Causal Inference Under Network Interference

TL;DR

This paper reviews recent advances in causal inference in connected populations where treatments can influence others, discussing experimental design, inference targets, and the impact of network structure on causal conclusions.

Contribution

It provides a comprehensive overview of methods and frameworks for causal inference under network interference, highlighting the importance of network structure and contrasting inferential paradigms.

Findings

Expected outcomes depend on network structure.

Network interference affects causal effect estimation.

Different inferential frameworks offer varying generalizability.

Abstract

We review and conceptualize recent advances in causal inference under network interference, drawing on a complex and diverse body of work that ranges from causal inference, statistical network analysis, economics, the health sciences, and the social sciences. Network interference arises in connected populations when the treatment assignments of units affect the outcomes of other units. Examples include economic, financial, and public health interventions with spillover in connected populations, reinforcement learning in connected populations, and advertising on social media. We discuss the design of experiments, targets of causal inference, interpretations and characterizations of causal effects, interference tests, and design- and model-based estimators of causal effects under network interference. We then contrast inferential frameworks based on fixed networks (finite population inference) and random networks (super population inference) and the generalizability afforded by them. We demonstrate that expected outcomes can depend on the network structure (e.g., the absence or presence of superstars and communities) and could be different if another network were observed, highlighting the need to understand how network structure affects causal conclusions. We conclude with a selection of open problems.