Content uploaded by Mingyang Du
Author content
All content in this area was uploaded by Mingyang Du on Apr 14, 2023
Content may be subject to copyright.
1
Spatial Variation of Ridesplitting Adoption Rate in Chicago
Mingyang Dua,b, Lin Chengb,*, Xuefeng Lia, Qiyang Liuc, Jingzong Yangd,*
aCollege of Automobile and Traffic Engineering, Nanjing Forestry University, Nanjing,
Jiangsu 210037, China
bSchool of Transportation, Southeast University, Nanjing, Jiangsu 211189, China
cSchool of Urban Planning and Design, Peking University Shenzhen Graduate School,
Shenzhen, Guangdong 518055, China
dSchool of Big Data, Baoshan University, Baoshan, Yunnan 678000, China
*Corresponding author at: School of Transportation, Southeast University, Nanjing, Jiangsu
211189, China.
E-mail addresses: dumingyangseu@foxmail.com (M. Du), gist@seu.edu.cn (L. Cheng),
lixuefengseu@foxmail.com (X. Li), tsql@pku.edu.cn (Q. Liu), yjingzong@foxmail.com (J.
Yang).
Abstract
Ridesplitting, a form of ride-hailing service where passengers with similar travel routes
are matched to the same driver, can reduce the negative effects of solo ride-hailing trips and
bring various environmental and social benefits. However, limited efforts were made to
examine the spatial variation of ridesplitting trips, which was not conducive to the formulation
of ridesplitting policies. To fill the gap, this work investigates the spatial variation of
ridesplitting adoption rate (the proportion of ride-hailing trips with shared trip authorized, RAR)
and its association with built environment and socio-economic factors at the census tract level,
using the ride-hailing trip data in Chicago. To addressing the spatial heterogeneity,
geographically weighted regression models are established to detect the factors influencing the
RAR during different time periods, such as weekday, weekend, weekday morning peak and
evening peak. Modeling results show that GWR models outperform the traditional global
models in terms of model fit. The census tract level factors including subway station density,
frequency of transit, land use mix, homicide density, percent female, the share of nonwhite,
and percent zero-vehicle households have impacts on RAR, and the coefficient estimates of
each explanatory variable vary across regions. The research results can help urban planners and
transportation network companies develop refined policies to promote shared ride-hailing trips.
Keywords: Ridesplitting adoption rate; Spatial variation; Built environment; Socio-economic
factor; Geographically weighted regression
2
1. Introduction
Benefitting from the rapid development and popularization of information technology and
mobile payment, the world has witnessed an explosion in ride-hailing services (Wang and Yang,
2019). Based on smart phone applications, transportation network companies (TNCs), such as
Uber, Lyft, Grab, DiDi Chuxing, can integrate the travel information promptly and match
passengers and vehicles accurately, which effectively alleviates the problem of the difficulty in
getting a taxi and improves passengers’ travel experience (Shen et al., 2020; Du et al., 2020;
Yang et al., 2021). Although ride-hailing services have numerous social advantages, the studies
have shown that it increases vehicle miles traveled and traffic congestion, and causes many
negative effects due to the increasing empty vehicles travel on the road, i.e., deadheading
(Morris et al., 2019; Kang et al., 2021; Anderson, 2014; Henao and Marshall, 2019). One
potential and promising way to alleviate these problems is to encourage more ride-hailing users
to utilize a more sustainable travel mode, i.e., ridesplitting service (also called shared ride-
hailing or ride-pooling), such as Uber Pool, Lyft Shared, DiDi ExpressPool, etc. (Chen et al.,
2018; Li et al., 2020; Li et al., 2019). It is a form of ride-hailing service in which passengers
with similar origins and destinations are matched to the same driver in real time (Shaheen et
al., 2016), and the travel cost are shared by the customers in a vehicle (Wang et al., 2019;
Shaheen and Cohen, 2019; Martin and Shaheen, 2011). Encouraging shared over solo ride-
hailing trips could offer a win-win-win proposition for TNCs, ride-hailing riders, and cities.
For riders, through the use of empty seats in the vehicles (increase vehicle occupancy), the
ridesplitting service makes it easier for customers to get a ride with fewer costs, especially
under the condition of limited vehicle supply (Wang et al., 2019; Brown, 2020). For TNCs,
sharing rides with other ride-hailing users during rush hours can increase the order response
rate and ridesplitting matching success rate, and improve the operation efficiency of the
platform and the attraction of the ridesplitting system (Agatz et al., 2011; Alonso-González et
al., 2020). For cities, compared with solo ride-hailing trips, ridesplitting can reduce the use of
vehicles (Alonso-Mora et al., 2017), and thus reduce traffic congestion, traffic accidents, and
parking demand on the roads (Morris et al., 2019). This service can also bring various
environmental benefits, such as reducing energy consumption and harmful gas emissions (e.g.
CO, SOx, NOx) (Yan et al., 2020b; Amey et al., 2011; Chen et al., 2018). However, in actual
operation, the significant ridesplitting adoption rate (RAR, the proportion of ride-hailing trips
with shared trip authorized, which reflects the willingness of ride-hailing riders to share rides)
has yet to become a reality. For example, the RAR in Chicago is 23.7% (Hou et al., 2020), and
that in Toronto is only 14.8% (Young et al., 2020), the proportion of ridesplitting trips in total
ride-hailing trips is 19% and less than 7% in Hangzhou and Chengdu, respectively (Chen et al.,
2018; Li et al., 2019). Given that the benefits of ridesplitting are encouraging, it is crucial to
understand ride-hailing riders’ willingness to utilize it and what factors that promote or deter
this willingness.
Previous studies have suggested that the spatial heterogeneity analysis of the impacts of
3
built environment and socio-economic attributes on the ride-hailing demand can be useful for
TNCs and policy makers to implement differentiated and customized management strategies
and planning policies for different local regions (Yu and Peng, 2019; Wang and Noland, 2021).
However, to our best knowledge, there is a paucity of literature that examined the spatial
variation of ridesplitting trips, especially the spatial heterogeneity of the effect of various
factors on RAR, and its differences during different time periods, such as weekday and
weekend, morning peak and evening peak on weekday, which limits the relevant policy
formulation and investment decisions. Better understanding these issues can bring the
following benefits. For TNCs, knowing the spatio-temporal variation of RAR could inform
TNCs that when and where to implement ridesplitting incentive strategies, such as discount
pricing and coupons, if they aim to promote shared trips, so as to improve existing ridesplitting
services. For policy makers: (i) identifying the relationship between RAR and built
environment and socio-economic factors can support urban planners to develop more refined
and flexible ridesplitting policies, i.e., location-based and time-based ridesplitting incentive
measures. For example, the government can cooperate with TNCs to promote ridesplitting/
micro-transit service in areas where fixed-route transit services are deficient or not available
during peak hours. (ii) It is helpful for policy makers to better balance the relationship between
ridesplitting and other more sustainable travel modes, such as transit and active modes of
transport, and better position the function and role of this service. (iii) The findings can help
cities prepare for the emergence of shared autonomous vehicles with dynamic ride-sharing
among strangers in the future (Xu et al., 2021; Merlin, 2017; Fagnant and Kockelman, 2018).
The contributions of our work are as follows. First, to capture the spatial nonstationarity
of the influences of built environment and socio-economic factors on RAR, geographically
weighted regression (GWR) models are applied to explore the spatial variation of RAR and its
influential factors at census tract level, based on the ride-hailing trip data in Chicago. Second,
we examine the differences of influential factors of RAR during four time periods, i.e., weekday,
weekend, weekday morning peak and evening peak. Finally, some policy suggestions and
implications are proposed to promote shared ride-hailing trips from the perspectives of policy
makers and TNCs.
2. Literature review
This section reviews the literatures regarding the ridesplitting. Section 2.1 summarizes the
studies of the ridesplitting willingness analysis based on the questionnaire data. Section 2.2
reviews the research on operational characteristics of the ridesplitting service. Section 2.3
focuses on the influences of built environment and socioeconomic attributes on ride-hailing
and ridesplitting trips based on the trip record data of TNCs, and section 2.4 discusses the
research gaps.
2.1. Willingness to use ridesplitting
Exploring passengers’ willingness to use the ridesplitting is significant for TNCs to
formulate ridesplitting incentive strategies (Brown, 2020). Recently, numerous studies have
4
explored the impacts of individual attributes, travel attributes and attitude perceptions on
passengers’ willingness to utilize the ridesplitting based on questionnaire data (Kang et al.,
2021; Lavieri and Bhat, 2019). For example, Sarriera et al. (2017) analyzed the social and
behavioral factors that influenced dynamic ride-sharing’s use based on the survey data in the
US. They revealed that the young (under 30), the individuals without cars, and unmarried
people tended to use the dynamic ridesharing. The main purpose of using this service was for
leisure, airport trips, and commuting to school or work. Shorter travel time, cheaper travel cost
and comfortability compared to transit and walking were main motivations for using this
service, while being paired with unpleasant passengers, the uncertainty of the trip duration, the
privacy sensitivity and security issues (especially for women) were the barriers to use it. Lavieri
and Bhat (2019) studied the attribute characteristics of ridesplitting users in the Dallas-Fort
Worth metropolitan area. The results showed that individuals with low income were more
inclined to adopt pooled ride-hailing, and privacy sensitivity was a concern, especially for non-
Hispanic Whites and the wealthiest population. Kang et al. (2021) developed a joint revealed
preference-stated preference model for the choice between pooled and private ride-hailing
service, using the survey data in Austin, Texas. The model results indicated that well-educated,
employed individuals, and those living in high density urban areas preferred the pooled ride-
hailing service, while the old people, women, and non-Hispanic/non-Latino whites tended not
to choose this service. Wang et al. (2019) investigated how ridesplitting services influenced the
travel behavior of young people, based on the survey data in Hangzhou, China. They compared
two ridesplitting services, e.g. DiDi Hitch and DiDi ExpressPool, the former was a ride-pooling
platform that helped commuters share rides, while the latter was served by operation drivers of
DiDi Express service. The results demonstrated that ridesplitting users tended to be married,
young and well-educated individuals without private cars. The purpose of choosing this service
was mainly for work or home-based trips, and the travel cost and parking problems were two
major motivations for users to utilize this service. Alonso-Gonzalez et al. (2020) explored the
determinants of shared ride-hailing use in the Netherlands. They discovered that ridesplitting
users mainly considered the trade-off between time and cost rather than the discomfort
associated with ridesplitting. The ridesplitting price discounts were usually inadequate owing
to the low matching success rate, the willingness of individuals to share rides depended on the
number of paired passengers.
