acceptodds
Under review as a conference paper at ICLR 2027

Amortizing Causal Sensitivity Analysis via Prior Data-Fitted Networks

Abstract

Causal sensitivity analysis aims to provide bounds for causal effect estimates in the presence of unobserved confounding. However, existing methods for causal sensitivity analysis are per-instance procedures, meaning that changes to the dataset, causal query, sensitivity level, or treatment require new computation. Here, we instead present an in-context learning approach. Specifically, we propose an amortized approach to causal sensitivity analysis based on prior-data fitted networks. A key challenge is that the sensitivity bounds are not directly available when sampling training data. To address this, we develop a general prior-data construction that is applicable across the class of generalized treatment sensitivity models. Our construction involves a Lagrangian scalarization of the objective to generate training labels for the bounds through a tradeoff between causal effect min/max-imization and sensitivity model violation, which avoids model-specific analytical derivations. We further show that, under standard convexity and linearity conditions, our objective recovers the full Pareto frontier of solutions. Empirically, we demonstrate our amortized approach across various datasets, causal queries, and sensitivity levels, where our approach achieves a test-time computation that is orders of magnitude faster than per-instance methods. To the best of our knowledge, ours is the first foundation model for in-context learning for causal sensitivity analysis.

Primary area:
causal reasoning
Keywords:
Causal Inference, Causal Sensitivity Analysis, Unobserved Confounding, Prior-Data Fitted Networks, In-Context Learning, Amortized Inference, Partial Identification, Generalized Treatment Sensitivity Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JTU5OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers the problem of causal sensitivity under different levels of hidden confounding, rather than point estimation. Traditional methods for causal sensitivity require large amounts of computation for any changes in the dataset, causal query, or level of confoundedness of interest. Hence, this work builds on prior work that trains foundation models using prior data-fitted networks for causal effect estimation, and considers this problem for causal sensitivity. However, in causal sensitivity, labels are not immediate given a structural causal model or data generating process, so the paper considers a new approach to getting upper and lower bound labels given a ground-truth DGP and causal query, and who theoretical strengths of their method. Empirically, the paper shows that their approach performs well across different held-out DGPs, causal queries, and sensitivity levels. Overall, the paper claims this is the first foundation model for in-context automatic causal sensitivity analysis.

Critical strengths
  1. This is a new and extremely well-motivated problem. There has been more interest in recent work about how prior data-fitted networks could work for causal effect estimation, but no one has tackled the harder problem of understanding how to do this for causal sensitivity analysis. The paper lays out the motivation for this work well and describes an extremely sensible approach to achieving this.

  2. The label engineering choice is neat, and the choice of warm starting was smart and was proven correct both through theory and empirical results. I appreciate that the theory was able to build a realistic mechanism for quick label definitions that were used for training.

Critical weaknesses
  1. The primary weakness is the limited evaluation of their methods. Primarily, I have three issues with the evaluation:
  • The label construction mechanism is never validated. The paper should consider whether they can evaluate how good their label construction technique is to the closed form solution with MSM.
  • For MSM, there should be stronger baselines that directly estimate the nuissance functions required for the closed-form solution. One option would be TabPFN, which the paper notes as related work, which could be used to obtain quantiles for the conditional outcome distribution. In general, the baselines are quite weak with respect to what else could be done in the simpler MSM case.
  • For the KL sensitivity model, the procedure for generating labels during training mirrors that for the labels used for evaluation (though evaluation has a higher budget). Though no data is shared, it seems quite circular to assert the model is better given the labels were estimated the same way for both. I am not super familiar with the field, but are there alternative, potentially high-compute methods that are not their proposed method to create labels for the held-out eval set for this model, so that we can ensure the high performance is not simply the alignment in how labels are created?
  1. The model is only trained for a limited set of covariates (10), and only evaluated on synthetic datasets that the paper has generated. The paper should consider using semi-synthetic data like IHDP or ACIC, and introducing hidden confounding in some way so that causal sensitivity is needed. This would also stress test the method that was built for a limited context size and limited set of covariates to a somewhat more realistic set of covariates.
Questions for authors
  1. What is the potential for better performance to be a result of a shared label-construction approach during training and eval? Are there other ways to construct labels separate from how they were build for training that can be used for the eval?
  2. How well do the learned labels from the proposed method compare to the MSM closed-form solution labels?
  3. How does your approach fare against baselines that estimate nuisance parameters of the closed-form solution for the MSM model?
  4. How do the learned models scale to new, unseen, realistic sets of covariates and datasets beyond the DGPs created synthetically?
5 comments

Thank you for recognizing the motivation and label-construction contribution, and for your detailed evaluation suggestions. These motivated additional checks against independent references and comparisons with stronger baselines. Changes to PDF are marked in blue.

