You are on page 1of 11

Transportation Research Part A 121 (2019) 1–11

Contents lists available at ScienceDirect

Transportation Research Part A


journal homepage: www.elsevier.com/locate/tra

Do new bike share stations increase member use: A quasi-


T
experimental study

Jueyu Wang , Greg Lindsey
Humphrey School of Public Affairs, University of Minnesota, Twin Cities, 301 S 19th Ave S #130, Minneapolis, MN 55414, United States

A R T IC LE I N F O ABS TRA CT

Keywords: As the number of bike share programs across the world has grown, studies of bike programs and
Bike share user behavior operations have proliferated. Most empirical studies of bike share demand have included ana-
Annual members lyses of station use while a limited number of studies have investigated member behavior.
Difference in difference models Moreover, a limitation of this research is that the most research designs have been cross-sectional
Natural experiment
and therefore unable to establish causality. To address this limitation, we employ a quasi-ex-
perimental, difference-in-difference modeling approach using a six-year panel data set of mem-
bers’ bike share trips from 2010 to 2015 in Minneapolis-St. Paul, Minnesota. This research design
takes advantage of changes in the bike share network over time to establish treatment and control
groups and test the significance of effects of changes in accessibility on the frequency of in-
dividual member’s use of bike share. Improvements in accessibility are measured as a reduction
in distance to stations resulting from placement of new stations or relocation of old stations. We
find a significant negative impact of distance on the frequency of use and that the effects of
increasing bike share accessibility are larger in areas with denser bike share services. Specifically,
members for whom access improved (i.e., distance decreased) were significantly more likely to
increase the frequency of use than members for whom access remained the same. Moreover, by
developing different models, we show the effects of distance are heterogeneous and vary with
different built environment contexts. Members who live in areas with higher population density
and a higher percentage of retail land use tended to increase their bike share use more. Our
results indicate that improvements in physical accessibility may not result in practically mean-
ingful changes in the frequency of use in all cases and imply that multi-faceted strategies for
increasing use may be needed.

1. Introduction

Bike share programs have potential to solve the first mile and last mile problems associated with transit use, reduce traffic
congestion by diversifying transportation options, provide environmental benefits by reducing greenhouse gas emissions, and bring
health benefits by increasing physical activity (Shaheen et al., 2010). Because of these benefits, bike share programs have proliferated
throughout the world, with more than 1000 cities now operating bike share programs (Meddin and DeMaio, 2017). As the number of
bike share programs has grown, many studies have explored socio-demographic characteristics of bike share users in different bike
share systems (Buck et al., 2013; Fishman et al., 2015; Murphy and Usher, 2011; Shaheen et al., 2012). Some, but fewer have
examined the behaviors of bike share users (e.g., Ogilvie and Goodman, 2012). Although some bike-share members use bike share


Corresponding author.
E-mail addresses: wang5931@umn.edu (J. Wang), linds301@umn.edu (G. Lindsey).

https://doi.org/10.1016/j.tra.2019.01.004
Received 6 December 2017; Received in revised form 25 September 2018; Accepted 3 January 2019
Available online 09 January 2019
0965-8564/ © 2019 Elsevier Ltd. All rights reserved.
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

daily or more often, many bike share members appear to be infrequent users with low frequencies of use (Bachand-Marleau et al.,
2012, Buck et al., 2013). The goal of this study is to add to this literature and increase our understanding of both the demand for bike
share and the behaviors of existing users.
Researchers (Buck and Buehler, 2012; Faghih-Imani et al., 2014) have employed a variety of research designs and estimated a
variety of models to explore correlates of bike share demand. The findings of these studies generally have been consistent and have
shown that bike share use is influenced by user characteristics (e.g., Bachand-Marleau et al., 2012), accessibility to both bike share
stations and destinations (e.g., Fishman et al., 2015), and the built environment (e.g., Wang et al., 2015). Most studies have employed
cross-sectional designs, which can confirm correlation but not establish causality. Besides, most studies (e.g., Noland et al., 2016;
Wang et al., 2015) have been aggregate station-level analyses that are unable to provide insights into the sociodemographic and
behavioral characteristics of individual bike share users. Fewer studies (e.g., Ogilvie and Goodman, 2012) have focused on member
behavior, particularly in response to changes in system configuration and accessibility. However, these existing studies have not fully
explained how the impacts of changes in accessibility to bike share stations vary in different built environment settings.
To address some of the limitations of prior research, we employ a quasi-experimental, difference-in-difference (DID) modeling
approach using a six-year panel data set of members’ bike share trips from 2010 to 2015 in Minneapolis-St. Paul, Minnesota. This
research design takes advantage of changes in the bike share network over time to establish treatment and control groups and test the
significance of effects of changes in accessibility on the frequency of use of bike share by individual members. The changes in
accessibility result from the placement of new stations or relocations of old stations. Our treatment variable, accessibility to bike
stations, is measured by network distance from each member’s residence to the nearest bike share station. Members for whom
distance does not change over time are used as the control group.
We use this DID approach to answer two questions:

(1) How does an improvement in accessibility to bike share stations influence the frequency of bike share use by annual members?
(2) How does the impact differ in different settings, specifically, in relation to different built environment contexts?

We contribute to the literature on bike share demand in two ways. First, by using multi-period DID models, we minimize the effect
of some confounders (such as calendar event effects). Second, by developing separate models, we demonstrate the heterogeneous
effects of the treatment variable across different built environment contexts. Overall, we find a significant negative impact of distance
on the frequency of use and that the effects of increasing bike share accessibility on use are larger in areas with denser bike share
services. We show that members who live in areas with higher population density and a higher percentage of retail land use tend to
increase their bike share use more. A limitation is that we do not control for all potentially confounding factors such as improved
quality of service through more efficient re-balancing of bikes.

