Do Causal Concentration-Response Functions Exist?

Concentration-response (C-R) functions relating concentrations of pollutants in ambient air to mortality risks or other adverse health effects provide the basis for many public health risk assessments, benefits estimates for clean air regulations, and recommendations for revisions to existing air quality standards (Schwartz et al. 2002). The assumption that C-R functions relating levels of exposure and levels of response estimated from historical data usefully predict how future changes in concentrations would change risks has seldom been carefully tested. This chapter critically reviews literature on C-R functions for fine particulate matter (PM2.5) and mortality risks, building on the ideas and methods from Chaps. 2 , 7 and 8 (especially that regression coefficients do not necessarily have valid causal interpretations) and Chaps. 9 , 10 , 11 , 12 , and 13 (especially that Bayesian networks and partial dependence plots can provide useful additional information about dependence relationships that help to identify, constrain, and quantify potential direct and total causal relationships, modelled via invariant causal conditional probability tables). We find that most of the historically influentialpapers in the literature on PM2.5 and mortality risks only describe historical statistical associations rather than valid causal models for predicting effects of interventions that change concentrations. The few papers that explicitly attempt to model causality rely on unverified statistical modeling assumptions, casting doubt on their predictions about effects of interventions. Modern causal inference algorithms for observational data (Chap. 9 ) have been little used in C-R modeling to date. Applying these methods to publicly available data from Boston and the South Coast Air Quality Management District around Los Angeles shows that C-R functions estimated for one do not hold for the other. Changes in month-specific PM2.5 concentrations from one year to the next do not help to predict corresponding changes in average elderly mortality rates in either location. Thus, the assumption that estimated statistical C-R relations predict effects of pollution-reducing interventions may not be true. Better causal modeling methods are needed to better predict how reducing air pollution would affect public health.