acceptodds
Under review as a conference paper at ICLR 2027

Causal Mamba for Estimating Counterfactual Outcomes over Time

Abstract

Accurately and efficiently estimating counterfactual outcomes over time from observational data is essential in various applications. However, existing methods suffer from computational inefficiency and over-balancing, limiting their practical use. In this paper, we propose Causal Mamba (CM), a novel end-to-end framework for efficient counterfactual estimation over time. Its GCSI-Mamba backbone introduces Gated Cross-Sequence Interaction (GCSI) and follows an interact-then-propagate paradigm, first modeling dependencies among treatment, covariate, and outcome streams and then using stream specific Mamba branches to capture their respective temporal dynamics, thereby preserving stream heterogeneity while retaining linear complexity with respect to sequence length. The Multi-Task Collaboration (MTC) module further coordinates outcome prediction, treatment domain confusion, and covariate reconstruction to reduce treatment specific bias while preserving relevant information for counterfactual prediction, avoiding information loss caused by over-balancing. Experimental results on synthetic and real-world datasets demonstrate that CM not only outperforms baselines but also achieves greater runtime efficiency.

Primary area:
causal reasoning
Keywords:
Causal Inference, Longitudinal Treatment Effect Estimation, Mamba, Representation Learning, Multi-Task Learning

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 3 reviews, above 27% of the others.
Reviewer_nCRWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose Causal Mamba (CM) for estimating counterfactual outcomes over time, addressing two key challenges in longitudinal causal effect estimation: computational inefficiency in modeling temporal dependencies and information loss caused by over-balancing. Specifically, CM introduces a GCSI-Mamba backbone to efficiently capture cross-stream interactions and temporal dynamics, and the MTC module to mitigate over-balancing while preserving predictive information. Experimental results on synthetic, semi-synthetic, and real-world datasets demonstrate the effectiveness and computational efficiency of CM.

Critical strengths
  1. The proposed CM method effectively combines cross-stream interactions and temporal modeling through GCSI-Mamba, while incorporating the MTC module to balance treatment-related information and preserve predictive information.

  2. The authors conduct extensive experiments on synthetic, semi-synthetic, and real-world datasets, including comparisons with representative baselines, ablation studies, and efficiency analyses.

Critical weaknesses
  1. CM focuses on multi-step counterfactual prediction, but Proposition 2 only establishes a risk bound for one-step counterfactual estimation, without providing theoretical guarantees for multi-step predictions.

  2. Although MTC is designed to mitigate over-balancing, the experiments provide limited direct evidence that covariate reconstruction effectively preserves outcome-relevant information.

  3. MTC reconstructs all covariates without distinguishing their relevance to treatment assignment and outcome prediction. However, preserving treatment-only information may weaken representation balancing and compromise counterfactual estimation accuracy.

  4. The comparisons with recent methods are limited, and the improvements over some baselines are relatively small.

  5. In the synthetic TG dataset, the covariate stream is absent and the reconstruction weight is set to zero. Therefore, the results on this dataset cannot directly validate the contribution of covariate reconstruction to mitigating over-balancing.

Questions for authors

See above weakness.

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

The paper tackles longitudinal counterfactual estimation from observational data. Existing methods (CRN, Causal Transformer) suffer from over-balancing — forcing treatment-invariant representations can throw away prognostic information — and from the O(T^2) cost of Transformers. Their approach uses separate Mamba branches for treatment, covariate, and outcome streams, with gated cross-stream attention before each selective scan. A multi-task loss combines outcome prediction, adversarial debiasing, and covariate reconstruction to balance confounding adjustment against information preservation.

Critical strengths

Applying Mamba to longitudinal counterfactual estimation is timely. The prior SOTA is Transformer-based, and linear-time alternatives are useful for longer clinical trajectories. The GCSI design correctly respects information asymmetry during projection: future covariates are masked, only planned treatments and autoregressive outcomes are visible. Benchmark coverage is reasonable: synthetic, semi-synthetic MIMIC-III, and real-world factual forecasting, compared against five baselines including the recent IGC-Net.

Critical weaknesses

The contribution feels incremental. It's mostly a backbone swap: replace Transformer with Mamba, add a reconstruction auxiliary task. The cross-stream interaction is just gated attention restricted to same-time-step pairs. None of these is a fundamentally new idea. GCSI barely matters — the ablation shows removing it only adds +0.01 to +0.03 RMSE. The biggest ablation gain is from residual connections (+0.10 to +0.36), not from anything Mamba-specific or causal-specific. Real-world validation is limited: the real MIMIC-III experiment only does factual forecasting, you can't verify counterfactual accuracy on real data.

Questions for authors

How does CM compare to an LSTM baseline with the same MTC loss? Is the gain really from Mamba or from the multi-task objective? Can you isolate the Mamba vs MTC contribution by running CT's CDC loss on top of a Mamba backbone?

