Identification and Estimation of Treatment and Interference Effects in Observational Studies on Networks

Laura Forastiere, Edoardo M. Airoldi, Fabrizia Mealli

Abstract

Causal inference on a population of units connected through a network often presents technical challenges, including how to account for interference. In the presence of interference, potential outcomes of a unit depend on their treatment as well as on the treatments of other units, such as their neighbors in the network. In observational studies, a further complication is that the typical unconfoundedness assumption must be extended—say, to include the treatment of neighbors, and individual and neighborhood covariates—to guarantee identification and valid inference. This paper proposes new estimands that define treatment and interference effects and derives analytical expressions for the bias of a naive estimator that wrongly assumes away interference. We propose an extended unconfoundedness assumption that accounts for interference and develop new covariate-adjustment methods based on a generalized propensity score (GPS). We carry out simulations, calibrated using friendship networks from a nationally representative longitudinal study of adolescents (Add Health), to explore finite-sample performance in realistic settings.

Introduction

In standard causal inference, researchers typically rely on the Stable Unit Treatment Value Assumption (SUTVA), which posits that an individual's outcome is affected only by their own treatment status. However, in social networks, infectious disease studies, or marketing, this assumption is frequently violated. If a student’s friend receives a flu shot, that student’s own risk of infection decreases regardless of their own vaccination status. This phenomenon is known as interference or spillover.

In observational studies, where treatment is not randomly assigned, interference introduces significant complexity. Social networks often exhibit homophily, where individuals form connections with others who share similar characteristics. These shared traits can influence both the likelihood of receiving a treatment and the final outcome, creating a complex web of confounding. If researchers ignore these network dependencies, their estimates of a treatment's effectiveness can be significantly biased. This paper provides a rigorous framework to define, identify, and estimate these effects when experimental control is unavailable.

The SUTNVA Framework and Estimands

The authors propose a relaxed version of SUTVA termed SUTNVA (Stable Unit Treatment on Neighborhood Value Assumption). This framework acknowledges that interference exists but assumes it is local—meaning an individual’s outcome depends on their own treatment and a summary of their immediate neighbors' treatments (e.g., the proportion of friends treated).

Under this framework, the paper defines two primary estimands:

  1. Overall Main Effect: The effect of an individual's own treatment on their outcome.
  2. Overall Spillover Effect: The effect of the neighbors' treatment status on an individual's outcome.

Unlike previous literature focusing on hypothetical population-wide interventions, these estimands are designed to describe effects within the observed distribution of the network, making them highly applicable to real-world observational data.

Methodology: The Joint Propensity Score

To address confounding, the authors introduce the Joint Propensity Score (GPS). While a standard propensity score calculates the probability of an individual receiving treatment, the GPS calculates the joint probability of an individual receiving a specific treatment and being exposed to a specific level of treatment among their neighbors, given both individual and neighborhood-level covariates.

The estimation strategy follows a two-step semi-parametric approach:

  • Subclassification: Units are grouped based on their individual propensity scores.
  • Dose-Response Adjustment: Within these groups, a model-based approach adjusts for the "dose" of treatment received from the neighborhood.

This method emphasizes that adjusting for individual characteristics is insufficient; researchers must also account for neighborhood covariates (e.g., the demographics of a student's peer group) and network structure (e.g., the number of connections or "degree") to achieve unbiased results.

Empirical Evaluation

The authors evaluated their methods using a simulation study calibrated with real-world data from the Add Health dataset (National Longitudinal Study of Adolescent Health). The study utilized friendship networks from 29 schools, encompassing 16,410 students. By using real student data such as race and grade level, the simulations reflected realistic social patterns like homophily.

The researchers tested four scenarios of increasing complexity, ranging from independent treatment assignments to cases where an individual's treatment was highly correlated with their neighbors' treatments.

Key Findings

  • Bias Mechanics: The paper derives explicit formulas showing that a "naive" estimator (one that ignores the network) is not necessarily biased by interference alone. Rather, bias arises from the interaction between the level of interference and the correlation between individual and neighborhood treatments.
  • GPS Performance: The proposed GPS-based estimator remained essentially unbiased across all four simulation scenarios. Even in the most challenging cases where treatments were highly correlated across the network, the GPS method successfully removed the bias that caused standard regression models to fail.
  • Spillover Detection: The method successfully estimated the "dose-response" curve of neighbor treatments, demonstrating how increasing the proportion of treated neighbors (e.g., those with health insurance) progressively improves an individual's outcomes.

Limitations and Future Directions

While the framework provides a significant advancement, it rests on the assumption of a known and fixed network structure. In many real-world applications, network data is noisy or dynamic. Furthermore, the SUTNVA assumption limits interference to immediate neighbors, potentially overlooking long-range "friend-of-a-friend" effects.

Finally, because units in a network are inherently connected, they do not satisfy the "independent and identically distributed" (i.i.d.) assumption. This complicates the calculation of standard errors and p-values. While the authors suggest bootstrapping, developing robust variance estimators for network data remains an active area of research.

Summary updated

Abstract

<b><i>Abstract–</i>Causal inference on a population of units connected through a network often presents technical challenges, including how to account for interference. In the presence of interference, for instance, potential outcomes of a unit depend on their treatment as well as on the treatments of other units, such as their neighbors in the network. In observational studies, a further complication is that the typical unconfoundedness assumption must be extended—say, to include the treatment of neighbors, and individual and neighborhood covariates—to guarantee identification and valid inference. Here, we propose new estimands that define treatment and interference effects. We then derive analytical expressions for the bias of a naive estimator that wrongly assumes away interference. The bias depends on the level of interference but also on the degree of association between individual and neighborhood treatments. We propose an extended unconfoundedness assumption that accounts for interference, and we develop new covariate-adjustment methods that lead to valid estimates of treatment and interference effects in observational studies on networks. Estimation is based on a generalized propensity score that balances individual and neighborhood covariates across units under different levels of individual treatment and of exposure to neighbors’ treatment. We carry out simulations, calibrated using friendship networks and covariates in a nationally representative longitudinal study of adolescents in grades 7–12 in the United States, to explore finite-sample performance in different realistic settings. Supplementary materials for this article are available online.</b>
BibTeX
@article{forastiere2020identification,
  author = {Laura Forastiere and Edoardo M. Airoldi and Fabrizia Mealli},
  title = {{Identification and Estimation of Treatment and Interference Effects in Observational Studies on Networks}},
  journal = {Journal of the American Statistical Association},
  year = {2020},
  doi = {10.1080/01621459.2020.1768100},
}
Figure