Spatial Modeling of Travel Demand Accounting for Multicollinearity and Different Sampling Strategies: A Stop‐Level Case Study

Abstract

Stop‐level ridership data serve as a basis for various studies toward increasing bus patronage and promoting sustainable land use planning. To address limitations found in previous studies, this study proposes a novel approach based on Geographically Weighted Principal Component Analysis (GWPCA) and Ordinary Kriging to predict the stop‐level boarding or alighting data along bus lines in São Paulo (Brazil), considering four different sampling methods. The main contributions are as follows: by accounting for the spatial heterogeneity of the predictor dataset, the GWPCA can identify the most important factor affecting transit ridership even in bus stops with no information on boarding and alighting; the spatial modeling of stop‐level ridership data using GWPCA components as explanatory variables allows visualizing the spatially varying effects from predictors on ridership, supporting the land use planning at a local level; GWPCA coupled with kriging simultaneously addresses the multicollinearity of predictor data, its spatial heterogeneity, and the spatial dependence of the stop‐level ridership variable, thus enhancing the goodness‐of‐fit measures of the transit ridership prediction in unsampled stops; and a balanced sample on predictor data and well‐spread in the geographic space might be preferred to accurately estimate missing stop‐level ridership data. In addition to solve the lack of stop‐level ridership data, supporting a reliable bus system planning, the proposed method indicates what predictors should be addressed by policymakers to stimulate a transit‐oriented development. The method can be successfully applied to other travel demand variables facing a lack of data such as traffic volume in road segments and mode choice at the household level.