In addition, social class discrimination and social trust also affected the usage of
ridesplitting. Based on the online survey data in the US, Moody et al. (2019) focused on the
associations between rider-to-rider race and social class discrimination attitudes and the use of
shared ride-hailing service. The results demonstrated that discriminatory attitudes discouraged
the use of this service. For the users who had used this service, these attitudes might reduce
their frequency of use; while for those who had not yet utilized this service, these attitudes
might lead them to completely avoid utilizing the shared ride-hailing. Amirkiaee et al. (2018)
performed a scenario-based survey to analyze emotional and psychological factors that
influenced the choice between shared and private ride-hailing service. The findings indicated
5
that in situations where transportation anxiety was high, such as construction on the road, social
trust would promote people to choose the ridesharing in the presence of time and economic
benefits.
2.2. Operation characteristics of ridesplitting
Analyzing operation characteristics of the ridesplitting based on the trip data can be helpful
for better evaluating this service (Li et al., 2019; Chen et al., 2018). Based on the trajectory and
order data of the ride-hailing service in Chengdu city, Li et al. (2019) analyzed the detours and
delays caused by the ridesplitting. They concluded that the ratio of ridesplitting was low,
accounting for 6-7%, which was related to the extra detour (1.55 km on average), delay (10
min on average), and the reduction of travel time reliability. Schwieterman and Smith (2018)
discussed the differences of travel time and cost when taking ridesplitting and transit in Chicago.
They concluded that the ridesplitting could reduce trip times and costs by 67.6% and
$0.38/minute on trips between neighborhoods, by comparison, it could reduce trip times and
costs by 13.7% and $1.29/minute on trips from or to the downtown area. Using ride-hailing
trip data of DiDi Chuxing and the survey data in Hangzhou, China, Chen et al. (2018) studied
the impacts of ridesplitting services, i.e. DiDi ExpressPool and DiDi Hitch, on vehicle
kilometers traveled (VKT). The results indicated that ridesplitting services could reduce VKT
by 58,124 km per day, and ExpressPool and Hitch contributed 55,949 km and 2,175 km,
respectively.
In addition, Chen et al. (2017) presented an ensemble learning approach for analyzing the
ridesplitting choice behavior (whether or not a customer would choose to split rides with other
passengers). A variety of important features were ranked by utilizing the ReliefF algorithm,
and the features of trip duration, surge pricing ratio, fare, trip distance, passenger waiting time
had the most important effects. The results also revealed that this method was better than other
classifiers, such as naive Bayes classification, support vector machines and logistic regression.
2.3. Impact of built environment on ride-hailing and ridesplitting trips
Understanding the influence of built environment on ride-hailing trips can not only
identify potential ride-hailing customers, but also improve the service quality of the ride-hailing
(Ghaffar et al., 2020; Marquet, 2020; Yan et al., 2020a). Using one month of ride-hailing trip
data in Chengdu, China, Zhang et al. (2020) examined the relationship between ride-hailing
trips and several types of points of interest (POIs) by establishing ordered logistic regression
models. The results demonstrated that the numbers of transport facilities and scenic spots had
the most impact on ride-hailing trips. The numbers of sports facilities and service facilities had
impacts on ride-hailing trips of pick-up locations, while the number of companies did not
influence ride-hailing trips significantly. Ghaffar et al. (2020) adopted random-effects negative
binomial regression models to study the determinants of the ride-hailing usage in Chicago. The
results showed that the ride-hailing demand was higher in areas with higher share of households
without vehicles, higher median household incomes, higher land-use mix, higher population
density and employment density, more restaurants, fewer parking facilities, and more
6
homicides. Marquet (2020) applied the truncated Poisson regression model to investigate the
correlation between walkability levels and ride-hailing trips in Chicago. The findings suggested
the ride-hailing usage was positively correlated with the walkability at both trip origin and
destination, and negatively associated with the factor of access to transit. Using the trip data of
Uber from 24 different areas in the US, Sabouri et al. (2020) explored how ride-hailing demand
was affected by the built environment. The results of multilevel modeling revealed that the
ride-hailing demand was positively associated with total population and employment, land use
entropy, transit stop density and activity density, while it was negatively associated with
destination accessibility variables and intersection density.
Ordinary regression models generally assume that explanatory variables are spatially
stationary across the study region and only provide global estimates. However, this assumption
could cause modeling biases as these models fail to consider the spatial heterogeneity of
influential factors (Qian and Ukkusuri, 2015; Yu and Peng, 2019), some scholars adopted local
GWR models to explore such characteristics. For example, Yu and Peng (2019) established
geographically weighted Poisson regression models to explore the spatial variation of ride-
hailing demand, using 2016-2017 ride-hailing trip data in Austin, Texas. The results indicated
that the regions with higher proportion of well-educated and young people, higher land use mix,
higher road network density and sidewalk density, lower level of balance between population
and employment, and higher levels of transit accessibility could promote the ride-hailing usage.
The modeling results also presented spatial variations of the effects of socio-economic factors
and the built environment on the ride-hailing demand. Using one month of DiDi trip data in
Chengdu, China, Wang and Noland (2021) applied GWR models to examine the effect of built
environment on ride-hailing trips for different time periods. The results indicated that road
density, housing price, floor-area ratio, population density, and the availability of public transit
had positive effects on the trip demand. They also discussed the spatial variation of the effect
of population density across the city specifically.
Few studies analyzed the impact of built environment on ridesplitting trips. Hou et al.
(2020) focused on the factors that affected ride-hailing users’ willingness to pool (WTP) with
considering several socio-economic, spatiotemporal and travel variables. Machine-learning as
well as multivariate linear regression models were applied to explore the determinants of WTP.
They found that income level, airport trips, population density and job density at both pick-up
and drop-off locations were important predictors of WTP. Using Lyft trip data in Los Angeles,
Brown (2020) investigated the factors affecting the ridesharing demand. The findings showed
that the proportion of the population between ages 15 and 34, the neighborhoods with higher
population density and transit stop density, the neighborhoods with a clear racial majority and
lower household income could promote the ridesharing use. Xu et al. (2021) studied the factors
that influenced the RAR using a random forest model. The results indicated that the proportion
of white population, education level, household income, travel distance and neighborhood
environment had the most important effects on the RAR. They also explored the nonlinear
relationships between RAR and several significant factors, such as income, the share of white
7
population and walkability.
2.4. Research gaps
As is evident from the above review, scholars mainly studied the ridesplitting based on
questionnaires and TNCs’ trip data. The former mainly explored the impacts of individual
attributes, trip characteristics and attitude perceptions on passengers’ willingness to use the
ridesplitting, while the latter mainly focused on the ridesplitting’s operational characteristics
and its impact on the environment. Although numerous studies have confirmed that the built
environment had influence on ride-hailing trips, only several literatures analyzed the
correlation between built environment and RAR, and these studies did not consider the spatial
non-stationary and heterogeneity of influential factors. As a result, we do not know how the
influences of explanatory variables vary in different spatial positions. In addition, to the best
of our knowledge, there is a paucity of literature that investigated the differences of influential
factors of RAR during different time periods, such as weekday and weekend, weekday morning
peak and evening peak. As mentioned, a better understanding of these issues could help urban
planners to develop detailed and targeted planning strategies to promote shared ride-hailing
trips, as well as provide guidance for TNCs to improve existing ridesplitting services. To fill
the knowledge gaps, this study applies local GWR models to explore the spatial variation of
RAR and its association with the built environment and socio-economic factors, and reveal the
spatial variation mechanism of RAR during different time periods.
3. Data and variables
Ride-hailing trip data was obtained from Chicago open data portal
1
, which provided all of
the ride-hailing trip records by TNCs in Chicago. For each trip, the following attributes are
included: trip start timestamp, trip end timestamp, trip duration, distance, fare, pick-up and
drop-off census tract, shared trip authorized (whether the ride-hailing passenger accepted a
shared trip with another passenger). To protect the individual privacy of transportation data in
public domain, the city of Chicago applied the aggregation technique in time, geographical
space and fares to reduce the risk of reidentification (Open Data Portal Team, 2019). The start
and end time of the trips were rounded to the nearest 15 minutes; fares were rounded to the
nearest $2.50; and actual pick-up and drop-off locations were not provided, instead the census
tract in which each trip started and ended was provided.
