Identifying Treatment and Spillover Effects Using Exposure Contrasts

2024-03-13

AI Summary

Exposure contrasts—commonly used in causal inference to estimate treatment and spillover effects—may yield estimates with signs opposite to the true unit-level causal effects, particularly under interference. Method: We systematically characterize the causal interpretability boundary of exposure contrasts within a nonparametric framework, formalizing exposure mappings via causal graphs. We derive verifiable necessary and sufficient conditions for sign consistency, ensuring that exposure contrasts robustly reflect the direction of causal effects—even under arbitrary assignment mechanisms (e.g., cluster-randomized trials, network experiments, or observational data with peer selection) and arbitrary interference structures. Contribution/Results: This work establishes, for the first time, a function-form–free theory guaranteeing sign preservation of exposure contrasts. It provides a rigorous foundation for reliable identification of spillover effects in social network analysis and policy evaluation, bridging theoretical causality and empirical practice without restrictive modeling assumptions.

Technology Category

• Reasoning under Uncertainty: Causality
• Machine Learning: Causal Learning
• Application Domains: Humanities & Computational Social Science

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
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

To report spillover effects, a common practice is to regress outcomes on statistics capturing treatment variation among neighboring units. This paper studies the causal interpretation of nonparametric analogs of these estimands, which we refer to as exposure contrasts. We demonstrate that their signs can be inconsistent with those of the unit-level effects of interest even under unconfounded assignment. We then provide interpretable restrictions under which exposure contrasts are sign preserving and therefore have causal interpretations. We discuss the implications of our results for cluster-randomized trials, network experiments, and observational settings with peer effects in selection into treatment.

Problem

• Identifies treatment and spillover effects using exposure contrasts
• Analyzes conditions to prevent sign reversals in effect estimates
• Applies to randomized trials, network experiments, and observational studies

Innovation

• Exposure contrasts estimate spillover effects nonparametrically
• Conditions prevent sign reversals in unit-level effects
• Applicable to cluster trials and network experiments