2. Literature review

2.1. Bike share user characteristics and behaviors

Researchers and bike share operators are interested in who uses bike share. Previous studies have examined gender, income, race,
age and educational status, among other characteristics. These studies have revealed some common trends. Bike share users are more
likely to be white (e.g., Buck et al., 2013), higher income (e.g. Fishman et al., 2015 ), younger (Murphy and Usher, 2011; Shaheen
et al., 2012) and more highly educated (Ricci, 2015), compared to the average population where bike share programs are im-
plemented.
In North American, bike share users in Washington DC (Buck et al., 2013) are more likely to be white, younger and have higher
incomes than the general population. Shaheen et al. (2012) reported that bike share users in Minneapolis, Toronto, Salt Lake City,
and Montreal are more likely to be in the 25–34 year-old age category. In Europe, in Dublin, Ireland, Dublinbike users are pre-
dominantly male, and young with higher income (Murphy and Usher, 2011). London BSP users also tend to be mainly white and
wealthier than the population at average (Goodman and Cheshire, 2014). In the Asian context, users in Beijing, Hangzhou, and
Shanghai have been found to have higher automobile ownership relative to the non-users (Yang et al., 2010; Fishman et al., 2013;
Ricci, 2015).
Several studies have explored the behavior of bike share member use, including members’ trip frequency, distance, and purpose
(e.g., Buck et al., 2013; Ogilvie and Goodman, 2012; Romanillos et al., 2018). Bike share members predominantly use bike share for
fulfilling commuting or utilitarian trips (Shaheen et al., 2012; Transport for London, 2014). However, many bike share members do
not use bike share frequently (Fishman et al., 2015). Ogilvie and Goodman (2012) found 25.4% users in London made no trip during
the study period (July 2010–March 2011). Half of the members of the London bike share program in 2014 reported no use of the
service in the prior month (Transport for London, 2014); similarly, Fishman et al. (2014) found that 46% of annual members in
Melbourne and Brisbane, Austria, also did not report use in the prior month. In Washington DC, 21% female users and 13% male
users reported zero monthly usage (Buck et al., 2013). The fact that many members ride relatively infrequently implies there may be
potential to increase their use of bike share through operational or system changes, promotional programs, incentives, or other
strategies.

2
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

2.2. Bike share demand

2.2.1. Bike share station demand analyses


Because of their system design, technology, and methods of charging users, most bike share programs generate comprehensive
databases that include, for each station, both check-out and check-in times for each bike docked at the station. Check-out data are
counts that generally represent trip origins, while check-in data are counts that generally represent trip destinations. Some check-out
and check-in data may represent rebalancing (i.e., relocation of bikes by system operators to match demand) and therefore may not
represent trips. Researchers may combine check-out and check-in counts into measures of station activity if their primary interest is in
which stations are used most, or they may analyze them separately if they are interested in whether stations are generators or
attractors of trips. Depending on their research question, researchers also may aggregate counts for different time periods, including
hourly (e.g., Hampshire and Marla, 2012), daily (e.g., Bachand-Marleau et al., 2012), or monthly (e.g., Rixey, 2013) use. For ex-
ample, researchers (e.g., Faghih-Imani et al., 2014) interested in how bike share is used for utilitarian purposes such as commuting
may analyze hourly patterns to see if particular stations serve as attractors during morning hours and generators during evening
hours.
Researchers also have adopted a variety of analytic approaches and used different statistical models. One approach to demand
analysis uses aggregate data to characterize spatial-temporal patterns of bike share use (Corcoran et al., 2014; O’Brien et al., 2014;
Vogel, et al., 2011). For example, O’Brien et al. (2014) measured aggregated system characteristics such as numbers of dock stations,
spatial characteristics, and temporal characteristics of bike share use of 38 bike share systems across the world, and classified these
different bike share systems based on their system use patterns. Another approach is to use different statistical models to model daily
or hourly bike share use, including count-based negative binomial models (e.g., Wang et al., 2015), fixed-effect or random effect
models (e.g., Hampshire and Marla, 2012), multi-level mixed models (e.g., Bachand-Marleau et al., 2012; Faghih-Imani and Eluru,
2016), and spatial statistical models that control for spatial autocorrelation (e.g., Noland et al., 2016). Most of these studies are at the
station level and provide no direct information about the behavior or demand of individual users or factors with individual behavior.

