Source-linked AI summary
When Interference Graphs Evolve: Doubly Robust Estimation of Dynamic Peer Effects
Xiaojing Du
TL;DR
The paper addresses causal effect estimation when peer-interaction graphs evolve and pre-assignment history, dynamic exposure, and post-assignment change have distinct roles. It defines controlled contrast profiles and develops DynaNet-DR, whose canonical estimator is doubly robust under stated assumptions; benchmarks show favorable full-profile accuracy, while the study does not generate counterfactual edges.
Problem
Evolving interaction graphs make it difficult to distinguish own-treatment responses from dynamic peer effects and post-assignment network change.
Method
The paper indexes potential outcomes by own treatment, temporally aggregated peer exposure, and a post-assignment evolution summary, then estimates the resulting controlled contrasts with DynaNet-DR.
Results
Among methods targeting the full (A, Z, M) profile, DynaNet-DR has the lowest unrounded mean RMSE in all 16 dataset-by-family cells, with a 25.9% mean cellwise reduction versus the better comparator.
Takeaways & Limitations
The paper supports summary-indexed controlled comparisons under its assumptions, while its benchmarks assess fixed graph sequences rather than counterfactual edge-generation policies.
Takeaways & Limitations
The estimand does not identify the effect of a particular edge-editing policy, and nominal interval coverage is reported empirically without claiming general validity under evolving-graph dependence.
Abstract
from arXiv · showhide
Peer effects are difficult to estimate when interaction graphs evolve because pre-assignment network history, dynamic peer exposure, and post-assignment network change have distinct causal roles. We introduce a controlled contrast framework that indexes potential outcomes by own treatment, temporally aggregated peer exposure, and a post-assignment evolution summary. Differences between the resulting means define own-treatment, peer-exposure, controlled network-evolution, and joint controlled contrasts rather than a mediation decomposition. We develop the Dynamic Network Doubly Robust estimator, DynaNet-DR, which combines a temporally factorized propensity with normalized augmentation. Under consistency, summary sufficiency, sequential exchangeability, positivity, nuisance convergence, and weak dependence, its canonical estimator is consistent when either the outcome regression or the propensity estimator is consistent. The reported implementation adds representative-score prediction, fixed clipping, and finite-sample stabilization. Semi-synthetic benchmarks on fixed real temporal graph sequences show favorable estimation accuracy among methods targeting the full profile. These benchmarks assess summary-indexed contrasts rather than counterfactual edge generation, and the MathOverflow study is an observational illustration under the stated assumptions.
1 Introduction
The paper frames evolving-network interference as a problem of separating own treatment, temporally structured peer exposure, and post-assignment network evolution. It proposes controlled contrasts and DynaNet-DR to estimate the resulting profile.
- 1 Introduction: Evolving platform networks create peer effects that classical no-interference analyses can conflate with responses to a unit’s own assignment.Network evolution and temporally changing neighborhoods make static contemporaneous exposure summaries potentially inadequate.
- 1 Introduction: The dynamic exposure mapping uses recent peer assignments, pre-assignment edge weights, and temporal decay, with interpretable intervention levels and overlap diagnostics.The mapping is designed to retain temporal structure omitted by static summaries.
- 1 Introduction: The framework treats pre-assignment history as adjustment and exposure-construction information, while post-assignment evolution is a separate controlled-intervention coordinate.Varying evolution while fixing own treatment and peer exposure defines a controlled network-evolution contrast, not a mediation effect.
- 1 Introduction: The benchmarks evaluate same-estimand accuracy on fixed real temporal graph sequences rather than counterfactual edge generation.MathOverflow is presented as an observational illustration.
- 1 Introduction: DynaNet-DR combines temporal graph features, a temporally ordered propensity factorization, and normalized augmentation to estimate controlled contrast profiles.The estimator’s canonical form is distinguished from the reported finite-sample stabilization.
2 Problem Definition
The problem definition indexes each potential outcome by own treatment, dynamic peer exposure, and a post-assignment network-evolution summary. Their marginal mean supports coordinate-wise and joint controlled contrasts without an additive decomposition.
