Causal Inference under Network Interference

This direction is merged into another direction: network interference. Its last version stays readable and citable here.

Version 1 of 1206 papers284 conceptsfrozen 2026-10-04

In the classical "gold standard" of experimental design, we treat participants as if they live on isolated islands. We assume that if we give a treatment to Person A, it has no impact on the outcome of Person B. In statistics, this is known as the Stable Unit Treatment Value Assumption (SUTVA).

But in a world of social networks, global marketplaces, and infectious diseases, these islands are connected by bridges. If your friend receives a vaccine, your own risk of infection drops, even if you weren't treated Gilbert and Hudgens. If a ride-sharing app offers a discount to some riders, it changes the availability of drivers for everyone else Luo et al.. This phenomenon is called network interference or spillover, and it makes traditional A/B testing dangerously misleading.

The Cost of Connection

When interference is present, standard estimators become biased Karwa. If a new feature on a social platform causes treated users to post more, their untreated friends might also engage more because they see more content. A simple comparison between the "treated" and "control" groups would underestimate the true impact of the feature because the control group was "contaminated" by the treatment Pouget-Abadie et al..

The challenge is to disentangle two distinct forces:

  1. Direct Effect: The change in my behavior caused by my treatment.
  2. Spillover Effect: The change in my behavior caused by your treatment.

Together, these sum to the Total Treatment Effect (TTE)—the metric that actually matters for a full product rollout Borgs and Chayes.

Designing Experiments for Interference

To measure these effects accurately, researchers have moved beyond simple randomization. One approach is Cluster Randomization, where entire communities or sub-networks are treated as a single unit Parkes et al.. By ensuring that treated individuals are geographically or socially distant from control individuals, we can minimize "leakage" between groups.

However, clustering is often inefficient. Modern techniques like the Independent-Set Design partition a network into "independent" units and "auxiliary" units Cai et al.. The auxiliary units are used to carefully control the "interference dose" that the independent units receive, allowing researchers to map out exactly how peer influence scales.

For platforms where fixed clusters are hard to define, Optimized Covariance Design (OCD) offers a more fluid solution Chen et al.. Instead of assigning treatments independently, OCD optimizes the correlations between assignments to balance the trade-off between bias (from interference) and variance (from small sample sizes).

Estimating the Unobservable

When we cannot perfectly control the experiment, we must rely on sophisticated models to "subtract" the interference.

Network Synthetic Interventions (NSI) extend the popular synthetic control method to networked data Agarwal et al.. By modeling the latent factors that govern how units interact, NSI can predict what would have happened to a user if neither they nor their neighbors had been treated.

Other researchers use machine learning to capture the complexity of social influence. IDE-Net uses Graph Neural Networks (GNNs) to model "Heterogeneous Peer Influence," recognizing that a close friend's actions likely matter more than a distant acquaintance's Adhikari and Zheleva. Similarly, Dual Weighting Regression (DWR) employs attention mechanisms to weigh the influence of neighbors, effectively "learning" the structure of the interference while estimating the treatment effect Bai et al..

Why It Works: The Geometry of Influence

The core intuition behind these methods is that network interference is not just "noise"—it is structured information. By treating the network as a first-class citizen in the causal model, we can use the graph's geometry to our advantage.

For instance, FLAME-Networks uses a matching strategy based on subgraphs Awan et al.. If we can find two people with nearly identical "neighborhood" structures—where one is treated and the other is not—we can isolate the treatment effect even in a highly connected system.

Even when the network structure is completely unknown, we can still estimate the TTE by observing how outcomes evolve over time Bayati and Shirani. By looking at the "echoes" of treatment as they propagate through a population, we can infer the aggregate impact of interference without ever seeing the individual links.

As we move from testing simple buttons to testing complex social and economic ecosystems, accounting for these connections is no longer optional. It is the difference between seeing a distorted reflection and seeing the true causal impact of our interventions.

Papers in this version

Showing 60 of 206 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_inference_under_network_interfere_a8c67c5f,
  title = {Causal Inference under Network Interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-under-network-interference/rdr_a8c67c5f827d4583ad7bada8652cbe08},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_inference_under_network_interfere_a8c67c5f_v1,
  title = {Causal Inference under Network Interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-under-network-interference/rdr_a8c67c5f827d4583ad7bada8652cbe08/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}