acceptodds
Under review as a conference paper at ICLR 2027

Estimating Causal Effect Trajectories under Dynamic Network Interference

Abstract

Dynamic network interference combines unknown propagation strengths and delays with treatment–outcome feedback. We propose DyNICL (Dynamic Network Interference Causal Learning) to estimate direct, peer, total, and interaction effect trajectories. Four fixed stochastic policies define the effects. Learned hop–lag exposure and history representations support neural backward-Q regression, with Q-based and assignment-weighted augmented estimation. Under sequential exchangeability, positivity, and exposure sufficiency, we identify the policy values and establish Q-based consistency under continuation convergence. For augmented estimation, we derive a cross-fitted orthogonal expansion separating continuation–assignment product error from conditional exposure drift. DyNICL-Q achieves lower peer and total effect RMSE than adapted NetEst and TNet across all nine synthetic neighbor-count/lag settings. On two semi-synthetic contact-network settings, DyNICL-Aug achieves the lowest peer and total effect RMSE, while DyNICL-Q achieves the lowest direct effect RMSE.

Primary area:
causal reasoning
Keywords:
Causal inference, Dynamic network interference, Exposure mapping learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eye8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies causal effect estimation under dynamic network interference. The setting considers peer treatments may affect outcomes across multiple hops and time lags, and outcomes may in turn influence future treatment assignments. The authors propose DyNICL, which learns hop-lag exposure and history representations and combines them with neural backward-Q regression to estimate effect trajectories. The paper provides identification results and consistency guarantees for the Q-based estimator. It also develops a cross-fitted augmented estimator with an orthogonal error expansion under sequential exchangeability, positivity, exposure sufficiency, and related regularity assumptions. Experiments on synthetic and semi-synthetic network data show that DyNICL generally improves peer and total effect estimation compared with baselines.

Critical strengths
  1. The paper studies a important and challenging setting. The setting considered in the paper is more complex than the standard static network causal estimation, and the proposed framework attempts to handle the difficulty in a unified way.
  2. The paper provides relatively complete theoretical results under its stated assumptions. It includes identification results, consistency analysis, and an orthogonal expansion for the augmented estimator.
Critical weaknesses

I have several concerns about the paper, which are summarized below.

  1. The theoretical guarantees rely on several (strong) assumptions. The analysis assumes assignment factorization and strong positivity (Assumption B.4), exposure sufficiency (Assumption B.5), and independent network trajectories (Assumption B.7). The stronger and more restrictive forms of these assumptions are primarily stated in Appendix B rather than being clearly discussed in the main text. The paper (main text and appendix) also provides limited justification for why these assumptions should be plausible in the intended applications, and offers insufficient discussion of how violations would affect the guarantees.
  2. I find out that there is a gap between the learned exposure module and the theoretical guarantees. The theory assumes that the learned exposure converges to a sufficient reference exposure ϕ⋆\phi^\star, but the paper does not establish that the proposed neural exposure-learning method satisfies this condition. Moreover, the augmented estimator is Neyman-orthogonal only with respect to the continuation and assignment nuisances (q,p)(q,p) at fixed ϕ⋆\phi^\star; exposure-learning error enters the expansion as a first-order Op(ϵϕ)O_p(\epsilon_\phi) term. Thus, the guarantees for one of the central learned components of the method remain conditional on an unverified assumption.
  3. The paper introduces many definitions in a short space, often without enough explanation of why they are needed (e.g., the reference exposure map ϕi⋆\phi_i^\star in Eq. (1), which is introduced rather abruptly with little intuition about why this abstraction is appropriate or how it connects to the learned exposure used later). This makes the framework and methodology hard to understand.
  4. The empirical advantage is not completely uniform. In some long-memory settings, the augmented estimator performs substantially worse than the Q-based estimator and can even underperform TNet. In addition, NetEst and TNet are adapted to the proposed longitudinal setting rather than being native competitors. Finally, the evaluation does not include a genuinely real-world causal outcome setting, which I consider a relatively minor concern.
  5. The discussion of related work and competing methods is rather limited. The related-work section mostly lists prior methods with only brief descriptions, without clearly explaining how DyNICL differs from them in terms of assumptions, estimands, and methodology. This is particularly noticeable for NetEst and TNet, which are the main external baselines in the experiments but receive only very limited discussion in the main text.

Overall, I think the paper has several important issues in the setting, results and presentation. I do not think it is ready for acceptance in its current form, so I lean toward rejection.

