Journal of Transport Geography 86 (2020) 102769

Contents lists available at ScienceDirect

Journal of Transport Geography

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

How does ridesourcing substitute for public transit? A geospatial perspective T

in Chengdu, China
⁎
Hui Konga,b, Xiaohu Zhangb, , Jinhua Zhaob
a
Singapore-MIT Alliance for Research and Technology, 1 Create Way, Singapore
b
Department of Urban Studies and Planning, Massachusetts Institute of Technology, Cambridge, MA 02139, United States

A R T I C LE I N FO A B S T R A C T

Keywords: The explosive growth of ridesourcing services has stimulated a debate on whether they represent a net substitute
DiDi for or a complement to public transit. Among the empirical evidence that supports discussion of the net effect at
Ridesourcing the city level, analysis at the disaggregated level from a geospatial perspective is lacking. Besides, it remains
Public transit unexplored the spatiotemporal pattern of ridesourcing's effect on public transit, and the factors that impact the
Substitution
effect. Using DiDi Chuxing data in Chengdu, China, this paper develops a three-level structure to recognize the
Complementary
potential substitution or complementary effects of ridesourcing on public transit. Furthermore, this paper in-
vestigates the effects through exploratory spatiotemporal data analysis, and examines the factors influencing the
degree of substitution via linear, spatial autoregressive, and zero-inflated beta regression models. The results
show that 33.1% of DiDi trips have the potential to substitute for public transit. The substitution rate is higher
during the day (8:00–18:00), and the trend follows changes in public transit coverage. The substitution effect is
more exhibited in the city center and the areas covered by the subway, while the complementary effect is more
exhibited in suburban areas as public transit has poor coverage. Further examination of the factors impacting the
relationship indicates that housing price is positively associated with the substitution rate, and distance to the
nearest subway station has a negative association with it, while the effects of most built environment factors
become insignificant in zero-inflated beta regression. Based on these findings, policy implications are drawn
regarding the partnership between transit agencies and ridesourcing companies, the spatial-differentiated po-
licies in the central and suburban areas, and the potential problems in providing ridesourcing service to the
economically disadvantaged population.

1. Introduction by abundant research and media coverage (Anderson, 2014; Bialik

The explosive growth of app-based, on-demand mobility service
providers (e.g., Uber, Lyft, Grab, DiDi Chuxing, etc.), aka transportation
network companies (TNCs), has put ridesourcing in the spotlight
(Hughes and MacKenzie, 2016; Meyer and Shaheen, 2017; Yu and Peng,
2019). Ridesourcing refers to on-demand services that use mobile de-
vices and applications to connect drivers—people who drive private
cars instead of commercial vehicles—with passengers. The term “ride-
sourcing” is commonly used by transportation researchers, while
practitioners describe themselves as “TNCs” or “mobility service pro-
viders” (MSPs), and media normally use the term “ride-hailing” or
“ride-sharing” (Shaheen et al., 2016). Examples of such services include
UberX, UberXL, Lyft, JustGrab, DiDi Express, DiDi Premier, etc.
The substitution effect of ridesourcing for taxis has been confirmed
et al., 2015; Glöss et al., 2016; Rayle et al., 2016). Besides competing
with the taxi industry, ridesourcing also pulls people away from public
transit. Existing research suggests that ridesourcing both substitutes for
and complements public transit (Jin et al., 2018), but it is not clear to
what extent such effects perform, how the effects evolve over time and
space, and what factors are impacting the effects.
This study uses data of DiDi Chuxing in Chengdu, China, to in-
vestigate the substitution and complementary relationship between ri-
desourcing and public transit. We focus on answering three questions:
(1) To what extent does DiDi substitute for public transit? (2) How does
the degree of substitution change over time and space? In other words,
what are the diurnal and spatial patterns of the relationship? (3) What
are the factors that impact the substitution effect of DiDi on transit? The
answers to these questions are significant for both transportation

⁎
Corresponding author at: Senseable City Lab, Department of Urban Studies and Planning, Massachusetts Institute of Technology, Cambridge, MA 02139, United
States.
E-mail address: zhangxh@mit.edu (X. Zhang).

