capture the characteristics associated with the target treatment T′T', the oracle classifier should recognize them as belonging to that treatment group with an accuracy comparable to its accuracy on real observations from the corresponding group.

    However, as discussed by De Sousa Ribeiro et al. (2025) [2], counterfactual effectiveness serves as a diagnostic measure of counterfactual soundness rather than a guarantee of counterfactual identifiability or causal validity. We therefore acknowledge that this metric provides only indirect evidence of counterfactual quality.

    To complement this evaluation, we also conducted experiments on a semi-synthetic PhysioNet dataset, where ground-truth counterfactual trajectories are available and counterfactual prediction errors can be computed directly. We have incorporated the corresponding references and a more detailed discussion of the evaluation methodology into the revised manuscript.

  2. Semi-synthetic evaluation.

    The ODE-based treatment effect was introduced to account for the longitudinal nature of the PhysioNet data. Specifically, modeling the treatment effect through a differential equation allows us to represent its continuous-time evolution and capture how treatment-induced changes in physiological measurements develop over time. This construction is intended to provide a more appropriate semi-synthetic evaluation setting for time-series counterfactual prediction. The results reported in the original submission correspond to a setting in which both covariates and treatment influence the observed outcomes, while treatment assignment is independent of the covariates. In the revised manuscript, we additionally consider a setting with covariate-dependent treatment assignment, which directly follows our causal graph in Figure 2. The corresponding results are presented in the table below, demonstrating that our method achieves performance comparable to that of CEVAE.

  3. Training stability and the ELBO objective.

    We did not observe training instabilities when employing the proposed multiple-shooting formulation. In contrast, training the model on complete trajectories without multiple shooting did not converge successfully. This issue is discussed in Appendix F, where we investigate the influence of shooting variables on model training and performance.

    Regarding the interaction between the ELBO objective and other model components, we did not introduce additional mechanisms specifically designed to prevent mode collapse or posterior collapse, as the proposed model exhibited stable training behavior without such interventions. Nevertheless, we carefully selected the hyperparameter λ\lambda, which controls the contribution of the ELBO term to the overall training objective, for each dataset, as described in Equation (5).

    Furthermore, the relatively small standard deviations reported across experimental runs provide empirical evidence of the robustness of our results to different random seeds.

  1. Counterfactual identification and the adversarial component.

    We thank the reviewer for highlighting this important distinction. We agree that identification has a precise technical meaning in causal inference and that our original formulation did not adequately distinguish counterfactual identification from improvements in counterfactual prediction.

    Our intended claim is that the adversarial component improves counterfactual prediction by encouraging the inferred initial latent state s0s_0 to contain less treatment-specific information. This objective is theoretically motivated by Lemma 1, which provides support for the proposed treatment-invariance regularization. However, we acknowledge that reducing treatment information in s0s_0 does not, by itself, guarantee counterfactual identifiability or the correctness of individual counterfactual predictions. To address this concern, we have revised the wording of contribution (ii) and the conclusions to clarify the scope of our claims.

    We would also like to emphasize that the manuscript includes an ablation study examining the contribution of adversarial regularization. Specifically, we compare the full CLL-NODE model with Causal-NODE, a latent neural ODE model without the adversarial component. In addition, Figure 4 illustrates the effect of the discriminator on the learned latent representations. Together, these results provide empirical evidence supporting the contribution of adversarial regularization to counterfactual prediction performance.

Table. Counterfactual prediction MSEs on semi-synthetic PhysioNet with covariate-dependent treatment assignment. Lower values are better.

MethodMSE
CLL-NODE1.483 ± 0.05
Causal-LNODE3.489 ± 0.1
VCI4.7 ± 1.0
CEVAE1.454 ± 0.02

In addition, in the revised pdf we also changed Figure 1 and included ablations regarding the observation sparsity and extrapolation. The revised text is marked with blue.

References

[1] Monteiro, M., De Sousa Ribeiro, F., Pawlowski, N., Castro, D. C., and Glocker, B. (2023). *Measuring Axiomatic Soundness of Counterfactual Image Models.

[2] De Sousa Ribeiro, F., Santhirasekaram, A., and Glocker, B. (2025). *Counterfactual Identifiability via Dynamic Optimal Transport.

