Causal Inference under Interference with Learned Exposure Mappings

Research investigates causal inference under interference where exposure mappings are learned rather than directly observed, focusing on environmental applications where transport processes induce these mappings. The study highlights how uncertainty in learned transport processes affects exposure mappings and subsequent spillover inference. It compares mechanistic transport models with operator-learning approaches (PDE, PINO, FNO, GeoPT) using both simulated and real-world environmental data.

Why it matters

This research underscores a critical challenge in causal inference for complex systems where observational data is used to derive relationships, particularly when unobserved processes influence key variables. Understanding how different modeling approaches impact the reliability of causal effect estimates is vital for robust decision-making and policy formulation in areas like environmental regulation, resource management, and public health interventions.

Key insights

• Exposure mappings in causal spillover analyses are often assumed known but frequently need to be learned from data, particularly in environmental contexts.
• Uncertainty in learned transport processes directly propagates into the derived exposure mappings and impacts downstream spillover inference.
• A comparison of mechanistic transport models with modern operator-learning approaches (PDE, PINO, FNO, GeoPT) was conducted.
• Despite similar pollution prediction accuracy across all four transport models in simulations, estimated spillover effects showed significant variance, ranging from 1.78 to 2.27.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.19224