Research Question and Hypothesis

Strava Metro data are being used in bicycling research but are critiqued for bias—both in demographic and trip type—related to who and how the app is used. How well Strava ridership reflects overall bicycle ridership depends on locational factors, for example, how busy a street is, if bicycling infrastructure exists, and safety (Livingston et al. 2020; Nelson et al. 2020). Bias can be addressed by using models that link Strava ridership to counts of all bicyclists, and adjusting for geographic covariates (Jestico, Nelson, and Winters 2016; Roy et al. 2019). Strava data also include a commute label, which may better align with overall ridership. However, there is limited understanding of how Strava trips labelled as commute versus recreation differ in terms of representativeness. Our goal is to understand how the commute label impacts data representativeness by comparing geographic covariates associated with Strava commute and recreational ridership. We hypothesize that commute and recreational ridership samples will be associated with different areas of a city and anticipate that, compared to recreational ridership, Strava data labelled as commute trips better represent overall ridership.

Methods and Data

Study area

The study area is the census metropolitan area (CMA) of Victoria, British Columbia, Canada, with a population of ~367,770 (Statistics Canada 2016) and a bicycle mode share of 6.6%, the highest among Canada’s CMAs (Statistics Canada 2017). Victoria has over 200 km of bicycling infrastructure including ~100 km of regional trails. These paths are the backbone of the bicycling network and are heavily used for both commuting and recreation.

Data and Analysis

We used spatial regression to identify associations between sociodemographic, network, and built and natural environment characteristics for each trip type, and for the pooled sample. The Strava Metro data are from January 1, 2016 - September 30, 2017, and include a spatial file representing the street network and tabular data with aggregate activity counts for each segment. An attribute provided by Strava Metro identifies the count of activities on each segment that were commute trips. We also had area-level summary statistics on the number of unique app users, age-gender distribution, and trip characteristics.

The geographical unit of analysis was the Statistics Canada Dissemination Area (DA; 400-700 people; n = 534). We operationalized Strava ridership as bicycle kilometers traveled (BKT), calculated by multiplying the activity count on each road segment by the segment length and summing the products for each DA (Hochmair, Bardin, and Ahmouda 2019). We standardized BKT by DA total road length and mapped results for each trip type compared to the pooled sample (Figures 1 and 2). Geographic covariates were identified based on previous studies using Strava data, relevance in bicycling studies using conventional data, or local importance to bicycling (Table 1).