In this study, a census tract is selected as the spatial unit, it provides a reasonable scale to
understand the relationships at the neighborhood level and provides rich socio-economic
information (Yu and Peng, 2019). The study area (the Chicago metropolitan region) contains
801 census tracts. The geographic data for census tract was obtained from the Chicago data
portal, as shown in Figure 1. We choose the month of March in 2019 as the study period, mainly
because March is suitable for ride-hailing and ridesplitting trips (the season with the largest
ride-hailing and ridesplitting trip counts in 2019, and the RAR is relatively high). We collected
1
https://data.cityofchicago.org/Transportation/Transportation-Network-Providers-Trips/m6dm-c72p
8
ride-hailing trip data for four consecutive weeks from March 3rd, 2019 to March 30th, 2019.
As for the data, all trips recorded either start or end in one of the 801 census tracts of the
Chicago metropolitan region (Hou et al., 2020). This study explores the spatial variation of
RAR and its influential factors at census tract level, and has outcome variables for pick-up
tracts, as utilized in other ride-hailing studies, with an underlying assumption that the joint
destination mode choice has been pre-selected when riders decide to open up the TNC
applications (Yu and Peng, 2019; Yu and Peng, 2020). After removing the trip records whose
pick-up locations outside the 801 census tracts (because the pick-up census tracts of these trips
were blank in the data set), a total of 6,715,121 trips were obtained. We retained the trip records
that started within the study area but ended outside the study area (although the drop-off census
tracts of these trips were blank, the pick-up census tracts were provided in the data set), these
trips account for less than 5%, even if a small proportion, we think they are worth exploring.
Table A1 lists which trips are included in this study and the reasons for them, according to the
pick-up and drop-off locations of trips. Referring to the relevant research, we also excluded
trips whose fares equal to $0 or larger than $100, trip duration was less than 1min, and trip
distance was less than 0.1miles (0.16 kilometers) (Marquet, 2020; Hou et al., 2020). Finally,
6,696,770 valid trip records were screened out for this study, of which 1,469,008 trips made
ridesplitting requests, accounting for 22%.
Fig. 1. Study area
3.1. Dependent variables
RAR is defined as the proportion of ride-hailing trips with shared trip authorized. This
9
indicator reflects the willingness of passengers to share rides in ride-hailing trips, regardless of
whether these customers are successfully paired with other passengers (Xu et al., 2021; Hou et
al., 2020). The calculation formula is as follows:
RAR = shared trip counts/ (shared trip counts + non-shared trip counts) (1)
In this paper, the dependent variables are the RARs of four time periods, i.e., weekdays,
weekends, weekdays morning peak and evening peak, at the census tract level. The RAR of
each census tract on weekday is the average RAR of this census tract for all weekdays. The
definitions of the RAR of each census tract on the other three periods are similar.
Figures 2a-b show the variation trend of RAR on weekday and weekend for all census
tracts. It can be seen that the RAR presents morning and evening peak characteristics on
weekday, while the trend is relatively stable on weekend. This difference may be caused by
different travel purposes: the trips during peak hours on workdays are mostly for commuting,
while the trips on weekend are mainly for recreation.
In addition, since the urban transportation system faces the greatest pressure during peak
hours for commuting (Wang and Noland, 2020), we also focus on the RAR during peak hours
on weekday. From Figure 2a, we can conclude that the RAR in the evening peak is higher than
that in the morning peak. Referring to the peak periods of Chicago Metropolitan Agency for
Planning (CMAP)
2
uses for their Trip-based travel demand model
3
and Activity-based travel
demand model
4
, the morning peak is defined as 7:00-9:00, and the evening peak is defined as
16:00-18:00.
(a) Weekday
(b) Weekend
Fig. 2. Time distribution of ridesplitting adoption rate (RAR)
3.2. Independent variables
Two types of explanatory variables are included in this work: built environment and socio-
economic variables, and they are also matched by the census tract.
2
CMAP, the local Metropolitan Planning Organization in Chicago. For specific information, please visit
the following website: https://www.planning.dot.gov/mpo/
3
https://www.cmap.illinois.gov/data/transportation/modeling#CMAPs_TripBased_Models_2017
4
https://www.cmap.illinois.gov/data/transportation/modeling#CMAPs_Activity_Based_Models_2017
10
Built environment variables are chosen based on “5D’s”, i.e., density, design, diversity,
distance to transit/ transit access, destinations accessibility (Ewing and Cervero, 2010). To be
specific, for density, we select population and job densities. Road network density is selected
to measure design variable. Diversity measure involves the number and the balanced degree of
different land uses in a given region. There are different ways to measure it, such as land use
entropy, jobs-to-population ratio and jobs-to-housing ratio (Sabouri et al., 2020). We select land
use mix and job/pop mix measures (Yu and Peng, 2019). Job/pop mix reflects the level of
balance between population and jobs in a neighborhood (United States Environmental
Protection Agency, 2021b). A higher index reflects a better balance between population and
jobs, while a lower index indicates the opposite. For land use mix, it measures the extent of
mixed land development in a neighborhood (Ewing and Cervero, 2010), we use the land-use
mix entropy method which was developed by Cervero and Kockelman (1997) to calculate the
land use mix of each census tract. The formula is expressed as:
( ) ( )
1
Land use mix=- *In In
n
ii
i
p p n
=
(2)
where
i
p
is the proportion of land use i for each census tract, and n is the number of land-
use categories (residential, commercial, institutional, industrial, transportation/communication,
and agriculture are considered in this study). An entropy value of 1 implies the highest degree
of land-use mix, whereas a value of 0 indicates a single land use.
Distance to transit/ transit access is usually measured as the distance to the nearest railway
or bus station. It can be also measured by the distance between transit stations, transit route
density, or transit station density (Ewing et al., 2015; Sabouri et al., 2020). Similar to some
ride-hailing studies (Ghaffar et al., 2020; Sabouri et al., 2020), subway station and bus stop
densities at the census tract level are selected to measure this D variable. For bus stops, this
study excludes the temporary service stops according to the station status field in the data set,
and only considers the regular service stops. In addition, we also consider the variable of
frequency of transit, i.e., aggregate frequency of transit service per square mile. Compared with
the subway station and bus stop variables (transit infrastructure), this one can better reflect the
transit supply and demand. Finally, transit access to jobs (the jobs within 45-minute transit
commute) is selected to measure destinations accessibility.
In order to explore the relationship between active travel and the RAR, we select the
variable of walkability index scores, which considers street connectivity, access to public
transit, and diversity of land uses, and is based on measures of the built environment that affect
the probability of whether people walk as a mode of transportation (United States
Environmental Protection Agency, 2021a). In addition, existing studies have revealed that
crime density has delayed and immediate effects on ride-hailing use (Ghaffar et al., 2020). In
order to test whether this variable affects the RAR, we collected crime data for five consecutive
years (2014-2018) for analysis. Referring to the relevant research of ride-hailing (Ghaffar et
al., 2020), the crime categories considered in this study are: assault, homicide, robbery, sex
11
offense and theft. When modeling, we treat the density of each crime type as a covariate in
order to better understand the relationship between crime type and the RAR. For the above
built environment variables, the data of land uses was from the CMAP’s Land use Inventory
for Northeastern Illinois
5
, the data of subway and bus stations, and crime data were downloaded
from Chicago data portal
6
, the other data was directly obtained or computed from United States
Environmental Protection Agency (EPA) Smart Location Database (SLD)
7
.
The relevant studies of ridesplitting based on questionnaires data have implied that socio-
economic attributes of individuals are important factors affecting RAR, we collected socio-
economic data at the census tract level from American Community Survey (ACS) 2014-2018
5-year estimate dataset
8
. The following variables are mainly considered in this study: percent
female, median age, percent nonwhite, percent bachelors, median household income, percent
zero-vehicle households (Yu and Peng, 2019; Marquet, 2020).
Figure 3 shows the spatial distribution of some explanatory variables. Population and job
densities show decreasing trends from the downtown to the suburban area, subway stations are
mainly distributed in the central and northern regions, and the distribution of bus stations is
relatively even in most of regions, except for some areas in the south of Chicago. The extent
of mixed land development is the highest in the central region, and the walkability score is
higher in the downtown and northern regions. Homicide crime and sex offense crime are
concentrated in the west central and south central regions. The areas with high percent zero-
vehicle households are mainly distributed in the transition area between the core area and the
suburb, and the groups with higher median income are mainly in the downtown and northern
region.
Table 1 presents the descriptive statistics of variables. It can be seen that, at census tract
level, the average RAR on weekday (36.13%) is higher than that on weekend (33.59%), the
average RAR in the evening peak on weekday (38.30%) is higher than that in the morning peak
(32.96%). As for socio-economic attributes, the average proportion of women (51.81%) is
slightly higher than that of men, and the proportion of non-whites is 52.31%. More than one-
third of the population over the age of 25 have at least a bachelor’s degree or above, and the
share of car-free households is 26.61%.
