acceptodds
Under review as a conference paper at ICLR 2027

Belief-Augmented Joint Models for Causal Inference of Time-Varying Treatment Effects from Sparse Observations

Abstract

Identifying and estimating time-varying treatment effects from observational healthcare data is challenging when clinicians act on beliefs about patient trajectories that are only partially observed. Calibration for observed covariates alone may leave residual confounding induced by these beliefs. We propose the Belief-Augmented Joint Model (BAM), which restructs natural recovery from treatment effects by incorporating a belief-dependent model of clinical decisions. BAM represents clinicians’ beliefs through a closed-form posterior over latent trajectory coefficients, derived from sparse, noisy observations under a linear-Gaussian trajectory model. We prove that under belief-driven confounding—the latent trajectory influences the decision only through this belief—the time-varying treatment effect is identified without estimating or parameterizing the underlying decision rule. In synthetic benchmarks, BAM outperforms eight standard models; its structural decision channels transfer to a rehabilitation cohort and MIMIC-III. These results establish a framework for estimating longitudinal treatment effects that explicitly accounts for the beliefs linking incomplete patient observations to clinical decisions, empowering clinicians to take a central role in high-risk, long-cycle rehabilitation information design.

Primary area:
causal reasoning
Keywords:
causal inference, time-varying treatment effects, unmeasured confounding, sparse longitudinal data, joint models, interpretability

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_mi2kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The author proposed the BAM framework, which reconstructs natural recovery trajectories by incorporating a belief-dependent process of clinical decisions. The model has two components: the recovery trajectory for each individual's recovery rate assessment and the decision model for whether to continue treatment. The BAM model can separate the effect of treatment and provide better insight regarding whether to continue treatment.

Critical strengths
  1. The work has various applications in biomedical field. It can support evaluation of recovery trajectory and future treatment planning. This could provide a more accurate assessment of the patient's recovery stage and inform whether to keep the treatment plan.

  2. The belief is interpretable and in closed form. The posterior mean explicitly represents the clinician’s belief about the patient’s recovery. In addition, the time-varying effect is explicit.

Critical weaknesses
  1. Table 2 reports only bias and RMSE, the author should report more results on variability, such as standard error. In addition, most tables did not report coverage. Given that difference between the proposed model and some baselines (eg, LMM) is relatively small, the results are not convincing to show better performance of the proposed model.

  2. The shared-parameter joint model is common in literature, and the decision model is a logistic regression model. The author should provide justification for using the slope for belief in logistic model and show how the current framework is different from existing models linking outcome and decision.

Questions for authors
  1. Could the author provide the necessary standard errors and credible intervals for the results to give a comprehensive comparison

  2. The performance of BAM and LMM is quite similar. How does the author justify this? Under what circumstances does the author expect the BAM to outperform the LMM

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

This paper proposes BAM, a joint framework that combines a linear-Gaussian longitudinal outcome model with a belief-mediated treatment-decision model to estimate time-varying treatment effects from sparse clinical observations. It derives identification and joint inference results under history-conditional ignorability and evaluates the framework through rehabilitation data, cohort-calibrated simulations, ablation studies, and MIMIC-III analyses.

Critical strengths
  1. BAM provides a clinically meaningful connection between latent recovery trajectories, inferred clinician beliefs, and time-varying treatment decisions.

  2. The closed-form Gaussian belief update and joint likelihood offer a computationally convenient framework that accounts for parameter dependence between outcome and treatment-decision components.

  3. The paper includes real-world clinical applications, cohort-calibrated simulations, robustness analyses, ablations, and cross-cohort experiments.

Critical weaknesses
  1. Assumption A2 already imposes treatment ignorability conditional on recorded history, while the inferred belief is a deterministic function of that history, making it unclear what additional identification capability BAM provides for unobserved confounding.

  2. BAM achieves nearly identical treatment-effect estimation accuracy to the outcome-only LMM (RMSE 0.0307 vs. 0.0312), suggesting that most improvements may come from the longitudinal outcome model rather than the proposed belief–decision coupling.

  3. The additive spline model allows effects to vary by treatment day but does not explicitly accommodate treatment-effect decay, delayed responses, or interactions across treatment days.

Questions for authors
  1. How were the spline basis, knot placement, and day-42 reporting point selected, and how sensitive are the results to these choices?

  2. How are irregular assessment schedules handled in practice, and could assessment timing depend on unrecorded patient conditions or clinical decisions?

  3. Is there any independent clinical evidence, such as physician assessments or decision records, supporting the interpretation of the inferred recovery slope as a proxy for clinicians' beliefs?

Reviewer_6RWROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper suggests a new method for time-varying treatment effect estimation in the presence of hidden confounders called the Belief-Augmented Joint Model (BAM). The method assumes a parametric model for both the outcome and treatment models (i.e., a linear Gaussian trajectory model), and the distribution of the latent confounders. The BAM then yields the predictions of time-varying effects in a Bayesian manner: It assumes the priors and then infers the exact posteriors for all the parameters. The authors compared their model with the state-of-the-art time-varying treatment effect estimators based on multiple synthetic benchmarks.

Critical strengths

I didn’t identify any substantive strengths.

Critical weaknesses

The main parts of the paper (Sec. 3 onward) read as if they were generated by a low-quality LLM: The paper suffers from multiple clarity issues. Due to this, I couldn’t fully understand the main novelty and contributions of the work:
- The setting isn’t clearly explained. In the related work, the authors mention deconfounding and proxy variable approaches; yet those are not included as baselines (and it’s not clear why the methods from experiments in Table 2 are even relevant to the setting, e.g., outcome-only LMM was never mentioned in the related work). Furthermore, assuming that the setting tackles the issue of hidden confounding, Assumption 2 reads as if there were actually no unobserved confounding.

  • Misleading, not properly introduced, and inconsistent notation. For example, I got lost already in Sec. 3. What is dd in Di(d)D_i(d)? Why is there division d/Td/T in Eq. 1 (and further)? Is the distribution of the latent vector bb a causal or estimation assumption? What does “Within an assesment day we order …” mean? It is not clear whether future or past interventions are considered in Yidˉ(t)Y_i^{\bar{d}}(t)? Is a pre-specified spline basis B(⋅)B(\cdot) an assumption or a modelling choice? What is β(.)\beta( .) in Eq. 1? Long story short, the math in the paper is very hard to follow: Almost every new equation introduces new notation that does not connect to the previous notation.
  • The whole narration is odd. In multiple places, the paper starts overdescribing the details of the methods where they are not expected (e.g., lines 131-132 in the related work). The theoretical statements (e.g., Theorem 2) operate with so many terms that were never mentioned before (e.g., “independent patient clusters”, “variation-independent nuisance decision laws”, “positive-definite limiting information”, “treated–untreated contrast equals the causal blip plus a Mills-ratio selection term”). Those inconsistencies make it almost impossible to read the paper.

Beyond the issues of clarity, I worry that the novelty of the paper is very limited: it suggests a simple linear Bayesian model (and those were already introduced previously; e.g., see [1] and all the downstream literature).

References:
- [1] Schulam, Peter, and Suchi Saria. "Reliable decision support using counterfactual models." Advances in neural information processing systems 30 (2017).

Questions for authors
  • In the asymptotic data regime, does the proposed method recover the point identification?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.