Table 1.Geographic covariates used to model Strava bicycle kilometers travelled (BKT) at the dissemination area (DA) level
Variable Operationalization Relevance Source
Sociodemographic
Bike commuters % of the population who bicycle to work Crowdsourced ridership patterns may be similar in areas where more people bicycle to work (Conrow et al. 2018) Statistics Canada 2016 Census
Male population % of the population who is male Men associated with higher levels of bicycling (Aldred, Woodcock, and Goodman 2016) and Strava app usership Statistics Canada 2016 Census
Median household income Quintile Higher and lower income associated with more bicycling (Pucher and Buehler 2006; Winters et al. 2007) Statistics Canada 2016 Census
Post-secondary education % of the population with a post-secondary degree or diploma Those with higher education more likely to bicycle (Winters et al. 2010) Statistics Canada 2016 Census
Under 15 years % of the population under age 15 Children and older adults associated with lower rates of bicycling (Aldred, Woodcock, and Goodman 2016; Pucher et al. 2011) Statistics Canada 2016 Census
Over 65 years % of the population over age 65 Statistics Canada 2016 Census
Visible minority % of the population who is a visible minority Higher neighborhood % of whites associated with more bicycling (Chen, Zhou, and Sun 2017); lower rates of bicycling investment in high % visible minority neighborhoods (Braun, Rodriguez, and Gordon-Larsen 2019) Statistics Canada 2016 Census
Network
Arterial, collector, local roads % of the total neighborhood road network in each functional road class Road class and density influence bicycling safety and route choice (Fraser and Lock 2011; Winters et al. 2010) Capital Regional District
On-street bicycle infrastructure % of the total neighborhood road network with on-street infrastructure Bicycling infrastructure associated with higher rates of bicycling (Fraser and Lock 2011); off-street paths associated with Strava ridership (Hochmair, Bardin, and Ahmouda 2019) Capital Regional District
Off-street bicycle infrastructure % of the total neighborhood road network with off-street infrastructure Capital Regional District
Regional trail Binary variable, 1 if regional trail network passes through DA Important cycling infrastructure in the Victoria region Capital Regional District
Intersection density Count of intersections in 1 km of DA centroid Higher intersection density associated with more bicycling (Winters et al. 2010) Canadian Active Living Environments (Can-ALE)
Built & natural environment
Bridge Binary variable, 1 if DA has a major bridge Bridges are important links between different parts of a street network (Boss et al. 2018; Hochmair, Bardin, and Ahmouda 2019) Capital Regional District
Distance to university Distance from DA centroid to nearest university Bicycling is higher around major destinations like universities (Krizek, Barnes, and Thompson 2009) British Columbia Open Data Catalogue
Steep slopes % of roads with maximum slope >5% (steep) Steep slopes impact route choice in utilitarian cycling and linked to more Strava bicycling (Winters et al. 2010; Lee and Sener 2019) GeoBC
Distance to shore Distance from DA centroid to shoreline Water bodies associated with more bicycling and Strava ridership (Chen, Zhou, and Sun 2017; Hochmair, Bardin, and Ahmouda 2019) Statistics Canada 2016 Census
Safety
Bicycle crash density # of crashes divided by road length Route choice affected by perceptions of safety (Winters et al. 2012) Insurance Corporation of British Columbia

We constructed three spatial error models using GeoDa 1.14 (Anselin 2019) and defined spatial neighbors using queen contiguity. The spatial error models use a Maximum Likelihood approach and treat spatially correlated residuals as a nuisance variable (Anselin 2009). The first and second models predicted DA commute and recreational BKT and the third predicted BKT for the pooled Strava sample.

Spatial analyses of areal data are often influenced by spatial autocorrelation (SAC)—the tendency for neighboring areas to have similar values. In ordinary least squares regression, spatial effects can lead to unreliable results as standard assumptions are violated (Anselin 2009). Spatial regression approaches include spatial lag and spatial error models, and model diagnostics indicate which is appropriate (Anselin 2009). We quantified SAC using Moran’s Ii, and used the Robust Lagrange Multiplier (LM) statistic to select the appropriate spatial regression model; the rule of thumb is to choose the model (lag or error) with the most significant LM test statistic (Anselin 2009). We also considered model fit (R2 and AIC).

Findings

In Victoria, there were 12,971 unique Strava app users and 315,200 activities; 49% (n = 155,252) of activities were identified as commutes. Men accounted for 74.9% (n = 9226) of app users, and 64% of users (n = 7,958) were under age 55. Table 2 shows BKT descriptive statistics. Notably, BKT for recreational trips comprised nearly two-thirds of the total BKT (63.5%); so, while there were approximately equal numbers of commute and recreational activities, the recreational trips tended to be longer in distance.

