Technical Summary: Estimating Heterogeneous Causal Effect on Networks via Orthogonal Learning

1. Problem Statement

The paper addresses the challenge of estimating heterogeneous causal effects in networked data, where the treatment of one unit influences not only its own outcome but also the outcomes of its neighbors (network interference). Two primary difficulties are identified:

1. Heterogeneity: Causal effects vary significantly across units and edges due to diverse node features and connection strengths. Standard methods often estimate only population-level averages, failing to capture these granular variations.
2. Confounding: Network homophily (the tendency of similar nodes to connect) and latent dependencies create complex confounding structures. Disentangling true causal spillovers from non-causal correlations requires adjusting for high-dimensional covariates and latent relational structures.

Existing approaches face a trade-off: classical methods provide rigorous inference for population-level averages but rely on restrictive structural assumptions, while recent graph neural network (GNN) approaches offer flexibility but often function as "black boxes," making edge-level effects difficult to interpret and uncertainty quantification (UQ) challenging.

2. Methodology

The authors propose a two-stage orthogonal learning framework designed to estimate heterogeneous direct and spillover effects while remaining robust to first-stage estimation errors.

2.1 Problem Setup and Identification

• Framework: The problem is formulated within the potential outcomes framework under interference.
• Assumptions:
  • Local Interference: Outcomes depend only on treatments within a closed one-hop neighborhood.
  • Neighborhood Unconfoundedness: Conditional on covariates from a larger $K$-hop neighborhood, local treatments are independent of potential outcomes.
  • Additive Local Interference (Key Structural Assumption): The expected outcome is modeled as a sum of a direct effect and pairwise spillover effects from neighbors. Specifically, $E[Y_i] = \mu_0 + \tau_i T_i + \sum_{j \in N_1(i)} \tau_{ij} T_j$. This reduces the exponential complexity of the interference space to a linear one, enabling edge-level identification.
• Target Estimands: The method targets the interference matrix $\Gamma$, where $\Gamma_{ii} = \tau_i$ (direct effect) and $\Gamma_{ij} = \tau_{ij}$ (spillover effect from $j$ to $i$). Node-level effects (IDE, ISE) and population-level averages (ADE, ASE) are derived as linear functionals of $\Gamma$.

2.2 Orthogonal Estimation Framework

The estimation follows a Neyman orthogonalization strategy (inspired by the R-Learner) to mitigate bias from nuisance parameter estimation:

1. Stage 1 (Nuisance Estimation):
   • Uses Graph Neural Networks (GNNs) to estimate the conditional outcome mean $m^*(X, A) = E[Y|X, A]$ and the propensity score $e^*(X, A) = P(T=1|X, A)$.
   • Employs cross-fitting (splitting the network into independent folds based on graph distance) to handle network dependence and prevent overfitting.
2. Stage 2 (Orthogonal Regression):
   • Residualizes the outcome and treatment using the Stage 1 estimates: $\tilde{Y}_i = Y_i - \hat{m}_i$ and $\tilde{T}_i = T_i - \hat{e}_i$.
   • Solves a least-squares problem to estimate the interference coefficients $\Gamma$ using the residualized variables.
   • Theoretical Guarantee: Under specific regularity conditions, the estimation error of the causal effects depends on the nuisance estimation error only at a higher order (e.g., $O(r^4)$). If nuisance estimators converge at rate $O(n^{-1/8})$, the causal estimator achieves the parametric rate $O(n^{-1/2})$, comparable to an oracle with known nuisance functions.

2.3 Attention-Based Interference Model

To capture heterogeneous and asymmetric spillovers, the interference matrix $\Gamma$ is parameterized via an attention-based model:

• Direct Effect: $\tau_i = \tilde{W}(X_i)$ via an MLP.
• Spillover Effect: $\tau_{ij} = \alpha_{ij} w_{ij}$, where $w_{ij}$ is a learned pairwise influence score based on concatenated features $(X_i, X_j)$, and $\alpha_{ij}$ are attention weights derived from a softmax function with a learnable temperature parameter $\beta$.
• This structure allows the model to interpolate between diffuse and concentrated spillovers and recover signed, edge-level effects.

2.4 Uncertainty Quantification

The paper develops a multiplier bootstrap procedure for the estimated spillover matrix:

• It perturbs the second-stage loss function using random multipliers.
• To avoid the computational cost of refitting the MLP for every bootstrap sample, a one-step linearization (using the Hessian of the loss) is employed.
• This enables the construction of pointwise and simultaneous confidence intervals for edge-level ($\tau_{ij}$) and node-level ($ISE_i$) effects.

3. Key Contributions

1. Orthogonal Learning Framework: A novel two-stage approach that combines flexible GNN-based nuisance modeling with an interpretable attention-based interference model, robust to first-stage estimation errors.
2. Edge-Level Heterogeneity: The ability to estimate and interpret signed, pairwise spillover effects ($\tau_{ij}$) rather than just aggregate averages, facilitating the detection of influential neighbors and heterogeneous interaction patterns.
3. Valid Uncertainty Quantification: A bootstrap-based procedure providing simultaneous inference for heterogeneous edge- and node-level effects, extending beyond the population-level focus of prior work.
4. Theoretical Guarantees: Proof that the estimator achieves oracle-like convergence rates under local additive interference and conditional neighborhood dependence assumptions.

4. Experimental Results

The method was evaluated on semi-synthetic data derived from real social networks (BlogCatalog, Flickr) and synthetic Stochastic Block Model (SBM) networks.

• Benchmark Performance: The proposed estimator ("Proposedest") significantly outperformed baselines (CFR, NetEst, TNet, GDML, etc.) in estimating node-level heterogeneous effects (IDE and ISE), measured by Precision in Estimating Heterogeneous Effects (PEHE). It matched or exceeded baselines on population-level metrics (ADE, ASE).
• Robustness to Nuisance Error: The performance gap between the proposed method and the "oracle" (using true nuisance functions) was minimal, validating the orthogonal design's ability to mitigate bias from GNN estimation errors.
• Interpretability:
  • Spillover Sign Recovery: The method achieved near-perfect accuracy in recovering the signs of pairwise influences, outperforming EdgeConv.
  • Influential Neighbor Detection: It successfully identified the top 20% of influential neighbors with high recall and NDCG, closely matching oracle performance.
• Uncertainty Calibration: Bootstrap confidence intervals achieved coverage rates close to nominal levels (90% and 95%) for both edge and node effects. Intervals were wider when using estimated nuisances compared to oracle nuisances, reflecting the cost of first-stage estimation, but converged as sample size increased.
• Real-World Application: A case study on air pollution (DAPSm dataset) demonstrated the method's ability to estimate direct and indirect effects of emission-control technology adoption, yielding positive estimates for both direct and spillover reductions in ozone.

5. Significance and Claims

The paper claims to bridge the gap between flexible, black-box representation learning and principled, interpretable causal inference in networks. By leveraging orthogonal learning, it enables the estimation of fine-grained, heterogeneous causal effects (down to the edge level) while maintaining statistical validity and interpretability.

The authors emphasize that their approach is particularly valuable for applications requiring downstream analysis, such as identifying influential nodes or understanding the directionality of spillovers (e.g., in political polarization or social media campaigns). They acknowledge that the additive local interference assumption is a first-order approximation; however, experiments suggest the method remains robust even when moderate treatment interactions are present, and the framework can be extended to include interaction terms if necessary.

The work positions itself as a step toward making network causal inference more granular and actionable, moving beyond average treatment effects to provide a detailed map of how interventions propagate through complex social structures.