Auto-G-Computation of Causal Effects on a Network.

Area of Science:

• Causal Inference
• Network Analysis
• Statistical Modeling

Background:

• Traditional causal inference methods assume no interference between units and independent group outcomes.
• These assumptions are often violated in complex, interconnected systems like social networks or biological systems.

Purpose of the Study:

• To develop new statistical methods for causal inference that relax traditional assumptions of no interference and group independence.
• To enable causal effect estimation in network settings with arbitrary interference and long-range dependence.

Main Methods:

• Developed new statistical methods for causal inference using a single realization of a network.
• Introduced the auto-g-computation algorithm, a network generalization of g-computation, to infer network causal effects.
• Leveraged a chain graph model assumption for tractability under network consistency and no unobserved confounding.

Main Results:

• The proposed methods allow for arbitrary forms of interference, where a unit's outcome can depend on interventions received by connected units.
• The approach accounts for long-range dependence, where outcomes of connected units may be dependent.
• Inference is made tractable through the chain graph model assumption and the auto-g-computation algorithm.

Conclusions:

• The developed methods provide a robust framework for causal inference in networked data, overcoming limitations of traditional approaches.
• The auto-g-computation algorithm offers a powerful tool for estimating causal effects in complex systems with interference and dependence.
• This work advances the field of causal inference by extending its applicability to network structures.