Abstract
Classical causal inference assumes treatments meant for a given unit do not have an effect on other units. This assumption is violated in interference problems, where new types of spillover causal effects arise, and causal inference becomes much more difficult. In addition, interference introduces a unique complication where variables may transmit treatment influences to each other, which is a relationship that has some features of a causal one, but is symmetric. In this paper, we develop a new approach to decomposing the spillover effect into unit-specific components that extends the DAG based treatment decomposition approach to mediation of Robins and Richardson to causal models that admit stable symmetric relationships among variables in a network. We discuss two interpretations of such models: a network structural model interpretation, and an interpretation based on equilibrium of structural equation models discussed in (Lauritzen and Richardson, 2002). We show that both interpretations yield identical identification theory, and give conditions for components of the spillover effect to be identified. We discuss statistical inference for identified components of the spillover effect, including a maximum likelihood estimator, and a doubly robust estimator for the special case of two interacting outcomes. We verify consistency and robustness of our estimators via a simulation study, and illustrate our method by assessing the causal effect of education attainment on depressive symptoms using the data on households from the Wisconsin Longitudinal Study.
Introduction
In classical causal inference, the Stable Unit Treatment Value Assumption (SUTVA) is a cornerstone, assuming that the treatment of one individual does not influence the outcome of another. In reality, social and biological systems are rife with spillover effects. For instance, a vaccination protects not just the recipient but also their contacts, and a spouse's educational success can positively impact their partner's mental health.
Existing methodologies for studying these effects often rely on mediation analysis, which necessitates a strict directed causal ordering. However, many interactions—particularly within households or social circles—are inherently symmetric and simultaneous. Traditional Directed Acyclic Graphs (DAGs) struggle to represent these "give-and-take" relationships where no clear "first mover" exists. This paper addresses this gap by introducing a framework for modeling stable, symmetric relationships using Chain Graphs.
Methods: From DAGs to Chain Graphs
The authors propose the use of Chain Graphs (CGs) to accommodate both directed causal edges and undirected edges representing symmetric interactions. A key innovation is Symmetric Treatment Decomposition, where a single treatment is split into hypothetical components (e.g., distinguishing between a vaccine's direct protection and its contagion-reducing properties).
As illustrated in Figure 1, the methodology evolves from standard DAGs (a) to models that decompose treatments (b, c) and finally to Chain Graphs (d, e, f) where outcomes $Y_1$ and $Y_2$ interact symmetrically. This allows every outcome in a block to act as both an outcome and a mediator simultaneously.
The paper introduces the Symmetric Mediation Formula, a new identifying functional that generalizes standard mediation to undirected network blocks. The authors justify this approach through two philosophical lenses: a system reaching a stable equilibrium and a "snapshot" of a complex network structural model. For estimation, they provide a Maximum Likelihood Estimator (MLE) and a Doubly Robust (DR) estimator, the latter of which remains consistent even if one of the nuisance models (treatment or outcome) is misspecified.
Empirical Evaluation
The framework was validated using both synthetic and real-world data:
- Simulation Studies: Using conditional Gaussian mixed interaction models, the authors demonstrated that the DR estimator successfully recovers true causal effects and maintains robustness against model misspecification.
- Wisconsin Longitudinal Study (WLS): Analyzing 1,768 spousal dyads, the authors investigated how a spouse’s educational attainment affects the other’s depressive symptoms. They found a significant direct spillover effect (size $\approx$ 0.048–0.056) from a spouse's four-year degree. Interestingly, the indirect effect—the portion of the spillover traveling through the first spouse's own mental health—was negligible. This suggests that the mental health benefits of a partner's education stem from shared resources or status rather than simple emotional contagion.
Discussion and Key Insights
A significant advantage of this symmetric framework is its falsifiability. Unlike many mediation models that must be assumed correct, the symmetric structure imposes testable restrictions on the observed data distribution.
While the method provides a robust tool for dyadic interactions, limitations remain. The "positivity assumption" requires that all combinations of treatment components be theoretically observable, which may be challenging in some contexts. Furthermore, while the current work focuses on "partial interference" (independent blocks like families), future research is needed to scale these methods to "full interference" scenarios involving large, interconnected populations.
Summary updated