Optimal Penalty Parameter Selection to Minimize the Impact of Exposure Measurement Error in 2-Stage Air Pollution Epidemiology Analyses

• Activity Number: 34
• Type: Contributed
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Silas Bergen*+ and Adam Szpiro
• Companies: University of Washington and University of Washington
• Keywords: Penalized regression models ; Penalty parameter selection ; Measurement error ; Air pollution epidemiology

Abstract

Air pollution epidemiology studies often implement a 2-stage process wherein a first-stage exposure model is built to derive exposure predictions at participant locations which are then used in a second-stage health model. Incorporating penalized regression in the exposure model provides flexibility in modeling the spatial structure of the unknown pollution surface and serves as a mechanism to avoid overfitting the available data. This context motivates selecting the penalty parameter to optimize inference in the health model rather than optimizing prediction accuracy in the exposure model. From a measurement error perspective, a larger penalty induces Berkson-like error from missing exposure surface characteristics; a smaller penalty induces classical-like error resulting from more variable exposure model coefficients. We present a methodology for quantifying the impact of the two types of measurement error, leading to a new criterion for optimal penalty parameter selection. A key feature of our methodology is that it does not rely on a correctly specified exposure model. We illustrate our methods in an analysis of PM2.5 and systolic blood pressure in the NIEHS Sister Study.