(RMSE pairs below report lower/upper endpoint errors in units of observed-outcome SDs.)

  • W1/Q2. Label construction versus analytical MSM bounds.

    We now validate the general Lagrangian label optimizer in MSM mode against analytical bounds on 16 fresh DGPs, comparing labels and analytical bounds at the same sensitivity level. The existing MSM PFN uses analytical training labels.

    In a matched four-DGP comparison, increased computation reduces label RMSE from 0.034/0.044 to 0.016/0.019. These results directly validate numerical label accuracy against known sharp bounds, with additional computation more than halving the remaining inward error.

    • Action: Added the experiment in Appendix D.10 (Validation of training labels: MSM analytical reference).
  • W2/Q3. Stronger MSM nuisance-estimation baselines.

    We now implement TabPFN nuisance estimation followed by analytical MSM bounds. TabPFN estimates propensity and outcome distributions through in-context inference; PI-FM predicts bounds directly. We also compare B-Learner and Dorn–Guo on their respective CATE and ATE targets.

    On matched in-prior CAPO predictions, PI-FM's RMSE is 0.296/0.296, versus TabPFN's 0.406/0.391. On ACIC-d10, TabPFN is stronger: 0.105/0.106, versus 0.218/0.223. The stronger baseline therefore confirms our model's in-prior accuracy advantage and identifies where nuisance estimation offers better transfer.

    • Action: Added Appendix D.13 (Specialized MSM baselines) and expanded the §6 Baselines paragraph.
  • W3/Q1. Independent KL validation and shared-solver error.

    We now add two references independent of the training solver. Exact Gaussian-family bounds give analytical ground truth: optimizer-label RMSE is 0.03–0.08 for Γ≤2.5\Gamma\le2.5.

    Separately, on 64 fresh Gaussian-conditional DGPs, the frozen PFN's outward interval contains the entire exact reference interval 90.2% of the time against a nominal 90% joint-containment target. Individual endpoint coverage remains below nominal 90% at 82.7%/83.6%. This supports outward reliability independently of shared-solver agreement.

    For IHDP, we compute reference bounds using an independent convex solver. Refining the numerical approximation changes checked endpoints by less than 0.001 observed-outcome SDs. These complementary checks substantially strengthen the KL evaluation by validating predictions against independent references in both analytical and real-covariate settings.

    • Action: Added Appendix D.11 (Independent KL references) and clarified reference provenance in Appendix D.2 (Evaluation details).
  • W4/Q4. Transfer to realistic covariates and higher dimensions.

    We now test frozen models on real ACIC-d10, ACIC-d82 and IHDP-d25 covariates, simulating treatment and outcomes with hidden confounding to retain computable reference bounds.

    MSM's nominal 90% endpoint coverage is 92.4%/90.4%, 70.7%/68.0% and 82.8%/81.8%, respectively. ACIC-d10 demonstrates near-nominal calibration on real covariates; the full-feature evaluations identify calibration limits. PI-FM also improves endpoint accuracy over fitted NeuralCSA on every matched MSM/KL cohort. Together, these results demonstrate useful transfer beyond synthetic training covariates.

    • Action: Added Appendix D.12 (Real-covariate evaluation) and Appendix D.17 (Practical warning signals); replaced the §6 Robustness paragraph.

Thank you for the new experiments and answering the questions I posed in the review. My answers have mainly been answered, but some of the answers bring up more concerns rather than clarification. I believe my questions around the KL validation and analytical MSM bounds have been answered.

However, the new results on the realistic covariates are quite concerning to me. First, the proposed TabPFN method performs substantially better on the realistic ACIC covariates compared to the proposed method, so NeuralCSA is not the right baseline. I agree that it seems the benefit to the proposed FM is for in-prior datasets. However, this limits the scope, novelty, and applicability of the approach significantly with respect to what the paper claims it can do. Specifically, applying this to unseen or held-out data distributions shows a large degradation compared to baselines, which limits the applicability of the model substantially. This is especially true given the fact that the DGP behind real covariates is unknown, so when applying to real data, we won't know whether the new dataset will be in-prior. Combined with the fact that it does not transfer well to datasets with more covariates, I worry that the scope of the method has been greatly diminished.

Hence, I will still keep my original score.

