📊 statistics

Causal Mediation Analysis for Network Data with Graph Neural Network

This paper proposes a nonparametric framework for causal mediation analysis in network data that utilizes graph neural networks to handle high-dimensional confounding and simultaneous treatment and mediator spillovers, enabling the identification of direct and indirect effects with doubly robust estimators and asymptotic normality under approximate neighborhood interference.

Original authors: Peikai Wu, Zhiguo Xiao

Published 2026-08-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Peikai Wu, Zhiguo Xiao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out why a rumor spreads through a school. You might think, "If I tell my best friend a secret, they will tell their friends, and eventually everyone knows." But what if your friend only tells the rumor because they heard it from someone else first? Or what if the rumor changes as it travels, becoming a different story by the time it reaches the other side of the playground? In the world of science, this is called causal mediation analysis. It's like trying to open a black box to see exactly how one thing (the "treatment") turns into another thing (the "outcome") by looking at the middle step (the "mediator"). Usually, scientists assume everyone is an island, acting only on their own choices. But in real life, we are all connected by invisible webs of friendship, family, and social media. When people are connected, one person's actions can spill over and affect their neighbors, making it incredibly hard to tell who did what to whom. This paper tackles the messy, tangled reality of those connected webs.

The authors, Peikai Wu and Zhiguo Xiao from Fudan University, are tackling a specific headache in data science: how to measure cause-and-effect when people are linked in a giant, complex network, and those links change how information flows. They are particularly interested in spillover effects, where a treatment (like a new teaching method) doesn't just help the person who gets it, but also their friends. They want to know: How much of the success comes from the person learning the material themselves (the direct effect), and how much comes from them passing that knowledge to their friends (the indirect effect)?

The problem is that most old-school math tools for this job assume people are isolated. If you try to use those tools on a connected group, you get confused results because the tools don't know how to handle the "noise" of friends influencing friends. To fix this, the authors built a new, super-flexible framework that doesn't force the network into a simple shape. Instead of guessing how friends influence each other, they let the data speak for itself using a special type of artificial intelligence called a Graph Neural Network (GNN). Think of a GNN as a detective that doesn't just look at one person's file, but walks through the entire friendship map, learning how information ripples through the whole group layer by layer.

In their study, the authors tested this new detective against older, simpler methods. They created fake networks of 1,000 to 4,000 people and programmed them with complex, non-linear rules—meaning the influence wasn't just a simple "friend A tells friend B," but a messy mix of second-degree friends and hidden patterns. The results were clear: the old methods, which tried to summarize the network by just averaging a friend's traits, got the math wrong. They were biased and missed the true effects. The new GNN method, however, successfully navigated the tangled web. It found the true answers with much less error and gave much more reliable confidence intervals.

The authors also applied their method to a real-world study about agricultural insurance in rural China. They wanted to know if a training session helped farmers buy insurance because the farmers actually learned about insurance (knowledge), or because they just thought more people were buying it (perception). The new method revealed that the training worked because it boosted actual knowledge, which then led to more sales. The "perception" angle, however, wasn't a real driver. This shows that their new tool can cut through the noise of social influence to find the true engine of change.

By combining advanced math with a flexible AI that understands how networks actually work, this paper provides a new way to untangle cause and effect in our connected world. It suggests that when we stop pretending people are islands and start using tools that respect the complexity of our social webs, we can finally see the true path of how actions ripple through a community.

Problem Statement
Causal mediation analysis traditionally relies on the Stable Unit Treatment Value Assumption (SUTVA), which posits that an individual's potential outcomes and mediators depend solely on their own treatment. This assumption is frequently violated in networked populations where interference (spillover) is prevalent. Existing literature on network causal inference often employs exposure mappings to aggregate treatments into low-dimensional vectors, but these mappings frequently conflate the definition of the causal estimand with the structural restriction of the interference mechanism. When the true interference mechanism does not align with a pre-specified mapping, identification results lose their causal interpretation. Furthermore, mediation analysis in network settings remains limited, often restricted to partial interference settings, linear structural equation models, or treating the network structure itself as the mediator. There is a lack of a framework capable of handling simultaneous treatment and mediator spillovers, high-dimensional network confounding, and nonparametric structural functions within a single large observed network.

Methodology
The paper proposes a nonparametric framework for causal mediation analysis on a single large observed network.