2.2.2. Correlates of bike share demand


Studies (e.g., Buck and Buehler, 2012; Faghih-Imani et al., 2014; Rixey, 2013; Wang et al., 2015) have examined the impacts of
several types of variables on bike share use. These variables generally can be grouped into four or five categories, depending on the
research approach. These categories typically include socio-demographic characteristics, urban form and land use mix or composi-
tion, transportation infrastructure, including street functional class and bicycle and transit facilities, weather, and temporal variables.
Socio-demographic variables. Socio-demographic variables often measure the socio-economic and demographic composition of
adjacent neighborhoods of bike share stations. Wang et al. (2015) modeled the annual average daily bike trips at bike stations in Twin
Cities and showed that the percentages of white and middle-aged populations in neighborhoods around stations were positively
associated with higher bike share station use. Rixey (2013) found that higher median income and lower percentages of non-white
populations were positively correlated with higher bike share station use. However, a member-level analysis by Ogilvie and Goodman
(2012) revealed that members in London living in areas of deprivation tend to use bike share more, after controlling station ac-
cessibility. The differences in results between station-level and member-level analyses indicate that additional study of members’
behavior using individual data is warranted.
Land use variables. Land use variables usually include population density, job density, percentages of different types of land use
(e.g., commercial), and uses that may represent trip-generators or destinations such as universities or parks. Researchers have shown
that bike share use tends to be higher at stations located in areas with higher population density (Faghih-Imani et al., 2014; Noland
et al., 2016), job density (Faghih-Imani et al., 2014; Wang et al., 2015), or close to businesses (Rixey, 2013). Bike share use also tends
to be higher at bike stations within Central Business Districts (Buck and Buehler, 2012; Faghih-Imani et al., 2014; Wang et al., 2015)
or at universities (Faghih-Imani et al., 2014).
Transportation infrastructure. Transportation infrastructure mainly has been modeled by incorporating measures of bicycling
infrastructure and transit service. Bicycling infrastructure usually has been measured as the lengths of different types of bike facilities.
Buck and Buehler (2012) modeled the impacts of bike lane and urban trail supply within a half mile of a bike share station and found
significant positive correlations with average daily bike share checkouts. These results were affirmed in research on the BIXI Bike
share program in Montreal, Canada (Faghih-Imani et al., 2014), the Nice Ride program in Minneapolis, MN (Wang et al., 2015), and
the bike share system in New York (Noland et al., 2016). Moreover, Rixey (2013) conducted a multi-city demand analysis and
reported that bikeway supply had a positive association with monthly bike share rentals in Minneapolis, MN, Montreal, Quebec,
Denver, Colorado, and Washington DC. Romanillos et al. (2018) used GPS data to explore the cycling route preferences of bike share
users in Madrid and found that bike share users tend to ride in streets with bicycling infrastructure. Station-level demand analyses
(e.g., Hampshire and Marla, 2012) also revealed that increasing the number of bike share stations near an existing bike share station,
which is one strategy for increasing connectivity within a network, is more likely to increase the use of that station.
Transit service is often measured as the availability of transit service. Previous studies on bike share demand analyses found that
bike share use tended to be higher at bike stations that are close to rail stations (Buck and Buehler, 2012; Faghih-Imani et al., 2014;
Noland et al., 2016; Wang et al., 2015). A study of the impacts of bike share service on bus ridership found that bike share stations
located close to bus stops could decrease bus ridership (Campbell and Brakewood, 2017).
Weather and temporal variables: Weather variables commonly in models of daily station use include temperature, precipitation,
and wind speed (Faghih-Imani et al., 2014; Rixey, 2013; Rudloff and Lackner, 2014). Temperature is positively correlated with use
(Faghih-Imani et al., (2014)) while precipitation is negatively correlated (Rixey, 2013). A detailed hourly weather study (Gebhart and

3
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

Noland, 2014), for instance, showed that weather was correlated with both the use of bikes and the duration of trips undertaken.
Faghih-Imani et al. (2014) studied the bike share program in Montreal and found that people tended to bicycle more on weekdays
than weekends. Moreover, built environment variables often were found to have different effects when interacting with temporal
factors. Faghih-Imani et al. (2014), for instance, found that built environment variables such as job density took opposite effects on
bike share use in AM and PM peak hours.

2.3. Accessibility and bike share use

Accessibility, which generally has been measured as network distance to a bike share station (Bachand-Marleau et al., 2012;
Fishman et al., 2012, 2015) has been found to have significant impacts on the probability that individuals will use bike share but its
correlation with frequency of use is less clear. Fishman et al. (2012) used an online survey data to examine the probability of
becoming a bike share member and found that proximity to bike share stations increases the probability of joining. Bachand-Marleau
et al. (2012) used data from an online survey and a binary logistic regression model to measure the probability of using BIXI and a
linear regression model of use frequency of bike share in Montreal, Canada. The results showed that the proximity to a BIXI station is
significantly correlated with the probability of using BIXI but has no significant correlation with use frequency. These studies used
sample survey data and potentially suffered from selection bias and the problem of representativeness.
The availability of user data offers new opportunities to the study of behavior at the member level. Ogilvie and Goodman (2012)
used user data from the London bike share program and found that residential proximity to bike share stations has a significant
association with the use frequency of London bike share users. The mixed results between accessibility and use frequency warrant
further studies of the relationship. Analysis of member data also may increase understanding of what bike share stations in some
neighborhoods are underutilized. Planners and bike share managers are eager to know what else could be done to encourage the use
of bike share, besides the provision of bike share stations. Numerous studies (e.g., Fishman et al., 2012, 2014) have identified barriers
to bike share use. However, these studies have not explored the heterogeneous impacts of accessibility on use in different built
environment contexts.
In sum, researchers have made considerable progress in modeling demand and identifying factors that are associated with station
use. As shown in this review, researchers consistently have found that bike share users tend to be male, white, younger, and more
educated with higher incomes. The presence of bicycle facilities and other elements of the built environment are associated with use
while the relationship between greater accessibility (i.e., shorter distances) to stations and higher levels of use is yet unclear. This
study will contribute to the existing literature in the following two ways. (1) The study used individual member level data and
adopted the DID research design to examine the impacts of accessibility on members’ use and provide a stronger link between
accessibility and bike share use. (2) The study explored the heterogeneous impacts of accessibility in different built environment
contexts, which will provide more practically-meaningful implications.

3. Data and methodology

3.1. Nice Ride Minnesota

The Nice Ride bike share system in the Twin Cities launched in the spring of 2010 with 65 stations and 1197 docks. These 65

Fig. 1. Nice ride stations in Twin Cities.

4
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

stations were concentrated in downtown Minneapolis, commercial areas, and the University of Minnesota Minneapolis campus. From
2010 to 2015, the Nice Ride system expanded from 65 stations to 190 stations in Minneapolis and St. Paul (Fig. 1), covering the
majority of areas within the Twin Cities. Nice Ride Minnesota operates from April to November each year due to the severe weather
in winter.

3.2. Research design

We use a multi-period DID research design to explore the relationship between accessibility to bike stations and frequency of bike
share use. A DID research design uses panel data to compare an outcome for one group that receives the treatment during the study
period to a control group that does not receive the treatment during the period in an attempt to quantify the impact of the treatment.
The phased expansion of Nice Ride between 2010 and 2015 allows us to adopt this research design. The treatment variable in our
study is physical accessibility to a bike share station, which we measure as street network distance to the nearest bike station from the
self-reported home address of each bike share member. Our treatment group comprises bike share members for whom the network
distance to a bike share station changed during the study period. Our control group comprises members for whom the distance to the
nearest bike station did not change. Because we measure the treatment as a continuous variable rather than a binary variable, we are
better able to assess the relative effects of changes in accessibility among members.
The general form of our model, or equation, is expressed as follows:
Yit = σDit + Xit γ + αi + ηt + νit (1)
where

• t in the study indicates year week.


