ArticlePDF Available

Do transportation network companies decrease or increase congestion?

Authors:
  • San Francisco County Transportation Authority

Abstract and Figures

This research examines whether transportation network companies (TNCs), such as Uber and Lyft, live up to their stated vision of reducing congestion in major cities. Existing research has produced conflicting results and has been hampered by a lack of data. Using data scraped from the application programming interfaces of two TNCs, combined with observed travel time data, we find that contrary to their vision, TNCs are the biggest contributor to growing traffic congestion in San Francisco. Between 2010 and 2016, weekday vehicle hours of delay increased by 62% compared to 22% in a counterfactual 2016 scenario without TNCs. The findings provide insight into expected changes in major cities as TNCs continue to grow, informing decisions about how to integrate TNCs into the existing transportation system.
Content may be subject to copyright.
ECONOMICS Copyright © 2019
The Authors, some
rights reserved;
exclusive licensee
American Association
for the Advancement
of Science. No claim to
original U.S. Government
Works. Distributed
under a Creative
Commons Attribution
License 4.0 (CC BY).
Do transportation network companies decrease or
increase congestion?
Gregory D. Erhardt
1
*, Sneha Roy
1
, Drew Cooper
2
, Bhargava Sana
2
, Mei Chen
1
, Joe Castiglione
2
This research examines whether transportation network companies (TNCs), such as Uber and Lyft, live up to
their stated vision of reducing congestion in major cities. Existing research has produced conflicting results
and has been hampered by a lack of data. Using data scraped from the application programming interfaces
of two TNCs, combined with observed travel time data, we find that contrary to their vision, TNCs are the
biggest contributor to growing traffic congestion in San Francisco. Between 2010 and 2016, weekday vehicle
hours of delay increased by 62% compared to 22% in a counterfactual 2016 scenario without TNCs. The findings
provide insight into expected changes in major cities as TNCs continue to grow, informing decisions about how
to integrate TNCs into the existing transportation system.
INTRODUCTION
Transportation network companies (TNCs) have grown rapidly in
recent years. In 2016, TNCs were 15% of all intra-San Francisco ve-
hicle trips, which is 12 times the number of taxi trips (1), while in New
York in 2016, TNC ridership equaled that of yellow cab and doubled
annually between 2014 and 2016 (2). TNCs are on-demand ride
services where rides are arranged through a mobile app to connect the
passenger with a driver, often a private individual driving their personal
vehicle (3). The current system is commonly viewed as a bridge tech-
nology that may be replaced by fleets of self-driving cars if and when
that technology is ready (4,5). TNCs are one form of shared mobility
and one form of Mobility-as-a-Service (MaaS). They have been re-
ferred to by several names, including ridesourcing, ride-hailing, and
e-hail. Ridesourcing is the preferred international standard (6), but
we refer to TNC throughout this text because it is the legal term used
in California, where this study was conducted.
Because they have the potential to reduce the reliance on private
cars, the TNCs themselves present a vision of the future in which
they reduce traffic congestion and allow roads to be repurposed to
other uses (7,8). There are several mechanisms by which TNCs could
reduce congestion. If TNCs are shared concurrently, a service known
as ridesplitting, they could reduce traffic if they replace a trip that
would otherwise be in a vehicle with fewer occupants. Simulations
show that ridesplitting has substantial potential to reduce congestion
(9). TNCs could induce travelers to shift trips from auto to transit by
providing better first- and last-mile connections to regional transit,
and there is some evidence to suggest that a small portion of travelers
may use TNCs in this way (10,11). Some have speculated that by
providing a convenient alternative to owning a car, TNCs could in-
centivize people to own fewer cars and, by extension, induce them
to shift other trips to transit or non-motorized modes, potentially re-
ducing their total vehicle travel (12,13).
Competing with these factors are several mechanisms by which
TNCs may increase traffic congestion. Deadheading, or out-of-service
movement, is the movement of a vehicle with no passenger. TNCs
and taxis deadhead to look for fares or reposition before or after a paid
trip. Out-of-service travel is estimated at about 50% of TNC vehicle
miles traveled (VMT) in New York (2) and 20% in San Francisco (1).
Whether a trip made by TNC adds traffic to the road also depends on
which mode would have been used for the trip if TNC was not avail-
able. Between 43 and 61% of TNC trips substitute for transit, walk, or
bike travel or would not have been made at all (10,11,14,15), adding
traffic to the road that otherwise would not have been there. TNC
pickups and drop-offs (PUDO) contribute to congestion on urban
streets by disrupting traffic flow inthecurblane,similartothecon-
gestioneffectsfoundinareasthat rely heavily on taxis (16).
Transportation planners and policy makers are interested in under-
standing the congestion effects of TNCs as they face decisions about
how to regulate TNCs and how to integrate them into the existing
transportation system (1719). However, studies assessing the net ef-
fect of TNCs on congestion have produced mixed results, concluding
that TNCs decrease congestion (20), TNCs add to VMT or increase
congestion (2,14,15), and TNCs did not drive the recent increase
in congestion(21), or have been inconclusive (10,11). There is a
need for further research to adjudicate these differences, but research
on the topic has been hampered by a lack of data (22,23). We enter
this debate to address the question: Do TNCs decrease or increase
traffic congestion?
We do this for the case of San Francisco while recognizing that the
results from a dense and transit-rich city may not translate into many
contexts. A data set scraped from the application programming inter-
faces of the two largest TNCs provides a unique insight into their
operations. These data were collected and processed as described by
Cooper et al. (23). We further processed the data to associate TNC
volumes, pickups, and drop-offs to each road segment in San Francisco
by time of day (TOD). These processed data are included in the Sup-
plementary Materials for use by other researchers.
This study is structured as a before-and-after assessment between
2010 conditions when TNC activity is negligible and 2016 conditions
when it is not, focusing on the change in average weekday conditions.
We derived measures of roadway conditions in both years from GPS-
based speed data licensed from INRIX. We estimated the relationship
between the change in TNC activity and the change in roadway travel
time, assuming zero TNCs in 2010.
To control for other factors that may also affect congestion over
this period, we used San Franciscos travel demand model, SF-CHAMP,
which produces estimates of traffic volumes on all roads in San Francisco
and is sensitive to changes in population and demographics, employ-
ment, transportation networks, and congestion. Since SF-CHAMPs
1
Department of Civil Engineering, University of Kentucky, 161 Raymond Building,
Lexington, KY 40506, USA.
2
San Francisco County Transportation Authority, 1455
Market Street, San Francisco, CA 94103, USA.
*Corresponding author. Email: greg.erhardt@uky.edu
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 1of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
initial development (24), it has been further enhanced (25,26), exten-
sively tested (27), and successfully applied to analyze policy and infra-
structure changes (28,29). The version of SF-CHAMP used in this
study was calibrated to 2010 conditions and does not account for
TNCs. This means that when the model is run for current-year inputs,
it represents a counterfactual case where TNCs do not exist.
The relationship between demand and traffic speed is nonlinear
such that adding vehicles in already congested conditions has a bigger
effect than adding them in uncongested conditions. Therefore, it is not
just the total VMT change that matters but when and where that
change occurs. We conducted our analysis directionally for segments
known as traffic messaging channels (TMCs), which average 0.3 miles
long. For each year, we aggregated all data to these TMC links and
averaged across days to represent average weekday conditions for five
TODs. These link-TOD-year combinations are more detailed than
past TNC studies, which are either more aggregate (2,13,20,21)or
based on smaller user surveys (1012,14,15)thatcannotbeexpanded
to the link level.
After estimating the relationships between the change in travel
times, TNCs, and control variables, we applied the estimated models
to evaluate network performance metrics for 2010, 2016, and a coun-
terfactual 2016 scenario with no TNCs. We compared the congestion
levels in these two scenarios to evaluate our research question. The dis-
cussion section of this paper addresses how our results relate to those of
the studies cited above, how the methods compare across these studies,
and the limitations of this study, focusing on other changes that may be
occurringoverthisperiod.
Observations and hypotheses
Like New York (2,21), San Francisco has experienced a notable in-
crease in congestion over the past few years (Fig. 1) (30). The speed
data used in this study confirm this trend, showing that the average
speed decreases from 25.6 miles per hour (mph) in 2010 to 22.2 mph
in 2016 and that the vehicle hours of delay (VHD) increase by 63%
over the same period. Delay is defined as the difference between the
congested travel time and the travel time under free-flow conditions.
This change corresponds to the period in which TNCs emerged.
Figure 2 shows the distribution of the TNC PUDO for an average
Wednesday in fall 2016. The data show that TNCs are concentrated
in the downtown area, consistent with findings elsewhere (11,13),
and in the locations where level-of-service deterioration is worst.
Several other changes may also affect congestion. Between 2010 and
2016, San Francisco population grew from 805,000 to 876,000 (31)and
employment grew from 545,000 to 703,000 (32). Important network
changes include a rebuild of the Presidio Parkway, the introduction
of turn restrictions on Market Street, several road diets,and bus im-
provements (33). We account for these changes through SF-CHAMP.