https://doi.org/10.1016/j.jtrangeo.2020.102769
Received 4 June 2019; Received in revised form 13 May 2020; Accepted 5 June 2020
Available online 22 June 2020
0966-6923/ © 2020 Elsevier Ltd. All rights reserved.
H. Kong, et al.
operation and urban planning.
The rest of the paper is organized as follows. The next section de-
fines substitution and complementary effects and reviews current re-
search on ridesourcing and its relationship with public transit. Section 3
explains the data and methodology used in this study, followed by a
section that presents and discusses the empirical results. Lastly, we
highlight the findings and policy implications and point out the lim-
itations of the current study and directions for future research.
2. Substitution vs. complementarity: research background
Ridesourcing can be a substitute for public transit as it provides an
alternative transport option. When taking ridesourcing, a rider does not
have to walk to a transit stop and wait for the bus/subway but can
instead request a service and be transported by a vehicle directly to the
final destination. Also, the in-vehicle environment is often more com-
fortable than that of public transit. Therefore, although ridesourcing
fares are typically higher, there is the potential for it to substitute for
public transit. On the other hand, ridesourcing can complement public
transit, as it increases the reach and flexibility of public transit's fixed
route, fixed schedule service by providing on-demand services (Hall
et al., 2017). In microeconomics, substitutes and complements are
distinguished by cross elasticity of demand EPA, QB which is formulated
as:
∂QB PA
E PA, QB = ∙ coverage is low. Through simulation (Basu et al., 2018) or matching
∂PA QB (1)
where PA is the price of good/service A and QA is the demand quantity
of good/service A. If EPA, QB > 0, meaning a price increase of A leads to a
demand increase of B, A and B are substitutes; on the other hand, if EPA,
QB < 0, they complement each other; and if EPA, QB = 0, A and B are
independent.
Some research discusses the net substitution or complementary ef-
fect of ridesourcing at the city level. However, similar to its variations
in different cities, the effect also varies spatially within a city area, and
few studies have explored the disaggregated level. This paper thus
proposes a methodology for investigating the substitution or com-
plementary effect in the geospatial context. For each origin-destination
(OD) pair, the effect of ridesourcing on public transit is either substitute
or complementary. We define ridesourcing as a potential substitute for
public transit if riders choose ridesourcing when public transit is ac-
cessible (within a comfortable walking distance to the transit station
and at comparable travel time cost). On the other hand, ridesourcing
complements public transit if it provides services when public transit is
not accessible (transit stations cannot be reached within an easy
walking distance or the trip duration by public transit is significantly
longer). The local effect in an area is determined by aggregating the
effects of all ridesourcing trips.
The limited literature that touched on the relationship between ri-
desourcing and public transit suggested that ridesourcing both sub-
stitutes for and complements public transit, yet insufficient quantitative
evidence has been provided to substantiate this observation. The ar-
guments of substitution were established based on surveys asking riders
questions such as “If ridesourcing is not available, what other trans-
portation modes would you use?” and previous surveys reached dif-
ferent substitution rates of ridesourcing for public transit, owing to
their different study areas, sampling strategies, and questionnaire de-
sign. Rayle et al. (2016) conducted surveys in San Francisco on 380
riders who had either just completed a ridesourcing trip or who had
used ridesourcing within the last two weeks, concluding that 33% of
ridesourcing trips replace public transit, and users do save about 10 min
on average by choosing ridesourcing over public transit. Henao (2017)
conducted a survey of 311 Uber/Lyft passengers in Denver while doing
participant observation as a ridesourcing driver himself and found that
22.2% of passengers would have used public transit for their trip if
ridesourcing was not an option. Gehrke et al.'s (2018) in-vehicle survey
2
Journal of Transport Geography 86 (2020) 102769
of 1000 riders in Metro Boston showed that 42% of respondents would
shift to public transit for their trips if ridesourcing was not available.
Clewlow and Mishra (2017) differentiated various types of public
transit and discovered that TNCs pull people away from the public bus
and light rail but enhance the ridership of heavy rail. The arguments on
complementarity come from both surveys and quantitative analysis.
The survey by the American Public Transportation Association
(Murphy, 2016) concluded that ridesourcing is more likely to substitute
for a trip by private car than for a trip by public transit, and it is more
popular when public transportation operates less (e.g., 8 PM – 4 AM).
Smith (2016) from the Pew Research Center revealed that public transit
usage is highly correlated with Uber usage (56% of regular ridesourcing
users also take public transit regularly) based on a survey of 4787
American adults.
Very few quantitative studies examined the relationship between
ridesourcing and public transit by studying the changes in their rider-
ship when a special case occurs or by conducting simulations of virtual
situations. Hoffmann et al. (2016) suggested that ridesourcing usage
increases over 30% when there is a subway service disruption. Hall
et al. (2017) applied statistical approaches to compare how transit ri-
dership changes in cities after Uber enters the market, and found that
Uber is complementary to transit agencies, as it increased transit ri-
dership by 5% within two years of its entry. Jin et al. (2018) compared
Uber pick-ups in NYC with public transit stops and found that Uber
substitutes for public transit most of the time, except when the transit
algorithms (Stiglic et al., 2018) of virtual systems, studies concluded
that public transit is irreplaceable, and the overall system efficiency
improves if ridesourcing serves as a feeder of the transit system.
Existing studies have revealed both substitution and complementary
relationships between ridesourcing and public transit. However, we still
lack empirical evidence to quantitatively measure the effects in the
geospatial context. The very few quantitative measurements of the re-
lationships only examine the overall substitution/complementary ef-
fect, without investigating whether each ridesourcing trip has the po-
tential to substitute for or complement public transit. In addition, the
spatiotemporal variation of such relationships and how factors impact
the relationship remain unknown. This paper fills in these gaps by: (1)
analyzing how each ridesourcing trip may substitute for/complement
public transit via a three-level structure; (2) revealing how the sub-
stitution varies over time and space; (3) examining how the substitution
is impacted by different factors through regression models.
3. Data and methodology
3.1. Study area
This study examines how DiDi Chuxing substitutes public transit in
Chengdu, China. Chengdu is the capital of Sichuan province in south-
western China. It consists of 11 districts, five county-level cities, and
four counties. The 11 districts constitute what is referred to as ‘Central
Urban Areas’, including five in the city center (Jinjiang, Qingyang,
Jinniu, Wuhou, Chenghua) and six in the suburban areas (Pidu, Xindu,
Qingbaijiang, Longquanyi, Shuangliu, Wenzhou). The rest five county-
level cities (Pengzhou, Dujiangyan, Chengzhou, Qionglai, Jianyang)
and four counties (Jintang, Dayi, Pujiang, Xinjin) constitute ‘Suburban
New Cities’.
Chengdu is one of the most important commercial, cultural, and
transportation centers in Western China. By the end of 2010, over 200
of Fortune 500 companies had set up branches in Chengdu. In 2016,
there were more than 400 bus lines in Chengdu with nearly 12,000
buses in total, 4 subway lines were in operation, and more than 20 lines
are planned to be implemented.
DiDi Chuxing is selected as a representative of ridesourcing services.
It is the largest transportation network company in China, and has
acquired Uber China in 2016. Studies have been focused on its impacts
H. Kong, et al.
Fig. 1. Chengdu city and the study area.
on taxi industry (Nie, 2017), the impacts of ride-splitting services (Chen
et al., 2018), traffic emissions (Sun et al., 2018; Xue et al., 2018), and
the impacts of restriction policies on demand (Sun and Ding, 2019).
Most of the DiDi trips data that we have access to are concentrated in
the 11 districts in the Central Urban Areas, so this study focuses on
these districts instead of the whole city region (Fig. 1).
3.2. Data
We collected five categories of data from various sources, explained
as follows:
(1) DiDi trip data.1 We used the DiDi trip data of November 1, 2016,
which includes 181,172 DiDi (DiDi Express + DiDi Premier) trips.
The data contain timestamps and coordinates of both the pick-up
and drop-off locations. The majority of trips were located in
counties near the central urban areas of Chengdu: City Center, Pidu,
Xindu, Qingbaijiang, Longquanyi, Shuangliu, and Wenjiang, so we
focus on analyzing 181,068 trips that fall inside these counties.
(2) Public transit stops and the operational period. Data of public transit
stops and the operational hours of each stop were obtained from
Baidu Map API, including 20,923 stops (112 subway stations
+20,811 bus stops).
(3) Travel time of public transit. The travel time of public transit was
estimated using the Google Distance Matrix API. For each DiDi trip,
we estimate the travel time of the same trip by public transit.