5
https://www.cmap.illinois.gov/data/land-use/inventory
6
https://data.cityofchicago.org/
7
https://www.epa.gov/smartgrowth/smart-location-mapping
8
https://data.census.gov/cedsci/
12
(a) Population density (people per
acre)
(b) Job density (jobs per acre)
(c) Subway station density (per sq. mile)
(d) Bus stop density (per sq. mile)
(e) Frequency of transit (per sq. mile)
(f) Land use mix
(g) Walkability index scores
(h) Homicide density (per sq. mile)
(i) Sex offense density (per sq. mile)
13
(j) Percent nonwhite (%)
(k) Median household income
($1000)
(l) Percent zero vehicles (%)
Fig. 3. Spatial distribution of some variables
14
Table 1 Descriptive statistics of variables
Variable
Description
Data source
Mean
Median
Std.
Min
Max
Dependent variables
Weekday RAR
Average RAR on weekday (%)
CDP
36.132
37.422
11.920
0.000
77.778
Weekend RAR
Average RAR on weekend (%)
33.586
35.172
14.571
0.000
83.333
Weekday morning peak RAR
Average RAR during weekday morning peak (%)
32.963
33.151
12.349
0.000
71.614
Weekday evening peak RAR
Average RAR during weekday evening peak(%)
38.299
39.655
13.661
0.000
91.667
Independent variables
Built environment
Density
Population density
Population density (people per acre)
EPA SLD
31.278
25.583
26.031
0.000
411.306
Job density
Job density (jobs per acre)
12.363
3.211
56.459
0.000
1220.882
Design
Road network density
Road network density (mile per sq. mile)
EPA SLD
30.283
30.098
7.359
6.388
68.662
Distance to transit/ Transit access
Subway station density
Subway station density (per sq. mile)
CDP
1.021
0.000
3.718
0.000
40.932
Bus stop density
Bus stop density (per sq. mile)
31.251
29.093
17.156
0.000
172.343
Frequency of transit
Aggregate frequency of transit service per hour during evening peak
period per square mile (per sq. mile)
EPA SLD
547.240
302.419
909.743
3.950
12585.526
Diversity
Land use mix
Land use mix
CMAP
0.489
0.486
0.172
0.000
0.889
Job/pop mix
The level of balance between jobs and population
EPA SLD
0.291
0.224
0.236
0.000
0.996
Destinations accessibility
Transit access to jobs
Jobs within 45-minute transit commute (1000)
EPA SLD
534.050
535.705
249.022
0.000
1499.203
Walkability
The walkability index scores
14.574
14.581
1.782
5.333
19.667
Crime density
Assault density
Crime density from 2014 to 2018 (per sq. mile)
CDP
308.927
232.777
260.395
0.000
1791.851
Homicide density
9.730
4.574
13.487
0.000
88.137
Robbery density
188.583
140.619
171.925
0.000
1382.251
Sex offense density
16.919
13.111
18.488
0.000
344.685
Theft density
1085.379
763.792
1542.984
0.000
24467.397
15
Socio-demographics
Percent female
Percent female (%)
ACS 2014-
2018
51.806
51.400
4.973
10.400
76.900
Median age
Median age
35.245
34.400
6.263
15.900
66.800
Percent nonwhite
Percent nonwhite population (%)
52.310
46.550
33.464
1.000
100.000
Percent bachelors
Percent population ages 25 and over with at least a bachelor’s
degree (%)
35.595
27.494
26.188
0.510
94.954
Median household income
Median household income ($1000)
57.298
47.943
32.258
9.787
178.750
Percent zero vehicles
Percent zero-vehicle households (%)
26.608
24.500
15.038
0.700
77.800
CDP: Chicago Data Portal;
EPA SLD: Environmental Protection Agency Smart Location Database;
CMAP: Chicago Metropolitan Agency for Planning;
ACS 2014-2018: American Community Survey 2014-2018 5-year estimate dataset.
16
4. Methods
4.1. Multicollinearity
Multicollinearity occurs when several explanatory variables show a strong linear
correlation with each other, which might cause bias in interpreting the impacts of other
explanatory variables (Wang and Chen, 2017). Two methods are utilized to eliminate this
phenomenon. The first one is Pearson correlation analysis. Based on empirical studies,
variables with correlation coefficients greater than 0.7 are assumed to be highly correlated and
should be removed from the models (Pan et al., 2020). The second one is variance inflation
factor (VIF), a common indicator for evaluating the severity of multicollinearity, is computed
by the ordinary least squares (OLS) models (Vandenbulcke et al., 2011). In general, variables
with VIF values higher than 10 are considered to be multicollinearity variables and removed
from the models (Yang et al., 2017).
4.2. Spatial autocorrelation
Spatial autocorrelation refers to the dependence of the value of a given variable on the
value of the same variable in neighboring positions. The most commonly utilized spatial
variability test is Moran’s I test, which can determine whether a variable has spatially
autocorrelation and the correlation degree (Vandenbulcke et al., 2011; Yang et al., 2017).
The range of Moran’s I statistic is between -1 and +1. A positive value indicates spatial
aggregation, a negative value indicates spatial dispersion, and a near zero value indicates a
spatially random distribution. The null hypothesis of Moran’s I test is that the independent
variables are spatially independent, and the test statistic can provide the confidence level to
reject the null hypothesis (Moran, 1950).
4.3. Geographically weighted regression
OLS is a linear regression model, which estimates the regression coefficient by minimizing
the sum of squares of residuals (Brunsdon et al., 1996). This method has been criticized for
neglecting the spatial variations of the data. GWR is proposed to deal with spatial data
regression, allowing the regression coefficients to vary in spaces. It can be seen as an extension
of the OLS model by linking independent variables with geographical positions (Brunsdon et
al., 1996), the calculation formula is as follows:
0( , ) ( , )
= + +
i i i i ik i i ik i
k
Y u v u v X
(2)
where i denotes the ith census tract;
( , )
ii
uv
are the coordinate of the centroid point of
census tract i;
i
Y
is the RAR of census tract i;
ik
X
is the kth independent variable;
i
is
the error term of census tract i;
0( , )
i i i
uv
represents the intercept; and
( , )
ik i i
uv
is the
regression coefficient between RAR and the explanatory variable.
The parameters are calibrated in the way that an observation will have greater influence
on location i (the ith cecus tract in this work) if the distance in between is closer, and the extent
17
of this influence can be determined by kernel function. In this work, we utilize the Gaussian
kernel with an adaptive bandwidth since the census tracts are denser in downtown regions while
sparser at the suburban areas. The function of Gaussian kernel is as follows:
( )
2
exp 0.5 ,
0, otherwise
−
=
ij ij
ij
d b d b
w
(3)
where wij is the allocated weight between the centroids of two census tracts; dij is the
distance between census tract j and census tract i; parameter b is the bandwidth, which is
utilized to exclude observations that exceed the distance threshold. The optimal bandwidth is
determined by finding the corresponding value that result in the minimum corrected Akaike
information criterion (AICc), which is utilized to evaluate the modeling results and avoid the
over-fitting phenomenon.
5. Results and discussions
5.1. Spatial analysis
Figures 4a-b show the spatial distribution of RAR on weekday and weekend at the census
tract level. Similar distribution patterns are observed: in the west and south of Chicago, RAR
is relatively high; in contrast, RAR is relatively low in the downtown and northern regions. The
explanation is that in the western and southern regions, subway and bus facilities are relatively
scarce (see Figures 3c-d), and the proportion of car-free households is relatively high (see
Figure 3l), ridesplitting can just offset the disadvantages of low accessibility in these areas.
Figures 4c-d present the spatial distribution of RAR in the morning and evening peaks on
weekday. It can be seen that RAR in the morning peak is lower than that in the evening one
across the region, especially in the western and southern areas. It is likely that the time budget
for commuting trips in the morning peak is smaller (time sensitivity is higher) than that in the
evening peak, and the extra delays and detours caused by the ridesplitting could reduce the
reliability of travel (Li et al., 2019).
18
(a) Weekday
(b) Weekend
(c) Weekday morning peak
(d) Weekday evening peak
Fig. 4. Spatial distribution of RAR
Figures 5a-d show the distribution curves of RAR at the census tract level for the four
time periods. It can be seen that the distribution of RAR during the four time periods
approximately obeys the normal distribution. In order to explore how built environment and
socio-economic factors influence RAR in different periods, we apply the global OLS and the
local GWR to perform in-depth analysis in the next section.
19
(a) Weekday
(b) Weekend
(c) Weekday morning peak
(d) Weekday evening peak
Fig. 5. Distribution curves of RAR
5.2. Model results
Before building the models, three census tracts of the 801 census tracts in the study area
were removed in the final dataset because they lacked values of several socio-economic
variables. The final models are to test influential factors of the RAR of 798 census tracts during
four time periods, i.e., weekday, weekend, weekday morning peak and evening peak, in
Chicago.
First, Pearson correlation analysis is performed to test multicollinearity, the results are
presented in Figure 6. It can be seen that most of the correlation coefficients are less than 0.7,
except for population density and frequency of transit (0.81), assault density and robbery
density (0.84), assault density and homicide density (0.7), percent bachelors and median
household income (0.79). We exclude the variable of percent bachelors and retain the variable
of median household income, because previous researches suggested that the income was an
important predictor of the RAR (Hou et al., 2020). We exclude assault density because it is
highly correlated with two variables of robbery density and homicide density. Further, we
exclude population density and retain frequency of transit because we are concerned about the
relationship between transit supply and the RAR.
20
Fig. 6. Pearson correlation coefficient for independent variables
Then, OLS models are calibrated to explore significant factors that affect the RAR during
four time periods (Qian and Ukkusuri, 2015; Li et al., 2021; Pan et al., 2020). The results,
including global coefficients, significance results and VIF values, are presented in Table 2.
