acceptodds
Withdrawn from ICLR 2027

Robust Propagation-Integrated Inference for Observational Studies with Network Interference

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In observational studies with network interference, the treatment-propensity error can affect exposure probabilities through the network, creating additional challenges for point estimation and uncertainty quantification. We propose Propagation-Integrated Inference (PIN), which constructs observable propagation directions through cross-fitting and uses the propagation directions for both point estimation and uncertainty quantification. PIN incorporates the propagation directions into a first-order estimator and then uses second-order propagation correction to adjust the remaining propensity–outcome product error. PIN further constructs a propagation-integrated standard error that simultaneously accounts for network dependence and the adjustment used in point estimation. We derive an upper bound on the approximation error of the propagation directions, prove that second-order propagation correction does not increase the upper bound of the remaining propensity–outcome product error, and establish asymptotic normality of the final PIN point estimator and conservative coverage of confidence intervals. Across three treatment-propensity shift simulations, PIN maintains 97–98% average coverage, with 5–12-point higher coverage than the state-of-the-art baseline and only 5–9% longer intervals, while reducing absolute bias by 32–78%. In a real-network study under high treatment-propensity shift, PIN achieves 97% average direct-effect coverage and reduces absolute bias by approximately 85–89% relative to the baseline.

Primary area:
causal reasoning
Keywords:
Network Interference, Observational Studies

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tE7fOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies exposure-effect estimation in observational networks, where estimated node-level treatment propensities induce exposure-probability errors that propagate through the network and affect both point estimation and uncertainty quantification. It proposes Propagation-Integrated Inference (PIN), which uses nested cross-fitting and treatment-prediction residuals to construct observable propagation directions, incorporates these directions into a first-order augmented estimator, adds a pairwise second-order correction for the remaining propensity-outcome product error, and constructs a network-aware standard error for the corrected estimator. Under local interference, exposure positivity, nuisance-rate, and complexity conditions, the paper establishes direction approximation, asymptotic normality, and conservative interval coverage. The method is evaluated in three propensity-shift simulations and a real-network semi-synthetic study.

Critical strengths

S1. The paper identifies an important inferential failure mode that is specific to observational networks: node-level propensity error is transformed by the exposure mapping and enters both inverse-exposure point estimation and network-dependent uncertainty quantification. Using the same observable propagation directions across propensity calibration, outcome augmentation, and variance construction gives PIN a coherent structure rather than treating undercoverage as a standard-error-only problem.

S2. The paper provides substantial technical and empirical development. It gives explicit decompositions of the propensity-outcome product error, derives asymptotic normality and conservative coverage under stated conditions, and evaluates three distinct propensity shifts, two outcome mechanisms, multiple regression specifications, mechanism ladders, and several observational-network baselines. The consistent coverage improvements under severe shifts are informative finite-sample evidence.

Critical weaknesses

W1. The theory and experiments do not establish the incremental necessity of the second-order propagation-integrated construction. Proposition 4.2/B.1 only shows that projecting the outcome-side error onto the propagation-direction span cannot increase an upper bound because the projection class includes the zero vector; Appendix B explicitly notes that this does not order the signed, absolute, or realized estimation errors. The final normality theorem still requires the standard product-rate condition a+b>1/2a+b>1/2. Moreover, Proposition 4.1 shows that the score-completed propensity itself approaches the true propensity, but the experiments do not compare PIN with a matched WLS+/AIPW procedure that directly uses e^complete\hat e_{\mathrm{complete}} or π^complete\hat\pi_{\mathrm{complete}}. Without this ablation, the reported ladder cannot distinguish the benefit of ordinary propensity recalibration from the proposed propagation-direction and pairwise corrections.

W2. The empirical validation is not aligned with the scope of the theoretical and applied claims. A main outcome model is linear-in-means with nonzero network feedback, which the paper acknowledges is not covered by the fixed-radius interference assumption without an additional approximation argument. The “real-network” study retains only the observed topology and pretreatment variables; treatments and outcomes are synthetic, so it does not validate inference with real observational outcomes. Appendix E also states that it will define the synthetic potential outcome but does not provide that equation, preventing reconstruction of this study from the manuscript. Finally, at high shift the reported PIN intervals are approximately 37% and 73% longer than WLS+ for the one- and two-dimensional mappings, respectively, so the coverage improvement is less favorable than the main simulation summary suggests.