Departure time is defined as the pick-up time of the DiDi trip. The
API returns the public transit duration as the sum of in-vehicle time,
walking time to and from the public transit stop, and, if there are
transfers, the transfer time. The in-vehicle time is given by con-
sidering both historical traffic conditions and live traffic. The public
transit travel time does not include the waiting time for the first trip
segment, as travelers are assumed to be able to plan their trips
1
Data source: Didi Chuxing GAIA Initiative (https://gaia.didichuxing.com)
3
Journal of Transport Geography 86 (2020) 102769
according to the public transit schedule to minimize waiting time.
This neglect is consistent with the DiDi data, since the DiDi travel
time in our data also ignores the waiting time, whereas the relative
travel time is compared by calculating the transit travel time for all
DiDi trips using Google Map API.
(4) Housing price. The housing price data in July 2018 of 9266 com-
munities were obtained from Lianjia (https://cd.lianjia.com/), a
major real estate trading platform in many Chinese cities.
(5) Point of Interest (POI). POI data were obtained from the 58-POI
website (http://www.poi58.com/search/s/download/chengdu_
pois.html). There are, in total, 124,449 POIs in the study area.
(6) GIS layers. GIS layers were extracted from OpenStreetMap, in-
cluding administrative boundaries and road networks.
3.3. Methods
3.3.1. Quantifying the degree of substitution
3.3.1.1. Identifying the substitution/complementary effect of individual
trips. Based on our definition in the geospatial context, whether a
ridesourcing trip has the potential to substitute for or complement
public transit is examined by how accessible the trip is on public transit:
a ridesourcing trip has the potential of substituting for public transit if
the rider uses ridesourcing when public transit is accessible; otherwise,
it has the potential of complementing public transit, as it provides an
alternative in areas with poorer public transit services. We consider
three criteria to measure public transit accessibility: transit coverage,
travel time difference, and service quality difference, upon which we
propose a three-level structure (Fig. 2) to examine the substitution/
complementary effect of ridesourcing on public transit. This section
proposes the measurement of all three levels. However, due to the data
limitations, this study only conducts analysis to the depth of level 1 and
level 2.
Level 1: transit coverage. At this level, a ridesourcing trip has the
potential of substituting public transit if both pick-up and drop-off
locations are within the transit coverage. This criterion is
H. Kong, et al.
Fig. 2. Three-level structure to differentiate the substitution/complementary
effect between ridesourcing and public transit.
formulated as
S1 = {t ∈ S | ot ∈ bn (p)&dt ∈ bn (p) } (2)
where S1 is the set of all ridesourcing trips identified as potential sub-
stitution, t is an individual ridesourcing trip, S is the collection of all
ridesourcing trips, otand dt are the pick-up and drop-off locations, and
bn(p) is the n-meter buffer of transit stops p.
Transit coverage can be measured using buffer analysis, and areas
inside the buffer are considered “covered” by the transit system (Hawas
et al., 2016). This study measures the transit coverage in terms of both
‘spatial coverage’ (where the transit services are available) and ‘tem-
poral coverage’ (when the services are available).
For ‘spatial coverage’, a 400 m buffer is used to measure transit
coverage in this paper, as it has been recognized as a comfortable
walking distance in previous studies (Demetsky and Bin-Mau Lin, 1982;
Murray et al., 1998; Wu and Murray, 2005; Hawas et al., 2016). Al-
ternatively, some studies use 400 m for subways and buses and 800 m
for suburban railways, as people might be willing to walk a longer
distance to take the suburban railway (Smith and Taylor, 1994; Jin
et al., 2019). In this paper, as the public transit in Chengdu only con-
tains bus and subway, and all the subway lines are in the central areas,
we use 400 m as a threshold distance for all the transit stations.
Regarding ‘temporal coverage’, since public transit service changes
throughout the day (different public transit stops have different oper-
ating hours), we construct the transit coverage for every 10 min, given
that public transit riders' typical wait time is 8 to 10 min (Watkins et al.,
2011). For example, for a DiDi trip that picks up passengers at 9:05 AM
and drops off passengers at 9:25 AM, the trip is classified into S1 only if
its pick-up location is covered by public transit stations operating
during the 9:00–9:10 AM period, and its drop-off location is covered by
public transit stations operating during the 9:20–9:30 AM period.
Level 2: travel time difference between public transit and ridesourcing.
The dichotomy based on transit coverage may overestimate the
substitution: the ridesourcing trips that have either the pick-up or
the drop-off outside the transit coverage are definitively categorized
complementary in level 1; however, the trips within transit coverage
should be deemed complementary to public transit if the travel time
saved by ridesourcing is significant. Therefore, this level of analysis
includes travel time difference as an additional criterion to identify
potential substitution. For ridesourcing trips defined as the potential
substitution in level 1 (t ∈ S1), we estimate the travel time of the
same trip by public transit. The ridesourcing trip is considered a
potential substitution for public transit only if the travel time
4
Journal of Transport Geography 86 (2020) 102769
difference between public transit and ridesourcing is shorter than a
pre-defined threshold, since it indicates that the rider chooses ri-
desourcing when public transit is not significantly time-consuming,
or, in other words, accessible. Formula (3) depicts this criterion as:
S 2 = {t ∈ S1 | TtPT − TtRS ≤ τT } (3)
2
where S is the set of all ridesourcing trips identified as the potential
substitution at level 2, TtRSis the travel time of the ridesourcing trip, TtPT
is the travel time of the same trip should public transit be used, and τT is
the threshold of travel time difference.
Level 3: service quality. Level 3 uses service quality to further dif-
ferentiate the ridesourcing service: if ridesourcing greatly improves
service quality, the trip can be considered complementary. Service
quality is determined by many factors, e.g., the crowdedness of the
vehicle, the in-vehicle environment, privacy, safety, fare, etc. For
example, if the bus/subway is too crowded, using ridesourcing
greatly improves the comfort of users, thus the trip can be con-
sidered complementary. Fare is another important factor to con-
sider. If the fare difference of ridesourcing and public transit is too
large, the ridesourcing trip could be considered as complementary.
Formula (4) depicts this criterion as:
S 3 = {t ∈ S 2 | QtRS − QtPT ≤ τQ} (4)
where S3 is the set of all ridesourcing trips identified as the potential
substitution at level 3, QtRS is the service quality of the ridesourcing
trip, QtPT is the service quality of the same trip should public transit be
used, and τQ is the threshold of service quality difference. The analytics
of Level 3 is not included in this paper, as the data used in this study
does not support the measurement of the service qualities.
3.3.1.2. Quantifying the substitution effect at the aggregate level. The
degree of substitution at the aggregate level could be calculated by:
|Si2 |
Ri =
|Si | (5)
2
where Ri is the substitution rate in area i, Si is the set of ridesourcing
trips being identified as substitution in area i, and Si is the complete set
of ridesourcing trips in area i.
3.3.2. Modeling impacts on the substitution effect
We divide the study area into regular grid cells to facilitate the
measurement of factors influencing the substitution and com-
plementary effects of ridesourcing and public transit. Since 400 m is
adopted as a threshold distance in the analysis of transit coverage, we
use 400 × 400 m grid cells to be consistent with the previous analysis.
Three regression models are used in this paper. The Ordinary Least
Squares (OLS) model is the base model, the spatial autoregression
model is used to accommodate spatial variation, and the zero-or-one
inflated beta regression model (ZOIB) is used to examine impacting
factors, considering the effect of the 0 and 1 values of the dependent
variable.
The OLS model is constructed as follows:
Ri = ei + bi + δi + ε (6)
where Ri is the substitution rate of all ridesourcing trips originating in
cell i; ei are socio-economic factors measured by average housing price
at cell i, which is used as a proxy of average wealth level; bi are built
environment factors; δi represents control variables, including the count
of ridesourcing trips in cell i, average travel time of all ridesourcing
trips in the cell and count of bus stops in the cell; and ε is the un-
observed error term.
The built environment has been considered to influence travel be-
havior (Cervero and Kockelman, 1997; Ewing and Cervero, 2001,
2010). In this study, the built environment consists of four measures:
H. Kong, et al.
POI density, land use diversity, road density, and distance to subway
station. POI density is measured as the number of POIs per cell; road
density is measured by the total length of roads per km2 in cell i; and
distance to subway station is measured by the distance of the cell
centroid to the nearest subway station. The diversity of land use is
measured by the mixture level of different POI categories. We reclassify
the POIs into eight categories (see Table A1 in Appendix A) and cal-
culate the Shannon entropy index to represent the level of diversity
(Shannon, 1948):
n off locations inside the transit coverage area. As discussed in Section
H = − ∑ pj logn pj
j (7)
where H is the entropy (ranging from 0 to 1), pj is the percentage of the
jth category of POI, and n is the number of categories. An entropy of 1
indicates extreme diversity of land use, whereas a value of 0 means the
least diversity (only one category of POIs in the cell).
We further adopt spatial autoregressive models to accommodate
spatial variation in the relationship being modeled. If there is any
spatial dependency presented, the ‘uncorrelated error terms’ and ‘in-
dependent observations’ assumptions in the OLS model are violated,
raising the necessity of using either spatial lag or spatial error models. If
there are omitted spatially correlated covariates influencing the re-
lationship over space, a spatial error model should be used to capture
the correlation of the error terms across different spatial units:
Ri = ei + bi + δi + λWεi + μi (8)
where λ represents the spatial error coefficient and μiis the unobserved
error term. W is the spatial weights matrix, defined by Queen contiguity
neighborhoods.
If the spatial lag exists, suggesting that the dependent variable is
also affected by the neighboring independent variables, a spatial lag
model should be used to incorporate the additional effect of neigh-
boring attribute values:
Ri = ei + bi + δi + ρWRi + μi (9)
where ρ represents the spatial lag coefficient and μi is the unobserved
error term. W is the spatial weights matrix, defined by Queen contiguity
neighborhoods.
Considering that the dependent variable is measured in percentage
and has a high concentration of zeros and ones, it is likely that the
impacts of factors come from the effects of 0 or 1 values. To exclude the
effects of 0 and 1 values, we also applied the zero-or-one inflated beta
regression model (ZOIB), to examine the impacting factors without
considering the extreme samples (Ferrari and Cribari-Neto, 2004;
Ospina and Ferrari, 2012).
4. Results and discussion
4.1. A graphical representation of DiDi and public transit services
Fig. 3 shows the distribution of DiDi pick-ups (Fig. 3a) and public
transit stops (Fig. 3b). Both modes are highly concentrated in the city
center.
To explore the diurnal variation of the two modes, we plot the
transit coverage ratio and count of DiDi trips based on their pick-up
time (Fig. 4). Transit coverage is measured by the 400 m buffer of op-
erational transit stops, and the transit coverage ratio is calculated as the
sum of transit coverage areas over the total study area. Due to data
limitations, we didn't eliminate areas where transit will not occur, such
as parks, forests, and water bodies. In general, the coverage ratio is
below 50% in Chengdu, and it declines significantly in the evening
(after 6:00 PM), while the DiDi pick-ups, although showing a downward
trend in the evening, do not decline as dramatically as the transit
coverage ratio. Also, most public transit lines don't operate from mid-
night to the early morning hours (around 12:00 AM to 6:00 AM), while
5
Journal of Transport Geography 86 (2020) 102769
DiDi remains active. These observations imply poor public transit ser-
vice during late-night and early morning hours.
4.2. Relationship between DiDi and public transit
As specified in Section 3.3, we differentiate the ridesourcing trips
that are potential substitution and complementary for public transit at
levels 1 and 2 (based on transit coverage and travel time differences).
For the analysis at level 1, we identified the trips with pick-up and drop-
3.3.1, 400-m buffer is used to measure the ‘spatial coverage’ as it has
been recognized as a comfortable walking distance in previous studies.
We also conduct the sensitivity analysis that calculates the percentage
of trips being recognized as potential substitution (overall substitution
rate) when different buffer radius are used (Fig. A1). It is obvious that,
when the threshold is smaller than 400 m, the overall substitution rate
increases as the buffer radius is larger, but the substitution rate remains
relatively flat when the threshold is greater than 400 m. For ‘temporal
coverage’, transit coverage is created for each 10-min time slot, and the
DiDi trips that may substitute for or complement public transit are
classified based on the criteria of whether both pick-up and drop-off
locations are within the transit coverage. Fig. 5(a) plots the overall
substitution rate and the public transit coverage of the whole study
area. Not surprisingly, at midnight and during the early morning hours
(12:00 AM to 6:00 AM), most DiDi trips have both pick-up and drop-off
locations outside the transit coverage. From 7:00 to 20:00, the sub-
stitution effect of DiDi for public transit is evident: around 90% of trips
have both pick-ups and drop-offs inside the transit coverage, while after
20:00, the complementary effect increases and becomes dominant
again (Fig. 5a). Regarding the relationship in different spatial regions,
the central area exhibits the highest transit coverage ratio (82.13%),
and 79.75% trips are identified as potential substitutions; while in the
Qingbaijiang county, whose transit coverage ratio (27.51%) is the
lowest among all counties, more than half of the trips are considered
complementary (Fig. 5b).
Further, we move on to level-2 analysis and differentiate potential
substitution and complementary trips based on the travel time differ-
ence. For all DiDi trips whose pick-ups and drop-offs are both inside the
transit coverage, we estimate the travel time if they were taken by
public transit (Appendix B). If the travel time by public transit is sig-
nificantly longer, we consider the DiDi trip as complementary instead of
substitution.
Based on transit coverage and travel time difference, we identify the
DiDi trips that may substitute for public transit by formula (3). The
substitution rate (percentage of DiDi trips that has the potential to
substitute for public transit) is calculated using formula (5) contingent
on the threshold (τT) of the travel time difference (Fig. 6). If we set τT as
10 min, which means a DiDi trip is considered as potential substitution
if the travel time difference is less than 10 min, then only around 5% of
the total DiDi trips have the potential to substitute for public transit. As
the benchmark of travel time difference increases, the substitution ef-
fect increases. When τT approximates 43 min, the substitution rate is
greater than 50%, indicating that the substitution effect surpasses the
complementary effect. We also conduct the sensitivity analysis con-
sidering the change of buffer radius threshold and travel time difference
threshold (Fig. A2). Based on the fact that ridesourcing saves about
20 min travel time than public transit (Schwieterman, 2019), and that
Google Distance Matrix API tends to overestimate the travel time on
public transit, we choose τT = 30 min to further examine the spatial
and temporal patterns of the relationship. Under this condition, we
identify that 33.1% of DiDi trips have the potential to substitute for
public transit.
4.3. Spatiotemporal variation of the relationship
To depict the spatiotemporality of the relationship between DiDi
H. Kong, et al.
Fig. 3. Spatial distribution of (a) DiDi pick-ups; (b) public transit coverage.
and public transit, we plot the substitution rate over time (Fig. 7) and
over space (Fig. 8) with τT = 30 min.
Identifying the substitution trips only by transit coverage results in a
very high substitution rate during the day, but the substitution becomes
less after considering the travel time difference (Fig. 7). Before 6:00, the
complementary effect is dominant, but the substitution effect increases
afterward as most buses and subways start operating. From 8:00 to
18:00, the substitution effect fluctuates around 40% and decreases
gradually after 18:00. The percentage of DiDi trips that has the po-
tential of substituting for public transit reaches two small peaks from
8:00 to 9:00 and around 18:00, which is in accordance with the smaller
travel time difference as shown in Fig. B1.
Compared to the public transit coverage, the overall trend of the
substitution rate follows changes in transit coverage. However, a tem-
poral lag is found from 18:00 to 20:00 in the substitution rate identified
from transit coverage, as the public transit coverage drops from around
45% to 25%, while the substitution rate remains high. This lag effect
could be explained by the spatial distribution of transit stops that end
operation during this time period (see Fig. A3 in Appendix A). The
Fig. 4. Public transit coverage ratio and count of DiDi trips, November 1, 2016.
6
Journal of Transport Geography 86 (2020) 102769
transit stops that end operation from 18:00 to 20:00 are mainly located
in the periphery, while the stops at the city center continue operating.
As most of the DiDi trips occur in the city center, they still highly
overlap with the transit coverage. After 20:00, the operation of most
transit stops in the city center ends; thus, the complementary effect of
DiDi becomes significant. The lag effect is less significant in the sub-
stitution rate identified from both transit coverage and travel time
difference, as the end of operation of transit stops in the periphery still
affects the overall travel time difference.
To explore the spatial pattern, we plot the pick-ups of all DiDi trips
(Fig. 8). Most of the substitution trips aggregate in the city center, while
the peripheral areas are mainly occupied by complementary trips. This
high substitution rate in the city center could be attributed to the higher
public transit coverage and relatively shorter trips, as people are more
likely to use ridesourcing for short-distance rides than for long-distance
ones. In addition, we discover a higher substitution rate in the north-
west-southeast direction and from the city center to the southern per-
iphery. These are two major development belts of Chengdu with hea-
vier traffic and are overlapped with the two oldest subway lines. All the
H. Kong, et al. Journal of Transport Geography 86 (2020) 102769

