Abstract
Traditionally, statistical inference and causal inference on human subjects rely on the assumption that individuals are independently affected by treatments or exposures. However, recently there has been increasing interest in settings, such as social networks, where individuals may interact with one another such that treatments may spill over from the treated individual to their social contacts and outcomes may be contagious. Existing models proposed for causal inference using observational data from networks of interacting individuals have two major shortcomings. First, they often require a level of granularity in the data that is infeasible in practice to collect in most settings and, second, the models are high dimensional and often too big to fit to the available data. We illustrate and justify a parsimonious parameterization for network data with interference and contagion. Our parameterization corresponds to a particular family of graphical models known as chain graphs. We argue that, in some settings, chain graph models approximate the marginal distribution of a snapshot of a longitudinal data-generating process on interacting units. We illustrate the use of chain graphs for causal inference about collective decision making in social networks by using data from US Supreme Court decisions between 1994 and 2004 and in simulations.
Introduction: Beyond the Independence Assumption
In traditional causal inference, researchers often rely on the Stable Unit Treatment Value Assumption (SUTVA), which posits that one individual's treatment does not affect another's outcome. In social contexts—ranging from public health interventions like vaccinations to the spread of information—this assumption is frequently violated. When "contagion" (outcomes influencing outcomes) and "spillover" (treatments influencing others' outcomes) occur, standard statistical tools break down.
Modeling these dependencies typically requires either high-frequency longitudinal data, which is rarely available, or high-dimensional models that suffer from the curse of dimensionality as network size increases. This paper by Ogburn, Shpitser, and Lee proposes a middle ground: using Chain Graphs to model social network dynamics from a single "snapshot" of observational data.
Methods: The Chain Graph Approach
The authors introduce Chain Graphs (CGs) under the Lauritzen, Wermuth, and Frydenberg (LWF) interpretation as a parsimonious solution for network data.
The Hybrid Structure
Chain graphs are unique because they utilize both directed ($\rightarrow$) and undirected ($-$) edges. In the context of social networks:
- Directed Edges: Represent external exposures or treatments (e.g., the specific legal issue of a court case) that influence individual responses.
- Undirected Edges: Represent the symmetric, contagious interactions between individuals' outcomes (e.g., justices influencing each other's final opinions during deliberation).
The Approximation Hypothesis
A key theoretical contribution is the "Approximation Hypothesis." The authors argue that a chain graph, which represents a system in equilibrium, can serve as a valid approximation for a complex Directed Acyclic Graph (DAG) that evolves over time. By using a log-linear parameterization (similar to the Ising model), the researchers keep the model complexity polynomial rather than exponential, making it feasible to fit to real-world data.
Data and Validation
The model was validated through two distinct lenses:
- U.S. Supreme Court Analysis: The researchers analyzed 893 cases from the "Second Rehnquist Court" (1994–2004). This period provided a stable network of nine justices, allowing the court to be treated as a fixed social structure where opinions are formed through collective deliberation.
- Simulations: Synthetic datasets were generated based on 3-node and 9-node networks. These simulations tested whether the chain graph could accurately recover causal effects from data generated by a multi-step temporal process.
Key Insights and Results
The study yielded several significant findings regarding the utility of chain graphs in causal inference:
- Snapshot Utility: The authors demonstrated that if a social process reaches a "collective decision" or equilibrium, a single snapshot of the network state contains sufficient information to estimate causal effects, provided it is modeled as a chain graph.
- Blocking Paths: A technical breakthrough involves the discovery that conditioning on neighbors' outcomes in a chain graph can "block" confounding paths from the past, capturing the essential conditional independencies required for valid inference.
- Empirical Success: In the Supreme Court analysis, the model identified that cases involving "Judicial Power" increased the probability of a unanimous conservative decision by 13% compared to other issue areas.
- Simulation Accuracy: The chain graph model estimated potential outcomes with remarkably low bias (average absolute bias between 0.002 and 0.055), even when the underlying "ground truth" was a complex temporal DAG.
Limitations and Future Work
While promising, the chain graph approach has specific requirements:
- Slow Evolution: The approximation is most effective when the social behavior evolves slowly toward equilibrium. Rapid or erratic changes may not be captured accurately.
- Latent Variables: The current framework assumes correlations arise from direct social ties. It does not yet account for latent variable dependence, where shared unobserved environments cause similar behaviors.
- Network Structure: The model assumes the network structure is known. Future research aims to integrate network structure learning directly into the causal framework.
By providing a tractable way to handle interference and contagion, this work opens new doors for causal analysis in the increasingly connected world of social networks.
Summary updated