Local modeling as a solution to the lack of stop-level ridership data

Transit ridership modeling at the bus stop level is an important tool for bus network planning and transit-oriented development. However, many cities, especially in developing countries, face a lack of boarding and alighting data due to the high costs of collection. Solutions based on smartcards often rely on assumptions that negatively affect the data accuracy. Noting that previous studies suggest the existence of spatial heterogeneity and dependence in factors affecting stop-level ridership, the present paper proposes the application of Geographically Weighted Negative Binomial Regression (GWNBR) to modeling the transit ridership along bus lines in São Paulo – SP (Brazil) under missing data conditions. Four important topics are analyzed: 1) whether the spatial variation of predictors' effects is statistically significant; 2) the consistency of parameter estimates; 3) the prediction power sensitivity to missing data scenarios; and 4) the use of network distances replacing the traditional Euclidean ones. Of the five predictors that explained the transit ridership better, overlapping and frequency proved to have coefficients with statistically significant spatial variation. Goodness-of-fit measures indicated that GWNBR is an effective tool to the transit ridership estimation in uncounted bus stops, even when the availability of data is low. GWNBR in missing data scenarios could, in fact, reproduce the spatial pattern of effects shown in the complete database model, for some explanatory variables. Network distances may better represent the spatial relationship between transit ridership and some of its predictors. In addition, GWNBR models were able to address the spatial dependence found in the Negative Binomial Regression.