Questions for authors
  1. Can the authors provide sufficient conditions under which the proposed learned exposure module converges to a sufficient reference exposure ϕi⋆\phi_i^\star?
  2. The root-RR result appears to require the learned exposure map to converge to the reference exposure faster than R−1/2R^{-1/2}. Can the authors explain under what conditions the proposed neural exposure-learning procedure can satisfy such a rate?
  3. The theory and cross-fitting procedure rely on independent network trajectories. How essential is this assumption? Could the analysis be extended to a single long network trajectory under temporal mixing or other dependence conditions?
  4. Since NetEst and TNet are adapted to the proposed longitudinal setting, could the authors provide more detail on what information, temporal history, and model capacity are given to each baseline? How sensitive are the empirical conclusions to these adaptation choices?
Reviewer_4rkGOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies a causal effect estimation problem and proposes a method (DyNICL) that, in experiments, achieves the lowest RMSE for the four estimated effects. However, because the paper piles up terminology, lacks definitions and illustrative examples, and lacks motivation, its readability is poor, and the reviewer cannot evaluate the technical novelty, or even correctness, based on the information currently provided in the paper.

Critical strengths

This paper studies a causal effect estimation problem and proposes a method (DyNICL) that, in experiments, achieves the lowest RMSE for the four estimated effects. However, because the paper piles up terminology, lacks definitions and illustrative examples, and lacks motivation, its readability is poor, and the reviewer cannot evaluate the technical novelty, or even correctness, based on the information currently provided in the paper.

Critical weaknesses

Poor writing:

  • The abstract gives only one sentence of background. The second sentence already introduces the method, and no connection to the first sentence is visible. The whole abstract is saturated with terminology (hop-lag, exposure, assignment, exchangeability, positivity, ...).
  • The Introduction is too short: the first paragraph states that networked systems have spatiotemporal coupling; the second paragraph piles up terminology (estimation, connections, treatments, outcome, exposure, ...); the figure introduced there does not explain the second paragraph, and the figure itself is not explained either.
  • The problem statement is unclear: §3.1 gives the composition of the observed trajectories, but the symbols used are not defined, e.g., A, X, Y. The definition of the fixed design D has the same problem. Likewise, §3.1 is full of undefined terminology: proxy skeleton, reference propagation, structural parameters, pre-generated exogenous covariates, and pre-selected focal/actionable sets.... Finally, there are symbols that are defined but never used, e.g., \tilde{S}_i in §3.2; and the symbols in Eqs. (1) and (2) are not defined (\phi^\star, p, g).
  • The method description is unclear: a fairly complex Figure 2 is presented without any explanation (the only citation appears to be the first sentence of §4), and its content is full of custom modules. Many terms inside the figure never appear in the paper (atom bank \mathcal{A}, Atom mean, Masked operator, ...).
  • The experimental description is unclear: necessary background and dataset descriptions are missing.
Questions for authors
  1. Could the authors improve the accessibility of the manuscript by providing a concrete motivating example and explicitly defining the key variables, concepts, and assumptions before using them? For reference, related works such as [1] and [2] provide examples of how network interference problems can be introduced with clearer definitions and more intuitive explanations. Ref [1] Chen, Weilin, et al. "Causal effect estimation under networked interference without networked unconfoundedness assumption." IEEE Transactions on Pattern Analysis and Machine Intelligence (2026). [2] Huang, Qiang, and Jin Tian. "Modeling Interference for Treatment Effect Estimation in Network Dynamic Environment." International Conference on Learning Representations. Vol. 2026. 2026.
Reviewer_snmuOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies causal effect estimation under dynamic network interference, where treatment effects propagate through networks with unknown strengths and time delays, while accounting for treatment–outcome feedback. The authors propose DyNICL (Dynamic Network Interference Causal Learning), a framework for estimating time-varying direct, peer, total, and interaction effects under four predefined stochastic treatment policies. The paper establishes identification results through backward recursion under sequential exchangeability, positivity, and exposure sufficiency assumptions. To estimate these effects, the authors develop a learned-exposure framework combining neural backward-Q regression with Q-based and augmented estimators. They further establish theoretical properties, including consistency of the Q-based estimator and an orthogonal expansion for the augmented estimator that characterizes different sources of estimation error. The proposed method is evaluated through synthetic and semi-synthetic experiments.

Critical strengths
  1. The paper addresses an important problem in causal inference under dynamic network interference, where treatment effects may propagate through networks with unknown strengths and time delays.
  2. The proposed DyNICL framework combines learned network exposure representations with neural backward-Q regression and augmented estimation. This approach offers a flexible way to model complex and potentially nonlinear relationships between historical treatments, network exposures, and outcomes.
Critical weaknesses
  1. The proposed framework appears to rely on observing the network structure and connections among units throughout the dynamic treatment process. However, in many real-world applications, network connections may be partially observed, missing, or evolving over time, making it challenging to obtain accurate network information across multiple time periods.
Questions for authors

I have a question about temporal dependence in treatment assignment. Does the treatment assignment at time (t) depend on the previous treatment assignment at time (t-1) or the historical treatment trajectories of neighboring units? It would be helpful for the authors to clarify how temporal dependence in treatment assignment is modeled and whether the proposed DyNICL framework can accommodate sequentially dependent treatment assignment mechanisms.

