Causal Inference under Low Order Network Interference

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

Version 1 of 18 papers15 conceptsfrozen 2026-10-04

In traditional A/B testing, we rely on the Stable Unit Treatment Value Assumption (SUTVA): the idea that Alice’s outcome depends only on whether Alice received the treatment. In social networks or marketplaces, this assumption is almost always violated. If Alice receives a new communication feature, her behavior changes, which in turn affects her friend Bob, even if Bob himself was in the control group.

This phenomenon, known as network interference, makes estimating the Total Treatment Effect (TTE)—the difference between a world where everyone is treated and a world where no one is—notoriously difficult.

The Curse of Connectivity

The standard way to handle interference is to look for "pure" instances: cases where a user and all of their neighbors happen to receive the same treatment assignment. However, as Hao Yin notes in Randomized Graph Cluster Randomization, the probability of this happening by chance in a dense network is vanishingly small.

If you have a neighbor with 100 friends, the odds that all 100 friends are assigned to the treatment group in a 50/50 split is $0.5^{100}$. This leads to an "exponential variance explosion." To get a reliable estimate, you would need a sample size that grows exponentially with the number of connections (the degree) of the nodes. In modern social networks, this makes unbiased estimation practically impossible.

The Low-Order Interaction Insight

Recent work by Mayleen Cortez-Rodriguez, Christina Lee Yu, and Matthew Eichhorn provides a way out of this trap. In Exploiting neighborhood interference with low order interactions under unit randomized design, they propose that we don't need to model the influence of the entire neighborhood at once.

Instead, they assume that interference is limited to low-order interactions. This means that while Alice is influenced by her neighbors, she is influenced by small subsets of them (e.g., pairs or triplets) rather than the complex, simultaneous combination of all 100 friends.

By modeling the potential outcome as a polynomial function of the neighbors' treatment statuses, the complexity of the problem shifts. The order of the polynomial ($k$) represents the complexity of these social synergies. If $k=1$, Alice is affected by each neighbor individually. If $k=2$, she is affected by interactions between pairs of neighbors.

From Exponential to Polynomial Variance

The mathematical breakthrough of this "low-order" assumption is the change in variance scaling. In Exploiting neighborhood interference with low-order interactions under unit randomized design, the authors demonstrate that if interactions are limited to order $k$, the variance of the estimator scales polynomially with the network degree ($d^k$) rather than exponentially ($2^d$).

This makes causal inference tractable even in dense, well-connected graphs. The bottleneck for accurate estimation is no longer how many friends a person has, but rather the "depth" of the social synergy. As long as the interactions remain relatively simple (low $k$), we can estimate global effects using standard randomized designs.

Designing for Interference

While the low-order assumption helps with estimation, the design of the experiment—how we assign treatment—still matters.

  1. Randomized Clustering: Hao Yin argues that moving from fixed clusters to randomized clustering distributions can further stabilize estimators, allowing for unbiased global treatment effect estimation in large-scale networks.
  2. Two-Stage Rollouts: In Combining Rollout Designs and Clustering for Causal Inference under Low-order Interference, Cortez-Rodriguez and colleagues suggest a hybrid approach. By using clustering to restrict treatment to specific sub-populations during a rollout, researchers can drastically reduce variance. This introduces a small amount of bias, but it makes the estimator far more practical for real-world A/B tests where data is limited.

By shifting the focus from "who is connected to whom" to "how many people interact at once," this research direction provides a theoretical and practical framework for measuring the true impact of interventions in an interconnected world.

Papers in this version

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_low_order_network_fc6e6de4,
  title = {Causal Inference under Low Order Network Interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-under-low-order-network-interference/rdr_fc6e6de425ac4c46bec3a2f11405db22},
  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_low_order_network_fc6e6de4_v1,
  title = {Causal Inference under Low Order Network Interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-under-low-order-network-interference/rdr_fc6e6de425ac4c46bec3a2f11405db22/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}