The U.S. Clean Air Act Amendments of 1990 mandated emission reductions during a period ending in 1995, with the objective of reducing concentrations of atmospherically transported pollutants in ambient air.  Assessment requires estimation of trend (defined as percent total change) in sulfur dioxide concentrations during the reference period on a regional scale. This paper proposes a two-stage statistical methodology for estimation of trends and associated standard errors, using data for the period 1989-95 from 35 rural monitoring locations across the Eastern United States.  The first stage provides site-specific models for estimating trend and site-specific standard errors, based on the logarithm of weekly ambient sulfur dioxide concentration, adjusted (by means of  GAM methodology) for seasonal and meteorological effects.  The second stage involves estimation of regional trends based on applying kriging methodology using maximum likelihood estimation to estimated site-specific trends and standard errors and site-to-site covariances.  A Bayesian analysis was conducted to assess and account for error introduced by the specification of covariance parameters.   Results were conformal with reported reductions in sulfur dioxide emissions.

LAWRENCE H. COX
U.S. Environmental Protection Agency
National Exposure Research Laboratory (MD-75)
Research Triangle Park, NC 27711,  USA
Cox.Larry@epamail.epa.gov
 
 

Many of the great advances within the scientific community are the result of deep yet simple concepts.  For instance, Newton's three laws replaced Kepler's rather complicated but useful laws of planetary motion.  The scientific community has always replaced complicated solutions with equally good or better simple solutions as soon as they become available.  In this paper, we present a new calibration method for the analysis of scientific data, called the parallel method, that is simple and often better when compared to standard calibration methods.  We compare the parallel method to standard methods such as classical least-squares and partial least-squares on two scientific data sets and on one computer-generated data set.  This new method shows better or comparable results for these data sets in terms of mean squared error but more importantly this method is much simpler than the popular partial least-squares approach and does not require a user-defined selection of latent variables.

CLIFF H. SPIEGELMAN
Department of Statistics
Texas A&M University
College Station, TX 77843-3143, USA
cliff@stat.tamu.edu