Fig. 5. The relationship identified from transit coverage: (a) the change of overall substation rate in the study area over time; (b) the percentage of potential
substitution and complementary trips over space.

Fig. 6. Substitution rate with the change of travel time difference threshold (τT).

7
H. Kong, et al. Journal of Transport Geography 86 (2020) 102769

Fig. 7. The change of overall substitution rate in the study area over time.

Fig. 8. Spatial distribution of DiDi trips that have the potential to (a) substitute or (b) complement public transit (τT = 30 min).

evidence shows that the relationship between DiDi and public transit is
correlated with both socio-economic factors and the built environment,
and the substitution effect is more significant as it gets closer to the city
center.

4.4. Impacting factors of the relationship

We employ regression analysis to explore the impact of different

factors on the substitution and complementary effects. To be consistent
with the 400 m buffer we used to measure transit coverage, we divide
the study area into 400 m × 400 m grid cells and aggregate the DiDi
trips to cells based on the pick-up locations. Cells that do not have DiDi
trips or are not covered by public transit are removed. The substitution
rate (Fig. 9) and factors are calculated for all the cells, and cells without
housing price data are interpolated using the ordinary Kriging method.
According to Shaheen et al. (2017), there are different potential ap-
plications of ridesourcing different urban built environment. Therefore,
we assume that the variations in percentage of substitution trips per cell
could be explained by socioeconomic factors, the built environment,
and other spatial factors, and based on our data availability, select
several variables for analysis as listed in Table 1. Descriptive statistics
of all variables are represented (Table 1).
The results are shown in Table 2. Firstly, based on the OLS model, Fig. 9. Substitution rate at 400 × 400 m grid cells for regression.
all factors show significant associations with the substitution rate.
However, in the diagnostics for spatial dependence, high values of the
OLS model (R2 = 0.3815). In addition, the spatial pseudo R2 of the
Moran's I index suggest a positive global spatial autocorrelation, in-
spatial lag model is 0.4357, which is a relatively less optimistic as-
dicating that the spatial autoregression model should be applied to take
sessment of the model fit (Anselin, 2014), still provides a slightly better
into account the spatial heterogeneity.
fit than the OLS model (adjusted pseudo R2 = 0.3803).
The (Robust) Lagrange Multiplier (LM) tests (Table 2) show that the
Regarding the results of spatial lag regression, the coefficients of
spatial lag model fits better in our case. The goodness-of-fit (i.e., pseudo
housing price and built environment factors are all statistically
R2) is 0.6780, indicating that the spatial lag model fits better than the