• Y denotes bike share use frequency of annual member i at year week t.
it

• D is the treatment variable indicating the network distance to the nearest bike station from the self-reported home address of
it
annual member i at year week t.
• σ is the estimated effect of the network distance to bike share on bike use frequency.
• X is a vector consisting of time and individual variant variables such as the provision of light rail service and bike infrastructure
it
supply.
• γ are the estimated effects of X on bike share use frequency.
it

• α is the time-invariant error term and refers to the individual-level fixed effects. α is usually used to control the unobserved
i i
individual characteristics that are associated with the weekly frequency of bike share use and do not vary with time. The in-
dividual-level fixed effects in this model include race, income, gender and other time-invariant characteristics of member i.
• ηt is the common unobserved temporal trend. This vector is usually used to control the temporal trend in specific t. In this study, ηt
includes temperature, precipitation, wind speed and other temporal characteristics in the year week t.
• vit is the error term with standard properties.

An important assumption of the DID design is that the locations of choices to install bike share stations are random after con-
ditioning on the time-invariant and individual-invariant variables. However, this assumption likely is violated in our application
because the bike share stations were installed with knowledge of the existing system, presumably to maximize overall use.
Specifically, Nice Ride might have prioritized installation of stations close to residential locations of users who use Nice Ride more
often. The limitation implies that our models may suffer from endogeneity, which means that the effects of increases in accessibility
could be overestimated. However, Wang and Lindsey (2018) studied the locational characteristics of Nice Ride stations and showed
that Nice Ride mainly serves employees and job-rich areas, which implies prioritization of station location based on the preferences of
employees and commuters instead of the residential addresses of population. To address the potential problem of endogeneity, we
used linear OLS models to test whether there were significant differences before treatment in the frequency of use between members
in the treatment group and members in the control groups by adding pretreatment interaction terms in the regression equation. The
results (not reported here) showed that the coefficients were not significant; indicating that the bias potentially associated with
endogeneity is not a major concern.

3.3. Modeling approach

Our dependent variable is the number of uses per week by annual members. These counts are not normally distributed, so
ordinary least square estimation is inappropriate. Given the non-normal distribution of these count data, we often have two options, a
Poisson regression model or a negative binomial regression model. Poisson regression models assume equi-dispersion while count
data frequently suffer from over-dispersion, which causes large standard errors and low p-values. The negative binomial model
generally is preferred because it relaxes the assumption about equi-dispersion. However, other factors must be considered. Our DID
research design adopted the fixed effect model to control unobserved individual-invariant and time-invariant effects. Research has
shown the conditional likelihood for the negative binomial distribution required for fixed effects models is problematic (Allison,
2005). Therefore, despite the potential limitations, we use fixed effects Poisson models. Furthermore, additional error terms in fixed
effects models (Rabe-Hesketh and Skrondal, 2008) could help to address the over-dispersion problem by allowing for sources of
variability not included in a Poisson model.

5
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

To interpret the practical effects of factors on the frequency of weekly use, we report the Incidence Ratio Rate (IRR), which takes
the exponential of the regression coefficients. The IRR indicates the percentage change in the dependent variable that is induced by a
one unit change in an independent variable. An IRR is usually compared with one: an IRR larger than one indicates a positive
correlation of the independent variable with the dependent variable. In contrast, an IRR smaller than one indicates a negative
correlation.
We hypothesize:

• Members who experience an increase in accessibility to a bike share station (i.e., members in the treatment group) increase
frequency of use more than members for whom accessibility does not change or decreases (i.e., members in the control group);
• Members who experience an increase in accessibility in areas that are bicycle-friendly (i.e., with greater densities of bicycle
facilities) increase frequency of use more than members in areas with lower densities of bicycle facilities; and
• The effects of changes in accessibility are heterogeneous and are associated with differences in the built environment.
With respect to the third hypothesis, most of built environment variables change little over the period of analysis. Thus, to account
for the heterogeneous treatment effects, we follow the established econometric procedure of incorporating interaction terms between
built environment variables and the treatment variable. However, because we estimate models using a non-linear fixed effect Poisson
model, we cannot make any statistical inferences about the impact of interaction terms on the frequency use by using the sign and
significance of estimated coefficients (Karaca-Mandic et al., 2012). Fixed effect models eliminate the constant term at the annual
member level (fixed-effect). It, therefore, is impossible to compute the marginal effects for both single and interaction term variables.
The only approach to compute marginal effects is that assuming that all the fixed effects are zero, which violates the aim of the
research design.
Thus, to enable interpretation of the heterogeneous treatment effects of the built environment, we augment our base models with
a set of models that use dummy variables to represent the built environment variables of interest. Specifically, we create a set of
dummy variables that are equal to one if the value of the built environment variable is larger than the mean, and zero if it is equal to
the mean or less.
Annual members reside across the Twin-Cities Metropolitan Area. Many members live in suburbs without bike share stations near
their homes, but purchase memberships to use while in the cities for work or other purposes. However, by using the network distance
to nearest bike stations to measure bike share accessibility, and by measuring the built environment around the home addresses of
annual members, we focus on members who have the greatest potential to experience improvements in station accessibility. We
therefore only consider annual members who live within the city boundaries of Minneapolis and St. Paul, where a bike share station
was more likely to be installed. We censored from the dataset members who live outside of the Minneapolis and St Paul municipal
boundaries because changes in accessibility from their homes were unlikely to affect their behavior. This analytic choice helps to
guarantee that these members have the potential to both (a) have bike share accessibility improved and (b) use bike share for home-
based trips.
Furthermore, the service area of a bike station is often defined as a quarter mile buffer around a bike share station; analysts have
used both network distance and airline distance to form buffers (Clark and Curl, 2016; Noland et al., 2016; Rixey, 2013). We use this
cut-off value in three models that explore the different impacts of accessibility on bike share on weekly use:

• Model 1 is for all annual members living within the city boundaries of Minneapolis and St. Paul.
• Model 2 is only for annual members whose network distances to nearest bike stations were always shorter than a quarter mile
during their memberships. For these members, installation of a new station has the effect of densifying bike share service.
• Model 3 is for annual members whose network distances were larger than a quarter mile at least at a time t.
3.4. Data

We obtained trip log data from Nice Ride Minnesota website (https://www.niceridemn.org/data/). The station installation dates
were derived from trip log data. We identified the date when the first trip occurred at each station and used it as the installation date.
Our trip analysis is at the member level. The member-level trip log data from 2010 to 2015 also was obtained from Nice Ride
Minnesota.
Between 2010 and 2015, Nice Ride had 13,569 unique annual members (six members registered the membership in the last week
of the 2015 operational year and were censored from the dataset because they had no meaningful opportunity to establish patterns of
use. Our decision to limit analyses to members within the city boundaries of Minneapolis and St. Paul resulted in 9046 members being
included in the analyses. The values of the dependent variable (trips per week) in the raw dataset range from 0 to 166. By checking
the original dataset, we identified these counts that were erroneous or outliers and censored them from the dataset: we removed 15
observations with trips per week larger than 70, In the final dataset used for analysis, the dependent variable, trips per week, ranges
from 0 to 70. The mean of trips per week is 1.7, the standard deviation is 3.5 and the median is 0.
The treatment group (i.e., those members who network distances changed during their membership) includes 1249 unique in-
dividuals. A total of 98,940 weekly observations (i.e., counts of trips per week) are available for analysis. Among the 1249 unique
members, 225 lived within a quarter mile of bike share stations for the duration of their subscriptions (Model 2) while 1024 members
lived outside a quarter-mile buffer for at least once at a time t (Model 3) (see Table 1).
Built environment variables were derived from ArcGIS (Table 2) and created specifically around members’ self-reported homes.

6
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

Table 1
Number of observations in the treatment and control group.
Model 1 Model 2 Model 3

Twin Cities Dit ≤ ¼ miles ¼ miles < Dit


Total N (NT) 9046 (431,432) 5031 (233,811) 4015 (197,621)
N (NT) in the treatment group 1249 (98,940) 225 (16,230) 1024 (82,710)
N (NT) in the control group 7797 (332,492) 4806 (217,581) 2991 (114,911)

Notes: N: the number of annual members; (NT): the total number of observations.

Table 2
Descriptive statistics of variables.
Variable Description Mean Std. Dev.

Treatment Variable (N = 431,432)


Distance (0.1 miles) Network distance to the nearest bike stations 4.50 6.66

Weekly observation level (N = 431,432)


Bike lane length (0.1 miles) Bike lane length within a quarter mile of home addresses of members at time t (year level data) 7.28 7.56
LRT (1/4 mile) A dummy variable, equal to 1 if there is a light rail station within a ¼ mile at time t 0.07 0.26
LRT (1/2 mile) A dummy variable, equal to1 if there is a light rail station within a ½ mile at time t 0.28 0.45

Member level (N = 9046)


Female (N = 8971)* A dummy variable, equal to 1 if the member is female 0.45 0.50
Pop density Total population per square meter in the block where members reside 0.008 0.008
Job density Number of jobs per square meter in the block where members reside 0.007 0.036
# Intersections Number of intersection within a ¼ mile buffer of each member’s home address 28 7
% Recreation Percentage of recreation land use within a ¼ mile buffer of each member’s home address 0.21 0.23
% Retail Percentage of retail land use within a ¼ mile buffer of each member’s home address 0.41 0.44
% Office Percentage of office land use within a ¼ mile buffer of each member’s home address 0.07 0.13
% Industrial Percentage of industrial land use within a ¼ mile buffer of each member’s home address 0.17 0.33

Notes:
* Some missing values for the gender of members.

Measures of bicycle infrastructure are at the year level and change over time. We incorporate bikeway length as the measure of bike
infrastructure. There are two light rail lines (LRT), the Green Line and the Blue Line, operating in Minneapolis-St. Paul. A dummy
variable is created to indicate whether a light rail station is within a quarter-mile of member i’s homes at year week t. We use the one-
quarter mile criterion for LRT accessibility in our base models to maintain consistency with distances used to create buffer zones for
other built environment variables. However, to test the reliability of results, and because some analysts measure pedestrian acces-
sibility to light rail stations using a ½-mile network walking distance, we also estimated models using a ½-mile LRT variable. Land
use variables are created to measure the built environment around each annual member, including population density, job density,
and proportions of different land uses.

4. Results

4.1. Impacts of bike share accessibility on use frequency

We present three basic models (M1, M2, M3) in Table 3. All three models are fixed-effect models to eliminate the effects of time-
invariant and individual-invariant variables. All three models incorporate three variables that vary over time and by individuals: the

Table 3
Basic DID model estimation results.
Model 1 Model 2 Model 3
Twin Cities Dit ≤ ¼ miles Dit > ¼ miles

IRR p IRR P IRR p

Distance 0.997 0.000 0.889 0.000 0.999 0.075


Bike lane length 1.015 0.000 1.023 0.000 0.998 0.290
LRT (1/4 Mile) 0.761 0.000 0.736 0.000 1.110 0.070

Year-week fixed effect Yes Yes Yes


Member fixed effect Yes Yes Yes
Prob. > Chi2 0.0000 0.0000 0.0000
Log likelihood −715491.93 −420856.04 −294127.67

