acceptodds
Under review as a conference paper at ICLR 2027

Reducing distribution drift over time for longitudinal counterfactual regression

Abstract

Estimating counterfactual outcomes under treatment sequences from longitudinal observational data is essential for time-series personalized decision-making. In this setting, the counterfactual regression model is trained on observations from preceding time steps and used to predict outcomes at future time steps. Therefore, it is crucial for the regression model to generalize across time steps. Existing methods typically address time-varying confounding by balancing representations of different treatment groups, while overlooking the temporal distribution drift of learned representations, which can hinder the generalization ability of the regression model over time. Motivated by this, we propose a longitudinal counterfactual regression method to reduce the temporal distribution drift of learned representations. Specifically, we characterize the temporal distribution drift using the Wasserstein distance between representations at consecutive time steps. Based on this, we establish that the future counterfactual outcome estimation error can be upper bounded by the current factual error and the distribution discrepancies measured across both treatment groups and time steps. Such a theoretical result naturally motivates us to jointly mitigate confounding bias and temporal drift. To achieve this, we propose to learn balanced representations by simultaneously reducing the distribution discrepancies across treatment groups and consecutive time steps. We conduct experiments on benchmark datasets to demonstrate the effectiveness of our method for longitudinal counterfactual regression.

Primary area:
causal reasoning
Keywords:
Longitudinal counterfactual regression, Temporal distribution drift, Distribution alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UFL8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ReCT for estimating counterfactual outcomes under treatment sequences from longitudinal observational data. The method combines an autoregressive encoder-decoder with adversarial treatment balancing and a temporal alignment penalty based on Wasserstein distances between representation distributions at consecutive time steps. The paper presents error bounds intended to relate future counterfactual prediction error to factual prediction error and distribution discrepancies across treatment groups and time steps. Evaluation includes synthetic tumor-growth data, semi-synthetic MIMIC-III data, and factual prediction on real-world MIMIC-III data. Ablation studies and experiments incorporating temporal alignment into CRN and CT assess the contribution and applicability of the proposed regularizer.

Critical strengths

1.The temporal alignment objective can be incorporated into existing longitudinal prediction models. The experiments with CRN and CT in Tables 5 and 6 show improvements at longer prediction horizons after adding the temporal alignment loss.

2.The synthetic and semi-synthetic experiments provide evidence of improved counterfactual prediction at longer horizons.

Critical weaknesses

1.The central theoretical guarantees rely on temporal consistency conditions that are not explicitly stated or justified. In Eq. (30), the proof replaces ℓk−1a\ell_{k-1}^{a} with ℓka\ell_k^{a}. Eq. (33) makes a corresponding replacement for the treatment-averaged losses at consecutive time steps. The stated Lipschitz assumptions constrain variation across representations and do not justify these replacements across time. Changes in the treatment probabilities πk\pi_k also need to be accounted for. Furthermore, Eq. (37) invokes a temporal consistency condition from Theorem 3, although that condition is absent from the theorem statement. Identical representation distributions do not, by themselves, imply identical outcome response functions or prediction losses. Therefore, the main bounds are not established under the assumptions as stated. Appendix B.4 additionally assumes a fixed model independent of the sampled trajectories, leaving the applicability of Theorem 4 to a model learned from those trajectories unclear. These gaps weaken the theoretical justification for the proposed objective.

2.Hyperparameter selection and implementation details are insufficiently specified.

Questions for authors
  1. What precise temporal consistency assumptions justify the replacements in Eqs. (30) and (33), and the equality in Eq. (37)? How do these assumptions account for changes in both outcome response functions and treatment probabilities? If temporal invariance is not assumed, can the bounds be corrected to include the corresponding drift terms?

  2. Theorem 4 in Appendix B.4 assumes a fixed model independent of the sampled trajectories. Does the guarantee extend to ReCT when its representations and predictor are learned from those same trajectories? Please provide the supporting argument or clarify the scope of the result.

