We addressed our research question and analyzed the relationship between land use and station characteristics around BART by using a series of regression and visualization tools. We addressed each subquestion first (collecting, cleaning up, and analyzing each “characteristic”), then created a multivariable regression.
Mode share data for BART riders was collected by the Bay Area Rapid Transit district as part of the Station Profile Study, the largest survey of BART riders ever conducted. Mode share refers to the percentage of travelers using a particular type of transportation. The data includes walk, bike, transit, carpool, dropped off, drove alone, taxi, and motorcycle modes. Data on SFO and West Dublin was not available, so our model shows how well station characteristics can predict how riders access BART for home-based trips, for 42 out of 44 BART stations. The data was aggregated to create a few mode categories that we wanted to look at more closely in relationship to the characteristics we are analyzing: walk mode, walk/bike/transit modes, car modes, and SOV mode. Car modes include drive-alone, dropped off, and carpooled. SOV, or “single-occupancy vehicle” mode includes only those who drove alone to the station; SOV mode does not include carpoolers or those who were dropped of at the station.
We obtained parking space data from BART. For gaps in this data (e.g. MacArthur), we used Google Earth to estimate the number of parking spaces. For walkability, we used Walk Score as an indicator (we collected data via walkscore.com, using the Walk Score from each BART station address). We analyzed employment in terms of “job containment,” or the percentage of residents who live within 1/2 mile of a BART station and also work within 1/2 mile of that BART station. We extracted this data from LEHD. Population density was determined through 2010 US Census data. The population within each block within a 1/2 mile radius of a BART station was added to determine the total population within 1/2 mile of each BART station. This information was then divided by the total land area to calculate population density.
We cleaned up the data using a combination of Excel and Python, ran preliminary analyses in Python, qGIS, and UrbanCanvas, and present final visualizations here through Google Charts and CartoDB.
Results for each subquestion – parking, Walk Score, job containment, and population density – can be explored on the website. Of all factors, parking availability is the strongest predictor of non-SOV mode share to BART, followed by population density.
A multivariate regression was preformed to analyze the relationship between SOV mode share and all four station characteristics observed—number of parking spaces, population density, job containment, and Walkscore. While each characteristic individually is a weak-to-moderate predictor of SOV mode share, the combined characteristics create a strong predictor, with an R-squared of 0.852.
We can use this analysis to give an estimated prediction of the non-SOV mode share for BART stations, using only the four variables combined. This same analysis can be done using a number of economic, planning, and urban design elements that directly influence a station area—for example, the number of cafes within ¼ mile, the number of trees within ¼ mile, the number of bus routes terminating at the station, etc. Our analysis is a beginning step to placing a rider’s complete trip in context, and can inform planners and guide policy decisions about how station area characteristics affect mode choice and region-wide sustainability.