Characterizing direct and indirect causal effects when outcomes are dependent due to treatment spillover and outcome spillover

2025-04-08

AI Summary

This paper addresses the challenge of disentangling direct and indirect causal effects in networked populations, where treatment spillovers and outcome spillovers are confounded. We propose the first identifiable and estimable causal framework for this setting. Methodologically, we introduce the first rigorous decomposition of the total treatment effect into three orthogonal components: the individual treatment effect, the treatment spillover effect, and the outcome spillover effect. Leveraging a low-rank random interference graph model, we integrate causal graphical models with asymptotic estimation theory to construct a unified estimator applicable to both sparse and dense networks. Theoretically, we establish the first convergence rate guarantees for estimators of direct and indirect effects under outcome-dependent interference and simultaneous presence of both spillover types. Our work substantially advances network causal inference in terms of interpretability, statistical rigor, and scalability.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Social Networks and Social Media: Computational social science
• Economics, Online Markets and Human Computation: Incentives in network design for Web infrastructures and ecosystems

Abstract

We provide novel insight into causal inference when both treatment spillover and outcome spillover occur in connected populations, by taking advantage of recent advances in statistical network analysis. Scenarios with treatment spillover and outcome spillover are challenging, because both forms of spillover affect outcomes and therefore treatment spillover and outcome spillover are intertwined, and outcomes are dependent conditional on treatments by virtue of outcome spillover. As a result, the direct and indirect causal effects arising from spillover have remained black boxes: While the direct and indirect causal effects can be identified, it is unknown how these causal effects explicitly depend on the effects of treatment, treatment spillover, and outcome spillover. We make three contributions, facilitated by low-rank random interference graphs. First, we provide novel insight into direct and indirect causal effects by disentangling the contributions of treatment, treatment spillover, and outcome spillover. Second, we provide scalable estimators of direct and indirect causal effects. Third, we establish rates of convergence for estimators of direct and indirect causal effects. These are the first convergence rates in scenarios in which treatment spillover and outcome spillover are intertwined and outcomes are dependent conditional on treatments, and the interference graph is sparse or dense.

Problem

• Disentangling direct and indirect causal effects from treatment and outcome spillover
• Developing scalable estimators for causal effects in interconnected populations
• Establishing convergence rates for estimators under intertwined spillover effects

Innovation

• Disentangling treatment and spillover effects via network analysis
• Scalable estimators for direct and indirect causal effects
• Convergence rates for sparse or dense interference graphs