  3. What values of λtem\lambda_{\mathrm{tem}}, λadv\lambda_{\mathrm{adv}}, and the Sinkhorn regularization coefficient were used for each dataset and backbone, and how were they selected? Please also specify the architecture, training configuration, and Sinkhorn differentiation settings used to obtain the reported results.

3 comments

Weakness 1 and Question 1: Explanation of temporal consistency and treatment probability shifts.

Answer: Thank you for your helpful comment. The temporal consistency assumptions underlying the replacement in Eq. (30) are that the representation captures the information needed for conditional mean potential outcome prediction and that the outcome mechanism given the representation and treatment is time-invariant. These assumptions are consistent with prior work: [a] relates conditional invariance to conditioning on "the complete set of corresponding direct causal predictors,'' while [b] assumes a shared response model across time while allowing "time dependencies and arbitrary changes in the distribution of XtX_t.'' Under our temporal consistency assumptions, the true response function is shared across time: $\mu_k^a(z)=\mu_{k-1}^a(z)=\mu^a(z)$. Together with the shared prediction head gθ(z,a)g_\theta(z,a), this implies ℓka(z)=ℓ(gθ(z,a),μa(z))=ℓk−1a(z)\ell_k^a(z) = \ell\bigl(g_\theta(z,a),\mu^a(z)\bigr) = \ell_{k-1}^a(z), justifying the pointwise loss replacement in Eq. (30).

For treatment probabilities, when representation sufficiency and mechanism invariance also hold for treatment assignment, the conditional assignment function ea(z)e_a(z) is shared across time. However, the marginal probability $\pi_k(a)=\int e_a(z)dP_k(z)$ still depends on the representation distribution at each time step. Therefore, the treatment-averaged replacements in Eqs. (33) and (37) must additionally account for changes in πk\pi_k. With pointwise losses bounded in [0,Cℓ][0,C_\ell], this contribution can be bounded by ${C_\ell}\mathrm{TV}(\pi_k,\pi_{k-1})=\frac{C_\ell}{2}\sum_{a\in\mathcal{A}}\left|\pi_k(a)-\pi_{k-1}(a)\right|$. If each eae_a is additionally LaL_a-Lipschitz, then ${C_\ell}\mathrm{TV}(\pi_k,\pi_{k-1})\leq\frac{C_\ell}{2}\left(\sum_{a\in\mathcal A}L_a\right)W(P_k,P_{k-1})$.

Thus, this additional contribution can also be bounded by temporal representation discrepancy, supporting Our method's (ReCT) temporal distribution alignment objective. We will explicitly state these assumptions and revise the corresponding bounds and derivations.

Weakness 2 and Question 3: Hyperparameter selection and implementation details.

Answer: Thank you for your helpful comment. We have expanded the implementation details to clarify hyperparameter selection, model architecture, training configuration, and Sinkhorn differentiation.

Hyperparameter selection. The selected values of λadv\lambda_{\mathrm{adv}}, λtem\lambda_{\mathrm{tem}}, and the Sinkhorn regularization coefficient ε\varepsilon across datasets and backbones are reported in Table~3. The hyperparameter search spaces, including learning rate, batch size, and network dimensions, are provided in the additional table. Hyperparameters are selected based on the mean factual validation RMSE over prediction horizons 1--10.

Model architecture. ReCT employs a GRU-based encoder--decoder architecture with MLP-based representation mapping, outcome prediction, and treatment classification modules. The detailed architecture is summarized in the accompanying table.

Training configuration. ReCT jointly optimizes outcome prediction, adversarial treatment balancing, and temporal distribution alignment. We use AdamW for the representation and outcome prediction networks, and SGD with momentum 0.9 for the treatment classifier. Early stopping is performed based on factual validation MSE, and all results are reported as the mean and standard deviation over five random seeds.

Sinkhorn implementation and differentiation. We use uniform empirical marginals and dimension-normalized squared Euclidean transport costs between representations. The transport plan is computed using 50 log-domain Sinkhorn iterations, with gradients backpropagated through all iterations. The temporal alignment loss is computed as the inner product of the resulting transport plan and cost matrix, while ε\varepsilon controls the entropy regularization during transport-plan computation.

