acceptodds
Withdrawn from ICLR 2027

LongiBand: Covariance-Robust Confidence and Conformal Prediction Bands for Longitudinal Treatment Effects

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Repeatedly randomized longitudinal studies pose two distinct uniform-inference tasks: estimating a population treatment-effect surface and predicting a new subject’s latent conditional-mean treatment-effect curve. We introduce LongiBand, a framework that explicitly separates the targets, sampling units, and inferential guarantees associated with these tasks. For population inference, LongiBand combines treatment-specific series regressions fitted under working independence with a subject-block CR3 sandwich estimator and subject-level Gaussian-multiplier calibration. Under explicit approximation, leverage, moment, and grid conditions, the resulting simultaneous confidence band is asymptotically valid even when the within-subject covariance is misspecified. For personalized prediction, pooled functional principal component analysis estimates a joint two-arm eigensystem and uses the new subject’s observed history to predict the latent scores. Because the calibration curves are unobserved, subject-level recovery sets yield feasible conformal scores, while a truncation adjustment accounts for omitted components and extends protection to the full latent curve. Under explicit recovery-event conditions, the resulting band achieves marginal simultaneous coverage for an exchangeable future subject. For covariate-dependent FPCA, we further establish an exact-label finite-grid guarantee and show that recovered proxy labels require an additional recovery-gap condition. Simulations show that longer histories reduce prediction error and model-based band width. They also reveal substantial undercoverage when omitted components are ignored. In one representative setting, coverage increases from 0.776 under naive truncation to 0.940 after truncation adjustment. End-to-end recovery-set conformal bands achieve conservative coverage under Gaussian innovations and either Gaussian or heavy-tailed random effects, although they are substantially wider than infeasible latent-label bands, quantifying the cost of recovering unobserved calibration curves. An analysis of HeartSteps illustrates the population-level procedure.

Primary area:
causal reasoning
Keywords:
Conformal prediction, Causal Inference, Functional Data Analysis, Simultaneous confidence bands.

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_5dwfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper develops two complementary methods for uniform inference in longitudinal randomized studies: a covariance-robust confidence band for the population treatment-effect surface and a conformal prediction band for a new subject’s latent treatment-effect curve. The population method combines working-independence regression, subject-block CR3 covariance estimation, and Gaussian multiplier calibration, while the personalized method combines FPCA, history-based score prediction, and recovery-set conformal calibration. Simulations and a HeartSteps analysis suggest that the population band approaches nominal coverage as sample size increases, whereas the protected personalized bands achieve conservative coverage but can be substantially wider.

Critical strengths

None

Critical weaknesses

Apparently, one of the paper’s many weaknesses is its logical flow.

  1. Setting aside technical correctness and novelty, I have substantial concerns about the manuscript's readability and practical framing, even as a researcher with expertise in healthcare and clinical trials. The introduction opens with "Repeatedly randomized longitudinal studies contain information at two levels," without first explaining the study design or providing a concrete motivating example. The term "repeatedly randomized longitudinal studies" should be defined and its relationship to the micro-randomized trial setting used in the application clarified. Conference page limits do not justify omitting the context readers need to understand the problem; additional background could also be provided in the supplementary material.

  2. Given the limited introductory context, I turned to Section 4.4, "HeartSteps Population Analysis," to understand the practical motivation. However, this section again begins with the number of participants and decision points, followed immediately by numerical results, rather than clearly stating the substantive research question. Before presenting these results, the authors should explain what they are trying to learn from HeartSteps and how the findings should be interpreted in the application's context.

  3. The supplementary HeartSteps analysis provides additional design and implementation details, along with some visualization, but it still does not adequately connect these elements to a clearly articulated practical problem. A reader-oriented explanation would help establish connection.

  4. Overall, my judgement is that, the manuscript has a substantial problem with narrative flow: it prioritizes technical constructions and numerical results before establishing a compelling practical motivation. Readers need to understand what problem is being solved, why it matters, and what existing methods fail to provide before they can meaningfully assess the proposed contribution.

  5. Technical rigor remains essential, but extensive derivations and proofs, whether developed with or without AI assistance, cannot substitute for a clearly motivated research question and logical writing. In its current form, the manuscript has not adequately communicated the practical significance of its contribution, and this is a substantive concern rather than a minor stylistic issue.