In addition, we reviewed a list of active construction projects during the
2016 analysis period to evaluate whether they were associated with dis-
proportionate speed decreases, and did not find that they were.
The data do not show the share of ridesplitting in San Francisco, but
it is between 13% and 20% elsewhere (14,15), with some of those trips
carrying no additional passengers (3,15). Rail ridership grows substan-
tially over this period and bus ridership does not (34), consistent with
otherfindingsthatTNCsmaycomplementrailandcompetewithbus
(11,35). We do not observe a meaningful change in car ownership, with
an average of 1.08 cars per household in 2010 and 1.10 cars per house-
hold in 2016 (36).
In addition to the 20% of TNC VMT that is out-of-service, 70% of
San Francisco TNC drivers live outside the city (1).Whilewedonot
explicitly track it in this study, the driverscommutes into the city
may add more VMT to the network. Our data do not provide a direct
Fig. 1. The p.m. peak period roadway level-of-service (LOS) in Sa n Francisco (30). (A) 2009 conditions; (B) 2017 conditions. LOS grades roadways by vehicle
delay, from LOS A representing free flow to LOS F representing bumper-to-bumper conditions. Data and an interactive mapping tool are available at congestion.
sfcta.org.
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 2of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
observation of what TNC users otherwise would have done, so they can-
not speak directly to modal substitution. The data do allow us to infer
the PUDO locations and associate those locations with specific direc-
tional roadways.
Some argue that TNCs have little effect on traffic operations because
they occur in the evening when congestion is less severe (12,13). Our
data show not only that 43% of TNC VMT occurs between 6:30 p.m.
and 3 a.m. but also that 26% of TNC VMT occurs inthe 3-hour a.m. or
p.m. peak periods compared to 40% for 4-hour peaks in Boston (15).
Given these observations, we suggest that the gap between the
background changes predicted by SF-CHAMP and the observed
change in travel times is an indicator of TNC impact. Specifically,
we hypothesize:
1) If TNCs have no effect on congestion, the background changes
should reasonably predict the observed travel time changes.
2) If TNCs decrease congestion, then the observed change in travel
time should be better than the background changes would predict.
3) If TNCs increase congestion, then the observed change in travel
time should be worse than the background changes would predict. We
expect the gap to be biggest for times and locations with high levels of
TNC activity.
MATERIALS AND METHODS
To test these hypotheses, we structured this study as a before-and-after
assessment between 2010 conditions where TNC activity is assumed
to be negligible and 2016 conditions when they are not. For each year,
an estimation data file is compiled with one observation on each road
segment and TOD combination. The data represent average weekday
conditions in the fall of each year. Fixed-effects panel data models (37)
are estimated where the dependent variable is a transformed version
of the observed travel time, and the descriptive variables include the
background traffic levels, TNC volumes, and TNC PUDO. We con-
verted the observed travel times to implied volumes using volume-
delay functions (VDFs). This time-implied volume is the models
dependent variable, and the conversion ensures that it is linearly re-
lated to the background and TNC volumes. The fixed-effects models
estimate coefficients based on the change between 2010 and 2016
Fig. 2. Daily TNC pickups and drop-offs for an average Wednesday in fall 2016 (1). Darker colors represent a higher density of TNC activity. Data and an interactive
mapping tool are available at tncstoday.sfcta.org.
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 3of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
conditions. There is precedent for using both before-and-after analysis
and panel data models in transportation analysis, including to study
changes in congestion (38), TNC growth (22), and the effects of new
technology (39). The estimated coefficients are applied to produce a
modeled estimate of 2010 and 2016 network conditions, as well as a
2016 counterfactual scenario that excludes the effect of TNCs.
Data
The analysis relies on three sources of data: background traffic esti-
mates, TNC data, and speed data. Those data and their processing
are described below.
Background traffic estimates
To estimate the net effect of TNCs on congestion, it is necessary to
control for other factors that are also expected to change congestion
levels, including changes to population, employment, and road and
transit networks. To control for these changes, this research uses
San Franciscostraveldemandmodel,SF-CHAMP.
SF-CHAMP is an activity-based travel demand microsimulation
model that is sensitive to a broad array of conditions that influence
travelerschoices. The model predicts the typical weekday travel pat-
terns for approximately 7.5 million San Francisco Bay Area residents,
including choices of vehicle availability, activity participation, destina-
tions, travel modes, and travel times. The simulated travel patterns are
sensitive to changes in population and demographics, employment,
transportation networks, and congestion. The model incorporates
detailed information about demographics and land use, using block,
block group, and tract-level geographies, and six broad employment
sectors. It also incorporates a detailed representation of the entire Bay
Area multimodal transportation system including roadways, transit
routes, and non-motorized facilities, as well as information about
how these change by TOD. The core behavioral components are based
on detailed travel surveys and capture time and cost trade-offs and
otherfactorsthatinfluencetravelerchoices,suchastheeffectsofde-
mographics and the availability and quality of alternatives. The model
has been used extensively in practice for almost two decades to eval-
uate long-range transportation plans, transportation infrastructure in-
vestments, pricing policies, and land use development proposals.
SF-CHAMP uses a detailed representation of the road network,
including a link for every street and in the city, along with attributes
that include length, number of lanes, capacity, turn restrictions, and
facility type. The outputs include an estimate of the average weekday
traffic volume and congested travel time on each link for each of five
TODs: 3 to 6 a.m., 6 to 9 a.m., 9 a.m. to 3:30 p.m., 3:30 to 6:30 p.m.,
and 6:30 p.m. to 3:00 a.m.
The analysis uses version 5.2.0 of SF-CHAMP, run using 2010
and 2016 inputs. The model runs uses actual inputs, not forecasts,
avoiding inaccuracies associated with errors in the inputs. This ver-
sion of SF-CHAMP was calibrated to 2010 conditions and does not
account for TNCs. Normally, this would be a limitation, but in this
case, it is beneficial because it means that when the model is run for
2016 population, employment, and network inputs, it represents a
counterfactual case where TNCs do not exist.
TNC data
Complementing SF-CHAMP are the TNC data, which were collected
and processed as described by Cooper et al. (23). The raw data show
the locations and time stamps of out-of-service TNC vehicles collected
in 5-s increments for a 6-week period in fall 2016, totaling about
12 terabytes of raw data. When a driver accepts a ride, that vehicle
no longer appears in the traces, and after the driver drops off the
passenger, the vehicle reappears. This structure allows the analyst
to infer that a trip was made between those two points. The point at
which the driver disappears from the traceisinferredasthelocation
of a passenger pickup, and the point at which it reappears is inferred
Table 1. Estimated relationships between PTI80 and TTI.
Facility type g
1
g
2
R
2
Freeways and expressways 1.029 1.498 0.831
Arterials 1.101 1.361 0.862
Collectors and locals 1.131 1.440 0.762
Table 2. Fixed-effects panel estimation results with TNC variables.
Parameter estimates
Variable Parameter Standard error T-statistic
SF-CHAMP background volume 0.9172 0.0541 16.952
Presidio Parkway scaling factor 0.3648 0.0189 19.327
TNC volume 0.6864 0.0720 9.5387
Average impact duration of TNC PUDO on major arterials (s) 144.75 7.7195 18.751
Average impact duration of TNC PUDO on minor arterials (s) 79.486 12.114 6.5617
Model statistics
Number of entities 7081
Number of time periods 2
R
2
between groups 0.5819
R
2
within groups 0.2985
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 4of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
as the passenger drop-off location. There is some uncertainty asso-
ciated with the pickup location because the driver must travel from
her current location to the location where the passenger is waiting,
but given the density of TNCs in San Francisco, the passenger wait
time is usually short. City-wide, the average wait time is 3 min (40),
and in our experience, it is often 1 to 2 min in the core of the city.
Duplicate traces are removed to avoid double-counting drivers who
work for both TNCs and vehicles recorded by multiple clients.
The TNC data were further processed for this study in several
ways. The out-of-service TNC vehicles were attached to directional
SF-CHAMP road links by TOD using a spatial matching process that
accounts for the trajectory of points. The in-service TNC volumes were
attached to directional road links by assigning each to the shortest path
between the inferred PUDO locations, where the shortest path is
calculated on the basis of the congested SF-CHAMP networks. Last,
the PUDO locations were assigned to directional road links, allowing
their effect on congestion to be measured. The end result is a set of
SF-CHAMP road networks that include the background traffic volumes
and other link attributes, and are annotated with 2016 TNC activity.
These are for average weekday conditions, segmented by SF-CHAMPs
five time periods. To the extent that in-service TNC volumes substitute
for other auto trips, we expect some overlap between these and the
background SF-CHAMP volumes.
Speed data
We use archived speed data from INRIX, a commercial vendor, that is
available in 5-min increments for each day from 2010 through the pres-
ent, allowing both the average travel time and reliability metrics to be
calculated. Spatially, the data are available directionally for segments
known as TMCs, which in San Francisco average about 0.3 miles in
length, or about three city blocks. TMCs exclude many local roads
but otherwise provide good coverage throughout the city. Links asso-
ciated with TMCs carry about 70% of the total VMT in San Francisco.