ModelDatasetλadv\lambda_{\mathrm{adv}}λtem\lambda_{\mathrm{tem}}ε\varepsilon
ReCTSynthetic10.10.1
Semi-synthetic150.05
Real-world150.05
CRN + Ltem\mathcal{L}_{\mathrm{tem}}Synthetic10.50.1
Semi-synthetic10.50.05
CT + Ltem\mathcal{L}_{\mathrm{tem}}Synthetic15.00.05
Semi-synthetic10.10.05
HyperparameterSearch space
Learning rate{0.0001,0.001,0.005,0.01}\{0.0001, 0.001, 0.005, 0.01\}
Batch size{32,64,128}\{32, 64, 128\}
Hidden dimension{10,12,…,30}\{10, 12, \ldots, 30\}
Dropout rate0.10.1
λtem\lambda_{\mathrm{tem}}{0.1,0.5,1,5,10}\{0.1, 0.5, 1, 5, 10\}
Sinkhorn ε\varepsilon{0.02,0.05,0.08,0.1}\{0.02, 0.05, 0.08, 0.1\}
ModuleArchitectureInputOutput
Input feature mappingTwo-layer feed-forward network with SELUXt,Yt−1,At−1,VX_t, Y_{t-1}, A_{t-1}, VCtC_t
History encoderSingle-layer GRUht−1,Cth_{t-1}, C_thth_t
Autoregressive decoderSingle-layer GRUhk−1,Y^k−1,Ak−1,Vh_{k-1}, \hat{Y}_{k-1}, A_{k-1}, Vhkh_k
Representation mappingLinear layer with SELUhkh_kzkz_k
Outcome predictorTwo-layer feed-forward network with SELUzkz_kY^k\hat{Y}_k
Treatment classifierTwo-layer feed-forward network with SELUzkz_kA^k\hat{A}_k

Question 2: Applicability of Theorem 4.

Answer: Thank you for the insightful comment. The fixed-model condition in Theorem 4 concerns the independence between the learned model and the trajectories used for empirical approximation, rather than requiring the model to be specified before training. In our evaluation setting, ReCT is trained on the training set and evaluated with fixed parameters on independent patient trajectories. Conditional on the trained model, the concentration inequality in Eq. (36) therefore remains applicable, since the evaluation trajectories are independent of the training data. We will clarify that the empirical distributions in Theorem 4 are constructed from independent evaluation trajectories. If the same trajectories were used for both model training and empirical approximation, additional uniform convergence or stability analysis would be required.

[a] Causal Inference by Using Invariant Prediction: Identification and Confidence Intervals, Journal of the Royal Statistical Society Series B: Statistical Methodology 2016.

[b] Invariant Causal Prediction for Sequential Data, Journal of the American Statistical Association 2019.

Dear Reviewer UFL8,​

Thank you for your careful review and insightful suggestions regarding our paper. We have addressed the concerns raised in your review. If any questions remain, please feel free to let us know, and we would be happy to discuss them.​

If our explanations have resolved your concerns, we would greatly appreciate your support during the Reviewer–AC discussion period.​

Sincerely,

The authors

Reviewer_Vs9GOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies longitudinal counterfactual prediction and argues that existing methods focus mainly on balancing treatment groups while overlooking how learned representations shift over time. The authors propose ReCT, which combines adversarial treatment balancing with Wasserstein-based alignment of representations at consecutive time steps. They also derive an error bound to motivate reducing both treatment imbalance and temporal distribution drift.

The method is evaluated on synthetic tumor-growth data and semi-synthetic and real-world MIMIC-III datasets, with comparisons against CRN, Causal Transformer, G-Net, and other baselines. ReCT generally performs better at longer prediction horizons. The authors also provide ablations and experiments across different model backbones.

