Identifying low dose measurement error corrected effects of multiple pollutants using causal modeling

点击翻译按钮获取中文摘要

The Global Burden of Disease estimates that ambient air pollution is responsible for over 4 million deaths per year, yet regulators in the US, EU, India, and China have been reluctant to tighten standards, which can be costly. Those costs and the observational nature of the epidemiology studies suggesting a tightening of existing standards would be protective of the public’s health is a major reason for this reluctance. To date, separate standards have not been set for particle components, and health impact assessments rarely examine environmental equity because of the paucity of subgroup-specific concentration-response functions. Further, studies on the effects of temperature on mortality and morbidity have focused on risk associated with short-term exposure, and not longer-term effects which may be larger. We propose to address these gaps by using national data (US Medicare and Medicaid data, and all age Death Certificate data from multiple states geocoded to a census block group) on mortality and hospital admissions; to use causal modeling techniques robust to omitted confounders by design; to extend methods for environmental mixtures to large data settings and use them to assess nonlinear and interactive effects of exposures; to use state of the art models estimating daily air pollution and temperature exposure for the contiguous US on a 1km grid for 18 years; to use state of the art methods to estimate exposure error in the contiguous US, to use restriction and spline methods to address low dose effects, and to develop and use state of the art measurement error correction methods to account for exposure error when estimating these risks. Specifically, we will use quasi-experimental designs (difference in differences and self-controlled) that control for many unmeasured confounders, either by stratifying on subject (controlling for individual level fixed or slowly varying covariates) or by stratifying on neighborhood (controlling for fixed and slowly varying neighborhood level covariates), while continuing to control for measured covariates. For acute effects of exposures, we will use instrumental variables to adjust for unmeasured confounding. We will access large, ready-to-use datasets we have compiled, including national Medicare and Medicaid mortality and admissions, and state-level geocoded death certificate data. We will use highly accurate national models we have developed for daily pollution on a 1km grid, and increase resolution to 500 m. We will use a new mixture model, fast Bayesian Kernel Machine Regression (BKMR), to address pollution and temperature mixtures, identify interactions and nonlinearities, and identify which exposures are most important (including which particle components) for a given health endpoint. We will use state of the art measurement error correction approaches (SIMEX) to identify biases in the concentration-response relationship due to exposure error. We will supplement the BKMR approach with analyses restricted to observations below current standards, and spline methods with propensity scores to determine whether causal effects continue below current standards.