Geographically Weighted Nonlinear Regression for Cost-Effective Policies to Enhance Bus Ridership

This paper introduces a new geographically weighted nonlinear regression (GWNR) model to predict bus boarding more accurately. The proposed model, based on empirical data from selected bus routes in Chennai city, India, simultaneously accounts for spatial variations and non-linear relationships. The proposed GWNR model improves boarding forecast accuracy by increasing R 2 by 18.5% and reducing MAE by 15% compared to linear models. The results are used to identify best-fitting non-linear transformations for key variables such as bus and train station density, scheduled headway, and occupancy, thereby providing deeper insights and better interpretability. Unlike existing aggregate models, bus consideration probability is identified as a key predictor of bus boarding, thus reflecting non-users’ behavior. Without this effect, the influences of nearby bus and train stations show counterintuitive trends. Upon incorporating consideration probability, the presence of a single nearby train station increases bus boarding by improving accessibility, whereas multiple stations nearby reduce it due to competition effects. Finally, an illustrative policy application demonstrates the ability of the model to identify priority locations where scheduled headway changes are needed and to determine the optimal magnitude of adjustments. Such a targeted policy intervention is found to be twice as effective in increasing the ridership gain index compared to uniform area-wide policies.