Table 2.Descriptive statistics for Strava ridership volumes (bicycle kilometers traveled, BKT) in 534 dissemination areas (DA)
Activity Type Mean Median SD Min Max Total
Commute BKT 776.1 529.7 853.9 14.6 12,041.1 414,456.5
Recreational BKT 1,354.2 821.9 1,610.4 46.0 12,454.6 723,148.5
Pooled BKT 2,130.4 14,114.1 2,308.7 67.9 24,495.7 1,137,605.0
Figure 1
Figure 1.Strava commute BKT compared to the pooled sample (quintiles) in DAs (n = 534).
Figure 2
Figure 2.Strava recreational BKT compared to the pooled sample (quintiles) in DAs (n = 534).
Table 3.Spatial regression estimates of DA Strava ridership (bicycle kilometers traveled, BKT) for trips labeled commute, recreation, and the pooled sample (all data combined).
Outcome Commute BKT Recreational BKT Pooled BKT
Spatial dependence diagnostics
Moran’s Ii 8.7*** 13.9*** 12.2***
Robust Lagrange Multiplier (lag) 2.7 14.3*** 6.7**
Robust Lagrange Multiplier (error) 12.2*** 9.0* 13.2***
R2 (lag) 0.6 0.6 0.6
R2 (error) 0.6 0.6 0.6
AIC (lag) 8305 9019 9422
AIC (error) 8289 8986 9389
Geographic covariates
Sociodemographic
Bike commuters 2120.6*** 4861.2*** 7030.5***
Male population 1793.9 3542.4 5341.3
Median household income 37.6 27.6 69.0
Post-secondary education 616.8 916.4 1489.9
Under 15 years -1182.7 -237.5 -1297.0
Over 65 years 574.2 2015.9** 2623.0*
Visible minority -291.2 -850.6 -1152.6
Street network
Arterial road 608.0 3348.4*** 4071.2***
Collector road 211.0 1847.6** 2117.4*
Local road -245.2 277.1 39.2
On-street bicycle infrastructure 1572.9*** 585.6 2087.9*
Off-street bicycle infrastructure 2589.7*** 2904.0** 5488.3***
Regional trail network 759.9*** 734.7*** 1475.0***
Intersection density -3.2 -20.9*** -24.2***
Built and natural environment
Bridge 765.9*** 701.1* 1450.0**
Proximity to university -29.0* -65.4 -96.8*
Steep slopes (more hills) 214.1 746.0* 927.9*
Distance to shore -202.4*** -509.7*** -720.6***
Safety
Bicycle crash density 93.5*** 52.7 140.1*
Constant -637.3 -1087.3 -1697.2
Spatial error (Lambda) 0.5*** 0.8*** 0.7***

*** p < 0.001, ** p < .01, * p < 0.05

Strava activities labelled as commutes showed higher ridership levels in DAs that were closer to universities, had more on-street infrastructure, and higher bicycle crash densities. Recreational activities showed higher ridership in DAs with older populations, more hills and major roads (arterial and collector), and lower intersection density. Strava activities for commute or recreation showed similar levels of ridership in DAs with regional trails, more off-street infrastructure, higher bicycle mode share, bridges, and closer proximity to the ocean.

Using Strava data labelled as commute activities, instead of a pooled sample, may mitigate bias in Strava data and better represent ridership patterns of people of all ages and abilities. When using only the commute activities, we saw different spatial patterning in ridership and found that activities tended to be in areas with bicycle infrastructure, lower levels of safety, and around universities. Recreation activities were more common in areas with major roads, steep slopes, and lower intersection density—factors that are typically less preferred and less safe for bicycling (Teschke et al. 2012; Winters et al. 2010). Other important covariates were common across commute and recreational trips (regional trails, off-street infrastructure, bicycle mode share, bridges, and proximity to the ocean) and would be selected from either the commute or pooled sample. This suggests that if a planning exercise aimed to model overall ridership using Strava data and geographic covariates, a distinct set of covariates would be selected if only the commute data were used. In Victoria, commuting accounted for 49% of all Strava activities and 36.5% of the total distance (e.g., BKT). From other Strava analyses we learn the proportion of commute trips varies across cities (e.g., 21% in El Paso, Texas (Lee and Sener 2019), 85% in Milan, Italy (Sunde 2019). If interested in modeling all ages and abilities ridership, researchers and practitioners should consider using only commute data, at least in cities where the sample is sufficiently large.

Acknowledgements

The authors would like to acknowledge Strava for providing the data.