Accountability Research for Air Quality Regulations Using Principal Stratification
Francesca Dominici, Harvard School of Public Health 
Yun Wang, Harvard School of Public Health 
*Corwin Zigler, Harvard School of Public Health 

Keywords: Causal inference, hierarchical model, spatial statistics, principal stratification, mediation analysis

Air quality regulations are typically enacted with the goal of improving public health, and large-scale regulations that reduce ambient air pollution are often accompanied by improvements in mortality and morbidity. It is rarely clear whether observed improvements in health are due to a regulation or to concurrent trends in other risk factors. Quantifying health effects of actions taken to improve air quality, also known as accountability research, is challenging for large-scale efforts meant to improve air quality and health over a long period of time and over large geographical areas. We propose a potential-outcomes framework for accountability research that uses a Bayesian MCMC computational strategy. The approach adopts ideas rooted in principal stratification to stratify areas based on the estimated causal effect of an air-quality regulation on ambient air pollution and assess the causal effect of the regulation on health outcomes within these strata. As regulations may affect concentrations of many interrelated pollutants, we extend the use of principal stratification to a multi-pollutant approach that accommodates a continuously-scaled multivariate intermediate response vector. Furthermore, we make use of recent advancements in hierarchical modeling for point-referenced spatial data to capitalize on information contained in the geographic locations of the pollution measurement sites. Insofar as these methods rely on unobserved potential outcomes, they rely on assumptions regarding associations not identified based on observed data, indicating the importance of sensitivity analyses. We apply our method to examine whether the 1990 Clean Air Act Amendments causally affected Medicare mortality through affecting ambient concentrations of particulate matter and ozone.