Reviewer_1LMzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary
  • (S1) It addresses an important and underexplored setting: high-dimensional, irregularly sampled outcomes with continuous-time dynamics.
  • (S2) It identifies a real failure mode, where s0s_0 absorbs treatment information and the dynamics ignore TT; the no-adversary ablation clearly demonstrates it.
  • (S3) Lemma 1 makes the adversary's role explicit by linking it to I(T;s0)I(T; s_0).
  • (S4) The method outperforms all reported baselines on every benchmark.
Critical strengths
  • (S1) It addresses an important and underexplored setting: high-dimensional, irregularly sampled outcomes with continuous-time dynamics.
  • (S2) It identifies a real failure mode, where s0s_0 absorbs treatment information and the dynamics ignore TT; the no-adversary ablation clearly demonstrates it.
  • (S3) Lemma 1 makes the adversary's role explicit by linking it to I(T;s0)I(T; s_0).
  • (S4) The method outperforms all reported baselines on every benchmark.
Critical weaknesses
  • (W1) Good factual reconstruction does not show that s0s_0 is the right state to hold fixed: different splits between s0s_0 and the treatment dynamics could fit the same trajectory yet give different counterfactuals.
  • (W2) If baseline prognostic information in s0s_0 is correlated with XX, which drives treatment, minimizing marginal I(T;s0)I(T; s_0) may remove useful patient information even without treatment leakage. The paper's SCM avoids this by modeling s0s_0 independently of XX, an assumption that seems particularly strong for the PhysioNet setting.
  • (W3) No experiment cleanly tests the intended observational setting: treatment is randomized in Causal MNIST and semi-synthetic PhysioNet, while in Causal MNIST Groups the unobserved group drives both treatment and recovery and XX is only a noisy proxy, appearing to violate the paper's own ignorability assumption. Although Section 4.2 says this setup follows Fig. 2, where XX causes treatment, treatment is actually assigned from the unobserved group gg, and XX has no effect on the outcome.
  • (W4) Full-trajectory training fails due to ODE solver instability (Table 4), while the best result uses one shooting variable per datapoint. Yet counterfactual estimation discards the intermediate shooting states and performs an uninterrupted rollout from s0s_0, so inference relies on precisely the long-horizon dynamics that training avoids.
  • (W5) Seedat et al. [1] already combine continuous-time neural differential equations with adversarial balancing for counterfactual prediction under irregular sampling. The paper should discuss this work and clarify what is fundamentally new beyond adapting existing techniques to the proposed setting.

[1] N. Seedat, F. Imrie, A. Bellot, Z. Qian, M. van der Schaar. Continuous-Time Modeling of Counterfactual Outcomes Using Neural Controlled Differential Equations. ICML 2022.

Questions for authors
  • (Q1) Both potential trajectories are available in Causal MNIST. How do the s0s_0 inferred from yy and from y′y' of the same unit compare, given that the adversary only matches distributions across treatment groups?
  • (Q2) If s0s_0 can share prognostic information with XX, would I(T;s0∣X)I(T; s_0 \mid X), e.g. via a discriminator that also sees XX, be a more appropriate target than I(T;s0)I(T; s_0)?
  • (Q3) How does the method behave when treatment depends on an observed XX that also affects the dynamics, so that ignorability holds, and as treatment overlap shrinks?
  • (Q4) How does counterfactual error grow with rollout length from s0s_0?
5 comments

We thank the reviewer for their careful evaluation of our manuscript and for the constructive comments and suggestions. We address each point below.

W1. We fully agree that accurate reconstruction of factual trajectories does not necessarily imply accurate counterfactual predictions. This limitation is one of the primary motivations for introducing an adversarial discriminator into CLL-NODE. The discriminator acts as a regularization mechanism that encourages the inferred initial latent state to be invariant to treatment assignment. By reducing treatment-related information in the initial representation, the proposed approach aims to improve the estimation of counterfactual trajectories while preserving information relevant to individual-specific dynamics. Our experimental results demonstrate the importance of this component for counterfactual prediction accuracy.

W2. Thank you for commenting this. Under our modeling assumptions, the initial latent state s0s_0 is intended to capture individual-specific latent information that is not represented by the observed covariates xx. The covariate information enters into the generative model in latent space via the dynamics function fdynf_{\text{dyn}} that takes as input the latent state, covariates and treatment. As pointed out by the reiviewer, this design choice avoids the risk of adversarial discriminator to remove useful patient information yet allows latent variables and their dynamics to depend on covariates.

W3. We agree that the description and motivation of the Causal MNIST Groups dataset in the original submission may have been insufficient and unclear. With real world datasets, it is often difficult to ensure that all causal assumption in practice. Our objective was therefore also to assess the robustness of the proposed approach to potential deviations from the assumed causal structure. To address the reviewer's concern more directly, we conducted additional experiments using a modified version of the Causal MNIST Groups dataset in which a noisy proxy variable directly influences both treatment assignment and recovery. This modification provides a more explicit data-generating mechanism for the dependencies under consideration and follows the assumed structural causal mechanisms. The corresponding experimental results are presented below, and the revised manuscript includes a detailed description of the modified data-generating process. As can be seen from the table below, our method shows the best performance.

