Evaluating the Potential of Crowdsourced Data to Estimate Network-Wide Bicycle Volumes

This research integrated and evaluated emerging user data sources (Strava Metro, StreetLight, and hybrid docked/dockless bike share) of bicycle activity data with conventional "static" demand determinants (land use, built environment, sociodemographics) and measures (permanent and short-duration counts) to estimate annual average daily bicycle traffic (AADBT). We selected six locations (Boulder, Charlotte, Dallas, Portland, Bend, and Eugene) covering varied urban and suburban contexts and specified three sets of Poisson regression models: city-specific models, an Oregon pooled model, and all cities pooled. Static variables, Strava, and StreetLight complemented one another, with each additional data source tending to improve the model performance. Sites with lower volumes were more difficult to predict, with considerable error in even the best-performing models. City-specific models in general exhibited improved fit and prediction performance. Expected prediction error increased by a factor of about 1.4 when using Strava or StreetLight alone, but without static adjustment variables, to predict AADBT. Combining Strava plus StreetLight, but without static variables, increased error slightly less; by 1.3 times. We also found that transferring the model specifications from one year to the next without re-estimating the model parameters resulted in a 10% to 50% increase in error rates across models, so such transfer is not recommended. The findings from this study indicate that rather than replacing conventional bike data sources and count programs, old "small" data sources will likely be very important for big data sources like Strava and StreetLight to achieve their potential for predicting AADBT.