You are here:
METHODOLOGICAL ISSUES IN THE USE OF GENERALIZED ADDITIVE MODELS FOR THE ANALYSIS OF PARTICULATE MATTER; CONFERENCE PROCEEDINGS FOR 9TH INT'L. INHALATION SYMPOSIUM ON EFFECTS OF AIR CONTAMINANTS ON THE RESPIRATORY TRACT - INTERPRETATIONS FROM MOLECULES TO META ANALYSIS
Citation:
Neas, L M. METHODOLOGICAL ISSUES IN THE USE OF GENERALIZED ADDITIVE MODELS FOR THE ANALYSIS OF PARTICULATE MATTER; CONFERENCE PROCEEDINGS FOR 9TH INT'L. INHALATION SYMPOSIUM ON EFFECTS OF AIR CONTAMINANTS ON THE RESPIRATORY TRACT - INTERPRETATIONS FROM MOLECULES TO META ANALYSIS. Presented at 9th International Inhalation Symposium on Effects of Air Contaminants on the Respiratory Tract - Interpretations from Molecules to Meta Analysis, Hanover, GERMANY, June 09 - 14, 2003.
Impact/Purpose:
To identify statistical and computational issues concerning the fitting of Generalized Additive Models in epidemiological analyses of particulate matter
Description:
Open cohort ("time-series") studies of the adverse health effects of short-term exposures to ambient particulate matter and gaseous co-pollutants have been essential in the standard setting process. Last year, a number of serious issues were raised concerning the fitting of Generalized Additive Models (GAM) in these epidemiological analyses. After a review of the various approaches to modeling nonlinear relationships and fitting procedures for the estimation of model parameters, several statistical and computational issues were identified including the adequacy of default convergence criteria and the problems introduced by the concurvity of multiple nonlinear predictors. In response to these issues, the epidemiology research community quickly proceeded to the identification, assessment, and correction of these issues in a series of meetings and workshop: the EPA Clean Air Science Advisory Committee meeting, the Workshop on GAM Issues sponsored by the EPA National Center for Environmental Assessment, and the Health Effects Institute's peer-review panel. Finally, several new alternative techniques have recently been proposed to model nonlinear relationships while avoiding the issues of fitting some classes of Generalized Additive Models. On the whole, this troubling episode has been successfully resolved without compromising the validity of the principal findings of previous epidemiologic research.