W4. We thank the reviewer for raising this important point. Training the model using full trajectories without multiple shooting leads to numerical instability in the ODE solver, primarily due to the propagation of gradients through long integration intervals during optimization. The multiple-shooting formulation mitigates this difficulty by dividing the trajectory into shorter integration intervals and introducing intermediate shooting variables. In contrast, counterfactual estimation is performed during inference that requires only solving the latent ODE forward in time, i.e., no gradient propagation through the ODE solver is needed. Consequently, integrating the dynamics over the entire trajectory does not exhibit the same numerical stability issues observed during training. Furthermore, intermediate shooting variables cannot be reused for counterfactual prediction, as only the initial latent state s0s_0 is explicitly regularized to be treatment-invariant. The intermediate shooting variables are inferred from factual observations and therefore encode information specific to the observed treatment. Counterfactual trajectories must instead be generated by integrating the dynamics under the counterfactual treatment assignment, starting from the inferred treatment-invariant initial state s0s_0.

W5. We thank the reviewer for highlighting this relevant work. The study by Seedat et al. (2022) [3] was not discussed in the original submission, but we discussed several other related methods (Wu et al., 2025[4]; Bica et al., 2020[1]). We acknowledge the relevance of (Seedat et al., 2022[3]) and have incorporated the corresponding discussion into the revised manuscript.

The Treatment Effect Neural Controlled Differential Equation (TE-CDE) proposed by Seedat et al. [3] addresses longitudinal counterfactual estimation in continuous time, including irregularly sampled observations. Its problem formulation is closely related to those considered by Bica et al. (2020) [1] and Melnychuk et al. (2022) [2], where treatment assignments may vary over time and representations of observed patient histories are used to estimate future counterfactual outcomes.

Our work considers a different experimental setting. Specifically, we assume a single treatment assignment at baseline and do not require the availability of pre-treatment observations. This distinction is particularly important because methods relying on pre-treatment histories cannot be applied directly when such observations are unavailable. CLL-NODE addresses this setting by inferring a treatment-invariant initial latent state from the observed post-treatment trajectory and subsequently generating counterfactual outcomes through continuous-time latent dynamics. To the best of our knowledge, CLL-NODE is the first approach specifically designed to estimate counterfactual trajectories for high-dimensional longitudinal outcomes under this combination of assumptions.

Q1. Thank you for the insightful comment and suggestion. To assess whether the inferred initial latent state s0s_0 preserves individual-specific information while remaining invariant to treatment assignment, we conducted an additional experiment comparing distances between latent representations. Specifically, we inferred initial latent states s0s_0 and s0′s_0' from the factual and counterfactual trajectories of the same instance, respectively, and computed their Euclidean distance, ∣s0−s0′∣2|s_0-s_0'|_2. As a reference, we computed the corresponding distances between initial latent states inferred from randomly selected, distinct instances. The resulting distributions are presented in the figure in the revised pdf. The distances between factual and counterfactual representations of the same instance are much more concentrated near zero and exhibit lower variance than those between randomly selected instances. These results suggest that the inferred initial latent state primarly captures individual-specific information.

Q2. Under our modeling assumptions, the initial latent state s0s_0 is intended to capture individual-specific latent information that is not represented by the observed covariates XX. Please also see our response to W2 above. We agree that, under alternative assumptions in which the latent initial state and observed covariates share prognostic information, minimizing the conditional mutual information I(T;s0∣X)I(T;s_0\mid X) may be a more appropriate objective. We acknowledge this distinction and consider conditional adversarial regularization a promising direction for future work.

Q3. The corresponding results are presented in the table below. CLL-NODE continues to achieve the best performance among the evaluated methods. A detailed description of this mechanism and the corresponding experimental results are incorporated into the revised manuscript.

Q4. We conducted additional experiments to evaluate the extrapolation capabilities of CLL-NODE on the Causal MNIST Groups dataset. Specifically, we assessed counterfactual prediction accuracy at time points outside the range of factual trajectory. The results, presented below and included in the revised manuscript, demonstrate that CLL-NODE maintains strong predictive performance when extrapolating beyond the observed temporal interval.

In addition, we evaluated the sensitivity of the proposed approach to observation sparsity by reducing the number of observed time points per trajectory. The corresponding results are presented below and included in the revised manuscript. These experiments further demonstrate the robustness of CLL-NODE when counterfactual trajectories must be estimated from limited longitudinal observations.

We made other small changes to the revised pdf file, for example, modified Figure 1, and our modifications in text are marked with blue.

