A Novel Principal Component Analysis for Spatially Misaligned Multivariate Air Pollution Data

• Activity Number: 351
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Roman Jandarov*+ and Adam Szpiro
• Companies: University of Washington and University of Washington
• Keywords: Spatial statistics ; Dimension reduction ; Air pollution ; Principal component analysis ; Multi-pollutant exposure ; Partial least squares analysis

Abstract

Analysis of health effects of multi-pollutant, long-term air pollution exposure poses two problems: (i) interpreting the parameters of the health models and (ii) spatial misalignment of monitoring data. We propose a novel approach for dimension reduction in multivariate spatial data that resolves these issues. Our approach seeks to find sparse principal components that explain a large proportion of the variance in the data while also ensuring that mixtures derived from these components are predictable at health outcome locations. Predictions of these lower dimensional components can then be used in a health effect analysis. We note that our approach is preferable to the sequential two-step approach (dimension reduction followed by spatial prediction), which may result in principal components that are difficult to predict at unmeasured locations. We provide an efficient implementation of our method and compare it with alternative dimension reduction approaches. We illustrate the practical utility of our approach by applying it to national multi-pollutant air pollution data from the U.S. Environmental Protection Agency and a health effect analysis in the NIEHS Sister Study cohort.