This study uses INRIX speed data, at a 5-min temporal resolution,
for non-holiday weekdays for the 6-week period in November and
December 2016 when TNC data were collected, and for a compara-
ble 6-week period in November and December of 2010. The data are
providedforeachTMCsegmentwith day and time stamps. A reference
speed is also available in the dataset representing speed under uncon-
gested condition.
The speed data depend upon probe vehicles and therefore varies in
confidence scores depending upon the TOD and presence of vehicles on
each TMC link that provides these data. For the purpose of this study,
INRIX speeds pertaining only to the highest confidence score of 30 are
used to calculate a reliable estimate for link-resolved travel time. Further, a
comprehensive evaluation of the data was conducted, including a compar-
ison to speed data from San Franciscos Congestion Management Program
(30). TMC links with unreasonable speeds were excluded from the analysis.
For example, a surface street running parallel to a freeway showed un-
reasonably high speeds, which we suspect is the link picking up probe
vehicles from the adjacent freeway. Additional data assurance is performed
to identify and exclude data labeled with the wrong travel direction.
Some TMC segments are filler segments.Links lying between two
stop bars at a traffic signal or unsignalized intersections, links denoting
the change in direction of a roadway, etc. are some examples of filler
segments. Because these links are extremely short in length (typically,
shorter than 0.025 miles) and, more importantly, not representative of a
typical roadway segment, they are excluded from the analysis. In total,
23% of TMCs were excluded from the analysis, but these TMCs account
for less than 4% of the total TMC road length.
To incorporate the predicted volume obtained from the SF-CHAMP
model, as well as normalizing the growth in background traffic attrib-
utable to the typical non-TNC factors, it is required to create an asso-
ciation between the TMC network and the SF-CHAMP network. The
remaining TMC links are associated with the corresponding SF-CHAMP
links. In most cases, SF-CHAMP links aggregate to TMC links. In in-
stances when a CHAMP segment is longer than a TMC segment,
multiple TMC segments were merged together to form one composite
TMC segment and correspond to the said CHAMP segment. In a few
cases, such as in some of the more complex freeway interchanges, a clean
correspondencecouldnotbeidentifiedbetweentheSF-CHAMPlinks
and the TMC links. Those cases were excluded from the analysis.
The 5-min speed data were aggregated to average weekday mea-
suresforeachofthefiveSF-CHAMPtimeperiods.Duringthisaggre-
gation, several speed metrics were calculated, including the mean, the
Table 3. Network performance metrics.
Scenario
Network performance metrics
VMT
Based on modeled travel time Based on observed travel time
VHT VHD Average speed (mph) PTI80 VHT VHD Average speed (mph) PTI80
2010 4,923,449 205,391 64,863 24.0 1.83 204,686 64,158 24.1 1.83
2016 no TNC 5,280,836 230,642 79,449 22.9 1.94 N/A N/A N/A N/A
2016 with TNC 5,559,412 266,393 105,377 20.9 2.12 269,151 108,134 20.7 2.21
Scenario
Percent change from 2010
VMT
Based on modeled travel time Based on observed travel time
VHT VHD Average speed (mph) PTI80 VHT VHD Average speed (mph) PTI80
2010 0% 0% 0% 0% 0% 0% 0% 0% 0%
2016 no TNC 7% 12% 22% 4% 6% N/A N/A N/A N/A
2016 with TNC 13% 30% 62% 13% 15% 31% 69% 14% 21%
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 5of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
standard deviation, the 5th percentile, and the 20th percentile. The
highest observed average hourly speed on each TMC link over the
observation period was assigned as the free-flow speed for that link.
Examination shows that the free-flow speed on a segment remained
largely unchanged between 2010 and 2016.
Merging the data
The data were merged such that TMC links serve as the common spatial
units for the remainder of the analysis. When the data were aggregated
from the SF-CHAMP links to the TMC links, the link attributes were
also aggregated. Volumes and capacities were combined using a
length-weighted average. There are two measures of distance: one from
theSF-CHAMPlinksandonefromtheTMClinks.TheSF-CHAMP
links are more spatially accurate, so the sum of the SF-CHAMP link
length was used as the primary measure of length in the combined data
set.IntheeventwheremultipleTMCsegmentsneedtobeaggregated,
the space mean speed was estimated by dividing the combined TMC
length by the sum of travel time across all TMCs. The speed was then
applied to the length of the combined SF-CHAMP links.
All of this was done for both 2010 and 2016 scenarios. The 2010 and
2016 data were matched for each TMC segment, and if there were
missing data in one or the other, both records were dropped. This can
happen, particularly in the 3 to 6 a.m. time period, if there are insufficient
Fig. 3. Speed (mph) difference between 2016 scenario with TNCs and a counterfactual 2016 scenario without TNCs. D ata represent four times of day: (A) 6 to 9 a.m.;
(B) 9 a.m. to 3:30 p.m.; (C)3:30to6:30p.m.;and(D) 6:30 p.m. to 3:00 a.m. Data are provided in the Supplementary Materials.
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 6of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
probe vehicles to achieve the highest confidence score in the INRIX data.
The end result is a matched panel with 2010 and 2016 for a total of 7082
TMC linkTOD combinations. This corresponds to 1450 TMC links
with up to five TODs each. The resulting estimation files are included
in the Supplementary Materials as part of data S3.
METHODS
To estimate the effect of TNCs, we used a fixed-effects panel data regres-
sion model (37). The fixed-effects standardize the link-dependent un-
explained constancy or variance that might affect the regressed variable.
Some examples of link-specific characteristics are location of links near
high foot traffic, recreational areas, and special roadway geometry. The
temporal unit used by the panel is 2, warranted by the before-after
nature of the study. Each data point in the dataset is a unique combi-
nation of a TMC, TOD, and observation year. Because there are only
two points in time, this is equivalent to estimating an ordinary least
squares (OLS) model on the change on each TMC for each TOD.
Converting travel time to implied volume
A challenge in estimating these models is that they assume a linear
relationship between the dependent variable and the regressors, but
the relationship between volume and travel time is nonlinear. To deal
with this, the VDFs from SF-CHAMP were used to convert the ob-
served travel times into implied passenger car equivalent (PCE) volumes.
They take the form
T¼T01þaV
C

b
! ð1Þ
where Tis the congested travel time, T
0
is the free-flow travel time, Vis
the traffic volume in PCEs, Cis the link capacity, and aand bare cali-
brated parameters. Solving for V,weget
VI¼C
T
T01
a
!
1=b
ð2Þ
where the subscript on V
I
is used to designate a time-implied volume, as
derived from the travel times. The panel models use V
I
as their de-
pendent variable. It is in units of PCEs, so it is linearly related to the vol-
ume measures in the descriptive variables.
The analysis was conducted for five multi-hour time periods, so it is
importantthatallvolumesandcapacitiesareeitherhourlyorforthe
period as a whole. Here, we defined them for the period as a whole
and scaled the hourly capacities to the period total using the same peak
hour factors that were used by SF-CHAMP.
Congestion effects of PUDO
In considering the effect of TNC PUDO on congestion, it is useful to
consider other scenarios in which a vehicle movement has an effect on
congestion beyond simply driving on the roadway. Several examples
where this occurs include taxis (16), delivery trucks (41), and move-
ments into or out of on-street parking spaces (42,43). Wijayaratna
(44) provides a useful method for considering the congestion effect of
on-street parking that follows the capacity adjustment approach used
frequently in the Highway Capacity Manual (45). The approach scales
the capacity of the road lane adjacent to the on-street parking based on
the share of time that the lane is blocked. To model the effect of TNC
PUDO, we took a similar approach, but defined the PUDO effect in
PCEs so that it was in the same units as our dependent variable, and
expressed the effect as
bAvgDur PUDO PHF
3600 C
Lð3Þ
where PUDO is the number of PUDO in the period, PHF is the peak
hour factor to convert the PUDO to an hourly value, Cis the capacity of
the link, Lis the number of lanes, and b
AvgDur
is an estimated model
parameter. For simplicity, we expressed this term, excluding the esti-
mated coefficient, as V
AvgDur
.b
AvgDur
can be interpreted as the average
durationthateachPUDOblocksordisturbstrafficinthecurblane.In
congested conditions, this can be longer than the duration of the stop
itself, because it can take some time for a queue to dissipate if it builds up
behind a stopped vehicle and for traffic to recover to its pre-PUDO con-
dition. b
AvgDur
can also be shorter than the actual duration of a stop if
there is some probability that the stopping vehicle can pull out of traffic
or if volumes are low enough that the probability of a vehicle arriving
behind the stopped vehicle is low.