Table 1. Counterfactual prediction MSEs on Causal MNIST Groups. Lower values are better.

MethodRegularIrregular
CLL-NODE0.010 ± 0.0010.010 ± 0.001
Causal-LNODE0.022 ± 0.0010.022 ± 0.001
VCI0.021 ± 0.0010.022 ± 0.001
CEVAE0.034 ± 0.0010.034 ± 0.001
CF-NODE0.068 ± 0.0020.072 ± 0.005

Table 2. Counterfactual prediction MSEs on Causal MNIST Groups under observation sparsity and extrapolation. Lower values are better.

Method5 pts (Regular)5 pts (Irregular)10 pts (Regular)10 pts (Irregular)Extrapolation (Irregular)
CLL-NODE0.012 ± 0.0020.016 ± 0.0020.011 ± 0.0010.012 ± 0.0010.011 ± 0.002
Causal-LNODE0.023 ± 0.0010.020 ± 0.0010.021 ± 0.0010.023 ± 0.0010.025 ± 0.001
VCI0.025 ± 0.0010.025 ± 0.0010.025 ± 0.0010.023 ± 0.001--
CEVAE0.025 ± 0.0010.017 ± 0.0010.026 ± 0.0010.026 ± 0.001--
CF-NODE0.061 ± 0.0030.071 ± 0.0020.076 ± 0.0020.069 ± 0.0040.065 ± 0.002

References

[1] Bica, I., Alaa, A. M., Jordon, J., and van der Schaar, M. (2020). Estimating Counterfactual Treatment Outcomes over Time Through Adversarially Balanced Representations. [2] Melnychuk, V., Frauen, D., and Feuerriegel, S. (2022). Causal Transformer for Estimating Counterfactual Outcomes. [3] Seedat, N., Imrie, F., Bellot, A., Qian, Z., and van der Schaar, M. (2022). Continuous-Time Modeling of Counterfactual Outcomes Using Neural Controlled Differential Equations. [4] Yulun Wu, Louis McConnell, and Claudia Iriondo. Counterfactual generative modeling with variational causal inference. In The Thirteenth International Conference on Learning Representations, 2025

Reviewer_g4aJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLL-NODE, a latent Neural ODE framework for counterfactual prediction in high-dimensional longitudinal data. The method incorporates treatment- and covariate-dependent continuous-time dynamics, sparse Bayesian multiple shooting for stable trajectory modeling, and adversarial regularization to encourage treatment-invariant initial latent representations. Experiments on synthetic Causal MNIST variants and PhysioNet demonstrate improved counterfactual prediction performance compared with several deep learning baselines.

Critical strengths

The paper addresses a relevant problem in longitudinal counterfactual prediction, particularly when only post-treatment observations are available. Modeling treatment-dependent dynamics in continuous time is potentially useful for irregularly sampled data. The reported results on synthetic benchmarks show consistent improvements over the selected baselines, and the adversarial regularization appears to provide meaningful empirical benefits.

Critical weaknesses
  1. Limited methodological novelty. CLL-NODE primarily combines existing latent Neural ODEs with sparse multiple shooting (Iakovlev et al., ICLR 2023) and adversarial representation balancing (Bica et al., ICLR 2020). Conditioning the dynamics on treatment and covariates appears to be a straightforward extension. The paper should clarify its core technical novelty beyond integrating these established components.

  2. Insufficient justification of counterfactual identifiability. The model infers a shared initial latent state exclusively from post-treatment observations, without establishing when this state can be reliably recovered. Moreover, minimizing I(T;s0)I(T;s_0) encourages treatment-invariant representations but does not guarantee correct individual counterfactuals. Stronger theoretical justification or controlled validation is needed.

  3. Potential violation of causal assumptions. In Causal MNIST Groups, an unobserved group variable influences both treatment assignment and recovery, while only a noisy proxy is observed. This may violate the conditional ignorability assumption underlying the method. The paper should clarify whether its causal assumptions hold in this benchmark.

  4. Insufficiently fair baseline comparisons. VCI and CEVAE treat individual time points as separate samples, whereas CLL-NODE exploits complete trajectories. CF-NODE additionally requires artificially generated pre-treatment observations. These differences complicate fair comparison. Stronger sequence-aware baselines under matched information and model capacity are needed.

  5. Evaluation does not fully support the claims. The synthetic tasks involve relatively simple intensity transformations, while the real-world PhysioNet evaluation relies on a classifier-based proxy that does not directly measure counterfactual accuracy. Furthermore, systematic analyses of temporal extrapolation, observation sparsity, and computational efficiency are missing despite being central motivations of the method.

