Decomposing Causal Effects Under Network Interference

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

Version 1 of 111 papers19 conceptsfrozen 2026-10-04

In classical causal inference, we rely on a fundamental assumption: the Stable Unit Treatment Value Assumption (SUTVA). It posits that the outcome of any individual depends only on the treatment they personally receive. However, in a connected world, this assumption is frequently violated. If your friend receives a vaccine, your risk of infection drops even if you aren't vaccinated. If a neighboring school receives a grant, students from your district might transfer there.

This phenomenon is known as network interference or spillover. When SUTVA fails, traditional causal estimators become biased because they cannot distinguish between the effect of one's own treatment and the "ripples" caused by the treatments of others. Recent research has moved beyond treating interference as a statistical nuisance, instead developing frameworks to decompose these ripples into measurable, actionable components.

The Anatomy of a Networked Effect

To understand why an intervention works in a network, we must break the total effect into pieces. Vishesh Karwa in A systematic investigation of classical causal inference strategies under mis-specification due to network interference formalizes this by partitioning the total change in an outcome into three parts:

  1. Direct Effect: The change caused by the treatment applied to the individual.
  2. Interference (Spillover) Effect: The change caused by treatments applied to the individual’s neighbors.
  3. Interaction Term: The combined effect of both the individual and their neighbors being treated simultaneously.

This decomposition is crucial because it reveals the "social utility" of an intervention. As Matt Kusner, Chris Russell, and Joshua R. Loftus argue in Causal Interventions for Fairness, accounting for these spillovers allows policymakers to distribute resources more efficiently. An intervention might have a modest direct effect but a massive social impact if it triggers significant positive spillovers to neighboring units.

Disentangling Peer Influence via Mediation

A major challenge in network data is that "peer effects" are often a black box. If your neighbor’s behavior influences yours, is it because of the action they took (the treatment) or the result of that action (the outcome)?

Xiaojing Du, Jiuyong Li, and colleagues address this in Telling Peer Direct Effects from Indirect Effects in Observational Network Data. They introduce a framework called gDIS that uses causal mediation analysis to separate:

  • Direct Peer Effects: Your neighbor gets a discount code (treatment), which encourages you to shop.
  • Indirect Peer Effects: Your neighbor gets a discount code, buys a cool new jacket (outcome), and seeing the jacket encourages you to shop.

By using Graph Neural Networks (GNNs) to model these dependencies, researchers can identify the specific causal pathways of influence. This distinction is vital for policy: if a vaccine's benefit is primarily "indirect" (mediated by the fact that neighbors aren't getting sick), the strategy for deployment differs significantly from a treatment where the "direct" peer effect (e.g., social encouragement to get treated) is the primary driver.

Directionality and Long-Range Dependencies

Interference is rarely uniform. Fei Fang and Laura Forastiere highlight in Inward and Outward Spillover Effects of One Unit's Treatment on Network Neighbors under Partial Interference that spillover has a direction. Outward spillover measures how much one person's treatment influences their community, while inward spillover measures how susceptible an individual is to their neighbors. Even in undirected social networks, these two measures are mathematically distinct, representing "influencers" versus "conformists."

Furthermore, interference isn't always limited to immediate neighbors. Eric J. Tchetgen Tchetgen and Isabel Fulcher developed the auto-g-computation algorithm in Auto-G-Computation of Causal Effects on a Network to handle long-range dependencies. By modeling the network as a Markov Random Field, they can estimate how a treatment might propagate through multiple "hops" in a network, even when we only observe a single snapshot of the entire system.

From Analysis to Impact Remediation

The ultimate goal of understanding network interference is to design better interventions. In Impact Remediation: Optimal Interventions to Reduce Inequality, Lucius E.J. Bynum and Julia Stoyanovich move from "fair predictions" to "fair impacts."

Traditional fairness models often fail because they ignore geographic or social spillovers. For example, placing a new health clinic in one neighborhood might reduce the burden on clinics in adjacent neighborhoods. By incorporating interference into Structural Causal Models (SCMs), the Impact Remediation framework allows for the discovery of optimal resource allocations that actively reduce systemic inequality.

By treating the network structure not as a source of noise, but as a map of causal pathways, this research direction transforms how we measure influence and implement social change.

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_decomposing_causal_effects_under_network_9928c32d,
  title = {Decomposing Causal Effects Under Network Interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decomposing-causal-effects-under-network-interference/rdr_9928c32d9e184f41b3571ee1b3ae6bcb},
  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_decomposing_causal_effects_under_network_9928c32d_v1,
  title = {Decomposing Causal Effects Under Network Interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decomposing-causal-effects-under-network-interference/rdr_9928c32d9e184f41b3571ee1b3ae6bcb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}