Model estimation
To estimate the effect of other factors on the change in implied vol-
ume, we used a fixed-effects panel data regression model (37). The
fixed-effects standardize the link-dependent unexplained constancy
or variance that might affect the regressed variable. Some examples
of link-specific characteristics are location of links near high foot traf-
fic, recreational areas, and special roadway geometry. Because these
characteristics do not change between 2010 and 2016, their influence
is absorbed into the fixed effect, preventing them from biasing the other
parameter estimates. The temporal unit used by the panel is 2, war-
ranted by the before-after nature of the study. Each data point in the
dataset is a unique combination of a TMC, TOD, and observation year.
Becausethereareonlytwopointsintime,thisisequivalenttoestimat-
ing an OLS model on the change on each TMC for each TOD. The es-
timated model can be expressed as
VI:i;t¼b1VSFCHAMP:i;t
þb2VTNC:i;t
þb3FTMajArt:iVAvgDur:i;t
þb4FTMinArt:iVAvgDur:i;t
þb5PRESIDIOi;tVI:i;2010
þFEiþei;t
ð4Þ
where the entities iare TMC links by TOD and the time periods tare
either 2010 or 2016, and each is used to index the remaining varia-
bles. V
I:i,t
is the time-implied volume. V
SF CHAMP:i,t
is the volume
predicted by SF-CHAMP in PCE, giving some additional weight to
trucks and buses. V
AvgDur:i,t
is the average duration variable, as defined
above. FT
MajArt:i
is a binary facility type flag indicating whether or
not the link is a major arterial, and FT
MinArt:i
is a binary facility type
flag indicating whether or not the link is a minor arterial. These fa-
cility type flags do not change between the two years. PRESIDIO
i,t
is
a binary flag identifying links on the Presidio Parkway and Veterans
Boulevard, where there was major construction in 2010 but not in
2016. PRESIDIO
i,t
is defined to be zero in 2010 and one in 2016 such
that the effect of a change can be estimated. V
I:i,2010
is the time-implied
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 7of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
volume in period 1 (2010), which allows the effect of the construction
change to be proportional to the starting volume on the link, as opposed
to additive and the same on every link. FE
i
is the fixed-effect, which is
effectively a constant on each entity, and e
i,t
is a random error term. In
this specification, the Presidio flag (PRESIDIO
i,t
) and the TNC terms
(V
TNC:i,t
,V
AvgDur:i,t
) are zero in 2010, so the 2010 time-implied volume
is simply a function of the SF-CHAMP volume plus the fixed effect and
an error term.
Anumberofvariationsonthisspecificationwereattemptedbefore
arriving at the preferred model. For example, specifications were tested
that split the TNC volume into separate in-service and out-of-service
volumes or segmented the PUDO coefficients in different dimensions.
One notable variation relates to our hypothesis that TNCs have no effect
ontrafficcongestion.Ifthisweretrue,wewouldexpectthechangein
background volume alone to reasonably predict the change in time-implied
volume (V
I
). Estimating such a model reveals that the background volume
is highly correlated with V
I
, with a coefficient of 1.78. This suggests that
time-implied volumes are increasing by 78% more than SF-CHAMP
would predict. It appears that the employment, population, and network
changes do not fully describe the congestion changes observed during this
period, and more terms are needed to do so.
Model application
After the model was estimated, it was applied to all links to predict the
V
I:i,t
for 2010 and 2016. It was also applied to predict a 2016 counter-
factual scenario with no TNCs by setting V
TNC:i,2016
and V
AvgDur:i,2016
to zero and otherwise applying the model to 2016 data. These pre-
dicted PCEs were then used to calculate the travel times using the
VDFs (Eq. 1).
The non-PCE volume on each link is calculated as
Vi;t¼VSFCHAMP:i;tþb2VTNC:i;tð5Þ
where V
i,t
is the traffic volume in units of vehicles instead of PCEs
and V
SF CHAMP:i,t
is the SF-CHAMP volume. b
1
is excluded such that
we count the full SF-CHAMP traffic volume, not their estimated effect
on speed. The inclusion of b
2
(which is less than one) accounts for the
partial overlap between the TNC volumes and the background volumes.
These volumes were combined with the link lengths to calculate VMT,
and combined with travel times to calculate vehicle hours traveled
(VHT) and VHD. The average speed is calculated as VMT/VHT.
The same volumes were used in combination with observed travel times
to calculate observed VHT, VHD, and average speed. In addition, a set
of reliability metrics was calculated as described below.
Travel time reliability metric
This study uses planning time index 80 (PTI80) as the measure of
travel time reliability. It is defined as
PTI80 ¼T80
T0
ð6Þ
where T
80
is the 80th percentile travel time and T
0
is the free-flow travel
time.APTI80valueof1.5meansthatfora30-mintripinlighttraffic,
45minshouldbeplannedtoensureon-timearrival80%ofthetime.
PTI80 can be calculated directly using measured travel times, or
estimated as a function of the travel time index (TTI) (46), which is
theratiobetweentheaveragetraveltimeandthefree-flowtraveltime.
The estimated relationship for each observation itakes the form
PTI80i¼g1TTIig2ð7Þ
where g
1
and g
2
are estimated model parameters. These parameters
were estimated for this study from the observed travel time data from
both 2010 and 2016, with one observation for each TMC, TOD, and
year combination. The relationships are specific to each facility type.
Table 1 shows the results of that estimation. PTI80 was calculated for
each TMC link, TOD, and year combination, and was aggregated to
the network level using a VMT-weighted average.
RESULTS
Table 2 shows our model estimation results from the fixed-effects
models. The SF-CHAMP background volume parameter estimate is
0.92, not significantly different than 1. This is logical, because we ex-
pect that each vehicle added in background traffic should have an ef-
fect on congestion of adding one vehicle to the implied volume. The
Presidio Parkway scaling factor accounts for major construction that
was underway on those links in 2010 but not 2016, and is equivalent
to reducing the 2010 implied traffic volume by 36%.
We include two measures of time- and location-specific TNC activ-
ity. The TNC volume parameter measures the net effect of TNCs. If
TNCs purely substitute for other car trips, the estimated TNC param-
eter should be zero as they substitute for other vehicles already counted
in the background volumes. Negative values would be consistent with
TNCs reducing traffic, while a value of positive 1 would be consistent
with TNCs purely adding to background traffic. The estimated co-
efficient of 0.69 can be interpreted as an addition of one TNC vehicle,
partially offset by a subtraction of 0.31 non-TNC vehicles.
The PUDO parameters represent the average number of seconds
that a pickup or drop-off disrupts traffic in the curb lane. Locally col-
lected data show that the average time needed for a passenger to board
oralightfrompassengervehiclessuchasTNCsandtaxisisabout1min.
The higher average impact durations estimated in these models suggest
that the traffic disruption persists after the stopped vehicle departs be-
cause additional time is needed for traffic flow to recover to its pre-
PUDO condition.
We applied the estimated model to assess network-wide perform-
ance metrics for three scenarios:
1) 2010: reflecting observed 2010 conditions, when no TNCs were
present;
2) 2016 no TNC: represents a counterfactual scenario of what 2016
conditions would be if there were no TNCs;
3) 2016 with TNC: the full application of the model to 2016
conditions.
Table 3 presents network performance metrics for these three
scenarios. VMT grows by 13% between 2010 and 2016, with almost
half of the VMT increase attributable to TNCs. We calculate VHT,
VHD, and average speed using both modeled travel times and,
where available, observed travel times. Without TNCs, VHT would
be 12% higher in 2016 than in 2010, VHD would be 22% higher,
and average speed would be 4% lower. With TNCs, VHT is 30%
higher, VHD is 62% higher, and speeds are 13% lower.
In addition, travel time is becoming less reliable, as measured by
PTI80. PTI80 is the ratio between the 80th percentile travel time and
the free-flow travel time. It is a measure of the day-to-day variability of
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 8of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
travel time. A PTI80 value of 1.8 means that for a 10-min trip in un-
congested condition, 18 min should be planned to ensure on-time ar-
rival 80% of the time. Between 2010 and 2016, PTI80 increases by 15%
with TNCs or by 6% without TNCs.
The distribution of congestion effects is not uniform throughout
the network or throughout the day. Figure 3 maps the speed difference
between the TNC scenario and the no-TNC counterfactual for four
TODs. TNCs have a larger effect on congestion in the downtown area
and on arterial roadways. TNCs have a disproportionately large effect
on evening congestion, but they also increase congestion in the peak
periods: a 48 to 52% increase in VHD in the a.m. and p.m. periods
with TNCs versus an 18 to 23% increase for the no-TNC counter-
factual. Additional tables showing TNC effect by TOD, facility type,
and area type are in the Supplementary Materials.
