Causal Inference under Network Interference

1,145 concepts418 papers2 proposals

In classical causal inference and standard A/B testing, we rely heavily on a foundational assumption: the Stable Unit Treatment Value Assumption (SUTVA). SUTVA states that the treatment assigned to one individual does not affect the outcome of any other individual. Under this independent and identically distributed (i.i.d.) worldview, isolating a causal effect is straightforward because every participant lives in a vacuum.

In the real world, however, units are rarely isolated. They are nodes in social networks, participants in two-sided marketplaces, or entities sharing physical infrastructure. When you intervene on one node, the effects ripple outward to its neighbors. This phenomenon—where the treatment of one unit influences the outcomes of others—is known as network interference or spillover effects.

When SUTVA is violated, traditional statistical frameworks collapse. To design reliable systems and run accurate experiments in interconnected environments, we must understand why interference biases our estimates and how modern causal frameworks adapt to it.

The Intuition of Interference and Bias

To understand why network interference invalidates standard A/B testing, consider a social platform introducing a new interactive feature. As noted by Liberty et al., social features inherently require a user's local network to also have access to the feature to provide any real value. If a treated user interacts with an untreated friend in the control group, the control user's behavior changes. The control group is no longer a pure baseline; it has been "contaminated" by the treatment signal.

This contamination introduces severe measurement bias, which manifests differently depending on the structure of the system:

  1. Resource Competition: In two-sided marketplaces (like ridesharing platforms) or congested networks, units compete for finite resources. As Zhao et al. demonstrate, if a new algorithm helps the treatment group secure resources more efficiently, it directly starves the control group. This artificial suppression of the control group causes researchers to overestimate the global treatment effect, miscalculating what would happen if the feature were rolled out to everyone.
  2. Information Cascades: In social networks, influence travels beyond immediate neighbors. Zheleva et al. point out that traditional network experiments often assume interference stops at first-order neighbors, ignoring multi-hop cascades. When information or behavior propagates through multiple degrees of separation, it bypasses simple boundaries, diluting the observed difference between treatment and control.

Correcting for Network Confounding

In observational settings where we cannot randomize treatment, network interference introduces a dual challenge: network confounding. In a network, similar individuals naturally cluster together—a phenomenon known as homophily. This means a node's position and the characteristics of its neighbors influence both its likelihood of receiving a treatment and its ultimate outcome.

Simply throwing neighbor features into a standard regression model fails to mitigate this bias. Cai et al. show that effective estimation requires explicit mechanisms to balance feature distributions across treatment groups. Because homophily creates distinct, non-comparable neighborhoods for treated and untreated groups (network-level covariate shift), models must actively learn balanced representations to make fair counterfactual comparisons.

To decouple these complex dependencies, Khatami et al. combine the representational power of Graph Neural Networks (GNNs) with Double Machine Learning. By using GNNs to digest the surrounding network structure, the framework can isolate and strip away the confounding influence of neighbors, allowing researchers to accurately estimate both the direct treatment effect on a unit and the indirect spillover effect originating from its peers.

Standard Causal Inference:
[Treatment] -------------> [Outcome] (Assumes units are isolated)

Network Interference:
[Neighbor Treatment] ----> [Neighbor Outcome]
         \                      ^
          \                    /  (Spillover / Interference)
           v                  /
[Unit Treatment] --------> [Unit Outcome]

Designing Around Interference

When we can control the experimental design, we can structure the assignment of treatment to either contain or explicitly measure the spillover.

Cluster-Based Randomization

Instead of randomizing at the individual level, we can partition the network into tightly knit clusters and assign entire clusters to either treatment or control. Pouget-Abadie et al. frame this as a correlation clustering problem, where the goal is to maximize the variance of treatment exposure. By ensuring that a user's connections are either almost entirely treated or almost entirely untreated, the design traps the spillover effects within the clusters, preserving the integrity of the global comparison.

Switchback Experiments

In highly fluid environments like marketplaces where physical or social boundaries are porous, spatial clustering fails because resources are shared globally. To circumvent this, Wen analyzes the mechanics of switchback experiments. Instead of splitting users simultaneously, a switchback design alternates the treatment policy for the entire system over sequential windows of time. This temporal isolation prevents cross-unit contamination, though it requires careful optimization to balance carryover effects between time windows against reward autocorrelations.

Embracing Incomplete Network Data

A persistent hurdle in deploying these network-aware estimators is that a perfect graph of human or system interactions is rarely available. Fortunately, exact topological knowledge is not a strict prerequisite for valid causal inference.

The authors of A/B testing under Interference with Partial Network Information demonstrate that treatment effect estimation remains unbiased as long as researchers can identify a superset of potential neighbors. By leveraging coarse proxies—such as shared geographic metadata or broad behavioral categories—the framework can construct conservative boundaries that are guaranteed to capture the true influential connections. This insight bridges the gap between theoretical network causal inference and the messy, partial data realities of production environments.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_network_interference_073d265b,
  title = {network interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/network-interference/rdr_073d265b71f246228c13f7e1213c007f},
  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_network_interference_073d265b_v1,
  title = {network interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/network-interference/rdr_073d265b71f246228c13f7e1213c007f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake