CMStatistics 2017: Start Registration
View Submission - CMStatistics
B0383
Title: Modeling non-stationarity via multi-resolution basis functions and mixture priors Authors:  Veronica Berrocal - University of Michigan (United States) [presenting]
Abstract: A typical challenge in air pollution population epidemiological studies is the lack of information on ambient exposure for most subjects. To circumvent this problem, and derive point-level estimates of air pollution, several methods have been proposed, including spatial statistical models that rely on the assumption of stationarity. This assumption might not be appropriate for PM2.5, a mixture of air pollutants that include both long-range contaminants and pollutants from more localized sources. To address this issue, building upon the M-RA model introduced recently, we express the spatial field as a linear combination of multi-resolution basis functions, and we provide the basis function weights with resolution-specific mixture priors. Simulation experiments demonstrate the ability of our model to detect regions of non-stationarity. Additionally, an application to daily average PM2.5 concentration indicates that: (i) the pattern of the spatial dependence of PM2.5 is non-homogeneous and (ii) our model outperforms ordinary kriging and is comparable to a previous non-stationary model in an out-of-sample prediction setting.