DISCUSSION
Our results show higher VMT and more congestion in the 2016 TNC
scenario than in the no-TNC counterfactual. These results are
consistent with the subset of TNC rider surveys that were able to draw
a conclusion about the net VMT effect of TNCs (14,15). Both of these
studies were based on intercept surveys where the TNC driver asks the
passenger what mode they would have chosen had they not used a
TNC. Our results are also consistent with the most recent findings
in New York that TNCs add VMT and increase congestion (2). This
study is the most similar to our own, in that it is based on an assess-
ment of data on e-hail trips, which are required to be reported to the
citys Taxi and Limousine Commission. Our results provide comple-
mentary evidence to the subset of surveys that were inconclusive re-
garding the net effect of TNCs on VMT (10,11). The first such study
is based on a survey distributed at TNC hot spots in San Francisco
during May and June 2014 (10). This study is useful because it
provided an early view into the TNC market, although it is recognized
both that the market has evolved over the intervening 2.5 years and
that this intercept method provides for a sample that is spatially tied to
trips at those specific hot spots. The second is based on a survey con-
ducted from 2014 to 2016 in seven U.S. cities (11). It explicitly explores
the question of whether TNC users reduce their vehicle holdings, and
whether that results in reduced vehicle travel, finding Those who have
reduced the number of cars they own and the average number of miles
they drive personally have substituted those trips with increased ride-
hailing use. Net VMT changes are unknown.Ourfindingsdifferfrom
the conclusions of several other studies (9,12,13,20,21). The relation-
ship between our findings and those of other studies is discussed below.
A study by Li et al. finds reasonable evidence that the entry of
Uber significantly decreases traffic congestion in the urban areas of
the U.S.(20). This study estimates models of the change in annual
congestion in metropolitan areas from 1982 to 2014, as measured by
the Urban Mobility Report (47). It introduces a binary variable into the
model based on the year of Ubers entry into each market and uses the
negative coefficient estimate as the basis for their conclusion. There
are two issues with this approach. First, it does not reflect spatial detail
in the distribution of TNCs, which are heavily concentrated in down-
town areas, so the aggregate nature of the study may obscure the
underlying effect. Second, it does not capture the quantity of TNC
use, which varies between cities and continues to grow after entering
a market. Our study does better on both accounts.
The City of New York (21) used New Yorks travel demand model
to develop 2010 and 2020 VMT estimates and examined e-dispatch
trip records in comparison to those total VMT estimates. They based
their conclusion that TNCs did not drive the recent increase in con-
gestion on a projection that TNCs largely substitute for yellow taxi
trips and on a lack of evidence for congestion effects associated with
PUDO. Our results show that, at least in San Francisco, substitution
for taxis and cars only offsets a portion of the TNC volume, and they
provide evidence of a PUDO effect.
Simulations, such as the Portugal Study,showing large benefits
from ridesplitting assume full participation and centralized optimiza-
tion (9). These assumptions do not reflect the way in which TNCs
operate today. While our data do not include vehicle occupancy, other
survey data show a modest share of ridesplitting (14,15), and our
results suggest that it is not sufficient to offset the ways in which TNCs
add to congestion. Such simulations can be useful in establishing the
positive potential of ridesplitting if such a system were effectively man-
aged to achieve socially desirable outcomes, but do not imply that TNCs
will achieve those outcomes on their own.
Two notable studies by Feigon and Murphy (12,13)promotethe
idea of TNCs as a complement to public transit. These studies base
their conclusions primarily on data summaries generated from surveys
of shared mobility users. Feigon and Murphy conclude that because
TNC use is high in the evening and weekend periods when transit
service is less frequent, TNCs largely complement public transit and
enhance urban mobility (12,13). However, their own data show (13),
and ours confirm, that TNC use is also high during the peak periods
when congestion is worst and transit service is frequent. Feigon and
Murphy find that a greater use of shared modes is associated with more
frequent transit use (12). However, this finding should not be taken to
imply a directional relationship, as it could be that frequent transit users
are likely to switch some trips to TNC, adding traffic to the roads.
Feigon and Murphy also note that TNC use is associated with decreases
in respondentsvehicle ownership and private vehicle trips (13). While
this may be true of specific users, we do not observe aggregate changes
in vehicle ownership in San Francisco between 2010 and 2016. Further,
this finding only accounts for the subtraction of private vehicle trips, not
the addition of TNC vehicle trips. Our results indicate that the net effect
of TNCs is to add more vehicles to the road.
Some limitations of this study are worth noting. First, the analysis
relies on VDFs that are limited in their ability to capture the underly-
ing complexity of traffic flow (48). They should be viewed as a means
of understanding the aggregate relationships observed in the data, not
of the expected operations at a specific location.
Second, while the predicted background traffic changes account for
several important control variables, there remains a risk that our results
are confounded by another factor. For example, our analysis controls
fordemographicandsocioeconomicchangesoverthisperiod,butlike
all travel models, SF-CHAMP assumes that the relationship between
those inputs and the resulting travel behavior remains stable. If there
are major behavioral changes over this period, it could affect the result.
For example, representatives from Uber argue that growing congestion
may instead be due to growing freight deliveries or increased tourism
(49). We discuss each of these possibilities.
Regarding freight deliveries, our analysis reflects growth in truck
travel associated with growing employment, but it does not account
for structural changes such as a large shift from in-person to online
shopping. Such a shift could increase delivery truck volumes but de-
crease personal shopping trips (50). The net effect of this trade-off is
not clear and depends on factors such as how efficiently the delivery
vehicle can chain multiple deliveries together, what TOD the different
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 9of11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
trips would occur, and whether the deliveries are to commercial lo-
cations in the downtown area or to less congested residential areas.
Unfortunately, we lack the commercial vehicle data necessary to
evaluate that effect.
Regarding tourism, SF-CHAMP does include a visitor model, with
visitor travel representing 4.5% of intra-San Francisco person trips in
2010. The visitor model is influenced primarily by the number of hotel
rooms in the city, which have not increased significantly over this pe-
riod. Data from the San Francisco Council on Economic Development
(51) show that the number of visitors to San Francisco grew by 22%
between 2010 and 2016, with some of those visitors staying in lodging
options such as Airbnb, which are not reflected in our visitor model.
When this growth is applied to the base of 2010 visitor trips, it might
generate up to 1% more intra-San Francisco person trips, beyond what
is already included in the background growth. However, these visitor
trips only add to congestion if they are in a vehicle, with transit, walk,
and TNC being the most commonly used modes among visitors. Thus,
due to the growth in tourism, the total vehicle trips in 2016 may be a
fraction of a percent higher than we estimate in the background traffic
volumes. For comparison, TNCs are 15% of intra-San Francisco vehicle
trips in 2016.
As we consider the possibility of other uncontrolled factors, it is
worth keeping in mind a few aspects of our research. To have an
effect, any uncontrolled factors must be different between 2010 and
2016. Also, our estimation results show that congestion is growing
more than expected specifically on the links and in time periods with
high levels of TNC activity. The most problematic factors would be
those that are spatially and temporally correlated with TNCs,
occurring on those same links in the same time periods.
Third, the analysis presented here is specific to a single city with a
dense urban core and a rich transit system. The data show that TNC use
is heavily concentrated in the densest portion of that city, consistent
with evidence from other cities (13). While we may expect similar
results in other comparable cities, further research is needed to confirm
that expectation. The effects of TNCs may be quite different in smaller
cities, in less dense areas, or in places with very different populations or
regulatory environments.
Several extensions would complement this research: better under-
standing the contributors to background growth, assessing the TNC ef-
fect on transit ridership, and considering how worsening congestion
and travel time reliability affect transit operations. Last, the study should
be repeated elsewhere to understand how the results vary in cities of
different sizes and compositions.
CONCLUSIONS
This study examines the effect of TNCs on traffic congestion and
reliability in San Francisco. It is intended to adjudicate between
competing arguments about whether TNCs decrease or increase
congestion.
The results show that the observed changes in travel time are worse
than the background changes would predict. The estimated TNC vol-
ume and PUDO coefficients show that travel times get worse on roads
with more TNC activity than on roads with less TNC activity after
controlling for background traffic changes. This result supports the hy-
pothesis that TNCs increase congestion, at least in San Francisco.
The results show some substitution between TNCs and other car
trips,butthatmostTNCtripsareaddingnewcarstotheroad.The
estimated models show that TNC vehicles stopping at the curb to pick
up or drop off passengers have a notable disruptive effect on traffic flow,
especially on major arterials.
The model is applied to estimate network-wide conditions for 2016
and for a counterfactual scenario that estimates what conditions would
be in 2016 if there were no TNCs. Both are compared to a 2010 base-
line, before TNCs. VMT, VHT, and VHD increase by 13, 30, and 62%,
respectively, from 2010 to 2016. Without TNCs, those same metrics
would have increased by 7, 12, and 22%. Average speeds decrease by
13%, compared to a 4% decrease without TNCs. TNCs are associated
with worsening travel time reliability, thus requiring travelers to further
buffer their travel times if they wish to consistently arrive on time.
These results lead us to conclude that TNCs are the biggest factor
driving the rapid growth of congestion and deterioration of travel
time reliability in San Francisco between 2010 and 2016, exceeding
the combined effects of population growth, employment growth,
and network changes.