8
H. Kong, et al. Journal of Transport Geography 86 (2020) 102769

Table 1 significantly negative, indicating that the further away from subway
Descriptive statistics. stations, the substitution effect decreases (and the complementary ef-
Mean Std. dev. Minimum Maximum fect increases). This indicates that DiDi services overlap with the public
transit coverage, as DiDi drivers are more likely to choose areas with
Dependent variable higher density (e.g., downtown, around subway stations) in pursuit of
Substitution rate 0.12 0.19 0 1
more potential passengers. The spatial lag coefficient is positive and
Independent variables
Housing price (×10,000) 1.28 0.35 0.55 2.58
significant, indicating the agglomeration effect of the relationship, and
POI density (×100 POIs per km2) 1.45 1.82 0 27.76 is in accordance with our observations in Section 4.3 that substitution
Land-use diversity (Shannon 0.55 0.28 0 0.95 concentrates in the city center. Sensitivity analysis has been conducted
entropy) for the spatial lag modeling for different travel time thresholds (τT =
Road density (in km/km2) 8.85 5.67 0 43.12
20 min, τT = 40 min), and the findings do not have significant changes,
Count of bus stops 3.08 4.21 0 46.00
Distance to nearest subway station 3.80 3.85 0.04 26.40 indicating the robustness of the results (Table A2).
(in km) In the ZOIB model, after mitigating the impacts of the zeros and
Control variables ones, the effects of most built environment measures become insignif-
Count of DiDi trips (×10) 4.34 10.66 0.10 196.9
icant, indicating that the impacts of built environment on the sub-
Average travel time (in hours) 0.57 0.25 0.09 3.43
stitution mainly inflated by the effects of the 0 (perfectly com-
Note: Std. dev. = Standard Deviation. plementarity) and 1 (perfectly substitution) values. However, housing
price, count of bus stops, and distance to the nearest subway station are
Table 2 still significantly correlated with the substitution rate. The magnitude
Modeling results. of the impact of the distance to a subway station becomes greater in the
ZOIB model. This is in accordance with what we found in Fig. 9: there
OLS Spatial lag ZOIB
are more substitutions between DiDi and transit in areas covered by
Constant −0.098 ⁎⁎⁎
−0.059 ⁎⁎⁎
−1.270⁎⁎⁎ subway lines, and this effect becomes more significant after we ignore
SES factors the zeros and ones in the ZOIB model. Again, sensitivity analysis has
Housing price (×10,000) 0.128⁎⁎⁎ 0.045⁎⁎⁎ 4.154E-05⁎⁎⁎ been conducted for the ZOIB modeling for different travel time
Built Environment
thresholds (τT = 20 min, τT = 40 min), and also reflect robustness re-
POI density (×100 POIs per km2) 0.009⁎⁎⁎ 0.003⁎⁎ 1.526E-04
Land use diversity (Shannon entropy) 0.045⁎⁎⁎ 0.015⁎⁎ −0.121 sults (Table A3).
Road density (in km/km2) 0.002⁎⁎⁎ 0.001⁎⁎ 0.003
Count of bus stops 0.006⁎⁎⁎ 0.002⁎⁎⁎ 0.011⁎⁎ 5. Conclusions and future research
Distance to nearest subway station (in −0.006⁎⁎⁎ −0.002⁎⁎⁎ −0.268⁎⁎⁎
km)
Control variables
To examine the substitution and complementary effects of ride-
Trip counts (×10) 0.005⁎⁎⁎ 0.002 0.001⁎⁎⁎ sourcing on public transit at the disaggregated level from a geospatial
Average travel time (in hours) −0.045⁎⁎⁎ 0.021⁎⁎ −0.061 perspective, this paper develops a three-level structure to categorize
Spatial lag coefficient NA 0.696⁎⁎⁎ NA every ridesourcing trip as potential substitution or complementary
Summary of Statistics
based on transit coverage, travel time difference, and service quality.
Number of observations 4123 4123 4123
(pseudo) R2 0.3815 0.6780 0.1028 This methodology is applied to DiDi Chuxing in Chengdu, which pro-
Adjusted (pseudo) R2 0.3803 NA NA vides empirical data to the debate over whether ridesourcing and public
Spatial pseudo R2 NA 0.4357 NA transit are substitutes or complements.
Diagnostics for spatial dependence Our results corroborate that DiDi both substitutes for and comple-
Moran's I 0.4656⁎⁎⁎ NA NA
Lagrange multiplier (lag) 2543.03⁎⁎⁎ NA NA
ments public transit. In total, 33.1% of DiDi trips are identifies as po-
Robust Lagrange multiplier (lag) 524.86⁎⁎⁎ NA NA tential substitutions for public transit when the travel time difference
Lagrange multiplier (error) 2033.81⁎⁎⁎ NA NA threshold is set as 30 min. Based on this result, we gain a good under-
Robust Lagrange multiplier (error) 15.64⁎⁎⁎ NA NA standing of the relationship at fine spatial and temporal scales. During
the day, around 40% of DiDi trips have the potential to substitute for
Note: NA = not applicable.
⁎⁎⁎ public transit, and this substitution rate decreases in the evening when
p-value < .001.
⁎⁎
p-value < .01. the supply of transit decreases. In the spatial dimension, the substitu-
tion effect is more significant in the city center and in the more de-
significant. Housing price is positively correlated with the substitution veloped areas covered by subway lines, while the peripheral areas are
rate. An increase of 10,000 Yuan in the housing price leads to about dominated by complementary trips. This indicates that the substitution
4.5% increase in the substitution rate, holding all other variables con- effect of DiDi is highly concentrated and is correlated with socio-eco-
stant. Housing price is an indicator of the average wealth level locally, nomic and built environment factors. To examine how different factors
so our results indicate that wealthier people are more likely to use DiDi. impact the relationship between DiDi and public transit, we further
Considering the built environment, POI density, land use diversity, and apply the spatial lag model and ZOIB model to study how the sub-
road density are positively correlated with the dependent variable and stitution rate is influenced by housing price, built environment, and
are significant at a 99% confidence interval. The coefficient of density is spatial lag factors. The results prove that housing price, distance to the
0.003, indicating an increase of 0.3% in the substitution rate, given that nearest subway station, and spatial lag effects have significant effects
the number of POIs increases by 100. This reflects the instinct of ride- on the relationship, while the influence of the built environment is less
sourcing toward pursuing profit: they tend to provide service and significant.
substitute for public transit in areas with denser activities. The positive Our results provide implications for policymaking. First, transit
coefficient of diversity means that the mix of land use is correlated with agencies are suggested to collaborate with TNCs to improve the first/
the higher substitution rate. Road density has a positive correlation last mile connection. However, ridesourcing not only complements but
with the substitution rate, which could be explained by the higher PT also substitutes for public transit, and public agencies should take
coverage in areas with denser roads. The count of bus stops is positively precautions prior to partnering with TNCs to serve areas with poor
correlated with the substitution rate at a 99.9% confidence interval. transportation connectivity while not causing disruptive changes in
The coefficient of ‘distance to the nearest subway station’ is public transit. It is also essential for public agencies to improve the
integration of infrastructure, information, and fare, with the goal to

