Analyzing the effects of air pollution and mortality by generalized additive models with robust principal components

Abstract

In this paper, we propose generalized additive models (GAMs) based on the robust principal component analysis (PCA) methods, to quantify the association between daily mortality and air pollutant concentrations, especially PM10, CO, NO2, SO2 and O3, for confounding effects of long-term time trend, seasonality, weekday, and meteorological factors. The two PCA methods that will be applied into the GAM are: one is classic PCA (CPCA) and the other is robust PCA (RPCA) with minimum covariance determinant, called CPCA–GAM and RPCA–GAM, respectively. Comparing the analyses between GAM, CPCA–GAM, and RPCA–GAM, we can reach to the conclusions as follows: (1) results from CPCA–GAM and RPCA–GAM are consistent with each other; (2) RPCA is much more effective tool to detect outliners than CPCA; and (3) because PCA eliminates the collinearity between covariates, the coefficients of air pollutants have shown to be more significant than GAM without PCA.