Primary Submission Category: Causal Inference in Networks
Nonparametric Network Causal Inference for Continuous Exposures in Mobile Source Air Pollution
Authors: Salvador Balkus, Nima Hejazi, Rachel Nethery, Scott Delaney,
Presenting Author: Salvador Balkus*
Continuous exposures pose a challenge for traditional causal inference estimators: positivity is frequently violated, and the intervention of “setting every unit’s exposure to exactly X” produces unrealistic counterfactuals. Modified Treatment Policies (MTPs) – interventions that depend on each unit’s naturally observed exposure – resolve these issues. But what if the outcome of each unit depended not just on its own exposure, but also on the exposure of other units in a network? For example, in observational mobile source air pollution studies, the pollution in a given region depends not only on vehicles registered within it but also on vehicles that commute in from other regions. Our work connects recent theory in nonparametric causal inference under general network interference to MTPs. We review the mathematical construction of candidate estimators, compare multiple variance estimation strategies, and introduce an open-source software implementation in Julia. Estimators are evaluated under various simulation settings and applied to a mobile source air pollution case study.