Eventually, 14 variables are left after the selection of OLS regression, namely, job density,
subway station density, bus stop density, frequency of transit, land use mix, transit access to
jobs, walkability, homicide density, sex offense density, percent female, median age, percent
nonwhite, median household income, percent zero vehicles. Meanwhile, we calculate the VIF
value for each candidate explanatory variable by OLS models, the results show that VIF values
of all variables are below 10 (the largest VIF is 2.823), indicating that the independent variables
are properly selected and the multicollinearity can be avoided.
21
Table 2 OLS model results
Variable
Weekday Model
Weekend Model
Weekday Morning
Model
Weekday Evening
Model
Coef.
VIF
Coef.
VIF
Coef.
VIF
Coef.
VIF
Built environment
Job density
-0.012**
1.276
Subway station density
-0.144**
1.085
-0.173**
1.087
-0.252***
1.138
Bus stop density
0.045**
1.565
Frequency of transit
-0.001***
1.481
-0.001***
1.868
-0.001***
1.573
-0.002***
2.061
Land use mix
3.484**
1.119
3.18*
1.127
6.056**
1.303
Transit access to jobs
0.009***
1.667
Walkability
-0.504*
1.825
Homicide density
0.057**
1.644
0.068***
1.629
Sex offense density
0.046*
1.596
Socio-economic
attributes
Percent female
-0.168***
1.289
-0.189***
1.286
-0.121*
1.279
Median age
-0.17***
1.131
-0.155***
1.131
-0.291***
1.109
-0.194***
1.122
Percent nonwhite
0.16***
2.579
0.259***
2.823
0.162***
2.447
0.163***
2.656
Median household
income
-0.156***
2.202
-0.149***
2.211
-0.123***
2.632
-0.126***
2.617
Percent zero vehicles
-0.143***
1.961
-0.138***
1.971
-0.068**
2.461
-0.126***
2.013
Intercept
45.201
45.201
48.506
58.121
*** Significant at 0.01, ** Significant at 0.05, * Significant at 0.1.
In addition, Moran’s I test is performed to examine whether the candidate explanatory
variables are spatially autocorrelated. The statistical results are shown in Table 3. P-values are
all less than 0.01, implying that the 14 explanatory variables have significant spatial
autocorrelation (Qian and Ukkusuri, 2015). Moreover, Moran’s I values are positive, which
means that spatial distributions of all candidate variables are more likely to be spatially
aggregated. Therefore, it is appropriate to utilize GWR models to examine the spatial variation
of RAR (Qian and Ukkusuri, 2015; Wang and Noland, 2021; Pan et al., 2020).
Table 3 Moran’s I test results
Variable
Moran’s I
Z-score
P-values
Weekday RAR
0.695
139.225
***
Weekend RAR
0.675
135.306
***
Weekday morning peak RAR
0.470
94.275
***
Weekday evening peak RAR
0.449
90.053
***
Built environment
Job density
0.154
37.905
***
Subway station density
0.047
9.922
***
Bus stop density
0.124
25.246
***
Frequency of transit
0.314
65.209
***
22
Land use mix
0.101
20.506
***
Transit access to jobs
0.431
86.560
***
Walkability
0.328
65.934
***
Homicide density
0.322
64.904
***
Sex offense density
0.075
16.561
***
Socio-economic attributes
Percent female
0.203
40.920
***
Median age
0.183
36.900
***
Percent nonwhite
0.674
134.837
***
Median household income
0.618
123.861
***
Percent zero vehicles
0.305
61.235
***
*** Significant at 0.01, ** Significant at 0.05, * Significant at 0.1.
GWR models are established using the same explanatory variables as OLS models (Qian
and Ukkusuri, 2015; Li et al., 2021). The model results of GWR for weekday, weekend,
weekday morning peak and weekday evening peak are presented in Tables A2-A5 in Appendix,
respectively. We present the statistics of local coefficients for each explanatory variable,
including minimum, maximum, average, and median of the coefficients.
With regard to performance indicators of OLS and GWR models, we show the values of
AICc, R2, and adjusted R2. As seen from Table 4, for the four time periods, GWR models all
outperform OLS models in terms of model fit. Taking the weekday model as an example, the
AIC value of the GWR model (5109.858) is lower than that of the OLS model (5317.109), the
R2 and adjusted R2 values of the GWR model (0.783 and 0.760) are much greater than those of
the OLS model (0.682 and 0.679). In addition, the spatial distributions of standard residuals
for four GWR models are also examined, as presented in Figure 7. We can find that only a few
of the local regression models fail the residual tests (the census tracts with over 2.5 times the
standard deviation) (Zhao et al., 2020).
Table 4 Comparison results of GWR and OLS models
Weekday Model
Weekend Model
Weekday Morning Model
Weekday Evening Model
GWR
OLS
GWR
OLS
GWR
OLS
GWR
OLS
AICc
5109.858
5317.109
5364.843
5569.106
5544.302
5733.464
5888.687
5976.322
R2
0.783
0.682
0.803
0.710
0.655
0.502
0.570
0.449
R2 Adjusted
0.760
0.679
0.781
0.707
0.616
0.495
0.520
0.442
23
(a) Weekday
(b) Weekend
(c) Weekday morning peak
(d) Weekday evening peak
Fig. 7. Spatial distribution of standard residuals for GWR models
5.3. Discussions
5.3.1. OLS models
From the results of OLS models (see Table 2), it can be seen that for built environment
variables, subway station density is negatively correlated with the RAR in most periods except
for weekday evening peak, frequency of transit has a negative association with the RAR. These
indicate that census tracts with better public transit services have lower RAR. During weekdays
and weekday peaks, diversified land use will promote the RAR. The possible reason is that the
24
areas with greater land-use mix will provide more diversified services and activities (Yu and
Peng, 2019), thus producing a large number of ridesplitting trips. The increase in travel demand
density will increase the probability of successful matching and reduce the travel cost, which
may increase riders’ willingness to pool. The higher the walkability, the lower the RAR, and
this influence is only significant on weekday evening peak. One of the major reasons is that
the areas with high walkability are more suitable for active travel. Another reason may be that
the areas with high walkability tend to be more congested (Xu et al., 2021), which will reduce
riders’ willingness to share because congestion will lead to higher uncertainty in travel time.
For the crime, homicide and sex offense are the crime type variables retained in final models,
and other crime type variables are excluded due to multicollinearity and statistically
insignificant. Specifically, homicide density will significantly improve the RAR on weekdays
and weekends. For the weekday evening peak, the higher the sex offense density, the higher
the RAR.
In terms of socio-economic attribute variables, percent female are negatively correlated
with the RAR. This is mainly because ridesplitting means sharing a relatively closed space with
strangers, in such close proximity, the risk of harassment are important concerns for women
(Kang et al., 2021). As previous studies based on questionnaires survey have shown that many
women prefer to match with passengers of the same gender (Sarriera et al., 2017). Median age
and median household income are negative with the RAR, percent nonwhite are positive with
the RAR. This suggests that young people, individuals with low-income, and non-white people
tend to have higher willingness to pool, which is similar to the findings of previous studies
(Kang et al., 2021; Brown, 2020; Hou et al., 2020). Percent zero vehicles is negative with the
RAR, we will analyze the reasons in the following section.
5.3.2. GWR models
For GWR models (see Tables A2-A5 in Appendix), the median values of local coefficients
for explanatory variables are similar to the global coefficients in OLS models in terms of the
direction and magnitude. In addition, most of variables show mixed estimations, that is, the
range of the estimated coefficients varies from negative to positive values, which implies that
the impacts of built environment and socio-economic factors on the RAR have significant
heterogeneity in different census tracts. However, this information is not reflected in the global
models. In the following, we will analyze the spatial heterogeneity of the effects of some
important and policy relevant variables.
(1) Subway station density
Figure 8 presents the spatial distribution of the effect of subway station density. As seen
in Figures 8a and 8c, on weekday and weekday morning peak, in the northern region where
subway lines are densely distributed, this variable is negatively associated with the RAR; in
the southern area where subway stations are scarce, especially in the southeast region, subway
station density will promote the RAR. These suggest that in regions where the supply of subway
facilities is insufficient, ridesplitting plays an important role; while in the other regions, this
effect seems to be more limited. Therefore, transportation planners should comprehensively
25
measure the impact of subway station on RAR in different regions when they formulate the
ridesplitting measures. On weekend, as shown in Figure 8b, subway station density shows
negative relationships with RAR for most of the census tracts, except for some western and
southern regions. This difference between weekday and weekend may be caused by the
different travel activities: residents need to make long-distance commuting trips on weekday,
while there are more social and entertainment activities on weekend.
(a) Weekday
(b) Weekend
(c) Weekday morning peak
Fig. 8. Coefficient estimates of subway station density
(2) Frequency of transit
Figures 9a-d present the spatial distribution of the coefficient estimates for frequency of
transit in GWR models. It can be seen that the effect of this variable shows similar patterns for
weekday, weekend, weekday morning peak and evening peak. Most of the coefficient estimates
in the study region are negative, indicating that census tracts with higher frequency of transit
26
have lower RAR. In the southern peripheral regions where transit services are less frequent
(see the spatial distribution of transit frequency in Figure 3e), frequency of transit has a positive
association with the RAR. To some extent, these imply the potential substitutive and
complementary effects between ridesplitting and transit. In areas with high-quality transit
service, transit is more superior and competitive than ridesplitting service; while in areas with
low coverage of transit service, ridesplitting can play a useful supplement to transit.
