- 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 with . 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 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
-
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?
-
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.
-
What values of , , 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 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 , , and the Sinkhorn regularization coefficient 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 controls the entropy regularization during transport-plan computation.
| Model | Dataset | |||
|---|---|---|---|---|
| ReCT | Synthetic | 1 | 0.1 | 0.1 |
| Semi-synthetic | 1 | 5 | 0.05 | |
| Real-world | 1 | 5 | 0.05 | |
| CRN + | Synthetic | 1 | 0.5 | 0.1 |
| Semi-synthetic | 1 | 0.5 | 0.05 | |
| CT + | Synthetic | 1 | 5.0 | 0.05 |
| Semi-synthetic | 1 | 0.1 | 0.05 |
| Hyperparameter | Search space |
|---|---|
| Learning rate | |
| Batch size | |
| Hidden dimension | |
| Dropout rate | |
| Sinkhorn |
| Module | Architecture | Input | Output |
|---|---|---|---|
| Input feature mapping | Two-layer feed-forward network with SELU | ||
| History encoder | Single-layer GRU | ||
| Autoregressive decoder | Single-layer GRU | ||
| Representation mapping | Linear layer with SELU | ||
| Outcome predictor | Two-layer feed-forward network with SELU | ||
| Treatment classifier | Two-layer feed-forward network with SELU |
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
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 Xt.'' 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), this implies ℓka(z)=ℓ(gθ(z,a),μa(z))=ℓk−1a(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) 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. With pointwise losses bounded in [0,Cℓ], 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 ea is additionally La-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.