Critical strengths

The paper identifies a reasonable issue with longitudinal counterfactual models. Even after balancing treatment groups, the distribution of learned representations can still change over time, making long-horizon predictions difficult. I think this is worth investigating, especially because prediction errors can accumulate during autoregressive rollout.

The experimental results are encouraging. ReCT shows consistent improvements at longer horizons on both the tumor-growth and semi-synthetic MIMIC-III benchmarks. The ablations are also useful, since removing temporal alignment leads to worse performance, and the authors show that the regularization can benefit multiple backbone architectures.

Critical weaknesses

I have some concerns about the assumption that representation distributions should remain stable across time. In longitudinal causal prediction, changes in patient states or user behavior are often driven by the treatments themselves. These changes can be meaningful rather than unwanted distribution drift. Aligning consecutive time steps may remove information needed to predict treatment effects. The paper does not sufficiently discuss this trade-off or evaluate settings where the underlying state distribution genuinely changes over time.

The methodological contribution also feels somewhat incremental. Wasserstein-based distribution alignment and adversarial treatment balancing are both established techniques. Combining them is reasonable, but the paper does not clearly demonstrate why this particular formulation is better than other temporal regularization approaches. The theoretical bounds are useful for motivation, although they rely on Lipschitz assumptions and do not by themselves establish that minimizing representation drift improves causal estimation.

One issue with the theoretical argument deserves clarification: Theorem 2 uses the loss function at time k under both consecutive representation distributions. This is not necessarily the same as bounding future prediction error by the actual prediction error at time k-1, since the outcome relationship may also change over time.

I would also like stronger experimental evidence. The paper evaluates standard benchmarks, but the main results do not establish whether temporal alignment remains beneficial under genuine distribution shifts, rather than shifts that regularization can conveniently suppress. The real-world MIMIC-III experiment only measures factual prediction accuracy, so it cannot validate the counterfactual claims directly.

Questions for authors
  1. What happens when the temporal distribution shift is caused by a real change in the underlying outcome process or treatment response? Could aligning representations across time hurt counterfactual prediction in these cases?

  2. Can the authors clarify the interpretation of Theorem 2? The bound evaluates the time-k loss under the previous representation distribution, which seems different from the actual prediction error at time k-1. What additional assumptions are needed to connect the two?

  3. How does ReCT compare with simpler alternatives, such as temporal consistency regularization or MMD-based alignment? Since Wasserstein alignment is a central component, this comparison would help establish whether the chosen method is necessary.

5 comments

Weakness 1 and Question 1: Discussion of maintaining stable representation distributions across time.

Answer: Thank you for the insightful comments. Our method (ReCT) aligns population-level representation distributions across consecutive time steps, rather than individual representations. It therefore does not directly constrain treatment-induced changes in individual patient states.

The motivation is to maintain similar representation distributions for the shared prediction head across time, while allowing individual patient states to evolve with treatment histories. Prior works, such as AdaRNN [a] considers temporal distribution shift as a challenge to generalization in time-series prediction, WDGRL [b] studies learning transferable representations across domains with different distributions. These perspectives support population-level distribution consistency for predictive generalization. Our theoretical analysis further connects temporal representation discrepancies to counterfactual prediction-error bounds, motivating the joint control of distribution drift and prediction error.

Accordingly, ReCT jointly optimizes temporal alignment and outcome prediction: the alignment term regularizes population-level distribution drift, while outcome supervision encourages the preservation of genuine treatment-induced changes useful for prediction. This joint objective balances temporal representation consistency with retaining information necessary for counterfactual prediction.

Weakness 2 and Question 3: Theoretical motivation and selection for Wasserstein-based alignment.

Answer: Thank you for the insightful question. Under the Lipschitz condition, Kantorovich--Rubinstein duality allows the Wasserstein distance to bound changes in expected prediction loss across consecutive representation distributions. Our theoretical analysis further connects these temporal discrepancies to longitudinal counterfactual prediction errors. Since the bounds also involve factual prediction error and treatment-group imbalance, ReCT jointly optimizes outcome prediction, treatment balancing, and temporal alignment to improve causal estimation. Thus, the theory motivates this joint objective, whose effectiveness is evaluated empirically.