9
H. Kong, et al.
encourage public transit and TNCs integration (Shaheen et al., 2016).
In addition, both the spatiotemporal analytics of the relationship
and the examination of factors indicate spatial heterogeneity of the
substitution effect. Therefore, transportation planners may consider
adopting different strategies in different areas. In urban peripheries,
where public transit has poor coverage, the complementary effect of
DiDi indicates that there are travel demands the current transit system
does not satisfy. It should be noted that the complementary effect does
not always suggest planning more transit routes because travel demand
may not be high enough to support transit operation. Instead, lever-
aging ridesourcing services to connect them with the transit system
would improve the overall efficiency of the system. However, re-ex-
amining areas with high complementary effects is necessary because
some travel demands may point to potential transit routes. On the other
hand, in the city center, there is high transit coverage, and the sub-
stitution effect of DiDi on transit is significant. In this case, DiDi is
pulling people away from transit-dependence to more car-dependence.
A challenging question for policymakers, then, is whether we should
rely on TNCs to improve urban mobility or continue to invest in and
improve public transit networks. As public transit has been proven to
have advantages over private automobiles in serving disadvantaged
groups (Bullard, 2003; Dodson et al., 2004), relieving congestion
(Pucher, 2004), and consuming less energy (Garrison and Levinson,
2014), policymakers may want to lessen the substitution of DiDi for
transit by regulating the spatial distribution of DiDi services through
strategic pricing or spatial quotas.
Furthermore, as ridesourcing substitutes a large proportion of public
transit and becomes an important mobility mode, policymakers have to
rethink about transporation equity of ridesourcing services. After all,
most of the ridesourcing services rely on smartphone apps and credit
card fare-paying. As a result, the unbanked population and the popu-
lation that do not own a smartphone may not have access to ride-
sourcing services (Taylor et al., 2015). How to provide mobility service
to such population is something that policymakers have to figure out.
Despite the significance of our results, this study has several lim-
itations that necessitate further research. First, the DiDi data is on
November 1, 2016, but the bus/subway stop information was crawled
in August 2018.2 Since the public transit system does not change sig-
nificantly, the mismatch of time periods should only slightly affect the
results. Second, in the dichotomy based on transit coverage, there
should be two types of effect: feeder and independence. DiDi comple-
ments public transit as a potential feeder if it delivers riders to and from
transit stops to use transit services, but it is independent from public
transit if the trip only fills the gaps in transit service without connecting
riders to the transit system. Future research could differentiate feeders
and independent trips by combining rider surveys with trip record
analytics. Third, this study has not differentiated the impacts of ride-
sourcing on bus and subway services. Not differentiating bus and
subway is reasonable for this specific study, as there were only four
subway lines in Chengdu in 2016, which accounted for quite limited
transit service in the whole study area, and all the subway lines locate
around the city center (studies using a different distance threshold for
rail are based on the assumption that people are willing to walk a longer
distance to rail stations in the suburban areas). However, if the similar
approaches are applied to a different city, it should be considered
carefully whether to apply a different distance threshold for the bus and
subway coverage. Lastly, it is important to remember that a trip con-
sidered as a potential substitution for transit may actually substitute a
trip by private automobile, taxi, cycling, etc. Surveys could be con-
ducted, as has been done in many previous studies (Rayle et al., 2016;
2 Jin, S.T., Kong, H., Sui, D.Z., 2019. Uber, public transit, and urban transportation equity:
For the subway, we have removed the two lines operating after November
2016 (Subway Line 7 and Subway Line 10); for buses, we don't have further
information about whether each bus line has started operated in November
2016.
10
Journal of Transport Geography 86 (2020) 102769
Henao, 2017), to figure out the actual substitution of ridesourcing on
the other modes.
Acknowledgment
The research is supported by the National Research Foundation
(NRF), the CREATE Programme from the Singapore Prime Minister's
Office, and the Singapore-MIT Alliance for Research and Technology
(SMART) Centre, Future Urban Mobility (FM) Interdisciplinary
Research Group. The DiDi trip data is shared by DiDi Chuxing GAIA
Initiative (https://gaia.didichuxing.com). The authors would like to
thank Baichuan Mo for helping with the public transit travel time es-
timation.
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://
doi.org/10.1016/j.jtrangeo.2020.102769.
References
Anderson, D.N., 2014. “Not just a taxi”? For-profit ridesharing, driver strategies, and
VMT. Transportation 41 (5), 1099–1117.
Anselin, L., 2014. Modern Spatial Econometrics in Practice: A Guide to GeoDa,
GeoDaSpace and PySAL. GeoDa Press.
Basu, R., Araldo, A., Akkinepally, A.P., Nahmias Biran, B.H., Basak, K., Seshadri, R., ...
Ben-Akiva, M., 2018. Automated mobility-on-demand vs. mass transit: a multi-modal
activity-driven agent-based simulation approach. Transp. Res. Rec. 2672 (8),
608–618.
Bialik, C., Flowers, A., Fischer-Baum, R., Mehta, D., 2015. Uber is serving New York’s
outer boroughs more than taxis are. In: FiveThirtyEight, Available at: https://
fivethirtyeight.com/features/Uber-is-serving-new-yorks-outer-boroughs-more-than-
taxis-are/ (accessed 3 August 2018).
Bullard, R.D., 2003. Addressing urban transportation equity in the United States.
Fordham Urban Law J. 31 (5), 1183.
Cervero, R., Kockelman, K., 1997. Travel demand and the 3Ds: density, diversity, and
design. Transp. Res. Part D: Transp. Environ. 2 (3), 199–219.
Chen, X., Zheng, H., Wang, Z., Chen, X., 2018. Exploring impacts of on-demand rides-
