acceptodds
Under review as a conference paper at ICLR 2027

Confounding-Aware Counterfactual Prediction under Global Continuous Treatments

Abstract

Counterfactual prediction is difficult when treatment is both global and continuous. In settings such as macroeconomic stress testing, all units are exposed to the same treatment path at a given time, eliminating contemporaneous treated–control variation, while latent aggregate factors may jointly affect treatment and outcomes. As a result, observational predictive relationships need not identify responses to external interventions. We develop a confounding-aware framework for counterfactual prediction in this setting. Under a sensitivity model for unobserved aggregate confounding, we derive partially identified sets for counterfactual outcomes under alternative continuous treatment paths, characterizing how causal uncertainty varies with the strength of confounding. We combine these bounds with existing conformal prediction methods to account for finite-sample uncertainty. Experiments on semi-synthetic data and a large-scale stress-testing application show that standard predictive approaches can substantially understate counterfactual uncertainty, while our framework provides informative sensitivity-aware prediction sets.

Primary area:
causal reasoning
Keywords:
causal inference, partial identification, conformal inference, uncertainty quantification, counterfactual prediction, continuous treatment, sensitivity analysis

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

This paper studies causal inference when treatment is global and continuous. In this setting, all units are exposed to a common continuous treatment path, with treatment variation occurring over time rather than across units. The authors propose structural models for outcomes and macroeconomic dynamics and develop partial identification strategies to construct intervals for their causal estimands. Financial stress testing provides the motivating application.

Critical strengths
  1. The paper considers a challenging and important setting. The authors highlight the distinction between prediction and causality, emphasizing that good historical predictive performance does not necessarily imply reliable causal conclusions.

  2. The paper separates historical prediction, future prediction, and intervention outcomes, which is conceptually meaningful. This distinction helps clarify the different quantities of interest and the assumptions needed to connect them.

  3. Applying partial identification and sensitivity analysis to this setting is a useful contribution. Rather than treating predictive forecasts as causal estimates, the proposed approach explicitly acknowledges uncertainty arising from unobserved confounding.

Critical weaknesses
  1. A major conceptual weakness is that uncertainty about confounding strength is not uncertainty about the confounding model. The paper still relies on strong assumptions about how unobserved confounders affect treatment and outcomes, while its sensitivity analysis does not examine violations of those assumptions. In particular, the latent confounder is assumed to enter the outcome and macro equations additively, with constant coefficients.

A concrete example of potential misspecification is that credit tightening may be more damaging during a downturn than under normal economic conditions. An illustrative alternative is Y=m(I,A)+(a+bA)U+xi, where A represents unemployment, U represents tighter credit conditions, and b>0. In this model, the effect of the latent financial factor becomes stronger as unemployment rises. Varying the magnitude of a constant confounding coefficient does not address this departure from the assumed model. The authors should clarify the scope of the proposed robustness analysis and explain why the maintained restrictions on the confounding mechanism are credible in the motivating application.

  1. The paper’s calibration does not sufficiently establish plausible bounds on unobserved confounding. Assumption 3 allows the analyst to choose or calibrate upper bounds on \rho_A and \rho_Y. The substantive usefulness of the resulting interval depends on whether these bounds plausibly contain the relevant confounding strengths. More justification is needed for how the bounds are selected and why they should be regarded as credible limits on unobserved confounding, rather than illustrative choices. Without such justification, it is difficult to assess how informative the sensitivity analysis is in practice or which conclusions should be considered robust.

  2. The causal estimand is not sufficiently justified. The paper usefully distinguishes prediction from causality, but two different questions are relevant to the motivating application: What losses should we expect in an adverse economic episode in which unemployment follows the specified path? What losses would occur if unemployment were externally set to this path while the latent financial-condition process retained its specified intervention law? The paper targets the second question. I personally find the first more closely aligned with the idea of stress testing, where deteriorating financial conditions may be part of the adverse scenario of interest rather than a source of confounding to remove. The authors should explain more clearly why isolating the effect of unemployment while maintaining the specified law of the latent financial process is the appropriate objective for their application.

  3. There are substantive proof mistakes that need correction, particularly in Proposition 2 and Theorem 3.

  1. Proposition 2 changes the inferential target between calibration and testing. The calibration scores use errors against realized cross-sectional average outcomes, whereas Appendix D.1 defines the test score using error against the population predictive mean. The proof then assumes exchangeability of these differently constructed scores without establishing it.
  2. Theorem 3 incorrectly states that the relevant bias vanishes for affine outcome models. The proof compares a deterministic rollout of the structural function m with the predictive mean and treats their difference as a nonlinear mean-state bias. Under confounding, that difference need not vanish even when m is affine.
Questions for authors
  1. The paper describes its framework as “model-agnostic,” but it imposes structural assumptions on the outcome, macroeconomic, and latent confounding processes. How should “model-agnostic” be interpreted here? Does it refer only to flexibility in the choice of predictive learner, rather than flexibility in the confounding model or causal assumptions?