Wasserstein distance provides a useful connection between distribution discrepancy and generalization error, as established in optimal-transport-based domain adaptation theory [c]. In addition, we conducted experiments replacing Wasserstein-based temporal alignment with MMD-based alignment while keeping all other model components unchanged. The results are given as follows. Wasserstein alignment consistently achieves lower prediction errors across all prediction horizons on both the synthetic and semi-synthetic MIMIC-III datasets, providing empirical support for our choice of Wasserstein-based temporal alignment.

Synthetic Data (NRMSE)

Modelτ = 1τ = 2τ = 3τ = 4τ = 5
ReCT (MMD)1.09 ± 0.071.23 ± 0.161.35 ± 0.171.45 ± 0.181.55 ± 0.18
ReCT (Wasserstein)0.97 ± 0.111.02 ± 0.111.08 ± 0.111.15 ± 0.121.21 ± 0.12
Modelτ = 6τ = 7τ = 8τ = 9τ = 10
ReCT (MMD)1.64 ± 0.181.71 ± 0.161.78 ± 0.141.82 ± 0.121.85 ± 0.11
ReCT (Wasserstein)1.28 ± 0.121.34 ± 0.121.40 ± 0.121.45 ± 0.121.48 ± 0.12

Semi-synthetic MIMIC-III Data (RMSE)

Modelτ = 1τ = 2τ = 3τ = 4τ = 5
ReCT (MMD)0.38 ± 0.060.44 ± 0.040.50 ± 0.040.54 ± 0.050.58 ± 0.07
ReCT (Wasserstein)0.36 ± 0.040.42 ± 0.030.48 ± 0.030.52 ± 0.040.55 ± 0.05
Modelτ = 6τ = 7τ = 8τ = 9τ = 10
ReCT (MMD)0.61 ± 0.080.63 ± 0.090.66 ± 0.100.68 ± 0.100.70 ± 0.11
ReCT (Wasserstein)0.58 ± 0.060.60 ± 0.070.62 ± 0.080.64 ± 0.090.66 ± 0.10

Weakness 3 and Question 2: Explanation of temporal consistency.

Answer: We thank the reviewer for identifying the missing temporal consistency condition. We assume that ZkZ_k is sufficient for conditional mean potential outcome prediction and that the conditional outcome mechanism given ZkZ_k and treatment aa is time-invariant. This assumption is consistent with the invariant conditional outcome mechanisms considered in prior work: [d] describes invariance of the conditional outcome distribution given "the complete set of corresponding direct causal predictors.'' while [e] considers settings where "the structural dependence (predictor set and regression parameters) of YtY_t on XtX_t remains fixed.'' Under these temporal consistency assumptions, the true response function is shared across time: $
\mu_k^a(z)=\mu_{k-1}^a(z)=\mu^a(z) $. Together with the shared prediction head gθ(z,a)g_\theta(z,a), this implies $
\ell_k^a(z) =\ell\left(g_\theta(z,a),\mu^a(z)\right) =\ell_{k-1}^a(z) $.

Thus, the time-invariant response function μa\mu^a and shared prediction head imply ℓka=ℓk−1a\ell_k^a=\ell_{k-1}^a, justifying the replacement in Theorem 2. The bound then follows from the Lipschitz assumption. We will explicitly add this condition to the revised theorem and proof.

Weakness 4: Ablation experiment and real-world MIMIC-III experiment

Answer: Thank you for the insightful comment. Our theoretical analysis supports maintaining stable population-level representation distributions across time to improve predictive generalization, while allowing individual patient states to evolve with treatment histories. To evaluate temporal alignment in the presence of temporal drift, we compared the full ReCT model with its variant without temporal alignment on the synthetic tumor-growth and semi-synthetic MIMIC-III datasets. The results show that temporal alignment reduces long-term counterfactual prediction errors in these settings with evolving patient states and treatment histories.