plitting on mobility via real-world ridesourcing data and questionnaires.
Transportation 1–21.
Clewlow, R.R., Mishra, G.S., 2017. Disruptive Transportation: The Adoption, Utilization,
and Impacts of Ride-hailing in the United States. University of California, Davis,
Institute of Transportation Studies, Davis, CA (Research Report UCD-ITS-RR-17-07).
Demetsky, M.J., Bin-Mau Lin, B., 1982. Bus stop location and design. Transp. Eng. J. ASCE
108 (4), 313–327.
Dodson, J., Gleeson, B., Sipe, N.G., 2004. Transport Disavantage and Social Status: A
Review of Literature and Methods. Urban Policy Program, Griffith University,
Australia.
Ewing, R., Cervero, R., 2001. Travel and the built environment: a synthesis. Transp. Res.
Rec. 1780, 87–114.
Ewing, R., Cervero, R., 2010. Travel and the built environment: a meta-analysis. J. Am.
Plan. Assoc. 76 (3), 265–294.
Ferrari, S., Cribari-Neto, F., 2004. Beta regression for modelling rates and proportions. J.
Appl. Stat. 31 (7), 799–815.
Garrison, W.L., Levinson, D.M., 2014. The Transportation Experience: Policy, Planning,
and Deployment. Oxford University Press.
Gehrke, S.R., Felix, A., Reardon, T., 2018. Fare choice: a survey of ride-hailing passengers
in Subway Boston. In: MAPC Research Brief, Available at. http://www.mapc.org/
wp-content/uploads/2018/02/Fare-Choices-MAPC.pdf (accessed 21 August 2018).
Glöss, M., McGregor, M., Brown, B., 2016, May. Designing for labour: uber and the on-
demand mobile workforce. In: Proceedings of the 2016 CHI conference on human
factors in computing systems. ACM, pp. 1632–1643.
Hall, J.D., Palsson, C., Price, J., 2017. Is Uber A Substitute or Complement For Public
Transit? (No. tecipa-585).
Hawas, Y.E., Hassan, M.N., Abulibdeh, A., 2016. A multi-criteria approach of assessing
public transport accessibility at a strategic level. J. Transp. Geogr. 57, 19–34.
Henao, A., 2017. Impacts of Ridesourcing-Lyft and Uber-on Transportation Including
VMT, Mode Replacement, Parking, and Travel Behavior. University of Colorado at
Denver.
Hoffmann, K., Ipeirotis, P., Sundararajan, A., 2016. Ridesharing and the Use of Public
Transportation.
Hughes, R., MacKenzie, D., 2016. Transportation network company wait times in Greater
Seattle, and relationship to socioeconomic indicators. J. Transp. Geogr. 56, 36–44.
Jin, S.T., Kong, H., Wu, R., Sui, D.Z., 2018. Ridesourcing, the sharing economy, and the
future of cities. Cities 76, 96–104.
a case study in New York City. Prof. Geogr. 1–16.
Meyer, G., Shaheen, S., 2017. Disrupting Mobility. Lecture Notes in Mobility. Springer
International Publishing, Cham.
H. Kong, et al.
Murphy, C., 2016. Shared mobility and the transformation of public transit. In: Technical
Report, Available at: https://www.apta.com/resources/reportsandpublications/
Documents/APTA-Shared-Mobility.pdf (accessed 1 September 2018).
Murray, A.T., Davis, R., Stimson, R.J., Ferreira, L., 1998. Public transportation access.
Transp. Res. Part D: Transp. Environ. 3 (5), 319–328.
Nie, Y.M., 2017. How can the taxi industry survive the tide of ridesourcing? Evidence
from Shenzhen, China. Transp. Res. C 79, 242–256.
Ospina, R., Ferrari, S.L., 2012. A general class of zero-or-one inflated beta regression
models. Comput. Stat. Data Anal. 56 (6), 1609–1623.
Pucher, J., 2004. Chapter 8. Public Transportation. Guilford, pp. 199–236.
Rayle, L., Dai, D., Chan, N., Cervero, R., Shaheen, S., 2016. Just a better taxi? A survey-
based comparison of taxis, transit, and ridesourcing services in San Francisco. Transp.
Policy 45, 168–178.
Schwieterman, J.P., 2019. Uber economics: evaluating the monetary and travel time
trade-offs of transportation network companies and transit service in Chicago,
Illinois. Transp. Res. Rec. 2673 (4), 295–304.
Shaheen, S., Cohen, A., Zohdy, I., 2016. Shared Mobility: Current Practices And Guiding
Principles (No. FHWA-HOP-16-022). Federal Highway Administration, United States.
Shaheen, S., Cohen, A., Yelchuru, B., Sarkhili, S., Hamilton, B.A., 2017. Mobility on
Demand Operational Concept Report (No. FHWA-JPO-18-611). Department of
Transportation. Intelligent Transportation Systems Joint Program Office, United
States.
Shannon, Claude E., 1948. A mathematical theory of communication. Bell Syst. Tech. J.
27 (3), 379–423. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x.
(July–October).
Smith, A., 2016. Shared, Collaborative and On Demand: The New Digital Economy. Pew
Research Center Available at: http://www.pewinternet.org/2016/05/19/the-new-
11
Journal of Transport Geography 86 (2020) 102769
digital-economy/ (accessed 5 September 2018).
Smith, P.N., Taylor, C.J., 1994. A method for the rationalisation of a suburban railway
network. Transp. Res. A Policy Pract. 28 (2), 93–107.
Stiglic, M., Agatz, N., Savelsbergh, M., Gradisar, M., 2018. Enhancing urban mobility:
integrating ride-sharing and public transit. Comput. Oper. Res. 90, 12–21.
Sun, D.J., Ding, X., 2019. Spatiotemporal evolution of ridesourcing markets under the
new restriction policy: a case study in Shanghai. Transp. Res. A Policy Pract. 130,
227–239.
Sun, D.J., Zhang, K., Shen, S., 2018. Analyzing spatiotemporal traffic line source emis-
sions based on massive didi online car-hailing service data. Transp. Res. Part D:
Transp. Environ. 62, 699–714.
Taylor, B.D., Chin, R., Crotty, M., Dill, J., Hoel, L.A., Manville, M., ... Zafar, M., 2015.
Between public and private mobility: examining the rise of technology-enabled
transportation services. Spec. Rep. 319 Available at : http://online. pubs.trb.org/
onlinepubs/sr/sr319.pdf.
Watkins, K.E., Ferris, B., Borning, A., Rutherford, G.S., Layton, D., 2011. Where is my
bus? Impact of mobile real-time information on the perceived and actual wait time of
transit riders. Transp. Res. A Policy Pract. 45 (8), 839–848.
Wu, C., Murray, A.T., 2005. Optimizing public transit quality and system access: the
multiple-route, maximal covering/shortest-path problem. Environ. Plan. B 32 (2),
163–178.
Xue, M., Yu, B., Du, Y., Wang, B., Tang, B., Wei, Y.M., 2018. Possible emission reductions
from ride-sourcing travel in a global megacity: the case of Beijing. J. Environ. Dev. 27
(2), 156–185.
Yu, H., Peng, Z.R., 2019. Exploring the spatial variation of ridesourcing demand and its
relationship to built environment and socioeconomic factors with the geographically
weighted Poisson regression. J. Transp. Geogr. 75, 147–163.