Model and Identification: The authors define a structural model where the outcome, mediator, and treatment are generated by unknown functions depending on individual covariates, the adjacency matrix, and error terms. Crucially, the outcome function allows arbitrary dependence on the treatment and mediator vectors of all individuals. To define causal estimands without restricting the true interference mechanism, the authors introduce exposure mappings ( $T_i$ ) and mediator mappings ( $S_i$ ) that summarize the spillover from other individuals ( $D_{-i}, M_{-i}$ ), while keeping an individual's own treatment ( $D_i$ ) and mediator ( $M_i$ ) separate. This separation allows for the definition of "own" effects (controlled direct, natural direct, and natural indirect) distinct from spillover effects.
Identification Conditions: Standard sequential ignorability is insufficient due to network dependence. The authors propose strengthened conditional independence assumptions, requiring the potential outcome to be conditionally independent of the treatment vectors of all individuals and, for natural effects, the mediators of all individuals. Primitive sufficient conditions are provided in terms of the independence of structural error terms.
Estimation: The authors construct Augmented Inverse Probability Weighted (AIPW) estimators for the defined causal effects. These estimators are doubly robust for controlled effects and multiply robust for natural effects. The core challenge is estimating high-dimensional nuisance functions (outcome regressions and propensity scores) that depend on the entire network structure.
Graph Neural Networks (GNN): To address high-dimensional network confounding without pre-specifying feature aggregation, the authors employ Graph Neural Networks (specifically a Principal Neighborhood Aggregation architecture). GNNs take node features and the adjacency matrix as direct inputs, using layer-wise message passing to aggregate information from neighbors and higher-order neighbors. This allows the model to learn complex, nonlinear network dependencies automatically.
Asymptotic Theory: The paper establishes asymptotic normality for the AIPW estimators under approximate neighborhood interference and weak network dependence (modeled as a $\psi$ -dependent random field). It also proves the consistency of a network Heteroskedasticity and Autocorrelation Consistent (HAC) variance estimator, with bandwidths adapted to network size and density.

Key Contributions

Framework Development: The paper develops a framework for causal mediation analysis in a single large observed network where treatment and mediator spillovers occur simultaneously. It extends the distinction between defining causal estimands and restricting interference structures (as suggested by Sävje, 2024) to the mediation context, cleanly separating own effects from spillover effects.
Integration of GNN and Robust Estimation: It combines GNN-based adjustment for high-dimensional network confounding with robust AIPW estimation. The nuisance functions are learned nonparametrically from the entire network, and the resulting estimators possess double or multiply robustness properties. This addresses the gap where existing methods struggle with high-dimensional nonlinear confounding and exposure mapping misspecification.
Inference Theory: The paper provides asymptotic theory for these estimators under a single growing network. Unlike previous work relying on independent groups or parametric models, it allows for arbitrary-order network dependence and nonparametric nuisance estimation. It establishes conditions for asymptotic normality and constructs a consistent network HAC variance estimator, enabling valid large-sample inference.

Results

Simulations: The authors conduct simulations under random geometric and Erdős–Rényi graph models with complex, nonlinear network spillovers. The GNN-based estimator significantly outperforms Multilayer Perceptrons (MLP) and Random Forests (RF) built on hand-constructed neighborhood averages. The GNN estimator achieves lower bias, better standard deviation, and confidence interval coverage close to the nominal 95% level, particularly when the number of GNN layers matches the data-generating mechanism's dependency range. In contrast, conventional ML methods exhibit persistent bias and poor coverage due to their inability to capture higher-order network dependence.
Empirical Application: A reanalysis of an agricultural insurance experiment in rural China (Cai et al., 2015) is presented. The study finds that intensive sessions significantly increase insurance take-up. The analysis identifies "insurance knowledge" as a substantive mediating channel, whereas perception-based mediators (perceived take-up rate or subjective disaster probability) show essentially zero natural indirect effects. This transforms informal discussions of mechanisms into estimable and testable causal decompositions.

Significance
The paper claims to fill a critical gap in the literature by providing the first systematic study of identification, estimation, and inference for causal mediation analysis in a single large network under a general interference structure and a nonparametric model. By separating the definition of estimands from the structural restrictions of interference and leveraging GNNs for high-dimensional nuisance estimation, the proposed framework offers a robust solution for decomposing causal effects in the presence of complex network spillovers. The development of valid large-sample inference theory for these estimators makes the framework applicable to real-world observational network data where interference is the norm rather than the exception.

More like this