- 2 Problem Definition: Each unit-time observation includes a pre-assignment graph, treatment vector, subsequent graph, and post-evolution outcome horizon.The history H_it contains covariates, treatments, outcomes, and network structure available before assignment.
- 2 Problem Definition: Dynamic peer exposure Z_it aggregates other units’ treatments using recent network history, edge strength, and temporal decay into discrete intervention levels.These levels define target interventions and support diagnostics.
- 2 Problem Definition: M_i,t+1 summarizes local post-assignment graph evolution and is treated as a controlled-intervention coordinate rather than an ordinary pre-assignment confounder.The summary may represent tie formation, dissolution, or intensity change and is binary in the stated setup.
- 2 Problem Definition: The target mean μ(a, z, m) averages potential outcomes indexed by own treatment, peer exposure, and network evolution over target unit-times and pre-assignment histories.The framework targets a single-world controlled mean rather than requiring mediation counterfactuals.
- 2 Problem Definition: Coordinate-wise contrasts vary one intervention coordinate while fixing the other two, and the joint contrast compares two complete intervention levels without imposing additivity.The four contrast families form a profile of distinct state comparisons.
3 Identification
Identification relies on consistency, summary sufficiency, sequential exchangeability, and positivity to recover controlled means and their contrasts. These assumptions are substantive, especially for exposure summaries and unmeasured homophily.
- 3 Identification: Consistency links each observed outcome to the potential outcome indexed by the observed treatment, peer exposure, and evolution summary.This is the first condition in the stated identification argument.
- 3 Identification: Summary sufficiency assumes that, conditional on H_it, potential outcomes depend on assignments and network paths only through A_it, Z_it, and M_i,t+1.Graph configurations mapped to the same summary levels are treated as outcome-equivalent.
- 3 Identification: Sequential exchangeability separates confounding for joint treatment and peer exposure from confounding for the subsequent evolution summary.The two conditions are stated conditional on H_it and then on H_it, A_it, and Z_it, respectively.
- 3 Identification: Positivity requires target treatment, peer-exposure, and evolution-summary levels to have probabilities bounded away from zero conditional on histories.This ensures the relevant regression and weighting quantities are defined at every target level.
- 3 Identification: Summary sufficiency and exchangeability remain substantive assumptions, and unmeasured homophily can violate exchangeability or prevent separating peer correlation from contagion.The default benchmark includes a generated community-correlated trait in nuisance features, with a sensitivity analysis that withholds it.
- 3 Identification: Under Assumptions 1 through 4, the mean potential outcome and all contrasts of μ(a, z, m) are identifiable.The result follows by connecting observed and potential outcomes, sequentially removing confounding, and averaging over histories.
4 Method
DynaNet-DR estimates controlled contrasts under evolving-network interference using temporally ordered nuisance models and normalized augmentation. The reported implementation adds score-based prediction, propensity clipping, stabilization, and diagnostics, while the formal guarantees apply under stated conditions to the canonical estimator.
- Method: DynaNet-DR combines structured dynamic exposure, post-assignment evolution summaries, temporally held-out nuisance estimation, and normalized augmentation.The propensity is factorized into joint assignment and subsequent evolution-summary mechanisms.
- Method: The reported exposure implementation supplies a continuous score to the outcome learner, using observed scores for residuals and conditional-mean scores for target predictions.This representative-score plug-in is not generally equivalent to integrating nonlinear outcomes over within-level score distributions.
- Method: The evolution summary enters target and nuisance models as a controlled level, whereas the benchmark and MathOverflow application use different summary constructions.The estimand does not identify the effect of a particular edge-editing policy.
- Method: Nuisance models are trained on complementary contiguous temporal blocks and averaged on a later estimation window, rather than observation-level cross-fitting.Theoretical consistency therefore requires convergence on the estimation-window distribution.