In addition, as seen in Figures 9c-d, in the southern area, the regions with positive effect
in the evening peak are wider than those in the morning peak on weekday. This difference may
be explained by the following reason: the commuting trips in the morning peak on weekday
are more sensitive to time, the extra detour distance of online ridesplitting will affect the
reliability of the arrival time, which reduces the willingness to pool. On the contrary, the return
trips in the evening peak are less sensitive to time.
(a) Weekday
(b) Weekend
(c) Weekday morning peak
(d) Weekday evening peak
Fig. 9. Coefficient estimates of frequency of transit
27
(3) Crime density
Figures 10a-c present the coefficient estimates of crime density. As shown in Figures 10a-
b, on weekday and weekend, most of the estimates in the study region are positive, indicating
that areas with higher homicide density have higher RAR. This effect is more obvious in the
area with high homicide density in the central region (see the spatial distribution of homicide
density in Figure 3h). It is probably that riders in the regions with more homicides are less
likely to travel by walking, biking or even transit due to the dangers outside (Ghaffar, 2020).
On weekday evening peak, there is a positive correlation between sex offense density and
RAR, especially in the western and southern regions, as shown in Figure 10c. It is likely that
in the evening rush hours, ridesplitting travel (in car travel) can make travelers feel safer than
other travel modes which are exposed outdoors.
(a) Weekday (homicide density)
(b) Weekend (homicide density)
(c) Weekday evening peak (sex offense
density)
28
Fig. 10. Coefficient estimates of crime density
(4) Percent nonwhite
Figure 11 presents the coefficient estimates of percent nonwhite. As can be seen, a positive
link between percent nonwhite and RAR is observed in most of the study region, especially in
the southern and northern regions. Previous researches have indicated that individuals traveling
from neighborhoods with a larger proportion of non-white-majority population are more
inclined to share rides with other passengers (Blumenberg and Smart, 2014; Brown, 2020). The
south of Chicago is the area where non-white population is concentrated: the proportion of non-
whites is as high as 86% (see Figure 3j), thus this positive effect is obvious. In addition, based
on the existing survey, a high proportion of white ride-hailing users stated that sharing rides
with other passengers with different races could make them uncomfortable (Sarriera et al.,
2017). As shown in Figure 3j, the north of Chicago is the area where white population is
concentrated, accounting for about 80%. When the white users take ridesplitting services in
this region, the reduction in the probability of matching with other race groups may promote
the RAR to a certain extent.
(a) Weekday
(b) Weekend
29
(c) Weekday morning peak
(d) Weekday evening peak
Fig. 11. Coefficient estimates of percent nonwhite
(5) Percent zero vehicles
Figures 12a-d present the spatial distribution of coefficient estimates for percent zero
vehicles. Most of coefficient estimates in the study region are negative, implying that the
regions with a higher percent zero-vehicle households have a lower RAR. This seems counter
intuitive. One reason may be that, for the families without cars, their daily trips do not depend
on motorized travel modes, and this habit has been formed. Another explanation is that the
decision of car ownership may be based on the availability and convenience of other modes of
transportation, such as walking, cycling or bus (we have controlled the variable of median
household income in the models) (Marquet, 2020). During weekday morning and evening
peaks, in some southern regions, this variable is positively associated with the RAR. It is
probably that the ridesplitting is a cost-effective travel mode for travelers in peripheral regions.
(a) Weekday
(b) Weekend
30
(c) Weekday morning peak
(d) Weekday evening peak
Fig. 12. Coefficient estimates of percent zero vehicles
6. Policy implications
According to the research results, the following detailed and targeted policy suggestions
are proposed to promote shared ride-hailing trips and improve the ridesplitting service from
the perspectives of policy makers and TNCs.
(1) Encourage shared trips in the areas with insufficient transit supply
As the results show, in the southern area where subway facilities are scarce/ where transit
services are less frequent, subway station density/ frequency of transit has a positive association
with the RAR. It implies the potential complementary effects between ridesplitting and transit
in these regions. In addition, in the south of Chicago, the variable of percent zero vehicles
shows a positive relationship with the RAR. In view of these, TNCs can cooperate with the
government to provide ridesplitting service /low cost micro-transit service in urban peripheral
areas where fixed-route bus cannot be added/ are not available in the near future. Specifically,
multiple types of ridesplitting services, such as long-distance ridesplitting, large-capacity
ridesplitting, and reservation-type ridesplitting services, could be developed to compensate for
social justice issues caused by insufficient transit supply.
(2) Implement ridesplitting incentives during peak hours
The previous analysis shows that the RAR is high on weekday morning peak and evening
peak, especially on weekday evening peak. In view of this, TNCs can implement corresponding
ridesplitting incentives during peak hours to encourage more ride-hailing users to choose
ridesplitting service, such as reducing or exempting the ridesplitting order tax, assigning
ridesplitting orders with priority. It can not only increase the order response rate and
ridesplitting matching success rate, but also reduce the congestion effect and other negative
effects of solo ride-hailing trips.
(3) Promote shared ride-hailing trips for target regions
The results show that regions with high RAR are mainly located in the west and south of
31
Chicago. In the central region with high homicide density, crime density is positive with RAR.
In southern areas where non-white population is concentrated as well as in northern areas where
white population is concentrated, percent nonwhite is positively correlated with RAR.
Therefore, the government and TNCs can encourage shared over solo trips in these regions.
For example, the government charges higher taxes for non-shared vehicles, and TNCs issue
ridesplitting discount coupons to the users in these regions.
(4) Improve the ridesplitting safety for women
The model results indicate that percent female presents a negative effect on the RAR.
TNCs can optimize the matching algorithm, if allowed, they can add the ridesplitting option
settings for female passengers, such as the gender of matching passengers, seat preference, and
even allow female users to customize the attributes of matching passengers.
(5) Optimizing ridesplitting route algorithms
In addition to the above incentives for ridesplitting, route algorithms of TNCs can also be
optimized to encourage more ride-hailing users to share. In actual operations, the additional
delays caused by ride-pooling will reduce the reliability of the trips and thus reduce the
passengers’ willingness to pool. To this end, TNCs can improve their route optimization
algorithms to ensure the reliability of travel time.
7. Conclusion
This study explores the spatial variation of RAR, using the trip data of TNCs in Chicago
integrated with various built environment and socio-economic variables. OLS and GWR
models are applied to examine the factors that affect the RAR during four time periods, i.e.,
weekday, weekend, weekday morning peak and evening peak. OLS models show that factors
including subway station density, frequency of transit, land use mix, homicide density, percent
female, percent nonwhite, median household income, and percent zero vehicles have
significant impacts on RAR for most periods. The results of GWR models are consistent with
the global regression results, and can provide detailed information about the spatial variation
of explanatory variables’ impacts.
The important conclusions regarding the factors affecting the RAR can be summarized
as follows:
(1) Public transit and active travel
On weekday, in the southern area with insufficient supply of subway facilities, subway
station density is positively associated with the RAR, indicating that ridesplitting plays an
important role in these areas. During the four periods, there are potential substitutive and
complementary effects between ridesplitting and transit. In the areas with high-quality transit
service, transit is more superior than ridesplitting; while in areas with less frequent transit
service, ridesplitting could play a useful supplement to transit. On weekday evening peak, the
areas with higher walkability have lower RAR.
(2) Security problems
On weekday and weekend, the regions with higher homicide density have higher RAR
32
for most areas. On weekday evening peak, sex offense density is positively correlated with the
RAR for most regions. These could be explained by that the ridesplitting travel (in car travel)
can make travelers feel safer than other travel modes which are exposed outdoors. In addition,
percent female shows a negative association with the RAR.
(3) Racial attitudes
Percent nonwhite has positively associated with RAR in most of the study region,
especially in the southern regions where non-white population is concentrated, and the northern
regions where white population is concentrated. It implies that riders have a higher willingness
to share rides when the probability of matching the same race groups is higher.
(4) Car ownership
The census tracts with a higher percent zero-vehicle households have a lower RAR. The
reason may be the result of the availability and convenience of other modes of transportation,
such as walking, cycling or bus. In some southern areas, this variable has a positive association
with the RAR during weekday morning and evening peaks.
This work has several limitations, and the following aspects can be considered in future
studies. First, to maintain the privacy of transportation data, Chicago masked individual trip
records spatially, and pick-up and drop-off locations were aggregated to the census tract.
Therefore, this study explores the effects of built environment and socio-economic factors on
the RAR at the census tract level, which may obscure some individual behaviors. In the future
research, stated preference and intercept surveys with smaller sample sizes might be carried
out to understand the impacts of some individual attributes and travel attribute variables on the
RAR, such as travel purpose, travel frequency and substituted travel modes (Yu and Peng, 2019;
Li et al., 2022). Second, the data whose origins and destinations outside the city limits (the
Chicago metropolitan region) are unavailable, obtaining region- or statewide TNC data can
provide decision makers with the useful information to understand and improve the regional
transportation system. Finally, this paper explores the factors that influence the RAR during
different time periods, such as weekday, weekend, weekday morning peak and evening peak in
Chicago city, the influential factors of RAR in other time dimensions, such as different seasons
and years, and in other cities, such as Los Angeles, Chengdu city in China, can also be
examined.
