Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Causal inference with spatio-temporal data using process-informed stochastic interventions
Authors: Nathan Wikle, Corwin Zigler,
Presenting Author: Nathan Wikle*
Causal inference with spatio-temporal data is often challenging due to the presence of interference: outcomes for observational units depend on some combination of local and non-local treatment. This is especially relevant when an individual’s treatment exposure is determined, in part, by some underlying physical process. For example, air pollution exposure is a function of both the locations of emissions sources (e.g., power plants, roads, etc.) as well as the physical-chemical process governing pollution transport. In this talk, we propose causal estimands that are defined with respect to stochastic interventions informed by the physical process; importantly, these estimands can accommodate both interference and positivity violations. In particular, we estimate the expected change in the number of outcome events in a specific area under different stochastic exposure distributions, where the stochastic distributions correspond to counterfactual exposure levels simulated from a dynamic spatio-temporal model of the physical process. We develop an augmented inverse probability of treatment weighting estimator for spatio-temporal data with a desirable double robustness property, and propose methods to assess its sensitivity to unmeasured confounding, positivity violations, and uncertainty in the physical process model. Finally, we use the proposed methods to estimate the expected change in pediatric asthma rates in Texas under two competing air pollution emissions scenarios.