[Paper Review] Identification and estimation of treatment and interference effects in observational studies on networks

Laura Forastiere, Edoardo M. Airoldi+1 more|arXiv (Cornell University)|Sep 20, 2016

Advanced Causal Inference Techniques4 citations

TL;DR

This paper proposes a new framework for identifying and estimating treatment and interference effects in observational network studies, where units' outcomes depend on both their own treatment and their neighbors' treatments. It introduces an extended unconfoundedness assumption and a generalized propensity score method to adjust for individual and neighborhood-level confounding, enabling valid causal inference under interference.

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 local interference, for instance, potential outcomes of a unit depend on its treatment as well as on the treatments of other local units, such as its neighbors according to 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 indi- vidual 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.

Motivation & Objective

• Address the challenge of interference in causal inference on networks, where a unit's outcome depends on its own and its neighbors' treatments.
• Overcome the limitations of standard unconfoundedness assumptions in observational studies by extending them to include neighborhood treatments.
• Develop a new estimand that separates treatment effects from interference effects in networked units.
• Propose a generalized propensity score method that balances both individual and neighborhood covariates across different treatment exposure levels.
• Enable valid estimation of causal effects in realistic network settings using simulation calibrated to real adolescent friendship networks.

Proposed method

• Introduce a new estimand that defines treatment and interference effects by distinguishing individual treatment from exposure to neighbors' treatments.
• Propose an extended unconfoundedness assumption that includes individual treatments, neighborhood treatments, and individual and neighborhood covariates.
• Derive analytical expressions for the bias of naive estimators that ignore interference, showing dependence on interference levels and treatment association.
• Develop a generalized propensity score that balances individual and neighborhood covariates across different levels of individual treatment and exposure to neighbors' treatments.
• Use inverse probability weighting based on the generalized propensity score to estimate treatment and interference effects.
• Calibrate simulations using empirical data from a nationally representative U.S. longitudinal study of adolescents in grades 7–12, with real friendship networks and covariates.

Experimental results

Research questions

• RQ1 How can treatment and interference effects be formally defined and identified in observational network studies?
• RQ2 What is the bias of standard estimators that ignore interference in networked settings?
• RQ3 How can unconfoundedness be extended to account for interference in observational network data?
• RQ4 What covariate adjustment methods are effective for estimating treatment and interference effects in the presence of network dependence?
• RQ5 How do the proposed methods perform in finite samples under realistic network and covariate structures?

Key findings

• The bias of a naive estimator that ignores interference depends on both the level of interference and the degree of association between individual and neighborhood treatments.
• The proposed extended unconfoundedness assumption enables identification of treatment and interference effects under network interference.
• The generalized propensity score method successfully balances individual and neighborhood covariates across treatment exposure levels, improving estimation accuracy.
• Simulations calibrated to real adolescent friendship networks show that the proposed method reduces bias and improves coverage in finite samples.
• The method is robust to realistic network structures and covariate dependencies observed in longitudinal survey data.
• The framework allows for valid inference on both direct treatment effects and interference effects in networked observational studies.