A non-linear and interaction effect analysis of distance and transport accessibility on bicycle use: the example of the University of Lyon (France)

This study explores the individual and spatial level determinants of the adoption cycling as a commuting mode by university staff members using data from Lyon, France (the MobiCampus-UdL survey). The empirical approach of the study is centered on the use of a gradient boosting machine prediction implemented using the XGBOOST framework, followed by the use of an interpretable machine learning method, namely Shapley Additive exPlanations (SHAP). We uncover various complex interactive and nonlinear relationships among model features and a binary outcome of being or not being a bike user for commuting. Our main findings suggest that policies designed towards broadening individual access to bicycles through ownership or sharing, in addition to the provision of shared cycle networks within 7 km of major employment centres can increase the adoption of cycling by commuters. Furthermore, among other results, we also observe that promoting regular teleworking among university staff, particularly for those who live at a distance more than 5 km of their place of work, could encourage commuting by bike. We also observe that cycling and public transport become complementary modes when home-work distances are greater that about 7 km.