Questions for authors

My current questions concern the scientific motivation, data selection, and practical interpretation of the application. Clarifying these points would help me assess the contribution and formulate any further technical questions.

  1. What specific scientific question does the HeartSteps reanalysis answer, and how does it extend the question addressed by the original study?

  2. Starting from the public HeartSteps V1 data, please provide or identify a participant- and decision-level exclusion flow leading to the 37 participants and 5,777 decision points. In particular, what is the scientific justification for requiring observed outcomes and consistent randomization and delivery indicators? Do any exclusions depend on information obtained after treatment assignment, and how might this selection affect the population represented by the analysis and the interpretation of the fitted contrasts?

  3. Were the scientific framing and analysis choices evaluated with researchers experienced in micro-randomized trials, behavioral interventions, or clinical-trial statistics? Without identifying individuals, please describe any substantive input and how it informed the pooling of active and anti-sedentary suggestions, the evaluation grid, and the interpretation of the results. If no such consultation occurred, what study-specific evidence supports these choices? Or, is this purely academic research purpose?

  4. It seems that the HeartSteps application demonstrates only the population layer. What concrete real-study setting motivates the personalized prediction layer, and what evidence supports its feasibility there? Please explain how calibration subjects would supply the required disjoint history and recovery observations, and how investigators would assess whether the resulting bands are sufficiently informative for the intended scientific question. Please distinguish demonstrated feasibility from a proposed future application.

Reviewer_cE7FOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies simultaneous confidence and prediction bands for treatment effects in longitudinal randomized studies. For population-level inference, it combines some methods like series regression, CR3 covariance estimation, and Gaussian multiplier calibration. For individual-level prediction, it uses FPCA and conformal calibration with recovery sets for unobserved treatment effect curves. For theory, the paper provides coverage guarantees, and for experiment, it evaluates the methods through simulations and the HeartSteps dataset. Results show reasonable coverage for population bands, while individual conformal bands tend to be conservative and relatively wide.

Critical strengths
  1. The paper addresses uncertainty quantification for both population-level and individual-level treatment effects in longitudinal studies, which is a meaningful and practically well motivated problem.

  2. The paper provides coverage guarantees for the proposed confidence and prediction bands, including within-subject dependence and the difficulty of observing individual treatment effect curves.

  3. Experiment has multiple simulation settings and a real-data analysis using HeartSteps. The paper examines coverage, band width, and robustness under different settings, providing useful insights into the strengths and limitations of the proposed methods.

Critical weaknesses
  1. Unclear main contribution and overly complicated presentation: I found a large number of definitions, equations, intermediate results, and methodological branches, making it difficult to identify its central contribution. In particular, the population confidence bands, personal conformal prediction bands, Gaussian model-based bands, and covariate-dependent FPCA extension address different problems and rely on different assumptions. The presentation feels crowded, and the relationship between these components is not sufficiently clear.

  2. Some key assumptions carry much of the theoretical difficulty. For example, Theorem 2 assumes that the recovery sets for the unobserved individual treatment effect curves satisfy a prescribed coverage condition. However, constructing such valid recovery sets is itself a central challenge of the problem. Although sufficient conditions are discussed in the appendix, the main coverage guarantee depends heavily on this assumption, which makes the theoretical contribution less convincing. Moreover, Assumptions 3 and 4 directly impose convergence rates for the empirical design matrix and the CR3 covariance estimator.

  3. In Section 4.3, the protected conformal bands achieve coverage close to 1 despite a nominal target of 0.95. For example, with 128 recovery visits, the reported band width is 6.932 under Gaussian effects, compared with 3.337 for the infeasible latent-label benchmark. This substantial increase in width raises concerns about the practical informativeness of the resulting prediction bands.

  4. The paper reports that the rank-four covariate-dependent FPCA method does not outperform a structure-matched rank-eight negative control in band width. But it seems to me that the simulation data-generating process does not contain genuinely covariate-varying covariance?

  5. The HeartSteps application focuses on population-level inference using only 37 participants. The resulting confidence bands all contain zero, and the analysis is described as providing descriptive projections rather than established causal effects.