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

The paper introduces Causal Mamba (CM) for estimating counterfactual outcomes over time, assuming consistency, sequential ignorability and overlap. Its GCSI-Mamba backbone tracks treatments, covariates and outcomes in separate streams. Gated interactions exchange available information at each time step before the selective state-space updates. For future steps, the model masks unknown covariates and uses planned treatments together with its own outcome predictions. MTC combines outcome regression, treatment-domain confusion and covariate reconstruction, aiming to balance representations without discarding useful information. The theory characterizes a fixed-weight pooled population objective at a Bayes-optimal treatment classifier. It also gives a comparator-dependent joint bound and a one-step counterfactual-risk transfer result under policy-coverage and input-availability conditions. Experiments cover synthetic tumor growth, semi-synthetic MIMIC-III with counterfactual labels, and real-world MIMIC-III for factual forecasting only. Five-run comparisons, ablations, runtime/model-size measurements and a long-history experiment support the main empirical contribution: a useful accuracy-efficiency trade-off on the tested benchmarks.

Critical strengths
  1. The semi-synthetic results are my main reason for supporting acceptance. Table 1 reports one-, three- and six-step RMSE of 0.11, 0.28 and 0.51 for CM, compared with 0.16, 0.33 and 0.59 for CT, with five-run variability. CM-ERM and the same-backbone controls in Table 4 add useful evidence: removing GCSI or replacing MTC with CDC each raises six-step RMSE by 0.03. The synthetic and long-history experiments also support the overall result, though CM does not win in every setting in Table 12. The paper correctly presents the real-world MIMIC-III results as factual forecasting, not counterfactual validation.

  2. The architecture handles the difference between observed history and future prediction in a sensible way. It exchanges available stream information before the separate selective scans, rather than treating future covariates as if they were observed (Section 4.1; Appendix D). The implementation is also compact and fast in the reported comparisons: total runtime is 1.3 minutes for CM versus 6.6 for CT on the tumor benchmark (Table 3), and parameter counts are 27.1K versus 58.8K on the semi-synthetic benchmark (Table 15). Appendix M tests histories up to length 150, where CM/CT inference takes approximately 4.1/8.4 minutes. This is attractive when computation is limited. The timings compare selected pipelines; they do not isolate the architecture's contribution or establish a universal latency advantage.

Critical weaknesses
  1. My main concern is what the multi-step output estimates. Eq. (1) asks for a conditional potential-outcome mean under a treatment sequence. The rollout in Section 4 and Appendix G.4 instead feeds point predictions back into a nonlinear model, without future covariates. Even without confounding, evaluating a nonlinear function at a predicted mean generally differs from averaging over intermediate outcomes. Appendix E.1 establishes a one-step result for changing the current treatment on observed histories with a fixed input kernel; it does not establish that the recursive predictions match Eq. (1) on intervention-induced future histories. I am not disputing Proposition 2. What is missing is the link to the central multi-step task, either through additional closure/sufficiency assumptions or through a clearly stated plug-in approximation.

  2. The ablations support the combined system, but do not yet isolate its distinctive mechanisms. Table 4 includes CDC replacement and GCSI removal; the concern is that the small mean differences lack paired-run uncertainty and do not measure retained prognostic/covariate information alongside treatment balance. CM-ERM removes both regularizers, so it cannot isolate reconstruction. The tumor t-SNE experiment cannot answer that question either: it has no covariate stream and uses αx=0\alpha_x=0 (Appendix D.1; Table 10). The paper describes a CT+MTC comparison in Appendix H.4, but I could not locate its numerical results. A comparably tuned plain/concatenated Mamba baseline is also missing. Without those comparisons, I am more convinced by the full system's performance than by the attribution of its gains to GCSI/MTC rather than a lightweight backbone and ordinary auxiliary regularization.

Questions for authors
  1. Clarification of the current estimand and estimator. For τ≥2\tau \ge 2, is CM explicitly a plug-in approximation to Eq. (1), or are additional mean-closure/sufficiency assumptions intended? Please explain the role of nonlinear intermediate-outcome integration and how generated histories relate to Appendix E.1's fixed-kernel, observed-history one-step result. A defensible target interpretation or precise approximation claim would materially reduce the causal-scope concern.

  2. Generator/implementation clarification (verified against the text). Appendix I.2 labels YtY_t as after treatment and Eq. (61) includes contributions through i=ti=t, whereas Eqs. (1)/(5) treat YtY_t as history before AtA_t; I confirmed the printed equation places AtA_t contemporaneously inside YtY_t with no stated lag. Please provide the implementation-consistent window/lag expression for Eq. (61) and the generator-to-model time-index mapping. Does an explicit shift ensure that current-treatment-affected observations do not enter the pre-treatment input? A correct mapping could resolve this through clarification; an actual contemporaneous-input problem would require reassessing the strongest counterfactual benchmark and could change the recommendation.

  3. Clarification of announced results and targeted contribution validation, not open-ended future work. Please provide the CT+MTC results described in Appendix H.4 and paired-run uncertainty for the CDC/GCSI contrasts. Is there evidence measuring information retention together with treatment balance, and a comparably tuned simple Mamba control, that distinguishes the proposed mechanisms from generic auxiliary regularization and backbone replacement? Robust controlled gains would strengthen the contribution; parity with simpler controls would weaken it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.