7
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

network distance to the nearest bike station (i.e., accessibility), bike lane length, and proximity to LRT. The Wald Chi-Square test-
statistics tests on joint significances of variables indicate that our models are significant.
The estimation results for all models are similar. The treatment variable has significant and negative correlations with the weekly
frequency of bike share use in each model (i.e., the IRRs are smaller than 1). These results indicate that reducing the network distance
from the residence of annual members to the nearest bike station by installing a new bike station nearby increases their weekly use of
bike share. However, while statistically significant, the effect size of the treatment variable is very small, especially in Models 1 and 3.
In Models 1 and 3, the IRRs are both larger than 0.99, indicating that one unit (0.1 miles) increase in the distance to bike share
stations decreases weekly use by only 0.3% (1 − 0.997) * 100% and 0.1% (1 − 0.999) * 100%. The IRR in Model 2 is comparatively
larger (IRR = 0.889), indicating that increasing the distance to bike stations from members’ residents by one unit (0.1 miles) de-
creases weekly use by a factor of 0.889. Conversely, one unit (0.1 miles) decrease in distance to bike stations increases the frequency
of weekly use by 12.5% ((1/0.889 − 1) * 100%). This comparison may indicate that members who live in areas with lower-density
bike share service (Model 3) are more likely to use bike share for their non-home based trips. As a result, improvements in acces-
sibility to stations around their residences have a smaller effect on their frequency of use. However, members living in areas with
denser bike share service (Model 2) may make more home-based bike share trips, and the improvements in accessibility have larger
effects on their use. The specific mechanisms through which densification increases the frequency of use are not known; in addition to
better accessibility, it could be because densification results in a better connected, and therefore a more flexible and reliable overall
network.
Bike lane length has a significant and positive correlation with weekly use frequency in Model 1 and Model 2 while is not
significant in Model 3. The results indicate that bike facilities investment around members’ homes could help increase their bike share
use only when these members live in dense service areas of bike share and have good bike share accessibility. Specifically, a unit
increase (0.1 miles) in bike lane length close to the home addresses of annual members increases the weekly use by 1.5% (1.015 − 1),
and 2.3% (1.023 − 1) respectively, for Models 1, and 2. However, members not living in denser service areas are more likely to use
bike share for their non-home based trips. The provision of bike infrastructure close to their homes does not have significant impacts
on their use frequency.
The IRRs for the LRT variable are smaller than 1 in Model 1 and Model 2, indicating that members who live in close proximity to
rail transit to less likely to use bike share. The results may indicate that members are more likely to substitute their bike share trips by
using light rail service when it is reasonably available; specifically, members living in dense service areas of bike share would
decrease their bike share by 26% (1 − 0.736) if a LRT station is newly located within a quarter mile buffer of homes. However, the
IRR is greater than 1 in Model 3. The results indicate that members not living in a dense service area of bike share are likely to
increase their bike share use by approximately 7% if a new LRT station is installed close to them. One possible explanation for this
result may be that these members are more likely to combine the two modes to fulfill their daily travel needs (i.e., bike share is
complementary to, and not a substitute for, use of the LRT). As noted, we also estimated the three models using a ½-mile measure and
the criterions for accessibility to bike share. The results (not presented here) were essentially the same for the Distance and Bike lane
length in all three models. The effects of increasing the LRT accessibility measure to ½-mile were consistent in all three models:
frequency of bike share use by members decreased.

4.2. Heterogeneous treatment effects

We develop a different set of models to explore heterogeneous treatment effects across different built environment contexts. We
estimate the model results for all potentially treated annual members living within the city boundaries of Minneapolis and St. Paul
(i.e., the equivalent of Model 1; Table 3).
The treatment variable, the network distance to the nearest bike share station is negatively correlated in all models with the
frequency of weekly use of annual members. Similar to Model 1 in Section 4.1, the effect size of the treatment is small.
In terms of responsiveness to changes in distance, there are few differences between genders: the IRRs for females (0.998) and
males (0.996) are comparable (Table 4). However, male members are more responsive to changes in the provision of bike facilities.
One unit of increase in the length of bike lanes (0.1 miles) increases the frequency of use by 1.6% for males but only 1.1% for females.
The existence of an LRT station close to the residence of male members decreases their weekly use by 35%, indicating they are more
likely to use light rail service to substitute bike share when light rail service is available for them.
The treatment effect (i.e., distance) is significant when interacted with most variables of interest, that is, for annual members
living in areas with higher population density, a higher number of street intersections, and higher percentages of commercial,
recreational, office and industrial land uses. However, the effect sizes are relatively small: the IRRs are generally larger than 0.95. The
treatment effects are larger in areas with higher population density and higher commercial land use. Specifically, one unit decrease in
the distance (0.1 miles) increases weekly use by 3.4% ((1/0.966 − 1) * 100%) for annual members living in densely populated areas
and 3% ((1/0.971 − 1) * 100%) for annual members living in areas with higher commercial land use.

5. Discussion and conclusion

Through analyses of 6-year longitudinal data from Nice Ride Minnesota, we examined the impacts of accessibility, measured as
network distance, to bike share stations on members’ weekly frequency of use. We also explored the heterogeneous impacts of
physical accessibility across areas with different built environment characteristics. This study confirms the relationship suggested by
previous cross-sectional studies and, because we control for changes in accessibility experienced by individual members over time,

8
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

Table 4
Heterogeneous treatment effects.
Model Distance Bike lane length LRT (1/4 mile)

Gender
Female (N = 4039) 0.998* 1.011** 0.929**
Male (N = 4932) 0.996** 1.016** 0.651**

Population density
< Mean (N = 5854) 0.999** 1.005** 0.518*
≥Mean (N = 3192) 0.966** 1.022* 0.898**

Job density
< Mean (N = 7368) 0.996** 1.011** 0.896**
≥Mean (N = 1678) 1.000* 1.017** 0.507**

# Intersections
< Mean (N = 4940) 0.999 1.025** 0.813**
≥Mean (N = 4106) 0.995** 1.005* 0.745**

% Commercial
< Mean (N = 5648) 0.997** 0.994** 0.834**
≥Mean (N = 3398) 0.971** 1.017** 0.705**

% Recreational
< Mean (N = 5537) 1.001 1.003 0.642**
≥Mean (N = 3509) 0.990** 1.038** 1.021
% Office
< Mean (N = 6976) 0.997** 1.005** 0.656**
≥Mean (N = 2070) 0.993** 1.023** 0.944**

% Industrial
< Mean (N = 6764) 0.999 1.017** 0.706**
≥Mean (N = 2263) 0.985** 1.017** 0.796**