Questions for authors

See weakness

5 comments

We thank the reviewer for their careful evaluation of our manuscript and for the constructive comments and suggestions. We address each point below.

  1. We agree with reviewer that the proposed CLL-NODE method utilizes previous methodologies, specifically latent neural ODEs, multiple shooting, and adversarial regularization. In addition, CLL-NODE carefully extends the underlying latent dynamics model with subject-specific covariates and treament such that the high-dimensional counterfactuals can be predicted. We also formally establish adversarial balancing for deep latent variable dynamics models in the context of longitudinal design in Lemma 1. To the best of our knowledge, the proposed CLL-NODE is the first method that can predict counterfactual trajectories in high-dimensional, longitudinal setting.

  2. We would like to clarify that the inferred initial state is not shared across trajectories; rather, each trajectory has its own initial state. The term "common," used in the appendix, was intended to indicate that a single initial state is shared across all observations within a given trajectory.

    We thank reviewer for raising question of identifiability of the initial state. Our proposed CLL-NODE model builds on latent neural ODEs, which is a type of deep latent variable model. Formal identifiability results for deep latent variable models can be obtained by incorporating additional information and structural assumptions. In particular Khemakhem et al. (2020)[1] show that the latent variables are asymptotically (up to affine transformation) identifiable when the latent prior is conditioned on observed auxiliary information, (see Khemakhem et al. (2020)[1] for technical details). These results concern identification of latent generative structure. Transferring them to longitudinal counterfactual prediction additionally requires establishing that all observationally compatible models agree on the counterfactual trajectories of interest, which does not hold in general for multi-dimensional latent variables (Nasr-Esfahany and Kiciman (2023)[2]). Consequently, adopting an auxiliary conditioned prior alone would not resolve counterfactual identification question. We therefore have defined our goal as estimating counterfactual predictive distribution (see problem setup in Section 3.1). We acknowledge that our terminology was not completely consistent: we have therefore carefully revised our manuscript to avoid suggesting formal identifiability claims. We note that comparable deep latent variable methods for counterfactual estimation such as CEVAE (Rissanen and Marttinen (2021)[3]) and VCI (Wu et.al (2025)[4]) likewise do not provide such guarantees.

    We agree that minimizing I(T;s0)I(T;s_0) encourages treatment-invariant representations but does not guarantee accurate individual counterfactual predictions. More generally, a model trained exclusively on factual observations cannot guarantee the correctness of individual counterfactual predictions without additional assumptions. This limitation is commonly referred to as the fundamental problem of causal inference. In this work, we aim to mitigate this limitation by combining factual-data training with adversarial regularization, which encourages the learned initial-state representations to be treatment-invariant.

  3. We agree that the original description of the Causal MNIST Groups dataset may have been unclear. In practice, ensuring that a dataset strictly satisfies all assumptions encoded in a particular causal graphical model is challenging. Our objective was therefore to investigate whether the proposed model remains robust to potential violations of these assumptions.

    To address this concern more directly, we have conducted additional experiments using a modified dataset in which a noisy proxy variable directly influences both treatment assignment and recovery, as in our structural causal model in Figure 2. The corresponding results are presented below, and the revised manuscript provides a detailed description of the data-generating mechanism.

  4. To the best of our knowledge, existing time-aware methods, including those proposed by Bica et al. (2020)[5] and Melnychuk et al. (2022)[6], are primarily designed for settings involving sequential treatment assignment at multiple time points and assume the availability of observations before the treatment interventions. In contrast, our setting considers a single treatment assigned at baseline, with no pre-treatment observations available. This fundamental difference in the problem formulation limits the applicability of these methods as direct baselines for comparison.

  1. We thank reviewer for this comment. Our metric for PhysioNet is the instance of counterfactual effectiveness from [7]. Counterfactual effectiveness states that if we intervene to set the treatment to T′T', the resulting counterfactual outcome should reflect treatment T′T'. Since the true counterfactual is not observed, we check it with pseudo-oracle: a classifier trained on observed data to predict the treatment from trajectory. If the generated counterfactual trajectories adequately capture the characteristics associated with the target treatment T′T', the pseudo-oracle should recognise them as belonging to T′T' about as accurately as it recognizes real trajectories as belonging to their actual treatment. Our setup resembles the structure in De Sousa Ribeiro et al. (2025)[8] which uses the same intuition.

    We agree with reviewer that effectiveness is necessary condition rather than direct measure of counterfactual accuracy, and we modify paper accordingly to clarify this difference and add references. However as noted by Louizos,et. al, 2017 [9], evaluating causal inference methods on real data is challenging as ground truth causal outcomes are unavailable. Common solution to this problem is to construct semi-synthetic datasets in which real data are modified so that true counterfactual outcomes are known. Following this practice, we constructed a semi-synthetic version of PhysioNet and report counterfactual MSE in Table 3, which directly measures counterfactual accuracy.

