Classification and regression tree (CART) analysis to determine best performing walkability metrics for studies of physical activity and public health
BACKGROUND
Walkability is difficult to quantify consistently or clearly for public health research and practice, in part because prominent metrics are (1) produced using proprietary black-box algorithms (i.e. Walk Score); (2) lack enough resolution to characterize precise activity locations (e.g. Environmental Protection Agency Walkability Index [EPA Index] characterizes at the census block group level); and (3) either lack consistency over time (e.g. changes to Walk Score algorithm in 2014) or are unlikely to be available longitudinally (e.g. EPA Index).

In contrast, transit ridership is driven by built environment components traditionally included in walkability metrics (e.g. land use, residential density, and employment density). Available at the transit stop level, it also captures aspects of place to which people frequently walk that may not score as high by other walkability metrics (e.g. Park-N-Ride lots, or car-oriented thoroughfares between high-density neighborhoods with parking constraints). Furthermore, ridership data are regularly collected and updated by transit authorities, with growing real-time availability. We hypothesized that transit ridership may capture walkability as well as traditional metrics, while having practical advantages over those metrics.

OBJECTIVE
We assessed the relative performance of five walkability metrics (transit ridership, employment density, residential density, EPA Index, and Walk Score®) in predicting objectively measured outdoor walking activity.

METHODS
We assembled a dataset of 72,900 GPS points recorded by 567 participants enrolled in the Travel Assessment and Community Study (TRAC), which objectively measured mobility patterns of a sample of King County, Washington residents before and after the installation of a light rail system (1,2). Participants wore accelerometry and Global Positioning System (GPS) devices and recorded travel diaries for approximately 7 days. These data were processed to identify physical activity bouts, and every recorded GPS point was categorized as either walking or non-walking according to a previously published algorithm (3). We used classification and regression trees (CART) to model five walkability metrics for each GPS location, adjusting for four sociodemographic features of the participant who recorded that GPS point (age, sex, household income, and household vehicle count). Walkability predictors were compared for their relative utility to predict walking vs. non-walking activity occurring at an XY location (using Gini importance index).

RESULTS
Residential density performed best in predicting whether walking or non-walking movement is occurring at an XY point, according to an overall measure of variable importance. Walk Score ranked second, with EPA National Walkability Index and transit ridership in close third and fourth positions, respectively. Employment density ranked fifth.

CONCLUSION
Our results suggest residential density performs better than Walk Score for the prediction of whether someone is walking at an XY location in urban King County, Washington. Transit ridership was as predictive as EPA Index. Residential density and transit ridership should be considered as single-component walkability metrics for public health studies requiring readily accessible longitudinal metrics. Further research will explore the replicability of these findings in other urban settings and extend this conceptualization of walkability to explore which metrics are best predictive of walking duration.

Presenter: Ronit Dalmat
Agency Affiliation: University of Washington - Urban Form Lab

Presenter Biographical Statement: [biography]