Identifying environmental factors related to motorcyclist crash rates: variable selection using spatial Random Forest with network distance and barriers

Motorcycle crashes remain a major global safety concern, particularly in many Asian countries where motorcycles are the primary mode of transportation. While previous studies have identified factors associated with motorcycle crashes using traditional variable selection methods, they often overlook spatial patterns and contextual constraints (such as motorcycles being prohibited from freeways), leading to lower model accuracy. To address this limitation, we introduce a novel variable selection framework, designated as Spatial Random Forest with Network Distance and Barriers (spatialRF-NDBAR), which incorporates network distance and road barrier constraints to improve spatial variable selection. The study focuses on the Central Business District of Taipei City, where motorcycle access to certain roads is explicitly prohibited by road barriers. We utilize motorcycle crash data from 2016 to 2020, combined with Point of Interest (POI) data extracted from OpenStreetMap. We compare the performance of the proposed framework with that of the conventional spatialRF and spatialRF-ND (network distance constraint only) methods using both a traditional Ordinary Least Squares model and a Multiscale Geographically Weighted Regression (MGWR) model. Among the evaluated models, MGWR using the variables selected by spatialRF-NDBAR shows the best performance, yielding the lowest AICc (1801.212) and highest R² (0.867). Overall, our results show that incorporating spatial predictors reduces residual spatial autocorrelation, improving model performance. Furthermore, POIs such as marketplaces, restaurants, bars, hotels, banks, bus stops, and metro stations are associated with a higher crash risk. These findings provide valuable insights for local governments and support targeted traffic enforcement efforts around high-risk POIs to improve motorcycle safety.