Questions for authors

Q1. Assessment-changing evidence: What happens if the score-completed propensity produced in Step 1 is used directly in an otherwise matched WLS+ or network-AIPW estimator and variance calculation? Existing results for this baseline would determine whether PIN's gains come from propensity recalibration itself or from the proposed propagation-direction and second-order machinery.

Q2. Clarification: Beyond the non-increasing projection upper bound, is there an existing result showing that the second-order correction reduces actual bias, improves mean-squared error, or relaxes the nuisance-rate conditions relative to the first-order estimator? If not, please clarify which formal guarantee is uniquely supplied by the second-order component.

Q3. Assessment-changing clarification: What is the exact potential-outcome equation used in the real-network semi-synthetic study, and does it satisfy the fixed-radius dependence assumption used by Theorems C.1--C.2? Please also clarify the propensity learner, outer/inner fold counts, exposure-probability clipping, and network distance cutoff used in the reported experiments.

Reviewer_gefhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Propagation-Integrated Inference (PIN) to address bias and inaccurate uncertainty quantification caused by propagated exposure-probability errors in observational studies with network interference. By constructing observable propagation directions, a first-order estimator, a second-order propagation correction, and a propagation-integrated standard error, PIN improves point estimation and uncertainty quantification simultaneously. Experiments across multiple simulations and a real-network semi-synthetic study demonstrate that PIN reduces bias and improves coverage.

Critical strengths
  1. This paper focuses on a specific yet significant challenge in network inference, connecting point estimation with uncertainty quantification. The authors provide insightful theoretical analysis of the underlying causes and proposed solutions.

  2. The authors provide theoretical guarantees under the stated assumptions and conduct various experiments. The experimental results support PIN’s benefits in improving coverage and reducing bias.

Critical weaknesses
  1. There is a gap between the scope of the theoretical guarantees and some of the main experimental settings. While the theory requires local interference within a fixed network radius, the LIM model in the main experiments allows feedback to propagate beyond this radius. An approximation analysis would help clarify whether the theoretical guarantees apply to these settings.

  2. The range of PIN’s applicability needs further justification. The experiments mainly vary treatment propensities, with limited exploration of different network structures and sources of error. For example, if the constructed propagation directions fail to capture the dominant errors, can PIN still effectively reduce bias and provide reliable uncertainty quantification?

Questions for authors
  1. As mentioned in Weakness 1, could you clarify the connection between the local interference assumption and the LIM experiments? Could an approximation analysis bridge this gap?

  2. Please see the final question in Weakness 2.

Reviewer_94jGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies causal effect estimation and statistical inference in observational studies under network interference, focusing on the challenges arising from treatment-propensity estimation errors that propagate through network exposure probabilities. The authors propose Propagation-Integrated Inference (PIN), a framework that uses cross-fitting to construct observable propagation directions and incorporates them into both point estimation and uncertainty quantification. The proposed method combines a first-order estimator with a second-order propagation correction to reduce the remaining propensity–outcome product error and develops a propagation-integrated standard error to account for network dependence and estimation adjustments. The authors establish theoretical guarantees for approximation error bounds, asymptotic normality, and conservative confidence interval coverage under certain regularity conditions. The proposed method is evaluated through simulation studies and experiments on a real network.

Critical strengths
  1. The paper clearly identifies three key challenges: Direction, Bias, and Uncertainty, and develops a unified framework to address them. The proposed PIN method integrates propagation-aware bias correction with uncertainty quantification, providing a systematic approach to handling treatment-propensity estimation errors under network interference.

  2. The authors evaluate the proposed method under different treatment-propensity shift scenarios, demonstrating improvements in estimation bias and confidence interval coverage compared with existing baseline methods.

Critical weaknesses
  1. The proposed PIN framework relies on pretreatment features and treatment-prediction residuals to construct observable propagation directions. However, it is unclear how sensitive the proposed method is to the selection and quality of these features. In real-world observational studies, relevant pretreatment variables may be unavailable or imperfectly measured, potentially affecting the accuracy of the estimated propagation directions.
Questions for authors
  1. How does the proposed PIN framework account for potential confounding and homophily in observational studies on networks? Since treatment assignment and network formation may both depend on shared individual characteristics and unobserved factors, distinguishing causal peer effects from spurious associations can be challenging. It would be helpful for the authors to clarify the assumptions required to address these issues and discuss whether the proposed propagation correction can mitigate bias arising from unmeasured confounding or homophily, beyond treatment-propensity estimation errors.

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

Related papers

Discussion (0)

Sign in to comment.