CMStatistics 2018: Start Registration
View Submission - CFE
A1076
Title: Additive semiparametric framework for land use regression Authors:  Svenia Behm - University of Passau (Germany) [presenting]
Markus Fritsch - University of Passau (Germany)
Harry Haupt - University of Passau (Germany)
Abstract: Existing land use regression (LUR) approaches usually employ parametric assumptions to model the conditional distribution of air pollutant measurements. We propose a flexible data-driven additive semiparametric framework for modeling the annual mean nitrogen dioxide concentration across Germany which rests on the crucial assumption of additivity. Through exploratory analysis, we find considerable spatial variation and nonlinearities in the data and, therefore, model the spatial characteristics via bivariate splines and the structural characteristics via univaritate splines and in linear additive form. Our specification allows us to account for local heterogeneity, potential nonlinearities and spatial anisotropy in a flexible, data-driven way -- while avoiding imposing parametric assumptions a priori (i.e., without looking at the data). In- and out-of-sample metrics support the proposed model. Additive semiparametric models are a promising choice to analyse and predict the conditional distribution of air pollutant concentration. A straightforward extension of our approach is to model several characteristics of the conditional pollutant distribution such as quantiles or expectiles.