For real-world MIMIC-III, only outcomes under the treatments actually received are observed, preventing direct evaluation of counterfactual estimation. Following established longitudinal causal inference studies, such as CRN [f] and CT [g],we evaluate factual prediction performance to assess ReCT's applicability to real clinical trajectories. This complements our synthetic and semi-synthetic experiments, where ground-truth counterfactual outcomes are available for direct evaluation.

[a] Adarnn: Adaptive learning and forecasting of time series, CIKM 2021.

[b] Wasserstein Distance Guided Representation Learning for Domain Adaptation, AAAI 2018.

[c] Theoretical analysis of domain adaptation with optimal transport, ECML PKDD 2017.

[d] Causal Inference by Using Invariant Prediction: Identification and Confidence Intervals, Journal of the Royal Statistical Society Series B: Statistical Methodology 2016.

[e] Invariant Causal Prediction for Sequential Data, Journal of the American Statistical Association 2019.

[f] Estimating counterfactual treatment outcomes over time through adversarially balanced representations, ICLR 2020.

[g] Causal transformer for estimating counterfactual outcomes, ICML 2022.

Thank you for the detailed rebuttal and additional experiments. The Wasserstein vs. MMD comparisons and clarification of Theorem 2 address several of my concerns. My main remaining concern is that population-level temporal alignment may still suppress meaningful treatment-induced distribution shifts, which is not fully addressed by the rebuttal.

Overall, the rebuttal strengthens the paper, and I am inclined to raise my rating from Clear Reject to Borderline Reject.

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

The paper studies temporal representation drift in longitudinal counterfactual regression and proposes ReCT. The method combines adversarial treatment balancing with Wasserstein alignment between representation distributions at consecutive time steps. The paper provides error bounds connecting future counterfactual prediction error to factual error, treatment-group discrepancy, and temporal distribution discrepancy. Experiments use synthetic tumor-growth data, semi-synthetic MIMIC-III, and real-world MIMIC-III, with comparisons against MSM, RMSN, G-Net, CRN, CT, and Causal CPC.

Critical strengths

The method is motivated by an explicit theoretical connection between temporal distribution discrepancy and future prediction error. The experimental evaluation includes multiple established longitudinal causal baselines, different confounding strengths and prediction horizons, and ablations of both temporal alignment and treatment balancing.

Critical weaknesses

Aligning consecutive representation distributions can also suppress genuine outcome-relevant temporal changes. The paper does not directly quantify how much temporal discrepancy is removed or whether useful temporal information is preserved. On real-world MIMIC-III, only factual prediction error can be evaluated; therefore, this experiment does not directly validate counterfactual estimation.

Questions for authors

Refer to the weakness. How does ReCT distinguish undesirable distribution drift from genuine temporal evolution that is useful for outcome prediction? Do you directly measure the Wasserstein discrepancy before and after applying the temporal-alignment loss?

2 comments

Weakness 1 and Question 1: Discussion of temporal alignment and genuine temporal evolution.

Answer: Thank you for your helpful comment. Our method (ReCT) aligns population-level representation distributions across consecutive time steps, rather than requiring each individual’s representation to remain unchanged. Individual representations can still evolve with historical states and treatment sequences.

Prior works, such as AdaRNN [a] considers temporal distribution shift as a challenge to generalization in time-series prediction, WDGRL [b] studies learning transferable representations across domains with different distributions. These perspectives motivate reducing distribution discrepancies to improve predictive generalization. Our theoretical analysis further connects temporal representation discrepancies to counterfactual prediction error. Accordingly, ReCT jointly optimizes temporal alignment and outcome prediction: alignment regularizes population-level distribution drift, while outcome supervision encourages the preservation of genuine individual-level temporal evolution useful for prediction.

Weakness 2 and Question 2: Ablation experiment to measure temporal Wasserstein discrepancy.

