A tensor product quasi-Poisson model for estimating health effects of multiple ambient pollutants on mortality

Abstract

Background: People are exposed to mixtures of highly correlated gaseous, liquid and solid pollutants. However, in previous studies, the assessment of air pollution effects was mainly based on single-pollutant models or was simultaneously included as multiple pollutants in a model. It is essential to develop appropriate methods to accurately estimate the health effects of multiple pollutants in the presence of a high correlation between pollutants. Methods: The flexible tensor product smooths of multiple pollutants was applied for the first time in a quasi-Poisson model to estimate the health effects of SO 2 , NO 2 and PM 10 on daily all-cause deaths during 2005-2012 in Guangzhou, China. The results were compared with those from three other conventional models, including the single-pollutant model and the three-pollutant model with and without first-order interactions. Results: The tensor product model revealed a complex interaction among three pollutants and significant combined effects of PM 10 , NO 2 and SO 2 , which revealed a 2.53% (95%CI: 1.03-4.01%) increase in mortality associated with an interquartile-range (IQR) increase in the concentrations of all three pollutants. The combined effect estimated by the single-pollutant model was 5.63% (95% CI: 3.96-7.34%). Although the conventional three-pollutant models produced combined effect estimates (2.20, 95%CI, 1.18-3.23%; 2.78, 95%CI: 1.35-4.23%) similar to those of the tensor product model, they distorted the estimates and inflated the variances of the estimates when attributing the combined health effects to individual pollutants. Conclusions: The single-pollutant model or conventional multi-pollutant model may yield misleading results in the presence of collinearity. The tensor product quasi-Poisson regression provides a novel approach to the assessment of the health impacts of multiple pollutants by flexibly fitting the interaction effects and avoiding the collinearity problem.