Comments on the paper

6 comments
Reviewer_eye8OpenReview ↗

Thank you for the detailed rebuttal.

Thank you for the detailed rebuttal. While the authors have clarified several points, some of my concerns remain unresolved, particularly regarding the strong assumptions of exposure sufficiency and independent network trajectories. The rebuttal does not sufficiently justify their applicability. Therefore, I will maintain my original rating of 2.

AuthorsOpenReview ↗

Official Comment Part 1/3

[Independent network trajectories]

Our independence assumption refers to independent repetitions of the same network experiment. The network and data-generating rules remain fixed, while treatments and outcomes are generated anew in each run, with no interference across runs. This independence is ensured by the experimental design, and Assumption B.7 incorporates this setting into the theoretical analysis. Within each run, nodes may influence one another, and past outcomes may affect subsequent treatment assignments.

When it applies: The data contain multiple independent network processes under the same design. Each complete process constitutes one statistical sample, while allowing interference between nodes and temporal feedback within that process.

For example, when several communities with no interference between them are surveyed at the same time, each community constitutes a complete statistical unit. Residents within a community may influence one another; independence concerns the communities, rather than individual residents.

When the current guarantees do not apply: Only one continuously evolving network is observed, and dependent segments of its trajectory are treated as independent samples. Estimation guarantees for a single long trajectory require separate analysis and are not directly covered by the current theorems under B.7.

For example, if the same community is followed for a year and its state in one month affects the next, dividing the record into monthly segments does not make each month an independent network experiment.

AuthorsOpenReview ↗

Official Comment Part 2/3

[Exposure sufficiency]

Example 1: Infection prevention. In a model with additive infection hazards, protective measures taken by different contacts reduce the focal individual’s infection risk according to contact frequency and propagation delay. Suppose that, given the same health status, infection stage, and other relevant information, the same aggregate risk reduction produces the same effect on the focal individual, regardless of which contacts took the protective measures. This summary then satisfies the exposure-mapping sufficiency requirement.

Example 2: Peer training. Colleagues’ participation in training can improve a focal individual’s work performance through communication, potentially after a delay. Changes in performance can, in turn, affect that individual’s subsequent training participation. Here, exposure summarizes how peers’ past training affects individual performance, while the retained performance history captures how outcomes affect subsequent treatment. Together, they describe the dynamic process “peer training → individual performance → subsequent training,” explaining why we extend exposure mapping to a longitudinal setting.

Clarification: Our exposure sufficiency condition follows the principle of properly specified exposure mappings in Aronow and Samii (2017) and extends it to longitudinal settings. The original condition requires equal exposures to imply equal potential outcomes. We further require exposure and retained history to jointly determine the distribution of relevant subsequent states, thereby accommodating propagation delays and treatment–outcome feedback.

Applicability of exposure sufficiency: As illustrated in Examples 1 and 2, this condition can hold when exposure summarizes the relevant effects of peer treatments and the retained history contains the other information governing subsequent changes. Infection prevention and peer training provide concrete application settings; the key requirement is that the selected exposure and history capture the underlying mechanisms.

Novelty of DyNICL: Single-time-point methods for network effects compare outcomes under treatment configurations. Our method further evaluates treatment effects after multiple steps of propagation and feedback. For example, it can assess how encouraging particular contacts to take protective measures at earlier times affects the focal individual’s subsequent health. Under the exposure condition, sequential exchangeability, and policy support, these longitudinal policy effects are identifiable, whereas a single-time-point contrast cannot directly answer this question. In the SocioPatterns setting with long memory and feedback, DyNICL-Aug achieves peer-effect and total-effect RMSEs of 0.309 and 0.781, respectively, supporting its ability to estimate these dynamic effects.

Significance of DyNICL: Infection prevention requires accounting for who contributes an effect, how long it takes to emerge, and how changes in health affect subsequent protective behavior. Fixed propagation summaries have limited ability to distinguish these effects. In the SocioPatterns semi-synthetic experiment, fixed multi-hop-Aug yields peer-effect and total-effect RMSEs of 0.791 and 1.277. By learning propagation weights and time lags and using backward-Q recursion to evaluate subsequent effects, DyNICL reduces these errors to 0.309 and 0.781, corresponding to reductions of approximately 60.9% and 38.8%. For prevention decisions affecting large populations, more accurate evaluation of intervention targets and timing could improve the allocation of limited protective resources.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.