Causal Methods to Adjust for Confounding When Air Pollution Exposure Is Measured with Error

• Activity Number: 30 - Missing Data and Measurement Error
• Type: Contributed
• Date/Time: Sunday, July 28, 2019 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Danielle Braun* and Xiao Wu and Marianthi-Anna Kioumourtzoglou and Francesca Dominici
• Companies: Harvard University and Harvard University and Mailman School of Public Health, Columbia University and Harvard T.H. Chan School of Public Health
• Keywords: Causal Inference; Measurement Error; Generalized Propensity Score; Air Pollution

Abstract

Long-term exposure to air pollution has consistently been associated with adverse health outcomes. Most previous studies assume the exposure is error-free, but in reality the exposure is subject to measurement error. Our interest is in estimating the association of long-term air pollution exposure and mortality in US Medicare beneficiaries. For the entire US Medicare population (main study), long-term exposure to fine particles (PM2.5) is determined from a spatiotemporal model that uses different sources as input (meteorology, satellite data, etc.). PM2.5 exposure based on these predictions is inaccurate, but for a subset of zip-codes (validation study) we have actual PM2.5 concentrations measured at monitors (error-free exposure). We begin by describing the complexities of addressing measurement error in this context that arise due to the spatial and temporal nature of the data. Next, using the internal validation study, we propose various approaches to adjust for measurement error, some treating PM2.5 exposures as continuous variables while others categorizing them into ordinal variables. We evaluate the approaches in simulations and apply them to the US Medicare population.