Dear reviewer,

Thank you for your response. We would like to respectfully point out that the TabPFN baseline is only applicable for the MSM-model case, where an analytical form of the estimand is available. This is in contrast to the general case where the estimand is an implicit solution to an optimization problem (and so tabpfn baseline is not possible; see Table 10, page 34) - in the general case our method remains the only amortized solution that successfully leverages deep learning scaling.

More broadly, we would like to argue that the value of our contribution goes beyond chasing sota on every metric. We are the first to propose a foundation model for general causal sensitivity problems, showing competitive performance overall, and where not currently sota, our model is best positioned to benefit from further naive scaling of the underlying transformer (please note we do not possess the computational budget to compete with prior labs and do not believe that is the purpose of our work in the first place). We will adjust the framing of the paper to more clearly outline our contribution.

We would like to kindly ask the reviewer to reconsider raising their score. We are happy to make any final changes necessary to assist this.

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

The paper proposes the first foundation model for causal sensitivity analysis. It fits a PFN to amortize the task across datasets, causal queries, sensitivity levels and treatments, making inference at deployment much quicker as compared to standard approaches that need to be re-trained every time the context changes. The core technical contribution of the paper is generating sensitivity bounds as labels to train the PFN via Lagrangian scalarization. The proposed approach and its components are evaluated extensively on synthetic data.

Critical strengths

Sensitivity analysis is an important problem for real world applicability of causal methods. Standard causal estimators make strong assumptions, such as no unobserved confounding, in order to provide point estimates. Relaxing these assumptions to unobserved confounding that leads to partial identification and bounded estimates instead is much more applicable for real world tasks.

Strong technical contributions. Generating sensitivity bounds as PFN training labels is an important innovation that is backed by theoretical analysis (section 5), and which enabled building the foundation model for sensitivity analysis. Empirical results support that the proposed label construction method and the overall sensitivity analysis approach work well on tested scenarios.

Critical weaknesses

Lack of non-synthetic evaluation. The experiments are extensive but limited to synthetically generated data. While I understand that real world data do not come with ground truth sensitivity bounds, having results on synthetic data only is not very reassuring evidence that the proposed method will work well in practice.

Limited baselines. The experiments have only one baseline: NeuralCSA. While I understand that this method subsumes the other three types (MSM, f, Rosenbaum), it would be still useful to see how the proposed foundation model compares to type-specific methods (e.g. [1] for MSM, [2] for f-sensitivity). The already cited B-learner [3] would also be an interesting baseline. Furthermore, the existing experiments cover MSM and KL cases, but Rosenbaum's model seems to be missing (which NeuralCSA covers).

References

[1] Dorn, J., & Guo, K. (2023). Sharp sensitivity analysis for inverse propensity weighting via quantile balancing. Journal of the American Statistical Association, 118(544), 2645-2657.

[2] Jin, Y., Ren, Z., & Zhou, Z. (2026). Sensitivity Analysis Under the f-Sensitivity Model: A Distributional Robustness Perspective. Operations Research, 74(2), 860-878.

[3] Oprescu, M., Dorn, J., Ghoummaid, M., Jesson, A., Kallus, N., & Shalit, U. (2023, July). B-learner: Quasi-oracle bounds on heterogeneous causal effects under hidden confounding. In International Conference on Machine Learning (pp. 26599-26618). PMLR.

Questions for authors
  1. Would it be possible to evaluate the proposed method on semi-synthetic or real-world data? This would considerably strengthen the paper. Perhaps a similar experimental approach to that found in the NeuralCSA paper (section 6, ii and iii) would be possible?
  2. Would it be useful to add the Rosenbaum’s sensitivity model to the experimental work as a third case on top of MSM and KL?
  3. Would it be useful to add model-specific baselines to the experimental work on top of NeuralCSA? (see second paragraph in Weaknesses section above)
3 comments

Thank you for your encouraging assessment and concrete suggestions for strengthening the evaluation. We have now added real-covariate experiments, specialized baselines and a trained Rosenbaum model. Changes to PDF are marked in blue.

  • W1/Q1. Evaluation on real covariates from ACIC and IHDP.

    We now test our models on real covariates from ACIC-2016 and IHDP. Following NeuralCSA, we simulate treatment and outcomes with hidden confounding to retain computable reference bounds. The evaluation covers ACIC's 10-feature subset and all 82 encoded features, plus all 25 IHDP covariates.