These findings are of interest to transportation planners, to policy
makers, and to the general public in San Francisco and other large
cities. It is in the public interest that decisions about the regulation
of TNCs, the allocation of curb space and right-of-way, and the inte-
gration of new mobility services with existing transit operations be
based on independent and peer-reviewed analysis as presented here.
SUPPLEMENTARY MATERIALS
Supplementary material for this article is available at http://advances.sciencemag.org/cgi/
content/full/5/5/eaau2670/DC1
Supplementary Text
Fig. S1. Area type map on SF-CHAMP links.
Table S1. Fixed-effects panel model estimation results only accounting for background traffic.
Table S2. Network performance metrics by TOD.
Table S3. Network performance metrics by area type.
Table S4. Network performance metrics by facility type.
Data S1. Supporting data for Fig. 1.
Data S2. Supporting data for Fig. 2.
Data S3. Model estimation files.
Data S4. Model application results and supporting data for Fig. 3.
REFERENCES AND NOTES
1. San Francisco County Transportation Authority, TNCs Today: A Profile of San Francisco
Transportation Network Company Activity (San Francisco County Transportation
Authority, 2017).
2. B. Schaller, Unsustainable? The Growth of App-Based Ride Services and Traffic, Travel and
the Future of New York City (Schaller Consulting, 2017); http://www.schallerconsult.com/
rideservices/unsustainable.pdf.
3. Transportation Research Board, Between public and private mobility: Examining the
rise of technology-enabled transportation services(TRB Special Report 319,
Transportation Research Board, 2016).
4. J. P. Zmud, I. N. Sener, Towards an understanding of the travel behavior impact of
autonomous vehicles. Transp. Res. Procedia 25, 25002519 (2017).
5. D. J. Fagnant, K. M. Kockelman, Dynamic ride-sharing and fleet sizing for a system of
shared autonomous vehicles in Austin, Texas. Transportation 45, 143158 (2018).
6. SAE International, Taxonomy and Definitions for Terms Related to Shared Mobility and
Enabling Technologies (SAE International, 2018); http://standards.sae.org/J3163_201809.
7. Uber Newsroom; https://www.uber.com/newsroom/company-info/.
8. J. Zimmer, The Third Transportation Revolution (2016); https://medium.com/
@johnzimmer/the-third-transportation-revolution-27860f05fa91.
9. L. M. Martinez, J. M. Viegas, Assessing the impacts of deploying a shared self-driving
urban mobility system: An agent-based model applied to the city of Lisbon, Portugal.
Int. J. Transp. Sci. Technol. 6,1327 (2017).
10. L. Rayle, D. Dai, N. Chan, R. Cervero, S. Shaheen, Just a better taxi? A survey-based
comparison of taxis, transit, and ridesourcing services in San Francisco. Transp. Policy 45,
168178 (2016).
11. R. R. Clewlow, G. S. Mishra, Disruptive transportation: The adoption, utilization, and
impacts of ride-hailing in the United States(Research Report UCD-ITS-RR-17-07,
UC Davis, 2017).
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 10 of 11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
12. S. Feigon, C. Murphy, Shared mobility and the transformation of public transit(Transit
Cooperative Research Program Report 188, Transportation Research Board, 2016).
13. S. Feigon, C. Murphy, Broadening understanding of the interplay between public transit,
shared mobility, and personal automobiles(Transit Cooperative Research Program
Report 195, Transportation Research Board, 2018).
14. A. Henao, Impacts of ridesourcingLyft and Uberon transportation including VMT,
mode replacement, parking, and travel behavior,thesis, University of Colorado at
Denver (2017).
15. S. Gehrke, A. Felix, T. Reardon, Fare Choices Survey of Ride-Hailing Passengers in Metro
Boston (Metropolitan Area Planning Council, 2018).
16. I. Golias, M. G. Karlaftis, The taxi market in Athens, Greece, and its impact on urban traffic.
Transp. Q. 55,6371 (2001).
17. J. Kuhr, C. R. Bhat, J. Duthie, N. Ruiz, Ridesharing & public-private partnerships:
Current issues, a proposed framework and benefits, in Transportation Research Board 96th
Annual Meeting, Washington, DC, 8 to 12 January 2017.
18. M. Moran, P. Lasley, Legislating transportation network companies. Transp. Res. Rec. 2650,
163171 (2017).
19. A. Cohen, S. Shaheen, Planning for shared mobility(PAS Report 583, American Planning
Association, 2016), p. 110.
20. Z. Li, Y. Hong, Z. Zhang, An empirical analysis of on-demand ride-sharing and traffic
congestion, in International Conference on Information Systems (ICIS16), Dublin, Ireland,
11 to 14 December 2016.
21. City of New York, Office of the Mayor, For Hire Vehicle Transportation Study (City of
New York, Office of the Mayor, 2016).
22. R. Gerte, K. C. Konduri, N. Eluru, Is there a limit to adoption of dynamic ridesharing
systems? Evidence from analysis of Uber demand data from New York City, in TRB Annual
Meeting, Washington, DC, 2018.
23. D. Cooper, J. Castiglione, A. Mislove, C. Wilson, Profiling TNC activity using big data,
in TRB Annual Meeting, Washington, DC, 2018.
24. N. Jonnalagadda, J. Freedman, W. A. Davidson, J. D. Hunt, Development of
microsimulation activity-based model for San Francisco: Destination and mode choice
models. Transp. Res. Rec. 1777,2535 (2001).
25. G. Erhardt, B. Charlton, J. Freedman, J. Castiglione, M. Bradley, Enhancement and
application of an activity-based travel model for congestion pricing, in Innovations in
Travel Modeling Conference, Portland, OR, 22 to 24 June 2008.
26. L. Zorn, E. Sall, D. Wu, Incorporating crowding into the San Francisco activity-based travel
model. Transportation 39, 755771 (2012).
27. M. Outwater, B. Charlton, The San Francisco model in practice: Validation, testing, and
application, in Innovations in Travel Demand Modeling, Volume 2: Papers, Austin, TX, 21 to
23 May 2006.
28. J. Castiglione, R. Hiatt, T. Chang, B. Charlton, Application of travel demand
microsimulation model for equity analysis. Transp. Res. Rec. 1977,35
42 (2006).
29. E. M. Brisson, E. Sall, J. Ang-Olson, Achieving goals of San Francisco, California, for
greenhouse gas reductions in transportation sector: What would it take? Transp. Res. Rec.
2287,8997 (2012).
30. San Francisco County Transportation Authority, 2017 Congestion Management Program
(San Francisco County Transportation Authority, 2017).
31. U.S. Census Bureau, Annual Estimates of the Resident Population Table PEPANNRES.
32. U.S. Bureau of Labor Statistics, Quarterly Census of Employment and Wages.
33. SFMTA, Projects; https://www.sfmta.com/projects.
34. G. D. Erhardt, Fusion of large continuously collected data sources: Understanding travel
demand trends and measuring transport project impacts,thesis, University College
London (2016).
35. R. Mucci, Transportation network companies: Influencers of transit ridership trends,
thesis, University of Kentucky (2017).
36. U.S. Census Bureau, American Community Survey 2010 and 2016 1-Year Estimates
(U.S. Census Bureau, 2016).
37. W. H. Greene, Econometric Analysis (Prentice Hall, ed. 5, 2003).
38. R. Hanna, G. Kreindler, B. A. Olken, Citywide effects of high-occupancy vehicle restrictions:
Evidence from three-in-onein Jakarta. Science 357,8993 (2017).
39. L. Tang, P. Thakuriah, Ridership effects of real-time bus information system: A case study
in the city of Chicago. Transp. Res. Part C Emerg. Technol. 22, 146161 (2012).
40. San Francisco County Transportation Authority, Emerging mobility evaluation
report(Draft Report, San Francisco County Transportation Authority, 2018);
http://www.sfcta.org/sites/default/files/Emerging%20Mobility%20Evaluation%20Report
%2004242018.pdf).
41. N. Chiabaut, Investigating impacts of pickup-delivery maneuvers on traffic flow
dynamics. Transp. Res. Procedia 6, 351364 (2015).
42. S. Yousif, A study into on-street parking: Effects on traffic congestion. Traffic Eng. Control
40, 424427 (1999).
43. S. Biswas, S. Chandra, I. Ghosh, Effects of on-street parking in urban context: A critical
review. Transp. Dev. Econ. 3, 10 (2017).
44. S. Wijayaratna, Impacts of on-street parking on road capacity, in Australasian
Transport Research Forum, Sydney, Australia, 30 September to 2 October 2015,
pp. 115.
45. Transportation Research Board, Highway Capacity Manual: HCM 2010 (Transportation
Research Board, ed. 5, 2010).
46. Cambridge Systematics Inc., Texas A&M Transportation Institute, University of
Washington, Dowling Associates, Street Smarts, H. Levinson, H. Rakha, Analytical
procedures for determining the impacts of reliability mitigation strategies(SHRP 2
Report S2-L03-RR-1, Transportation Research Board, 2012).
47. D. Schrank, B. Eisele, T. Lomax, J. Bak, 2015 Urban Mobility Scorecard (Texas A&M
Transportation Institute and Inrix Inc., 2015); https://mobility.tamu.edu/ums/.