  6. Related to Weakness 1, the novelty of some theoretical components is not sufficiently distinguished from existing techniques. Several building blocks, including CR3 covariance estimation, Gaussian multiplier calibration, FPCA prediction, and conformal ranking, are established methods. While their combination may be useful, the paper does not always clearly explain which theoretical difficulties require fundamentally new arguments and which results follow from adapting existing tools. I think the paper would benefit a lot from a well presentation of how to position the paper.

  7. Too many auxiliary results obscure the main theoretical message. The paper includes several propositions for model properties, oracle guarantees, recovery-set constructions, and technical conditions. Some of these results appear to establish relatively standard properties or apply only to idealized settings and some are more like supporting results. The current presentation makes it difficult to distinguish the paper's main results from supporting technical details.

Questions for authors

Please refer to Weakness 2, 3 and 4.

In addition, I am confused that there are several different coverage guarantees, including asymptotic simultaneous coverage for population effects, marginal conformal coverage for individual latent curves, Gaussian model-based guarantees, and finite-grid oracle coverage for the covariate-dependent branch. Since these guarantees are not directly comparable, could the authors clarify how they relate to each other and which theoretical guarantees apply to the actual implemented methods?

Reviewer_P1wBOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper tackles the setting of uncertainty quantification for randomized longitudinal studies. For this the authors consider two estimate two targets: (i) A confidence band for the CATE function and (ii) a prediction band for the individual treatment effect of a new subject. For this the authors combine different existing statistical procedures and ideas from conformal prediction. They evaluate their methods for both targets on simulation studies and a small medical dataset and semi-synthetic extensions.

Critical strengths

Uncertainty quantification for randomized longitudinal studies seems like an interesting direction that has not received much attention in the ML for causal inference literature yet.

Critical weaknesses

The main weakness of this paper is its presentation which makes it extremely hard to follow the paper and judge the novelty and soundness of the paper. To be considered for acceptance at a proper venue, the paper should be restructured and rewritten from scratch to ensure readability, logical flow, consistent use of terms throughout the paper and proper definition before usage, proper comparison to existing work, etc. A short, non-exhaustive list of problems below:

  • In the abstract, the second and third sentence directly jump into a detailed description of the method without properly having described the problem setting or any intuition or relevance before. Also directly using unexplained usually not well known terms like “subject-block CR3 covariance estimator” etc without any intuition, abbreviations like “SCB analysis” are not explained. In the end of the abstract, results are reported without any explanation of the experimental setup or dataset or further intuition. Then, further unexplained confusing terms like “ separate rank-four covariate-dependent working predictor” are used without any intuiton or explanation. This ambiguous language and style is maintained throughout the whole paper, making it extremely hard to understand and follow.
  • The paper does not contain any proper comparison to existing methods/backgrounds or related works. E.g., no conformal prediction methods from the causal inference literature are discussed at all, where there are plenty. This makes it hard to judge the novelty and contribution of this paper, or to build on top of it for future research.
  • In sec. 2, also here the definitions are not clear, e.g. the assumptions for the causal interpretation are not formally described or even referenced just loosely mentioned, the connection between the time dependent Eq. (1) to the population target which is non-time dependent is already not clearly described. Also here, constructs and concepts are randomly used that were not introduced before. E.g., sec. 2.2 refers to “HeartSteps sensitivity analysis” with “M0 is the complete contrast dictionary, M1-M3[...]”, where neither HeartStpes nor M0 or anything related has been introduced before. At this point, I need to say I cannot judge the soundness of Sec. 2 and Sec. 3 at all, simply because I cannot follow argumentation style even after several tries.
  • In sec 4, the experiments follow again the same pattern, reporting results for experiments that are unclear from the main text (and might be extracted from searching through the appendix), but no motivation or intuition is clear here. Also for the visual presentation, all results are just embedded in the main text, with no plots or tables in the main paper, making it extremely hard to evaluate the method.

Besides the presentation, of what I could extract from the paper, the contribution also seems to be more like combining existing components, the experiments are restricted to simple simulation studies and partially semi-synthetic experiements and the real world HeartSteps data analysis that seems to show no clear findings, probably due to its limited sample size.

Questions for authors

At this point I have no questions that would change my assessment. I think the presentation of the paper needs to be entirely fixed from scratch before providing a feedback loop which can be really detailed with respect to the technical content.

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

Related papers

Discussion (0)

Sign in to comment.