Development and evaluation of land use regression models for black carbon based on bicycle and pedestrian measurements in the urban environment

Land use regression (LUR) modelling is increasingly used in epidemiological studies to predict air pollution exposure. The use of stationary measurements at a limited number of locations to build a LUR model, however, can lead to an overestimation of its predictive abilities. We use opportunistic mobile monitoring to gather data at a high spatial resolution to build LUR models to predict annual average concentrations of black carbon (BC). The models explain a significant part of the variance in BC concentrations. However, the overall predictive performance remains low, due to input uncertainty and lack of predictive variables that can properly capture the complex characteristics of local concentrations. We stress the importance of using an appropriate cross-validation scheme to estimate the predictive performance of the model. By using independent data for the validation and excluding those data also during variable selection in the model building procedure, overly optimistic performance estimates are avoided.