A data fusion approach for spatial analysis of speciated PM2.5 across time

PM2.5 exposure is linked to a number of adverse health effects such as lung cancer and cardiovascular disease. However, PM2.5 is a complex mixture of different species whose composition varies substantially in both space and time. An open question is how these constituent species contribute to the overall negative health outcomes seen from PM2.5 exposure. To this end, the Environmental Protection Agency as well as other federal, state, and local organization monitor total PM2.5 along with its primary species on a national scale. From an epidemiological perspective, there is a need to develop effective methods that will allow for the spatially and temporally sparse observations to be used to predict exposures for locations across the entire United States. Toward this objective, we have collected data from three separate monitoring station networks as well as output from a deterministic atmospheric computer model. We introduce a novel multi-level speciated PM2.5 model, which captures the following features: (1) it fuses data from three monitoring networks; (2) it simultaneously models each of the five primary components of PM2.5 from each network along with the computer model output; (3) it introduces species and network level measurement error models as well as total PM2.5 measurement error models, all varying around the respective latent true levels; (4) it incorporates an unobserved "other" species component as well as a sum constraint such that the total is physically consistent (i.e., total must be equal to the sum of the primary species and "other"), which is not always the case with the observed data.

Duke Scholars

Author Colin Rundel Statistical Science

Author Alan E. Gelfand Statistical Science