    MSM's nominal 90% predictive intervals cover the lower/upper reference endpoints in 92.4%/90.4% of predictions on ACIC-d10, 70.7%/68.0% on ACIC-d82 and 82.8%/81.8% on IHDP. Calibration therefore transfers well to ACIC-d10, while the higher-dimensional evaluations identify the limits of the current checkpoint. The framework has no fixed ten-covariate restriction; we expect larger models trained on higher-dimensional priors, beyond our current compute budget, to extend this range. PI-FM also has lower endpoint RMSE than fitted NeuralCSA in every matched MSM/KL comparison. These results support the model's usefulness beyond fully synthetic data.

    • Action: Added the ACIC/IHDP design and results in Appendix D.12 (Real-covariate evaluation).
  • W2/Q3. Comparison with specialized sensitivity-analysis methods.

    We now include all three requested baselines: Dorn–Guo quantile balancing, Jin–Ren–Zhou (JRZ) ff-sensitivity estimation and B-Learner. We also test TabPFN nuisance estimates followed by analytical MSM bounds.

    In prior, PI-FM has lower endpoint RMSE than TabPFN for CAPO, B-Learner for CATE and quantile balancing for ATE. On real covariates, TabPFN improves MSM accuracy, and linear quantile balancing improves ATE accuracy on ACIC-d10 and IHDP. Direct amortization also reduces computation: median synthetic task time is 0.13 s, versus 24.5 s for TabPFN on the same GPU after initialization. The expanded MSM comparisons therefore support direct amortization as a fast, competitive alternative to specialized estimators.

    • Action: Added Appendix D.13 (Specialized MSM baselines) and Appendix D.14 (ff-sensitivity comparison).
  • W2/Q2. Rosenbaum as a third trained sensitivity model.

    Following your suggestion, we have now trained a Rosenbaum PI-FM on 10,000 DGPs. On 16 untouched synthetic test DGPs, lower/upper endpoint RMSE is 0.305/0.295. Nominal 90% predictive intervals achieve 87–91% endpoint coverage across heads and Γ\Gamma bands. On matched IHDP predictions, lower/upper RMSE is 0.298/0.240, versus 0.893/0.875 for NeuralCSA fitted entirely from observed data. This demonstrates that the amortized approach successfully extends beyond MSM and KL to Rosenbaum sensitivity analysis.

    • Action: Added the label audit, training and evaluations in Appendix D.15 (Rosenbaum sensitivity analysis).

Many thanks for addressing all my comments. I recognize this revision took a substantial effort and I think the authors have done a great job.

I particulary welcome the experiments on ACIC and IHDP as they show reasonable generalizability outside synthetic data. I do notice the drop in performance on ACIC-d82, but my understanding is that this is due to scale of the trained model, and which could be addressed by larger-scale training. This is just to say that it is not an inherent limitation of the proposed method, but the result of deployment parameters.

Thank you also for adding more baselines, including the Rosenbaum model. I can see that certain specialized models can perform better under specific conditions, whereas the proposed approach remains consistently competitive across the board, which I think is exactly the selling point behind PFNs. With these additional baselines, the picture now should be more complete that this is indeed the case.

All my concerns have been addressed. I believe the paper is now considerably stronger and deserves a score of 4 (will update shortly).

Thank you very much for raising your score! We are glad we were able to address all your comments. Your review has been very helpful for improving our submission.

Reviewer_4DEGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an amortized approach to causal sensitivity analysis using prior-data fitted networks. A Lagrangian scalarization is used to generate sensitivity-bound training labels by tracing the Pareto frontier between causal-effect optimization and sensitivity-model violation. The resulting foundation model predicts lower and upper sensitivity bounds from a dataset, causal query, and sensitivity level in a forward pass. Experiments evaluate MSM- and KL-based sensitivity settings, compare against NeuralCSA, analyze posterior coverage and RMSE, and report substantial reductions in test-time computation. 30160_Amortizing_Causal_Sensit

Critical strengths

The paper converts repeated optimization-based sensitivity analysis into amortized inference while retaining uncertainty estimates, addressing a clear computational bottleneck. The evaluation includes calibration, comparison with an optimization-based baseline, out-of-prior robustness tests, runtime analysis, and detailed reproducibility information.

Critical weaknesses

The accuracy of the amortized estimator depends on the training prior. The paper reports degradation for context sizes and covariate dimensions substantially outside the training range. The predicted sensitivity bounds are learned approximations rather than optimization-derived bounds at deployment, so their reliability under substantial prior mismatch remains dependent on empirical calibration.

Questions for authors

Are the predicted lower and upper bounds guaranteed to be conservative under any stated conditions, or is their validity entirely empirical once the PFN is used at test time? Is there a practical diagnostic for identifying inputs that are substantially outside the training prior?