Acknowledgments: This work was supported by National Natural Science Foundation of
China (No. 51578150; No. 52172318; No. 52131203), and Scientific Research Startup Fund
for Advanced Talents of Nanjing Forestry University (No. 163106065).
Author Contributions: Mingyang Du: Data curation, Formal analysis, Writing-original draft,
Writing-review and editing, Visualization. Lin Cheng: Resources, Supervision. Xuefeng Li:
Data curation, Formal analysis, Writing-original draft, Writing-review and editing, Validation.
Qiyang Liu: Formal analysis, Writing-review and editing. Jingzong Yang: Formal analysis,
Writing-review and editing.
33
Conflicts of Interest: The authors declare no conflict of interest.
Appendix
Table A1 Trips included in this study and the reasons for them
Origin
Destination
Whether to
consider in
this study?
Reason
Within the 801
census tracts
Within the 801
census tracts
Yes
Within the 801
census tracts
Outside the 801
census tracts
Yes
Outside the 801
census tracts
Within the 801
census tracts
No
This study has outcome variables for
pick-up tracts, but the pick-up census
tracts of these trips were blank in the
data set
Outside the 801
census tracts
Outside the 801
census tracts
No
The data whose origins and
destinations outside the 801 census
tracts are unavailable
Table A2 GWR results on weekday
Variable
Min
Max
Mean
Median
Built environment
Subway station density
-0.210
0.508
0.009
-0.071
Frequency of transit
-0.007
0.009
-0.002
-0.001
Land use mix
-2.785
8.715
3.097
3.488
Homicide density
-0.073
0.122
0.039
0.039
Socio-economic attributes
Median age
-0.239
0.153
-0.063
-0.076
Percent nonwhite
-0.019
0.288
0.147
0.141
Median household income
-0.243
-0.060
-0.135
-0.124
Percent zero vehicles
-0.243
-0.021
-0.122
-0.125
Intercept
25.340
60.890
41.292
40.659
Table A3 GWR results on weekend
Variable
Min
Max
Mean
Median
Built environment
Subway station density
-0.339
0.151
-0.098
-0.099
Bus stop density
-0.071
0.108
0.037
0.036
Frequency of transit
-0.007
0.009
-0.001
-0.001
Homicide density
-0.153
0.182
0.044
0.052
Socio-economic attributes
34
Percent female
-0.497
0.232
-0.045
-0.042
Median age
-0.305
0.170
-0.081
-0.073
Percent nonwhite
0.067
0.372
0.227
0.233
Median household income
-0.297
-0.019
-0.137
-0.113
Percent zero vehicles
-0.255
0.064
-0.117
-0.119
Intercept
6.392
76.431
37.255
36.620
Table A4 GWR results on weekday morning peak
Variable
Min
Max
Mean
Median
Built environment
Job density
-0.032
0.047
-0.004
-0.008
Subway station density
-0.354
0.788
0.009
-0.102
Frequency of transit
-0.011
0.004
-0.002
-0.002
Land use mix
-4.819
6.988
2.975
3.369
Transit access to jobs
-0.016
0.014
0.000
0.000
Socio-economic attributes
Percent female
-0.424
0.160
-0.067
-0.054
Median age
-0.667
-0.016
-0.226
-0.169
Percent nonwhite
0.013
0.235
0.136
0.139
Median household income
-0.292
0.008
-0.105
-0.088
Percent zero vehicles
-0.162
0.051
-0.049
-0.050
Intercept
17.819
68.348
45.574
46.937
Table A5 GWR results on weekday evening peak
Variable
Min
Max
Mean
Median
Built environment
Frequency of transit
-0.008
0.010
-0.001
-0.001
Land use mix
-0.499
23.578
6.152
4.987
Walkability
-2.280
0.245
-0.593
-0.526
Sex offense density
-0.038
0.155
0.050
0.043
Socio-economic attributes
Percent female
-0.390
0.220
-0.027
0.000
Median age
-0.377
0.257
-0.036
-0.060
Percent nonwhite
-0.057
0.252
0.148
0.174
Median household income
-0.212
0.101
-0.107
-0.098
Percent zero vehicles
-0.256
0.094
-0.132
-0.142
Intercept
20.972
78.532
49.229
50.283
References
Agatz, N., Erera, A. L., Savelsbergh, M. W., Wang, X., 2011, Dynamic ride-sharing: A
simulation study in metro Atlanta. Transportation Research Part B: Methodological,
45(9): 1450-1464. https://doi.org/10.1016/j.trb.2011.05.017.
Alonso-González, M. J., Cats, O., van Oort, N., Hoogendoorn-Lanser, S., Hoogendoorn, S.,
2020. What are the determinants of the willingness to share rides in pooled ondemand
35
services? Transportation 1-33. https://doi.org/10.1007/s11116-020-10110-2.
Alonso-Mora, J., Samaranayake, S., Wallar, A., Frazzoli, E., Rus, D., 2017. On-demand high-
capacity ride-sharing via dynamic trip-vehicle assignment. Proceedings of the National
Academy of Sciences, 114(3), 462-467. https://doi.org/10.1073/pnas.1611675114.
Amey, A., Attanucci, J., Mishalani, R., 2011. Real-time ridesharing: Opportunities and
challenges in using mobile phone technology to improve rideshare services.
Transportation Research Record 2217(1), 103-110. https://doi.org/10.3141/2217-13.
Amirkiaee, S. Y., Evangelopoulos, N., 2018. Why do people rideshare? An experimental study.
Transportation Research Part F: Traffic Psychology and Behaviour, 55, 9-24.
https://doi.org/10.1016/j.trf.2018.02.025.
Anderson, D. N., 2014. “Not just a taxi”? For-profit ridesharing, driver strategies, and VMT.
Transportation 41, 1099–1117. https://doi.org/10.1007/s11116-014-9531-8.
Blumenberg, E., Smart, M., 2014. Brother can you spare a ride? Carpooling in immigrant
neighbourhoods. Urban Studies 51 (9), 1871-1890.
https://doi.org/10.1177/0042098013502825.
Brown, A. E., 2020. Who and where rideshares? Rideshare travel and use in Los Angeles.
Transportation Research Part A: Policy and Practice 136, 120-134.
https://doi.org/10.1016/j.tra.2020.04.001.
Brunsdon, C., Fotheringham, A. S., Charlton, M. E., 1996. Geographically weighted regression:
A method for exploring spatial nonstationarity. Geographical Analysis 28(4), 281-298. .
https://doi.org/10.1111/j.1538-4632.1996.tb00936.x.
Cervero, R., Kockelman, K., 1997. Travel demand and the 3Ds: Density, diversity, and design.
Transportation Research Part D: Transport and Environment, 2(3), 199-219.
https://doi.org/10.1016/S1361-9209(97)00009-6.
Chen, X., Zheng, H., Wang, Z., Chen, X., 2018. Exploring impacts of on-demand ridesplitting
on mobility via real-world ridesourcing data and questionnaires. Transportation 1-21.
https://doi.org/10.1007/s11116-018-9916-1.
Chen, X., Zahiri, M., Zhang, S., 2017. Understanding ridesplitting behavior of on-demand ride
services: An ensemble learning approach. Transportation Research Part C: Emerging
Technologies 76, 51-70. https://doi.org/10.1016/j.trc.2016.12.018.
Du, M., Cheng, L., Li, X, Yang, J., 2020. Acceptance of electric ride-hailing under the new
policy in Shenzhen, China: Influence factors from the driver’s perspective. Sustainable
Cities and Society 61, 102307. https://doi.org/10.1016/j.scs.2020.102307.
Ewing, R., Cervero, R., 2010. Travel and the built environment: A meta-analysis. Journal of
the American Planning Association 76 (3), 265-294.
https://doi.org/10.1080/01944361003766766.
Ewing, R., Tian, G., Goates, J. P., Zhang, M., Greenwald, M. J., Joyce, A., Kircher, J., Greene,
W., 2015. Varying influences of the built environment on household travel in 15 diverse
regions of the United States. Urban Studies 52 (13), 2330-2348.
https://doi.org/10.1177/0042098014560991.
Fagnant, D. J., Kockelman, K. M., 2018. Dynamic ride-sharing and fleet sizing for a system of
shared autonomous vehicles in Austin, Texas. Transportation 45, 143-158.
https://doi.org/10.1007/s11116-016-9729-z.
Ghaffar, A., Mitra, S., Hyland, M., 2020. Modeling determinants of ridesourcing usage: A
census tract-level analysis of Chicago. Transportation Research Part C: Emerging
Technologies 119, 102769. https://doi.org/10.1016/j.trc.2020.102769.
36
Henao, A., Marshall, W. E., 2019. The impact of ride-hailing on vehicle miles traveled.
Transportation, 46(6), 2173-2194. https://doi.org/10.1007/s11116-018-9923-2.
Hou, Y., Garikapati, V., Weigl, D., Henao, A., Moniot, M., Sperling, J., 2020. Factors
influencing willingness to pool in ride-hailing trips. Transportation Research Record
2674(5), 419-429. https://doi.org/10.1177/0361198120915886.
Kang, S., Mondal, A., Bhat, A. C., Bhat, C. R., 2021. Pooled versus private ride-hailing: A joint
revealed and stated preference analysis recognizing psycho-social factors.
Transportation Research Part C: Emerging Technologies, 124, 102906.
https://doi.org/10.1016/j.trc.2020.102906.