Answer: Thank you for your insightful suggestion. We conducted an ablation study comparing the full ReCT model with ReCT without temporal alignment to directly measure temporal Wasserstein discrepancy and assess prediction accuracy. The results in the following table. ReCT achieves smaller Wasserstein distances between consecutive representation distributions and lower prediction errors, suggesting that temporal alignment reduces distribution drift while preserving useful information for outcome prediction.

Synthetic Data (Wasserstein distance)

Modelτ = 1τ = 2τ = 3τ = 4τ = 5
ReCT w/o Ltem\mathcal{L}_{\mathrm{tem}}–0.2522 ± 0.03180.1255 ± 0.03800.0911 ± 0.02910.0690 ± 0.0226
ReCT–0.1399 ± 0.03710.0762 ± 0.02450.0532 ± 0.01640.0404 ± 0.0137
Modelτ = 6τ = 7τ = 8τ = 9τ = 10
ReCT w/o Ltem\mathcal{L}_{\mathrm{tem}}0.0566 ± 0.01850.0470 ± 0.01570.0401 ± 0.01400.0360 ± 0.01310.0320 ± 0.0119
ReCT0.0343 ± 0.01370.0298 ± 0.01420.0263 ± 0.01340.0237 ± 0.01220.0219 ± 0.0113

Semi-synthetic MIMIC-III Data (Wasserstein distance)

Modelτ = 1τ = 2τ = 3τ = 4τ = 5
ReCT w/o Ltem\mathcal{L}_{\mathrm{tem}}–0.3653 ± 0.07730.2846 ± 0.08330.2448 ± 0.09240.2242 ± 0.0879
ReCT–0.1921 ± 0.06050.1358 ± 0.03540.1137 ± 0.01660.0990 ± 0.0120
Modelτ = 6τ = 7τ = 8τ = 9τ = 10
ReCT w/o Ltem\mathcal{L}_{\mathrm{tem}}0.2114 ± 0.08500.2021 ± 0.08360.1969 ± 0.08200.1929 ± 0.08160.1905 ± 0.0788
ReCT0.0875 ± 0.00980.0800 ± 0.01410.0759 ± 0.01370.0736 ± 0.01840.0732 ± 0.0190

Weakness 3: Experiment of Real-World Factual Evaluation.

Answer: Thank you for your valuable comment. In real-world observational datasets, only outcomes under the treatments actually received are observed. Counterfactual outcomes are unavailable, preventing direct evaluation of counterfactual estimation. Following established longitudinal causal inference studies, such as CRN [c] and CT [d], we evaluate factual prediction performance on real-world MIMIC-III to assess ReCT's applicability to real clinical trajectories. This evaluation complements our synthetic and semi-synthetic experiments, where ground-truth counterfactual outcomes are available for direct evaluation.

[a] Adarnn: Adaptive learning and forecasting of time series, CIKM 2021.

[b] Wasserstein Distance Guided Representation Learning for Domain Adaptation, AAAI 2018.

[c] Estimating counterfactual treatment outcomes over time through adversarially balanced representations, ICLR 2020.

[d] Causal transformer for estimating counterfactual outcomes, ICML 2022.

Dear Reviewer ixyH,​

Thank you for your detailed review and valuable comments on our paper. We have provided responses to your concerns. If you have any remaining questions, please let us know, and we would be happy to discuss them further.​

If you feel our responses have addressed your concerns, we would be grateful for your support during the Reviewer–AC discussion period.​

Sincerely,

The authors

Comments on the paper

2 comments

Dear AC,

We highly appreciate the insightful reviews, which help us improve the quality of the manuscript. We summarize the reviews and responses as follows.

1. We sincerely thank the reviewers for highlighting the following strengths of our manuscript.

Reviewer ixyH: "The paper provides error bounds connecting future counterfactual prediction error to factual error, treatment-group discrepancy, and temporal distribution discrepancy." "The experimental evaluation includes multiple established longitudinal causal baselines, different confounding strengths and prediction horizons, and ablations of both temporal alignment and treatment balancing."