  2. Why is it appropriate to assume that the specified treatment path can be imposed while the latent financial-condition process retains its maintained law? In the motivating stress-testing application, why should financial conditions not deteriorate jointly with unemployment? What economic intervention or decision makes the proposed causal estimand the relevant target?

Reviewer_M1YrOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies counterfactual prediction when a continuous treatment affects all units simultaneously, making it difficult to construct a conventional control group. Financial stress testing is the main motivating application.

The authors propose a framework that combines partial identification under unobserved confounding with importance-weighted conformal prediction. The method uses an additive latent-confounding model to derive bounds on the causal mean, then combines these bounds with conformal intervals to account for prediction uncertainty and distribution shift.

The experiments include linear and nonlinear semi-synthetic settings based on historical unemployment data, as well as an application to Freddie Mac mortgage data. The proposed intervals generally achieve better coverage of the causal target than standard predictive intervals, although the real-data experiment cannot directly verify causal coverage.

Critical strengths

The distinction between predictive uncertainty and uncertainty from unobserved confounding is important and often overlooked in counterfactual prediction. A model can predict observed outcomes accurately while still giving misleading estimates under interventions. The paper explains this issue well.

I also like that the framework separates these two sources of uncertainty. The closed-form sensitivity bound makes the effect of assumed confounding relatively interpretable, and the mortgage stress-testing example gives the work a concrete application.

Critical weaknesses

My biggest concern is the reliance on fairly strong structural assumptions. The theoretical bounds depend on additive latent confounding, independence assumptions, and stability of the conditional outcome relationship over time. These assumptions seem particularly difficult to justify in the financial stress-testing setting, where macroeconomic shocks can change the relationships between variables. The paper acknowledges some of these limitations, but I think their implications for the practical validity of the intervals deserve more attention.

The experiments are also somewhat limited. Both semi-synthetic settings follow data-generating processes that are closely aligned with the proposed structural model. It would be more convincing to evaluate cases where the latent confounding is nonlinear, non-additive, or otherwise misspecified. The nonlinear experiment changes the observed outcome function but still retains the additive confounding structure.

Finally, I am not sure how informative the resulting intervals are in challenging settings. In the linear simulation, the full method reaches 100% coverage against a nominal 90%, which raises the question of whether the intervals are overly conservative. The Freddie Mac experiment is interesting, but the intervals become wide enough to contain all observed outcomes, while most stress horizons are flagged as poorly supported. I would like to see a more explicit analysis of interval width and practical usefulness, rather than primarily reporting coverage.

Questions for authors
  1. How robust are the causal intervals when the additive confounding assumption is violated? For example, what happens if the unobserved confounder interacts nonlinearly with the treatment or observed states?

  2. How sensitive are the results to misspecification of the confounding sensitivity parameters? In practice, these quantities are not directly observable, and the choice of proxy variables may substantially affect the resulting bounds.

  3. Could the authors report interval width alongside coverage across different confounding strengths? In particular, how often do the intervals become too wide to support meaningful decisions, especially when the target treatment path has limited historical overlap?

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

The paper mainly studies the difference between observational and experimental means under hidden confounding. It uses additional assumptions to bound this difference, then adds an uncertainty interval. The application is reasonable, but I am not convinced what is new beyond existing causal bounds and sensitivity analysis. Also, the main target is an intervention mean in the second layer, rather than a cross world counterfactual quantity.

Critical strengths
  1. A practical application in stress testing.
  2. Shows that good prediction does not mean the intervention estimate is correct.
  3. Separates finite-sample uncertainty from uncertainty due to confounding.
Critical weaknesses
  1. The main issue is the difference between observational and experimental distributions. Already a basic problem in causal inference. why existing bounds and sensitivity methods are insufficient needs more explanation.
  2. The title says counterfactual, but the target is an intervention mean. It belongs to the second layer. The terminology needs clarification.
  3. Strong assumptions for the computable bound: additive confounding, bounded confounding strength, Gaussian dynamics, and affine dependence on lagged outcomes. Not clear how reliable the result is when these assumptions fail.
Questions for authors
  1. What does the proposed bound add beyond existing partial identification and sensitivity analysis?
  2. Why describe the target as counterfactual when it is a second-layer intervention mean?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.