Generalized model for mapping bicycle ridership with crowdsourced data. (April 2021)

• Record Type: Journal Article
• Authors: Nelson, Trisalyn; Roy, Avipsa; Ferster, Colin; Fischer, Jaimy; Brum-Bastos, Vanessa; Laberee, Karen; Yu, Hanchen; Winters, Meghan
• Is Part Of: Transportation research. Volume 125(2021)
• Publication Date: 2021-04
• Subjects: Bias-correction -- LASSO -- Big data -- Bicycling ridership -- Exposure -- Strava; Transportation -- Periodicals; Transportation -- Technological innovations -- Periodicals; 388.011
• DOI: 10.1016/j.trc.2021.102981
• Languages: English
• ISSNs: 0968-090X

Abstract

Highlights: All bicycling ridership can be modeled with fitness app and geographic data. Variables needed to predict all ridership with Strava: income and safety. Model accuracy improves when official counts represent all bicycling conditions. Abstract: Fitness apps, such as Strava, are a growing source of data for mapping bicycling ridership, due to large samples and high resolution. To overcome bias introduced by data generated from only fitness app users, researchers build statistical models that predict total bicycling by integrating Strava data with official counts and geographic data. However, studies conducted on single cities provide limited insight on best practices for modeling bicycling with Strava as generalizability is difficult to assess. Our goal is to develop a generalized approach to modeling bicycling ridership using Strava data. In doing so we enable detailed mapping that is more inclusive of all bicyclists and will support more equitable decision-making across cities. We used Strava data, official counts, and geographic data to model Average Annual Daily Bicycling (AADB) in five cities: Boulder, Ottawa, Phoenix, San Francisco, and Victoria. Using a machine learning approach, LASSO, we identify variables important for predicting ridership in all cities, and independently in each city. Using the LASSO-selected variables as predictors in Poisson regression, we built generalized and city-specific models and compared accuracy. Our results indicate generalized prediction of bicycling ridership on a road segment in concert with Strava data should include the following variables: number of Strava riders, percentage of Strava trips categorized as commuting, bicycling safety, and income. Inclusion of city-specific variables increased model performance, as the R 2 for generalized and city-specific models ranged from 0.08–0.80 and 0.68–0.92, respectively. However, model accuracy was influenced most by the official count data used for model training. For best results, official count data should capture diverse street conditions, including low ridership areas. Counts collected continuously over a long time period, rather than at peak periods, may also improve modeling. Modeling bicycling from Strava and geographic data enables mapping of bicycling ridership that is more inclusive of all bicyclists and better able to support decision-making.