Lavieri, P.S., Bhat, C.R., 2019. Investigating objective and subjective factors influencing the
adoption, frequency, and characteristics of ride-hailing trips. Transportation Research
Part C: Emerging Technologies 105, 100-125. https://doi.org/10.1016/j.trc.2019.05.037.
Li, W., Pu, Z., Li, Y., Ban, X.J., 2019. Characterization of ridesplitting based on observed data:
A case study of Chengdu, China. Transportation Research Part C: Emerging
Technologies 100, 330-353. https://doi.org/10.1016/j.trc.2019.01.030.
Li, W., Chen, S., Dong, J., Wu, J., 2021. Exploring the spatial variations of transfer distances
between dockless bike-sharing systems and metros. Journal of Transport Geography,
92, 103032. https://doi.org/10.1016/j.jtrangeo.2021.103032.
Li, X., Du, M., Zhang, Y., Yang, J., 2022. Identifying the factors influencing the choice of
different ride-hailing services in Shenzhen, China. Travel Behaviour and Society, 29,
53-64. https://doi.org/10.1016/j.tbs.2022.05.006.
Li, X., Zhang, Y., Du, M., Yang, J., 2020. The forecasting of passenger demand under hybrid
ridesharing service modes: A combined model based on WT-FCBF-LSTM. Sustainable
Cities and Society, 62, 102419. https://doi.org/10.1016/j.scs.2020.102419.
Marquet O., 2020. Spatial distribution of ride-hailing trip demand and its association with
walkability and neighborhood characteristics. Cities 106, 102926.
https://doi.org/10.1016/j.cities.2020.102926.
Martin, E. W., Shaheen, S. A., 2011. Greenhouse gas emission impact of carsharing in North
America. IEEE Transactions on Intelligent Transportation Systems 12 (4), 1074-1086.
https://doi.org/10.1109/TITS.2011.2158539.
Merlin, L.A., 2017. Comparing automated shared taxis and conventional bus transit for a small
city. Journal of Public Transportation 20(2), 19-39. http://doi.org/10.5038/2375-
0901.20.2.2.
Moody, J., Middleton, S., Zhao, J., 2019. Rider-to-rider discriminatory attitudes and
ridesharing behavior. Transportation Research Part F: Traffic Psychology and
Behaviour 62, 258-273. https://doi.org/10.1016/j.trf.2019.01.003.
Moran, P. A., 1950. Notes on continuous stochastic phenomena. Biometrika 37, 17-23.
https://doi.org/10.2307/2332142.
Morris, E. A., Pratt, A. N., Zhou, Y., Brown, A., Khan, S. M., Derochers, J. L., Brown, A.,
Chowdhury, M. R., 2019. Assessing the experience of providers and users of
transportation network company ridesharing services.
https://doi.org/10.13140/RG.2.2.11680.84486.
Open Data Portal Team. How Chicago Protects Privacy in TNP and Taxi Open Data.
http://dev.cityofchicago.org/open%20data/data%20portal/2019/04/12/tnp-taxi-
privacy.html (accessed April 12th, 2019).
Pan, Y., Chen, S., Niu, S., Ma, Y., Tang, K., 2020. Investigating the impacts of built
37
environment on traffic states incorporating spatial heterogeneity. Journal of Transport
Geography 83, 102663. https://doi.org/10.1016/j.jtrangeo.2020.102663.
Qian, X., Ukkusuri, S. V., 2015. Spatial variation of the urban taxi ridership using GPS data.
Applied Geography 59, 31-42. https://doi.org/10.1016/j.apgeog.2015.02.011.
Sabouri, S., Park, K., Smith, A., Tian, G., Ewing, R., 2020. Exploring the influence of built
environment on Uber demand. Transportation Research Part D: Transport and
Environment 81, 102296. https://doi.org/10.1016/j.trd.2020.102296.
Sarriera, J. M., Álvarez, G. E., Blynn, K., Alesbury, A., Scully, T., Zhao, J., 2017. To share or
not to share: Investigating the social aspects of dynamic ridesharing. Transportation
Research Record 2605(1), 109-117. https://doi.org/10.3141/2605-11.
Schwieterman, J., Smith, C. S., 2018. Sharing the ride: A paired-trip analysis of UberPool and
Chicago Transit Authority services in Chicago, Illinois. Research in Transportation
Economics 71, 9-16. https://doi.org/10.1016/j.retrec.2018.10.003.
Shaheen, S., Cohen, A., 2019. Shared ride services in North America: Defnitions, impacts, and
the future of pooling. Transport Reviews 39(4), 427-442.
https://doi.org/10.1080/01441647.2018.1497728.
Shaheen, S., Cohen, A., Zohdy, I., 2016. Shared mobility: Current practices and guiding
principles. Federal Highway Administration, US Department of Transportation, United
States. http://www.ops.fhwa.dot.gov/publications/fhwahop16022/fhwahop16022.pdf.
Shen, H., Zou, B., Lin, J., Liu, P., 2020. Modeling travel mode choice of young people with
differentiated E-hailing ride services in Nanjing China. Transportation Research Part D:
Transport and Environment, 78, 102216. https://doi.org/10.1016/j.trd.2019.102216.
United States Environmental Protection Agency, 2021a. National Walkability Index
Methodology and User Guide. https://www.epa.gov/smartgrowth/national-walkability-
index-user-guide-and-methodology.
United States Environmental Protection Agency, 2021b. Smart location database technical
documentation and user guide. https://www.epa.gov/smartgrowth/smart-location-
database-technical-documentation-and-user-guide.
Vandenbulcke, G., Dujardin, C., Thomas, I., de Geus, B., Degraeuwe, B., Meeusen, R., Panis,
L. I., 2011. Cycle commuting in Belgium: Spatial determinants and ‘re-cycling’
strategies. Transportation Research Part A: Policy and Practice 45(2), 118-137.
https://doi.org/10.1016/j.tra.2010.11.004.
Wang, C. H., Chen, N., 2017. A geographically weighted regression approach to investigating
the spatially varied built-environment effects on community opportunity. Journal of
Transport Geography 62, 136-147. https://doi.org/10.1016/j.jtrangeo.2017.05.011.
Wang, Z., Chen, X., Chen, X. M., 2019. Ridesplitting is shaping young people’s travel behavior:
Evidence from comparative survey via ride-sourcing platform. Transportation Research
Part D: Transport and Environment 75, 57-71. https://doi.org/10.1016/j.trd.2019.08.017.
Wang, S., Noland, R. B., 2021. Variation in ride-hailing trips in Chengdu, China. Transportation
Research Part D: Transport and Environment 90, 102596.
https://doi.org/10.1016/j.trd.2020.102596.
Wang, H., Yang, H., 2019. Ridesourcing systems: A framework and review. Transportation
Research Part B: Methodological 129, 122-155.
https://doi.org/10.1016/j.trb.2019.07.009.
Xu, Y., Yan, X., Liu, X., Zhao, X., 2021. Identifying key factors associated with ridesplitting
adoption rate and modeling their nonlinear relationships. Transportation Research Part
38
A: Policy and Practice 144, 170-188. https://doi.org/10.1016/j.tra.2020.12.005.
Yan, X., Liu, X., Zhao, X., 2020a. Using machine learning for direct demand modeling of
ridesourcing services in Chicago. Journal of Transport Geography 83, 102661.
https://doi.org/10.1016/j.jtrangeo.2020.102661.
Yan, L., Luo, X., Zhu, R., Santi, P., Wang, H., Wang, D., Zhang, S., Ratti, C., 2020b.
Quantifying and analyzing traffic emission reductions from ridesharing: A case study
of Shanghai. Transportation Research Part D: Transport and Environment, 89, 102629.
https://doi.org/10.1016/j.trd.2020.102629.
Yang, H., Liang, Y., Yang, L., 2021. Equitable? Exploring ridesourcing waiting time and its
determinants. Transportation Research Part D: Transport and Environment, 93, 102774.
https://doi.org/10.1016/j.trd.2021.102774.
Yang, H., Lu, X., Cherry, C., Liu, X., Li, Y., 2017. Spatial variations in active mode trip volume
at intersections: A local analysis utilizing geographically weighted regression. Journal
of Transport Geography 64, 184-194. https://doi.org/10.1016/j.jtrangeo.2017.09.007.
Young, M., Farber, S., Palm, M., 2020. The true cost of sharing: A detour penalty analysis
between UberPool and UberX trips in Toronto. Transportation Research Part D:
Transport and Environment, 87, 102540. https://doi.org/10.1016/j.trd.2020.102540.
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. Journal of Transport Geography 75, 147-163.
https://doi.org/10.1016/j.jtrangeo.2019.01.004.
Yu, H., Peng, Z. R., 2020. The impacts of built environment on ridesourcing demand: A
neighbourhood level analysis in Austin, Texas. Urban Studies, 57(1), 152-175.
https://doi.org/10.1177/0042098019828180.
Zhang, B., Chen, S., Ma, Y., Li, T., Tang, K., 2020. Analysis on spatiotemporal urban mobility
based on online car-hailing data. Journal of Transport Geography 82, 102568.
https://doi.org/10.1016/j.jtrangeo.2019.102568.
Zhao, P., Xu, Y., Liu, X., Kwan, M. P., 2020. Space-time dynamics of cab drivers’ stay
behaviors and their relationships with built environment characteristics. Cities 101,
102689. https://doi.org/10.1016/j.cities.2020.102689.