    To investigate the effects of observation sparsity and extrapolation, we have conducted additional experiments on Causal MNIST Groups with varying numbers of observations per trajectory, as well as an extrapolation experiment evaluating our model at time points outside the range of factual trajectory. The results, presented below and included in the revised manuscript, demonstrate that CLL-NODE maintains strong predictive performance when extrapolating beyond the observed temporal interval and is robust when counterfactual trajectories must be estimated from limited longitudinal observations.

    Finally, we discuss computational efficiency in the context of the multiple-shooting formulation and provide ablation experiments investigating the role of shooting variables in Appendix F.

    We made other small changes to the revised pdf file, for example, modified Figure 1 and added additional figure regarding initial latent state s0s_0 in the Appendix G. The revised text is marked with blue.

Table 1. Counterfactual prediction MSEs on Causal MNIST Groups. Lower values are better.

MethodRegularIrregular
CLL-NODE0.010 ± 0.0010.010 ± 0.001
Causal-LNODE0.022 ± 0.0010.022 ± 0.001
VCI0.021 ± 0.0010.022 ± 0.001
CEVAE0.034 ± 0.0010.034 ± 0.001
CF-NODE0.068 ± 0.0020.072 ± 0.005

Table 2. Counterfactual prediction MSEs on Causal MNIST Groups under observation sparsity and extrapolation. Lower values are better.

Method5 pts (Regular)5 pts (Irregular)10 pts (Regular)10 pts (Irregular)Extrapolation (Irregular)
CLL-NODE0.012 ± 0.0020.016 ± 0.0020.011 ± 0.0010.012 ± 0.0010.011 ± 0.002
Causal-LNODE0.023 ± 0.0010.020 ± 0.0010.021 ± 0.0010.023 ± 0.0010.025 ± 0.001
VCI0.025 ± 0.0010.025 ± 0.0010.025 ± 0.0010.023 ± 0.001--
CEVAE0.025 ± 0.0010.017 ± 0.0010.026 ± 0.0010.026 ± 0.001--
CF-NODE0.061 ± 0.0030.071 ± 0.0020.076 ± 0.0020.069 ± 0.0040.065 ± 0.002

[1]. Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. [2]. Arash Nasr-Esfahany and Emre Kiciman. "Counterfactual (non-) identifiability of learned structural causal models." [3]. Severi Rissanen and Pekka Marttinen. A critical look at the consistency of causal estimation with deep latent variable models. In Advances in Neural Information Processing Systems, 2021. [4]. Yulun Wu, Louis McConnell, and Claudia Iriondo. Counterfactual generative modeling with variational causal inference. In The Thirteenth International Conference on Learning Representations, 2025. [5] Bica, I., Alaa, A. M., Jordon, J., and van der Schaar, M. (2020). Estimating Counterfactual Treatment Outcomes over Time Through Adversarially Balanced Representations. [6] Melnychuk, V., Frauen, D., and Feuerriegel, S. (2022). Causal Transformer for Estimating Counterfactual Outcomes. [7] Miguel Monteiro, Miguel, et al. "Measuring axiomatic soundness of counterfactual image models." The Eleventh International Conference on Learning Representations (2023). [8] Fabio De Sousa Ribeiro, Ainkaran Santhirasekaram, and Ben Glocker. "Counterfactual identifiability via dynamic optimal transport."Advances in Neural Information Processing Systems 38 (2026): 147855-147903. [9]Christos Louizos,et. al. Causal effect inference with deep latent-variable models. Advances in neural information processing systems, 30, 2017.

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

The paper proposes CLL-NODE for high-dimensional longitudinal counterfactual estimation. It models continuous-time trajectories with latent neural ODEs conditioned on treatment and auxiliary covariates and uses adversarial regularization to reduce treatment information in the initial latent state. Experiments include regular and irregular Causal MNIST variants, a more confounded Causal MNIST Groups setting, and semi-synthetic and observational PhysioNet evaluations. CEVAE, VCI, CF-NODE, and a non-adversarial CLL-NODE variant are used as baselines. 14320_High_dimensional_counter

Critical strengths

The method jointly addresses high-dimensional observations, irregular sampling, continuous-time evolution, and counterfactual prediction.

