Methodological Issues and Statistical Analysis in Land Use Regression Modeling

Abstract

S-30B1-4 Background/Aims: Land use regression (LUR) models are increasingly used to provide estimates of spatial variation of outdoor air pollution. LUR models for particulate matter typically are based on monitoring that is not performed simultaneously at all sites, requiring adjustments for temporal variation at references sites using different methods. Traditional LUR models provide relatively crude estimates of traffic impacts, especially within urban areas. Methods: We will illustrate the improvement of models by characterizing street configuration and height above ground using GIS databases. In the TRAPCA study including 40 sites spread over the Netherlands, we observed that inclusion of street canyon and detailed traffic count data, improved the R2 of prediction of outdoor soot concentrations from 0.81 to 0.94. Results: We will also discuss the development LUR models (typically based on linear regression) using alternate statistical methods such as universal kriging in cases with significant spatial autocorrelation in the pollution data. In a recent EU-wide LUR model, universal kriging predicted concentrations substantially better than simple linear regression (eg, for urban NO2 the R2 was 0.51 for universal kriging vs. 0.33 for linear regression). In a recent study on LUR, we further illustrate co-kriging as a method that makes use of (correlated) patterns in different pollutants to improve models for the least intensive measured pollutant. Conclusion: Significant improvements of LUR models are possible using data that can be obtained with some effort. In large study areas, kriging methods may outperform standard linear regression methods.