48. Y.-C. Chiu, J. Bottom, M. Mahut, A. Paz, R. Balakrishna, T. Waller, J. Hicks, Dynamic Traffic
Assignment: A Primer (Transportation Research Circular E-C153, 2011).
49. C. Said, Uber, Lyft cars clog SF streets, study says,San Francisco Chronicle,
16 October 2018; https://www.sfchronicle.com/business/article/Uber-Lyft-cars-clog-
SF-streets-study-says-13309593.php.
50. F. Pettersson, L. W. Hiselius, T. Koglin, E-commerce and urban planningComparing
knowledge claims in research and planning practice. Urban Plan. Transp. Res. 6,121
(2018).
51. San Francisco Center for Economic Development, Tourism Statistics; http://sfced.org/
why-san-francisco/facts-figures/tourism/.
Acknowledgments: Data collection: A. Mislove, C. Wilson, and L. Chen; data quality
assurance and geographic information system: J. Brashear and A. Mucci; initiation:
D. Tischler; draft reviews: R. Souleyrette, L. Cagle, A. Erhardt, J. Walker, K. Konduri, and
anonymous reviewers. Funding: The San Francisco County Transportation Authority
funded this study. Author contributions: G.D.E.: design, direction, and final draft;
S.R.: analysis and first draft; D.C.: TNC data and SF-CHAMP; B.S.: visualization and software
quality assurance; M.C.: speed data and reliability; J.C.: initiation and direction; all
authors: weekly meetings, review methods, results, and interpretation. Competing
interests: The authors declare that they have no competing interests. Data and materials
availability: All data needed to evaluate the conclusions in the paper are present in
the paper and/or Supplementary Materials. Additional data and software may be
requested from the authors.
Submitted 23 May 2018
Accepted 1 April 2019
Published 8 May 2019
10.1126/sciadv.aau2670
Citation: G. D. Erhardt, S. Roy, D. Cooper, B. Sana, M. Chen, J. Castiglione, Do transportation
network companies decrease or increase congestion? Sci. Adv. 5, eaau2670 (2019).
SCIENCE ADVANCES |RESEARCH ARTICLE
Erhardt et al., Sci. Adv. 2019; 5: eaau2670 8 May 2019 11 of 11
on May 8, 2019http://advances.sciencemag.org/Downloaded from
Do transportation network companies decrease or increase congestion?
Gregory D. Erhardt, Sneha Roy, Drew Cooper, Bhargava Sana, Mei Chen and Joe Castiglione
DOI: 10.1126/sciadv.aau2670
(5), eaau2670.5Sci Adv
ARTICLE TOOLS http://advances.sciencemag.org/content/5/5/eaau2670
MATERIALS
SUPPLEMENTARY http://advances.sciencemag.org/content/suppl/2019/05/06/5.5.eaau2670.DC1
REFERENCES http://advances.sciencemag.org/content/5/5/eaau2670#BIBL
This article cites 16 articles, 1 of which you can access for free
PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions
Terms of ServiceUse of this article is subject to the
registered trademark of AAAS. is aScience Advances Association for the Advancement of Science. No claim to original U.S. Government Works. The title
York Avenue NW, Washington, DC 20005. 2017 © The Authors, some rights reserved; exclusive licensee American
(ISSN 2375-2548) is published by the American Association for the Advancement of Science, 1200 NewScience Advances
on May 8, 2019http://advances.sciencemag.org/Downloaded from
... The findings highlight the need for a critical examination of ride-sharing's impacts and the integration of TNCs into urban transport systems. Policymakers must address lifecycle costs, promote multimodal solutions, and balance the platforms' convenience with their environmental consequences to ensure genuinely sustainable urban mobility and consumption practices (Erhardt et al., 2019). ...
Article
Full-text available
Global economic growth is closely tied to energy consumption, with fossil fuels still dominating the global energy mix. The transportation sector, a major energy consumer, significantly impacts global primary energy demand and CO₂ emissions. This paper critically examines the Discourse surrounding technological innovations in transportation and their potential to reduce energy consumption. Using a critical Discourse analysis (CDA) framework, we explore the promotion of technologies such as intelligent automation, self-driving vehicles, and digitalization in transportation. The study highlights a techno-optimistic narrative that often overlooks broader structural challenges, unintended consequences, and the need for systemic changes, including policy innovation and behavior modification. Although technological advancements offer potential efficiency gains, their impact on energy consumption depends on implementation, governance, and socioeconomic context. This paper argues for a holistic approach, combining technological innovation with non-technological strategies to achieve sustainability in the transportation sector. Through case studies such as high-speed rail, ride-sharing, and automated vehicles, it examines how policies, user behavior, and structural barriers influence energy use. Findings stress the importance of adaptive governance and inclusive public engagement to foster sustainable, equitable mobility transitions in the evolving transport landscape.
... As such, researchers across the globe are now investigating as to how the RHS are changing the urban travel dynamics. Considering their disruptive nature and potential to alter the status quo, there has been a growing body of research, mostly in the global north, highlighting the unintended effects like induced travel (Erhardt et al. 2019), increase in trip length (Henao and Marshall 2019), decrease in transit ridership (Hall et al. 2018), increase in congestion levels (Agarwal et al. 2019), substitution of active modes (Young and Farber 2019), and increase in car ownership (Guo et al. 2020). Simultaneously, a notable degree of uncertainty persists regarding the differential impact of various factors that affects usage, or in many cases the non-usage of ride-hailing services. ...
Article
Full-text available
The study investigates the latent heterogeneity in travel behaviour among urban travellers, including ride-hailing service (RHS) users and non-users, by incorporating attitudes so as to reinforce conventional user-segmentation approaches. Simultaneously, prioritisation of ride-hailing specific attributes was carried out to assess how RHS will operate in a sustainable way. The study initially examines latent heterogeneity in travellers through a Latent Class Cluster Analysis (LCCA) model. Subsequently, it prioritises key RHS-specific attributes for each cluster using three established Multi Criteria Decision Making (MCDM) techniques. Three clusters were identified based on individuals’ attitudes and covariates (socio-demographics, travel habits, and built environment attributes). The largest cluster is the Tech-savvy ride-hailing-ready individuals (48%) with higher technological literacy, showing maximum acceptance towards ride-hailing. The second largest cluster comprises the Traditional active-mobility individuals (28%) who display the least proclivity towards RHS, probably due to their technological inhibition coupled with greater attachment to traditional travel alternatives. Lastly, the PV-loving multimodal individuals (24%) are primarily vehicle owners but prefer RHS for occasional trips. The final ranking obtained from the analysis has revealed that travel time, reliability, and flexibility are the motivators, while travel cost and waiting time are the deterrents, as perceived by the users, that influence RHS in the Indian context.
... Some of these studies addressed the proactive route guidance approach to avoid congestion [21], effects of congestion adaptive to travelers' choices on travel and accessibility [22], impact of public bicycle schemes on a congested multimodal transport network [23], and reducing congestion and travel costs in multimodal urban transport networks with limited parking and dynamic pricing [24]. Others investigated the association between mode choice scenarios and travel satisfaction [25], the correlation between the active mode of transportation chosen and the travel behavior exhibited by university students [26], the positive impact of using active transportation modes on public health, climate, and the environment [27], effects of transportation network companies on increasing or decreasing congestion [28], and impact of mixed modes on network efficiency and performance [29]. ...
... Many studies emphasized modeling scenarios and travel behavior, seeking to inform ride hailing regulation at the urban level by identifying challenges and opportunities for future policies. For example, some researchers promoted pooled ride hailing, analyzed efforts to reduce vehicle miles traveled, and highlighted connections between ride hailing and transit ridership (Clewlow & Mishra, 2017;Erhardt et al., 2019;Tirachini & Del Río, 2019;Young et al., 2020). More recently, scholars have observed a shift in the literature toward discussing the relationship between ride hailing practices, equity, and sustainability (McKane & Hess, 2023). ...
... There has been a vivid academic debate about the total cumulative impact of climate impact on ridehailing (summarized in Table 1). Several studies have attempted to measure ridehailing's impact on GHG emissions, and they overwhelming conclude that the industry increases emissions [43,105,106] by substituting trips that would otherwise have taken place in more sustainable forms of transportation (e.g., public transportation, cycling, and walking) [107], increasing VKT through having a driver come to your location [26], and adding to traffic congestion through induced demand [108]. Because of the rapid growth of the transportation-for-hire industry and the role it plays in our urban transportation market, some argue that vehicle and fleet standards ought to be tougher than those for personal-use vehicles. ...
