Spatio-Temporal Analysis of the Longitudinal PM Data with Missing Values

• Activity Number: 208
• Type: Contributed
• Date/Time: Tuesday, August 13, 2002 : 10:30 AM to 12:30 PM
• Sponsor: Section on Statistics & the Environment*
• Author(s): Stanislav Kolenikov*+ and Richard Smith
• Affiliation(s): University of North Carolina and University of North Carolina
• Keywords: particular matter ; kriging ; EM algorithm ; clustering ; variogram ; semiparametric

Abstract

This paper analyses the data on the particular matter of size of 2.5 microns or less (PM2.5) for the states of North Carolina, South Carolina, and Georgia. The spatio-temporal model features additive semiparametric temporal and spatial trends. Reduction in the number of nodes for the thin-plate component is achieved through the use of clustering routines that combine adjacent observations sites. The residual spatial covariance is modelled by an exponential-power family semivariogram with nugget effect. The parameters of the model are estimated by the generalized EM algorithm to account for the missing data. The relation between predictions produced by the EM algorithm and the universal kriging is discussed.

The resulting PM maps show that concentrations are lower off the coast and in sparsely inhabited Appalachians, although the standard errors of predictions for those regions rise sharply, too. The concentrations of the PM were found to have stayed around 10-15 mkg/m^3 for most of the year, and increase to 25-35 mkg/m^3 in late summer. It was also found that most of Georgia violates the federal regulations of 15 mkg/m^3 most of the time.