Reviewer Vs9G: "The paper studies longitudinal counterfactual prediction and argues that existing methods focus mainly on balancing treatment groups while overlooking how learned representations shift over time." "They also derive an error bound to motivate reducing both treatment imbalance and temporal distribution drift." "The paper identifies a reasonable issue with longitudinal counterfactual models. Even after balancing treatment groups, the distribution of learned representations can still change over time, making long-horizon predictions difficult. I think this is worth investigating, especially because prediction errors can accumulate during autoregressive rollout." "The experimental results are encouraging. ReCT shows consistent improvements at longer horizons on both the tumor-growth and semi-synthetic MIMIC-III benchmarks. The ablations are also useful, since removing temporal alignment leads to worse performance, and the authors show that the regularization can benefit multiple backbone architectures."

Reviewer UFL8: "The paper presents error bounds intended to relate future counterfactual prediction error to factual prediction error and distribution discrepancies across treatment groups and time steps." "The temporal alignment objective can be incorporated into existing longitudinal prediction models. The experiments with CRN and CT in Tables 5 and 6 show improvements at longer prediction horizons after adding the temporal alignment loss." "The synthetic and semi-synthetic experiments provide evidence of improved counterfactual prediction at longer horizons."

2. We have provided detailed responses in the rebuttal and uploaded a revised manuscript. We believe that the comments are addressed in our responses.

Reviewer ixyH: (1) For the temporal alignment and genuine temporal evolution (Weakness 1 and Question 1), in the rebuttal, we have explained the relationship between temporal alignment of representation distributions and the genuine temporal evolution.

(2) For measuring temporal Wasserstein discrepancy (Weakness 2 and Question 2), in the rebuttal and in Appendices D.1.3 and D.2.2 of the revised version (Tables 7 and 8), we have conducted an ablation study comparing ReCT with and without Ltem\mathcal{L}_{\mathrm{tem}} in terms of the Wasserstein distance between consecutive representation distributions.

(3) For the real-world MIMIC-III evaluation (Weakness 3), in the rebuttal, we have explained the reason for evaluating factual prediction performance on the real-world dataset.

Reviewer Vs9G: (1) For maintaining stable representation distributions across time (Weakness 1 and Question 1), in the rebuttal, we have explained the trade-off between temporal distribution alignment and preserving outcome-relevant changes through outcome supervision.

(2) For the selection of Wasserstein distance (Weakness 2 and Question 3), in the rebuttal, we have explained the connection between Wasserstein distance and the prediction error bound. In Section 5.5 of the revised version (Table 5), we have reported a comparison of Wasserstein and MMD alignment.

(3) For the temporal consistency of Theorem 2 (Weakness 3 and Question 2), in the rebuttal and in Appendix B.2 of the revised version, we have explained the temporal consistency condition that justifies using the same pointwise loss at consecutive time steps.

(4) For the ablation study and real-world MIMIC-III experiment (Weakness 4), in the rebuttal, we have discussed the ablation results on the synthetic and semi-synthetic benchmarks and explained the reason for evaluating factual prediction performance on real-world MIMIC-III.

Reviewer UFL8: (1) For temporal consistency and treatment probability shifts (Weakness 1 and Question 1), in the rebuttal and in Appendices B.2, B.3 and B.4 of the revised version, we have explained the condition under which the pointwise loss is shared across time steps and revised the bounds in Theorems 3 and 4 in revision to account for changes in marginal treatment probabilities.

(2) For hyperparameter selection and implementation details (Weakness 2 and Question 3), in the rebuttal and in Appendices E, F and G of the revised version, we have described the Sinkhorn implementation and training procedure, specified the model architecture, and reported the hyperparameter search spaces and selected values of λadv\lambda_{\mathrm{adv}}, λtem\lambda_{\mathrm{tem}}, and ε\varepsilon.

(3) For the applicability of Theorem 4 (Question 2), in the rebuttal, we have explained how the fixed-model assumption in Appendix B.4 applies when the trained model is evaluated on trajectories independent of the training data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.