Critical weaknesses
  1. The baseline comparison is difficult to interpret uniformly. CEVAE and VCI are adapted by treating individual trajectory elements as separate observations, while CF-NODE assumes a different observation setting and cannot be evaluated on PhysioNet. 14320_High_dimensional_counter
  2. For observational PhysioNet, counterfactual quality is assessed using the change in accuracy of a treatment classifier. The connection between this proxy and counterfactual prediction accuracy is indirect. 14320_High_dimensional_counter
Questions for authors

See the weakness. How should the comparison with VCI, CEVAE, and CF-NODE be interpreted given the different temporal and observational assumptions required by these baselines? What justifies using the treatment-classification accuracy difference as a proxy for counterfactual quality on observational PhysioNet?

4 comments

We thank the reviewer for their careful evaluation of our manuscript and for the constructive comments and suggestions. We address each point below.

  1. We agree with the reviewer that the considered baselines rely on different assumptions regarding the observational setting. However, we adapted these methods to our experimental setup, as detailed in Appendix E.

    Our choice of baselines is motivated by the fact that existing time-aware causal inference methods, including those proposed by Bica et al. (2020) [1] and Melnychuk et al. (2022) [2], are primarily designed for settings involving sequential treatment assignment at multiple time points and assume the availability of observations preceding treatment interventions. In contrast, our setting considers a single treatment assigned at baseline, without access to pre-treatment observations. These fundamental differences in the problem formulation limit the applicability of such methods as direct baselines for comparison.

  2. We agree that there is no formal mathematical guarantee that the proposed evaluation metric correlates with the true counterfactual prediction error. The metric we employ is related to counterfactual effectiveness, extensively discussed by Monteiro et al. (2023) [3]. Our evaluation procedure, which employs an oracle predictor, follows De Sousa Ribeiro et al. (2025) [4].

    Counterfactual effectiveness assesses whether generated counterfactual outcomes reflect the intended intervention. Specifically, if the generated counterfactual trajectories adequately capture the characteristics associated with the target treatment T′T', the oracle classifier should recognize them as belonging to that treatment group with an accuracy comparable to its accuracy on real observations from the corresponding group.

    However, as discussed by De Sousa Ribeiro et al. (2025) [4], counterfactual effectiveness serves as a diagnostic measure of counterfactual soundness rather than a guarantee of counterfactual identifiability or causal validity. We therefore acknowledge that this metric provides only indirect evidence of counterfactual quality.

    To complement this evaluation, we also conducted experiments on a semi-synthetic PhysioNet dataset, where ground-truth counterfactual trajectories are available and counterfactual prediction errors can be computed directly. We have incorporated the corresponding references and a more detailed discussion of the evaluation methodology into the revised manuscript.

    In addition, in the revised pdf we also changed Figure 1, included ablations regarding the observation sparsity and extrapolation and added additional setting for semi-synthetic Physionet. The revised text is marked with blue.

References

[1] Bica, I., Alaa, A. M., Jordon, J., and van der Schaar, M. (2020). Estimating Counterfactual Treatment Outcomes over Time Through Adversarially Balanced Representations.

[2] Melnychuk, V., Frauen, D., and Feuerriegel, S. (2022). Causal Transformer for Estimating Counterfactual Outcomes.

[3] Monteiro, M., De Sousa Ribeiro, F., Pawlowski, N., Castro, D. C., and Glocker, B. (2023). *Measuring Axiomatic Soundness of Counterfactual Image Models.

[4] De Sousa Ribeiro, F., Santhirasekaram, A., and Glocker, B. (2025). *Counterfactual Identifiability via Dynamic Optimal Transport.

Thank you to the authors for their response and clarification. I accept the points that the initial state is shared within the trajectory and that the claim has been adjusted to a predicted distribution. However, I still prefer to retain the original score: the new experiments mainly focus on Causal MNIST Groups and semi-synthetic PhysioNet, lacking larger-scale data validation, and their generalization is unclear; the core novelty and the essential differences from Bica et al. (2020) are also not fully explained. Therefore, I currently prefer to maintain the original score.

We would like also to clarify in details the difference between our method and CRN of Bica et al. (2020):

The key distinction between CLL-NODE and Bica et al. (2020) lies in the causal questions they address, rather than the architectural components they share. CRN estimates E[Yt+τ(aˉ)∣Hˉt]\mathbb{E}[Y_{t+\tau}(\bar{a}) \mid \bar{H}_t], i.e., the expected future outcome under a treatment plan aˉ=(at,…,at+τ−1)\bar{a}=(a_t,\dots,a_{t+\tau-1}), conditional on the observed history Hˉt=(Xˉt,Aˉt−1,V)\bar{H}_t=(\bar{X}_t,\bar{A}_{t-1},V), which comprises time-varying covariates, past treatments, and static features. Since this history contains only information available before the future treatment plan is applied, CRN does not condition on outcomes affected by the treatment whose effect it aims to estimate.