- Canonical estimator: The canonical estimator uses normalized augmentation with λ̂ = 1 and an unclipped propensity estimate, as covered by Theorem 2.Its bias correction removes plug-in contrast bias when the propensity estimator is consistent, while outcome-regression consistency makes residual corrections vanish.
- Reported stabilization: The reported implementation clips estimated joint propensities to [0.02, 0.98] and applies finite-sample stabilization based on effective sample size and clustered variance.These implementation choices are distinct from the canonical theorem, which does not cover an active fixed clip or representative-score approximation without additional conditions.
5 Experiments
The experiments evaluate same-estimand accuracy, restricted-design scope, nuisance-model stress tests, and descriptive interval coverage across semi-synthetic temporal network benchmarks and an observational MathOverflow illustration. DynaNet-DR performs favorably among full-target methods, while comparisons and coverage remain bounded by target mismatch and inferential assumptions.
- Experimental design: The benchmarks use fixed temporal graph sequences from CollegeMsg, email-Eu, Primary School, and High School 2012, retaining temporal structure, activity heterogeneity, and communities.Treatment, dynamic exposure, post-assignment evolution, and outcomes are generated on these observed graph sequences.
- Metrics and estimands: The evaluation reports seven contrasts across own-treatment, peer-exposure, controlled network-evolution, and joint controlled families, using family and overall RMSE against exact benchmark truths.Empirical coverage is the fraction of 70 node-clustered nominal 95 percent intervals containing the truth, and is descriptive rather than evidence of general validity.
- Main results: Among full-target methods, DynaNet-DR has the lowest unrounded mean RMSE in all 16 dataset-by-family cells, with a 25.9% mean cellwise reduction relative to the better of adapted TL and GML-DR.CollegeMsg PE is tied with adapted GML-DR at three-decimal precision; larger reported gains include email-Eu DE (0.026 versus 0.079) and PrimarySchool CNE (0.096 versus 0.192).
- Main results: Five restricted rows return CNE as zero and Naive-A returns PE as zero, while DynInt is lowest in two displayed cells but does not support full-profile comparison.Restricted methods target different estimands, and DynInt omits M and uses a representative continuous score.
- Coverage: DynaNet-DR’s nominal intervals attain empirical coverage from 0.91 to 1.00, but node clustering does not establish a central limit theorem under evolving-graph dependence.The reported intervals therefore have nominal rather than guaranteed general coverage validity.
- Ablations and stress tests: The adaptive-gate Full variant has lower mean overall RMSE than the fixed λ = 1 ablation on all four datasets, while correction lowers RMSE relative to plug-in prediction on three datasets.The comparison isolates the adaptive gate within the reported pipeline, but does not establish that the gate is optimal beyond the evaluated settings.
- Ablations and stress tests: Under nuisance restrictions, DynaNet-DR improves on plug-in estimation for the outcome restriction, does not uniformly outperform adapted TL or GML-DR, and performs best under the propensity restriction.These finite-sample stress tests do not verify the theoretical double-robustness result.
- Illustrative application: The MathOverflow application reports descriptive, representative-score, model-adjusted contrasts because it lacks ground truth and hidden-confounding sensitivity analysis; its estimates are not causal validation.Within-bin ordering of some variables relative to A and Z is unresolved.
6 Related Work
Network causal inference formalizes interference through exposure mappings and network experiments, while observational identification depends on strong adjustment assumptions. Graph-learning and dynamic methods extend exposure modeling to high-dimensional or evolving networks.
- Exposure mappings formalize interference, while network experiments address it through study design.
- Observational identification of network effects relies on strong adjustment assumptions.
- Graph-learning and dynamic methods learn exposure mappings or use evolving graphs as predictors.
7 Conclusion
DynaNet-DR estimates a controlled contrast profile using a temporally factorized propensity and normalized augmentation. Its canonical estimator is doubly robust under the stated conditions, while stochastic graph interventions and cross-node inference remain future work.
- DynaNet-DR estimates a controlled contrast profile using a temporally factorized propensity and normalized augmentation.
- The canonical estimator is doubly robust under the stated conditions.
- Stochastic graph interventions and cross-node inference remain future work.