Abstract
In interconnected systems, the treatment of one unit often affects the outcomes of others—a phenomenon known as interference or spillover. Estimating causal effects in such settings is challenging due to the high dimensionality of treatment combinations and the presence of confounding and homophily in observational data. This paper introduces a novel covariate-adjustment estimator based on a Bayesian Generalized Propensity Score (GPS). By assuming neighborhood interference and unconfoundedness, the proposed method uses a three-step Bayesian procedure to estimate the Average Dose-Response Function (ADRF). The approach incorporates penalized spline regression for flexibility and community detection to account for network-driven correlations.
1. Introduction
Standard causal inference typically relies on the Stable Unit Treatment Value Assumption (SUTVA), which posits that the potential outcome of any unit is independent of the treatment assignment of other units. However, in social, economic, and biological networks, this assumption is frequently violated. For instance, a vaccination program's success depends not only on individual uptake but also on the immunity of neighbors. Such "spillover effects" complicate the identification of both direct and indirect treatment effects.
Forastiere et al. address these challenges by developing a framework for observational studies on networks. Their approach focuses on estimating the Average Dose-Response Function (ADRF), which maps the expected outcome to different levels of individual and neighborhood treatment exposure.
2. Structural Assumptions
To make estimation tractable in a network of $N$ units, where the number of possible treatment vectors is $2^N$, the authors propose two critical structural assumptions:
- Neighborhood Interference: The outcome of a unit is assumed to depend only on its own treatment and a summary function of the treatments received by its immediate (first-degree) neighbors. This reduces the complexity of the interference from a global to a local scale.
- Unconfoundedness of Individual and Neighborhood Treatment: This assumption posits that all variables influencing both the treatment assignment (at the individual and neighborhood levels) and the outcome are observed. This is essential for addressing homophily—the tendency of similar individuals to form ties—which can otherwise be mistaken for causal spillover.
3. The Bayesian GPS Estimator
The core of the methodology is a three-step Bayesian procedure designed to estimate the GPS and the outcome model while properly propagating uncertainty.
3.1 Step 1: Individual Propensity Score
The first step involves modeling the probability of an individual receiving the treatment given their covariates. This follows standard propensity score modeling but serves as the foundation for the subsequent neighborhood-level analysis.
3.2 Step 2: Neighborhood Generalized Propensity Score
The "neighborhood treatment" is often a continuous or multi-valued "dose" (e.g., the proportion of treated neighbors). The authors calculate a GPS that balances neighborhood-level covariates. This allows for the comparison of units that had a similar probability of being exposed to a specific level of neighborhood treatment, effectively mimicking a randomized experiment within strata of the GPS.
3.3 Step 3: Outcome Modeling and Modularization
The final step models the outcome as a function of individual treatment, neighborhood treatment dose, and the estimated propensity scores. Two key features distinguish this step:
- Penalized Splines: Instead of assuming a linear relationship, the model uses penalized splines to capture non-linear dose-response curves.
- Modular Bayesian Inference: To prevent "model feedback"—where the outcome model might erroneously influence the estimation of the propensity scores—the authors use a modular approach. This ensures that the design stage (propensity modeling) remains distinct from the analysis stage (outcome modeling).
4. Handling Network Correlation
Units within a network are rarely independent. To account for the correlation between interacting units, the framework incorporates community detection algorithms. By identifying clusters or "communities" within the network, the authors include community-level random effects in the outcome model. This accounts for unobserved factors shared within local clusters, improving the precision of the causal estimates.
5. Empirical Evaluation
The authors validated the framework using extensive simulations and real-world data:
- Simulations: Using Stochastic Block Models (SBM) and Latent Cluster Models (LCM), the researchers demonstrated that the spline-based GPS estimator significantly outperforms linear models, especially when the underlying causal mechanisms are non-linear.
- Add Health Dataset: The model was applied to the National Longitudinal Study of Adolescent to Adult Health, a complex friendship network. This application showcased the model's ability to handle the "messy" topologies of real-world social structures while disentangling homophily from true spillover effects.
6. Key Insights and Results
The Bayesian framework provides several advantages over frequentist alternatives. First, it naturally accounts for the uncertainty inherent in the propensity score estimation, leading to more robust credible intervals. Second, the use of GPS effectively balances covariates across different levels of neighborhood exposure, a task that is notoriously difficult in network settings. Finally, the results indicate that accounting for community structure is vital for accurate inference in the presence of network-correlated outcomes.
7. Limitations and Conclusion
While the framework is a significant step forward, it is bounded by the assumption of first-degree interference. In systems where effects ripple through multiple degrees of separation, the neighborhood interference assumption may be too restrictive. Additionally, the reliance on unconfoundedness means that unobserved homophily remains a potential source of bias. Future work may explore scalable approximate inference for massive networks and extensions to higher-order interference.
In conclusion, the Bayesian GPS approach offers a rigorous and flexible toolkit for researchers seeking to understand causal dynamics in an interconnected world, providing a path to move beyond the limitations of SUTVA.
Summary updated