In contrast, CLL-NODE estimates the full counterfactual predictive distribution p(y′∣y,x,T,T′,t)p(y' \mid y,x,T,T',t) (Section 3.1), where yy is the trajectory observed after receiving treatment TT, and y′y' is the counterfactual trajectory under T′=1−TT'=1-T. Thus, unlike CRN, CLL-NODE conditions on observations affected by the received treatment. This distinction is particularly important when treatment is assigned at baseline and no pre-treatment observations are available. In this case, the history available to CRN reduces to static features, Hˉt=V\bar{H}_t=V, so subjects sharing the same VV receive the same conditional prediction. CLL-NODE, however, can still produce subject-specific counterfactual predictions by inferring the initial state s0s_0 from the observed post-treatment trajectory. Conditioning on post-treatment outcomes introduces an additional challenge, which also explains the different roles of adversarial training in the two methods.

In CRN, treatments are assigned repeatedly based on evolving patient histories. Adversarial balancing is therefore applied at each time step to learn representations of Hˉt\bar{H}_t that are not predictive of treatment assignment (Bica et al., 2020, Section 4). In CLL-NODE, treatment is assigned once at baseline, and confounding arises through the static covariates xx. Under our structural causal assumptions (Appendix A.1 and Figure 2), the initial state s0s_0 is unaffected by treatment and does not determine treatment assignment, while xx is the confounder of TT and subsequent latent states z(t)z(t). We account for this confounding by conditioning the dynamics fθ(z,T,x)f_\theta(z,T,x) on xx, under the ignorability assumption (Appendix A.1).

However, because the observed trajectory yy depends on the received treatment TT, the encoder may inadvertently encode treatment information into the inferred initial state s0s_0. To address this issue, CLL-NODE applies adversarial regularization specifically to s0s_0, minimizing I(T;s0)I(T;s_0) (Lemma 1) to encourage treatment invariance. Importantly, subsequent latent states z(t)z(t), for t>t0t>t_0, are deliberately allowed to depend on treatment, since they must capture the treatment effects that the model aims to predict.

To demonstrate the effectiveness of our method compared to CRN in this setting, we additionally evaluate CRN on Causal MNIST and Causal MNIST Groups. Following a strategy similar to that used for CF-ODE, we generate additional pre-treatment trajectories using the same data-generating mechanism as the original trajectories. These pre-treatment observations provide CRN with the historical information required for counterfactual prediction. The corresponding results are presented in the tables below:

Table. Counterfactual prediction MSEs. Lower values are better.

MethodCausal MNIST RegularCausal MNIST IrregularCausal MNIST Groups RegularCausal MNIST Groups Irregular
CLL-NODE0.009 ± 0.0020.013 ± 0.0020.010 ± 0.0010.010 ± 0.001
Causal-LNODE0.066 ± 0.0010.065 ± 0.0010.022 ± 0.0010.022 ± 0.001
VCI0.014 ± 0.0010.026 ± 0.0010.021 ± 0.0010.022 ± 0.001
CEVAE0.045 ± 0.0010.045 ± 0.0010.034 ± 0.0010.034 ± 0.001
CRN0.023 ± 0.0010.022 ± 0.0010.017 ± 0.0010.018 ± 0.001
CF-ODE0.065 ± 0.0030.068 ± 0.0020.068 ± 0.0020.072 ± 0.005

Table. Counterfactual prediction MSEs on Causal MNIST Groups under observation sparsity and extrapolation. Lower values are better.

Method5 points (Regular)5 points (Irregular)10 points (Regular)10 points (Irregular)Extrapolation (Irregular)
CLL-NODE0.012 ± 0.0020.016 ± 0.0020.011 ± 0.0010.012 ± 0.0010.011 ± 0.002
Causal-LNODE0.023 ± 0.0010.020 ± 0.0010.021 ± 0.0010.023 ± 0.0010.025 ± 0.001
VCI0.025 ± 0.0010.025 ± 0.0010.025 ± 0.0010.023 ± 0.001--
CEVAE0.025 ± 0.0010.017 ± 0.0010.026 ± 0.0010.026 ± 0.001--
CRN0.017 ± 0.0010.019 ± 0.0010.017 ± 0.0010.019 ± 0.0010.021 ± 0.001
CF-ODE0.061 ± 0.0030.071 ± 0.0020.076 ± 0.0020.069 ± 0.0040.065 ± 0.002

[1] Bica, I., Alaa, A. M., Jordon, J., and van der Schaar, M. (2020). Estimating Counterfactual Treatment Outcomes over Time Through Adversarially Balanced Representations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.