1 comment

Thank you for highlighting the computational benefits of our approach and for raising these important reliability questions. Your feedback helped us strengthen the evaluation of predictive uncertainty and practical diagnostics. Changes to PDF are marked in blue.

  • W1/Q1. Conservativeness of the predicted bounds.

    We now evaluate an outward interval using the lower endpoint's 5th and upper endpoint's 95th predictive percentiles. On 512 held-out in-prior MSM DGPs, it contains the entire analytical reference interval in 89.0% of query/arm/Γ\Gamma predictions, close to the nominal 90% joint-containment target (95% whole-DGP bootstrap interval: 88.3–89.7%). On 64 KL DGPs, containment is 89.4% against higher-budget numerical references on common support.

    We agree that a theoretical conservativeness guarantee would also be valuable. Existing statistical PFN theory [1,2,3] does not establish such a guarantee for predicted sensitivity bounds. Deriving one is beyond the scope of this work. The close agreement with nominal containment is encouraging and supports the usefulness of the model's uncertainty estimates in the evaluated in-prior regime.

    • Action: Added the outward-interval experiment in Appendix D.16 (Empirical reliability of outward intervals) and revised Appendix E (Limitations) to identify theoretical conservativeness guarantees as future work.

    [1] https://arxiv.org/abs/2305.11097

    [2] https://arxiv.org/abs/2505.11325

    [3] https://arxiv.org/abs/2603.12037

  • W1/Q2. Practical diagnostics for prior mismatch.

    For MSM, we now check whether context size, covariate dimension and Γ\Gamma fall within the assessed regime. We also test two empirical warning signals: low estimated treatment support and disagreement at Γ=1\Gamma=1 with an independent TabPFN outcome estimate. We set thresholds using separate in-prior DGPs, without OOD failure labels.

    Each screened prediction is one dataset/query/arm/Γ\Gamma combination. On held-out OOD data, the screen has 73% recall for harmful failures and a 36.8% flag rate. Joint containment is 87.2% overall and 91.8% among unflagged predictions; the in-prior false-positive rate is 19%. A harmful failure has total reference-set miss exceeding 0.1 outcome SDs. Among predictions passing the regime rule, the empirical signals detect 50% of remaining harmful failures while flagging 18.1%. A meaningful fraction of out-of-prior tasks can thus be detected with these diagnostics.

    • Action: Added Appendix D.17 (Practical warning signals), with the screening rules, threshold selection and held-out recall/false-positive results.
  • W1. Out-of-prior evaluation on ACIC and IHDP.

    We now test the pretrained models on real covariate data from ACIC and IHDP, outside the synthetic training prior and without retraining. We simulate treatment and outcomes with hidden confounding to retain computable reference bounds. MSM's nominal 90% lower/upper endpoint coverage is 92.4%/90.4% on ACIC-d10, 70.7%/68.0% on ACIC-d82 and 82.8%/81.8% on IHDP. The near-nominal coverage on ACIC-d10 demonstrates that calibration can transfer to real covariates, while the full-feature evaluations identify the current checkpoints' limits.

    PI-FM also has lower endpoint error than fitted NeuralCSA on matched MSM and KL predictions. Its accuracy advantage therefore extends to these real-covariate settings.

    • Action: Added ACIC/IHDP experiments in Appendix D.12 (Real-covariate evaluation) and summarized existing and new transfer results in the §6 Robustness paragraph.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers for their thoughtful reviews and actionable suggestions! We appreciate their recognition of the motivation and technical contributions of our work.

We have addressed all weaknesses and questions in our individual responses and made the following main improvements to the paper (changes are marked in blue in the updated PDF):

  1. Broader evaluation and scope. We added semi-synthetic evaluations using real ACIC and IHDP covariates and trained a Rosenbaum model, extending our experiments beyond MSM and KL.
  2. Stronger baselines. We added all three suggested specialized baselines—Dorn–Guo quantile balancing, Jin–Ren–Zhou and B-Learner—as well as TabPFN nuisance estimation followed by analytical MSM bounds.
  3. Independent validation and reliability. We added analytical and independent numerical reference checks, evaluated outward-interval reliability and tested practical prior-mismatch diagnostics. We also clarified the scope of conservativeness guarantees in Limitations.

These revisions substantially strengthen the empirical foundation and practical scope of our paper. We believe the revised work offers a valuable new direction for causal sensitivity analysis and is a strong fit for ICLR 2027.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.