Causal Inference for Network Autoregression Model: A Targeted Minimum Loss Estimation Approach

Yong Wu, Shuyuan Wu, Xinwei Sun, Xuening ZhuarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a novel Targeted Minimum Loss Estimation (TMLE) approach for causal inference in networks with "infinite interference," where treatment effects can propagate indefinitely through the network. The method effectively handles complex network structures and endogenous peer effects, leading to more accurate and efficient estimation of average treatment effects compared to existing methods.

What Problem Does This Paper Address?

Estimating the causal impact of an intervention (like a marketing campaign or policy change) is challenging when individuals are connected in a network. This is because the treatment applied to one person can influence others (interference). Traditional methods often assume "finite-order interference," meaning effects only propagate through a limited number of network connections. However, in many real-world scenarios, like social networks or online platforms, interference can be much more widespread, extending to "infinite interference" where effects can theoretically reach any node, albeit with diminishing strength over distance. This paper tackles the problem of accurately estimating causal effects in such complex network settings, specifically addressing the "endogenous peer effect" where individuals' behaviors are influenced by their neighbors' behaviors, which in turn are influenced by the treatment.

Methods

The core contribution is a Targeted Minimum Loss Estimation (TMLE) procedure adapted for network autoregression (NAR) models. TMLE is a powerful double-robust estimation framework that aims to eliminate bias by intelligently updating an initial estimator. Here's a breakdown of their approach:

  • Network Autoregression (NAR) Model: They model the outcome $Y_i$ for each node $i$ as a function of its own covariates $X_i$, its treatment $Z_i$, and a weighted sum of its neighbors' outcomes ($pWY$). This autoregressive term ($pWY$) captures the endogenous peer effect.
  • Infinite Interference: The model allows for distance-decaying influence, effectively modeling interference that can propagate throughout the entire network, not just to immediate neighbors.
  • Targeted Minimum Loss Estimation (TMLE):
    • They first derive an influence function (IF) for the estimand (the average treatment effect).
  • A key innovation is the development of a targeted update direction that incorporates the network autoregressive structure and assigns heterogeneous, network-dependent weights to units. This step is crucial for removing "plug-in bias" that arises from estimating model components.
  • They then find a targeted distribution that minimizes a loss function, effectively correcting the initial estimator and yielding a bias-free TMLE.
  • Novel Limit Theory: A significant theoretical contribution is the development of a new limit theory to establish asymptotic normality under these complex network-dependent scenarios, particularly when network degrees diverge. This is necessary because standard asymptotic theories often assume finite network structures.

Datasets

The paper uses two types of data for evaluation:

  • Synthetic Data: Used for simulation experiments to rigorously test the performance of their proposed method against existing ones under controlled conditions.
  • Live-streaming Data: A real-world dataset from Douyin (a Chinese short-video platform) is used to analyze how live-streaming strategies affect viewer engagement. This dataset includes information on over 2,415 live-streaming sessions.

Key Insights and Ideas

  • Infinite Interference is Real and Matters: The paper strongly argues that ignoring the potential for widespread interference in networks leads to biased and inefficient causal effect estimates. Their method explicitly accounts for this.
  • Endogenous Peer Effects are Tricky: The autoregressive term ($pWY$) introduces endogeneity, meaning the treatment and outcome are intertwined through network effects. Their TMLE approach is designed to handle this complexity.
  • Heterogeneous Network Weights are Key: The targeted update direction in their TMLE assigns weights to network units that depend on their position and connections within the network. This is crucial for correctly accounting for how influence propagates.
  • Diverging Network Degrees Require New Theory: Standard statistical theory often assumes fixed network sizes or degrees. The authors developed new theoretical tools to handle scenarios where network sizes and degrees grow, which is more realistic for large online platforms.
  • TMLE Offers Superior Efficiency: Their method consistently achieves smaller asymptotic variances than existing approaches, especially when covariates are i.i.d. and estimated using empirical distributions.

Key Results

  • Simulation Studies:
    • The proposed TMLE estimator consistently outperforms competing methods (Direct Estimation, Network Direct Interference, Approximate Neighborhood Intervention) across various settings, especially when the autoregressive parameter $\rho$ is non-zero.
  • TMLE yields the smallest standard errors (SE) in most settings, corroborating theoretical findings about its efficiency.
  • When $\rho \neq 0$, TMLE shows substantially smaller bias and better coverage proportions (CP) compared to other methods, even when the underlying model is correctly specified. For instance, in one scenario, TMLE achieved a bias of -0.006 with 95.6% CP, while a "correctly specified" Direct Estimation method had a bias of -0.097 with 78.6% CP.
  • Live-streaming Data Analysis:
    • The analysis of Douyin live-streaming data revealed a positive peer effect, consistent with findings in marketing research about positive spillovers.
  • The TMLE method was used to evaluate the impact of different live-streaming strategies (e.g., changing broadcast time, launching promotions) on viewer numbers. The results indicated that certain strategies could enhance broadcasting popularity, with regional differences observed.

Limitations and Future Work

  • Doubly Robustness: While the proposed TMLE method is shown to be robust, it is not "doubly robust" in the strictest sense because it assumes an infinite interference setting. The authors suggest developing doubly robust estimators for future work.
  • Model Specification: The validity of the TMLE method relies on good estimation of the peer effect parameter ($\rho$) and the function $g(\cdot)$. Investigating the consistency and convergence rates of these estimators would be valuable.
  • Temporal Dependence: The current framework primarily considers a cross-sectional network structure. Extending the framework to jointly accommodate temporal and network dependence is identified as an important direction for future research.

Summary updated

Abstract

We study estimation of the average treatment effect (ATE) from a single network in observational settings with interference. The weak cross-unit dependence is modeled via an endogenous peer-effect (network autoregressive) term that induces distance-decaying network dependence, relaxing the common finite-order interference to infinite interference. We propose a targeted minimum loss estimation (TMLE) procedure that removes plug-in bias from an initial estimator. The targeting step yields an adjustment direction that incorporates the network autoregressive structure and assigns heterogeneous, network-dependent weights to units. We find that the asymptotic leading term related to the covariates $\mathbf{X}_i$ can be formulated into a $V$-statistic whose order diverges with the network degrees. A novel limit theory is developed to establish the asymptotic normality under such complex network dependent scenarios. We show that our method can achieve smaller asymptotic variance than existing methods when $\mathbf{X}_i$ is i.i.d. generated and estimated with empirical distribution, and provide theoretical guarantees for estimating the variance. Extensive numerical studies and a live-streaming data analysis are presented to illustrate the advantages of the proposed method.
BibTeX
@misc{wu2025causal,
  author = {Yong Wu and Shuyuan Wu and Xinwei Sun and Xuening Zhu},
  title = {{Causal Inference for Network Autoregression Model: A Targeted Minimum Loss Estimation Approach}},
  year = {2025},
  eprint = {2511.06652},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.06652},
}
Figure