Notes:
** Significant at p < 0.01.
* Significant at p < 0.05.

provides stronger evidence of a link between accessibility and frequency of weekly use. The proximity to bike share stations from
residence to bike share stations has a significant impact on bike share use by members: Nice Ride members were more likely to
increase their weekly bike share use after accessibility to Nice Ride stations has been increased. The magnitude of this effect,
however, is very limited. For example, increasing accessibility by 0.1 miles increases weekly use, on average, by only 0.3%. This
means that site location strategies that reduce distance to bike share may increase total use by members only marginally. To further
increase member use, other strategies may be required. For example, accessibility has psychological as well as physical dimensions.
Users may be less likely to use stations with limited light at night, or in areas that are perceived to be less safe. Additional research
will be required to understand how psychological aspects of accessibility affect use and can be addressed.
Our results show that for areas with higher densities of bike share stations, the impact of installing a new station for users is larger,
verifying the importance of densifying bike share service. For areas without good, prior bike share accessibility, the effects of
improving accessibility on weekly use were not as great. Thus, installing new stations in neighborhoods with few stations without
forming/connecting them as part of a dense network system may not increase use significantly.
Consistent with previous studies (Buck and Buehler, 2012; Noland et al., 2016; Wang et al., 2015), the length of on-street bike
facilities is positively correlated with bike share use, providing additional evidence of the importance of investing bike facilities in
cities. Furthermore, the effect size of bike share facility is larger than that of bike share accessibility, illustrating the complementary
and synergistic effects of cycling infrastructure. However, cycling infrastructure is not significantly correlated with use by members in
areas without dense bike share service. This finding underscores the importance of joint investments of bike share and cycling
infrastructure. These results can be used to inform strategies to increase bike share use in particular neighborhoods. For example, in
North Minneapolis, which includes racially concentrated areas of poverty designated by the Metropolitan Council, bike share stations
are underutilized. The provision of additional bike share stations may not be sufficient to increase use unless corresponding in-
vestments in cycling infrastructure are made.
We also examined the heterogeneous treatment effects across different contexts. We found that the effects of increasing acces-
sibility to bike share were similar for female and male members, but that male members were more likely to increase use with greater
availability of bicycle facilities. Other findings are that the treatment effects are larger for annual members living in areas with higher
population density and a higher percentage of retail land use. The results re-confirmed the fact that provision of bike share stations or
increasing physical accessibility to bike share stations is not likely to increase bike share use substantially in all places. Characteristics
of the built environment also affect use and need to be considered. In the short term, placing a Nice Ride station in an area with
higher population density and a higher percentage of retail land use can increase bike share uses of members
Our research has several limitations that can inform future research. We only modeled the distance from the home residences of
members to stations; we did not study or model the distance from members’ places of employment to stations. It is probable that some

9
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

bike share members use bike share service to complete their last mile trips near their workplace locations. Future studies that
examine the accessibility of workplace to bike share stations would be helpful in developing a more complete understanding of bike
share use.
Another limitation is that some potentially confounding factors that may affect bike share use were not controlled in our research
design. Although our use of the DID model controlled for individual-invariant and time-invariant variables, other factors, such as a
growing and better system, or improved levels of service at particular stations, or very local site-specific changes in the built en-
vironment were not directly addressed. These factors potentially could cause bias in our estimation and as a result, result in over-
estimation of the effects of accessibility.
The scope of our study included current bike share members and factors that may affect their use. Another important approach to
increasing bike share use is to increase memberships. Additional studies employing similar DID research designs could help examine
the impacts of accessibility on the probability of becoming a bike share member. From a managerial perspective, we have focused
only on the supply of bike share stations. We have not incorporated these findings in more general benefit-cost analyses that would
help managers consider station location decisions in light of cost constraints. Finally, our analyses provide a framework for the
assessment of programmatic interventions to increase bike share use. That is, our modeling approach is well-suited to studying the
effects of educational, public health, and sales initiatives aimed at increasing ridership. Quasi-experimental research designs that
provide additional evidence to support inferences of causality can maximize the validity of these efforts.

Acknowledgements

The authors would like to acknowledge financial support from NSF Sustainability Research Networks (award 1444745):
“Integrated Urban Infrastructure Solutions for Environmentally Sustainable, Healthy, and Livable Cities.” The authors also thank the
operators of Nice Ride MN for making their ridership data available, and Mr. Mitch Vars to provide us the Nice Ride Member data.

Appendix A. Supplementary material

Supplementary data to this article can be found online at https://doi.org/10.1016/j.tra.2019.01.004.

References

Allison, P.D., 2005. Fixed Effects Regression Methods for Longitudinal Data Using SAS. SAS Institute.
Bachand-Marleau, J., Lee, B., El-Geneidy, A., 2012. Better understanding of factors influencing likelihood of using shared bicycle systems and frequency of use. Transp.
Res. Rec. J. Transp. Res. Board 2314, 66–71.
Buck, D., Buehler, R., 2012. Bike lanes and other determinants of capital bikeshare trips. In: Paper presented at the 91st Transportation Research Board Annual Meeting
2012, Washington, DC.
Buck, D., Buehler, R., Happ, P., Rawls, B., Chung, P., Borecki, N., 2013. Are bikeshare users different from regular cyclists? A first look at short-term users, annual
members, and area cyclists in the Washington, DC, region. Transp. Res. Rec. J. Transp. Res. Board 2387, 112–119.
Campbell, K.B., Brakewood, C., 2017. Sharing riders: how bikesharing impacts bus ridership in New York City. Transp. Res. Part A: Pol. Pract. 100, 264–282.
Clark, J., Curl, A., 2016. Bicycle and car share schemes as inclusive modes of travel? A socio-spatial analysis in Glasgow, UK. Soc. Inclus. 4 (3). https://doi.org/10.
17645/si.v4i3.510.
Corcoran, J., Li, T., Rohde, D., Charles-Edwards, E., Mateo-Babiano, D., 2014. Spatio-temporal patterns of a Public Bicycle Sharing Program: the effect of weather and
calendar events. J. Transp. Geogr. 41, 292–305.
Faghih-Imani, A., Eluru, N., El-Geneidy, A.M., Rabbat, M., Haq, U., 2014. How land-use and urban form impact bicycle flows: evidence from the bicycle-sharing system
(BIXI) in Montreal. J. Transp. Geogr. 41, 306–314.
Faghih-Imani, A., Eluru, N., 2016. Determining the role of bicycle sharing system infrastructure installation decision on usage: case study of Montreal BIXI system.
Transp. Res. Part A: Pol. Pract. 94, 685–698.
Fishman, E., Washington, S., Haworth, N., 2012. Barriers and facilitators to public bicycle scheme use: a qualitative approach. Transp. Res. Part F: Traff. Psychol.
Behav. 15 (6), 686–698.
Fishman, E., Washington, S., Haworth, N., 2013. Bike share: a synthesis of the literature. Transp. Rev. 33 (2), 148–165.
Fishman, E., Washington, S., Haworth, N., Mazzei, A., 2014. Barriers to bikesharing: an analysis from Melbourne and Brisbane. J. Transp. Geogr. 41, 325–337.
Fishman, E., Washington, S., Haworth, N., Watson, A., 2015. Factors influencing bike share membership: an analysis of Melbourne and Brisbane. Transp. Res. Part A:
Pol. Pract. 71, 17–30.
Gebhart, K., Noland, R.B., 2014. The impact of weather conditions on bikeshare trips in Washington, DC. Transportation 41 (6), 1205–1225.
Goodman, A., Cheshire, J., 2014. Inequalities in the London bicycle sharing system revisited: impacts of extending the scheme to poorer areas but then doubling prices.
J. Transp. Geogr. 41, 272–279.
Hampshire, R.C., Marla, L., 2012. An analysis of bike sharing usage: explaining trip generation and attraction from observed demand. In: 91st Annual meeting of the
Transportation Research Board, Washington, DC, pp. 12–2099.
Karaca-Mandic, P., Norton, E.C., Dowd, B., 2012. Interaction terms in nonlinear models. Health Serv. Res. 47 (1pt1), 255–274.
Meddin and DeMaio, 2017. < https://www.google.com/maps/d/viewer?ll=3.33795392692141%2C-42.890624&spn=143.80149%2C154.6875&hl=en&msa=0&
z=1&source=embed&ie=UTF8&om=1&mid=1UxYw9YrwT_R3SGsktJU3D-2GpMU > (accessed in July 2017).
Murphy, E., Usher, J., 2011. An analysis of the role of bicycle-sharing in a European city: the case of Dublin. Irish Transport Research Network (ITRN), August
31st–September 1st. University College Cork, Ireland IRTN.
Noland, R.B., Smart, M.J., Guo, Z., 2016. Bikeshare trip generation in New York city. Transp. Res. Part A: Pol. Pract. 94, 164–181.
Rabe-Hesketh, S., Skrondal, A., 2008. Multilevel and Longitudinal Modeling Using STATA. STATA Press.
O’Brien, O., Cheshire, J., Batty, M., 2014. Mining bicycle sharing data for generating insights into sustainable transport systems. J. Transp. Geogr. 34, 262–273.
Ogilvie, F., Goodman, A., 2012. Inequalities in usage of a public bicycle sharing scheme: socio-demographic predictors of uptake and usage of the London (UK) cycle
hire scheme. Prev. Med. 55 (1), 40–45.
Ricci, M., 2015. Bike sharing: a review of evidence on impacts and processes of implementation and operation. Res. Transp. Bus. Manage. 15, 28–38.
Rixey, R., 2013. Station-level forecasting of bikesharing ridership: Station Network Effects in Three US Systems. Transp. Res. Rec. J. Transp. Res. Board 2387, 46–55.
Romanillos, G., Moya-Gómez, B., Zaltz-Austwick, M., Lamíquiz-Daudén, P.J., 2018. The pulse of the cycling city: visualising Madrid bike share system GPS routes and
cycling flow. J. Maps 14 (1), 34–43.

10
J. Wang, G. Lindsey Transportation Research Part A 121 (2019) 1–11

Rudloff, C., Lackner, B., 2014. Modeling demand for bikesharing systems: neighboring stations as source for demand and reason for structural breaks. Transp. Res. Rec.
J. Transp. Res. Board 2430, 1–11.
Shaheen, S., Guzman, S., Zhang, H., 2010. Bikesharing in Europe, the Americas, and Asia: past, present, and future. Transp. Res. Rec. J. Transp. Res. Board 2143,
159–167.
Shaheen, S., Martin, E., Cohen, A.P., Finson, R., 2012. Public bikesharing in North America: Early operator and user understanding. Mineta Transportation Institute,
San Jose, CA.
Transport for London, 2014. Barclays Cycle Hire Customer Satisfaction and Usage Survey: Members Only. London: London. Retrieved from < http://www.tfl.gov.uk/
cdn/static/cms/documents/barclays-cycle-hire-css-and-usage-members-q3-2013-14.pdf > .
Vogel, P., Greiser, T., Mattfeld, D.C., 2011. Understanding bike-sharing systems using data mining: exploring activity patterns. Proc.-Soc. Behav. Sci. 20, 514–523.
Wang, X., Lindsey, G., Schoner, J.E., Harrison, A., 2015. Modeling bike share station activity: effects of nearby businesses and jobs on trips to and from stations. J.
Urban Plann. Dev. 142 (1), 04015001.
Wang, J., Lindsey, G., 2018. Measuring equity in bike share programs: a case study of the Twin Cities. Transportation Research Board Annual Meeting.
Yang, T., Haixiao, P., Qing, S., 2010. Bike-sharing Systems in Beijing, Shanghai and Hangzhou and their Impact on Travel Behaviour. Paper presented at the
Transportation Research Board Annual Meeting 2011, Washington, DC.

11

You might also like