Article
Full-text available
Given that most environmental regulation is performed on the international and national scale and that municipal powers are often limited, what are the necessary conditions for local government to take on leadership and policy innovation in this space? The transportation-for-hire industry provides an interesting case study as municipalities around the world recently updated their local regulatory frameworks because of the disruption caused by the advent of ridehailing. Over time, policy learning resulted in policy convergence of major components, including license requirements, vehicle standards, and rate standardization, across the industry. However, overlooked from these initial regulations, but now gaining traction is the regulation of greenhouse gas (GHG) emissions of ridehailing vehicles and the transportation-for-hire industry more generally. Because of how many transportation-for-hire vehicles are in use, activists argue that environmental fleet standards ought to be tougher than that for personal-use vehicles. This paper investigates the emerging regulatory frameworks for GHG emissions by examining four case studies of the first-mover cities (London, Amsterdam, Mexico City, and New York City) in this emerging policy space, untangling the complex web of multi-level regulatory governance to tease our policy learning outcomes. Through the innovative lens of examining multi-level governance for policy learning, we conclude that GHG emissions in the transportation industry at the local level only occurred after international and national frameworks were set, giving “permission” to local leaders to demonstrate innovation and leadership on how to achieve targets set by higher-order governments. Cities, assisted with policy-learning transfer through international not-for-profit organizations, acted via local government powers to fill in the gaps of international and national frameworks with policies implemented at higher orders of government in other places. These local environmental transportation-for-hire bylaws took a three-pronged form: (a) vehicle permit standards; (b) regulations of public space and infrastructure; and (c) public–private partnerships to assist in the transition to electric vehicles.
Article
Full-text available
Digitalisation is reshaping production and consumption practices across society. There is uncertainty around its net energy demand and related greenhouse gas emissions, owing to its complex and varied indirect impacts. This paper focuses on three indirect impacts of digitalisation: 1) efficiency; 2) rebound; and 3) substitution in the context of energy consumption associated with the use of digital innovations in daily life activities. Systemic conditions such as equitable access, trust, and control and agency interact with these domains of activity and determine the ultimate climate impacts of digitalisation in daily life. While digitalisation has the potential to be a game-changing tool in reducing energy consumption, in practice, this will require concerted efforts and policies to steer it in a desirable direction. We outline a research agenda that supports SHAPE (Social Sciences, Humanities and the Arts for People and the Economy) research on the environmental implications of digital engagement, interdisciplinary research bridging SHAPE and STEM (Science, Technology, Engineering, and Mathematics), further research on the indirect and systemic energy impacts of digitalisation, and the importance of factoring in digitalisation as a cross-cutting process within different activity domains. We conclude with policy recommendations focused on enhancing the positive climate impacts of digitalisation in daily life.
Preprint
Full-text available
As the worldwide ride-hailing (RH) industry grows, the integration of multi-modal transportation systems has become increasingly important to reduce traffic congestion and improve user mobility. Most RH studies used nominal scenarios with regular demand patterns. However, disruptions must be considered due to their negative impact on operational efficiency, such as longer travel times, higher costs, increased transfers, and more service delays. Therefore, this study investigates the resilience of RH fleet management strategies in multi-modal transportation systems against disruptions on train lines. Two fleet management strategies, centralized and decentralized, are introduced to evaluate their performance. A traffic simulation is conducted using a 900 km2 toy urban network that includes multiple transportation modes such as trains, metros, buses, and RHs. The results revealed that during the peak of disruption, the centralized strategy maintained a higher level of service but required a longer recovery time. In contrast, the decentralized strategy showed a lower service level but recovered more quickly. The centralized strategy showed higher performance during regular scenarios, while the decentralized strategy showed better resilience during disruptions. Comparing the two strategies provides critical insights for developing intermediate RH strategies to effectively manage disruptions and enhance the seamlessness of multi-modal transportation systems.
Thesis
Full-text available
The major transit systems operating in San Francisco are San Francisco Municipal (MUNI), Bay Area Rapid Transit (BART), and Caltrain. The system of interest for this paper is MUNI, in particular the bus and light rail systems. During the past decade transit ridership in the area has experienced diverging growth, with bus ridership declining while rail ridership is growing significantly (Erhardt et al. 2017). Our data show that between 2009 and 2016, MUNI rail ridership increases from 146,000 to 171,400, while MUNI bus ridership decreases from 520,000 to 450,000. Direct ridership models (DRMs) are used to determine what factors are influencing MUNI light rail and bus ridership. The DRMs predict ridership fairly well, within 10% of the observed change. However, the assumption of no multi-collinearity is voided. Variables, such as employment and housing density, are found to be collinear. Fixed-effects panel models are used to combat the multi-collinearity issue. Fixed-effects panel models assign an intercept to every stop, so that any spatial correlation is removed. A transportation network company, Uber and Lyft, variable is introduced (TNC) to the panel models, to quantify the effect they have on MUNI bus and light rail ridership. The addition of a TNC variable and elimination of multi-collinearity helps the panel models predict ridership better than the daily and time-of-day DRMs, both within 5% of the observed change. TNCs are found to complement MUNI light rail and compete with MUNI buses. TNCs contributed to a 7% growth in light rail ridership and a 10% decline in bus ridership. These findings suggest that the relationship TNCs have with transit is complex and that the modes cannot be lumped together.
Article
Full-text available
In planning for a future sustainable transport system, it is important to consider the likely effects of e-commerce. This paper analyses the understanding of impacts of e-commerce on urban planning as expressed in planning practice in a number of Swedish municipalities. Knowledge claims concerning the potential effects of e-commerce are studied through an analysis of planning documents and supplementing interviews. These knowledge claims in planning practice are contrasted with the findings presented and discussed in the research literature. The results indicate that the current understanding of e-commerce in Swedish planning practice is limited and that it is primarily viewed as an opportunity to address transport system issues by reducing passenger transport. However, the overview of the literature shows that e-commerce has so far only had limited effects on prevailing mobility habits and that freight transport has increased in parallel with increased e-commerce. Furthermore, e-commerce challenges city centres as commercial areas. Thus, from a planning perspective e-commerce should not be regarded as an easy solution for an unsustainable transport sector. To harness its potential benefits, there is a need to develop pro-active planning strategies that address the potential impacts of e-commerce.
Article
Full-text available
How to make traffic worse everywhere One policy aimed at improving traffic flows in large cities requires vehicles to carry two or three passengers, usually in lanes or roads set aside for high-occupancy vehicle (HOV) travel. Whether this actually increases speeds depends on how heavily the HOV roads are used and the availability of alternate routes. Hanna et al. took advantage of an abrupt policy change—the elimination of HOV rules—in Jakarta to collect the travel times from Google Maps for HOV and alternate routes before and after the change (see the Perspective by Anderson). They observed a serious worsening of traffic throughout the city. Science , this issue p. 89 ; see also p. 36
Article
Full-text available
This paper examines the changes that might result from the large-scale uptake of a shared and self-driving fleet of vehicles in a mid-sized European city. The work explores two different self-driving vehicle concepts – a ridesharing system (Shared Taxi), which emulates a taxi-like system where customers accept small detours from their original direct path and share part of their ride with others and a dynamic bus-like service with minibuses (Taxi-Bus), where customers pre-book their service at least 30 minutes in advance (permanent bookings for regular trips should represent most requests) and walk short distances to a designated stop. Under the premise that the “upgraded” system should as much as possible deliver the same trips as today in terms of origin, destination and timing, and that it should also replace all car and bus trips, it looks at impacts on car fleet size, volume of travel and parking requirements. Mobility output and CO2 emissions are also detailed in two different time scales (24 hr. average and peak-hour only). The obtained results suggest that a full implementation scenario where the existing metro service is kept and private car, bus and taxi mobility would be replaced by shared modes would significantly reduce travelled vehicle.kilometres and CO2 emissions.
Article
Full-text available
This study gathered empirical evidence on adoption patterns of self-driving vehicles, people's likely use of them, and how that might influence amount of travel, mode choice, auto ownership, and other travel behaviour decisions. Because self-driving vehicles are not yet on the market, a car technology acceptance model was applied to understand adoption and use. Researchers implemented a two-stage data collection effort. An online survey was conducted with 556 residents of metropolitan Austin to determine intent to use. Based on results, four “intent to use” categories were determined: (1) extremely unlikely = Rejecters (18%); (2) somewhat unlikely = Conservatives (32%); (3) somewhat likely = Pragmatists (36%); (4) extremely likely = Enthusiasts (14%). Individuals with a higher level of intent to use have any physical conditions that prohibit them from driving; use technology – smartphone, text messaging, Facebook, transportation apps – and are not concerned with data privacy about using online technology; think using self-driving vehicles would be fun, decrease accident risk, and easy to become skilful at using; and believe people whose opinions are valued would like using them. Among those who indicated intent to use, qualitative interviews were conducted to ascertain the impact on their travel behaviour. Most respondents would rather own self-driving vehicles (59%) than just use one (41%), like a Car2Go or Uber taxi. Additionally, respondents reported that using one would have no change on where people would choose to live in Austin (80%), no change to their annual VMT (66%), and no change to the number of vehicles owned (61%).
Article
Recent technological advances have paved the way for new mobility alternatives within established transportation networks, including on-demand ride hailing/sharing (e.g., Uber, Lyft) and citywide bike sharing. Common across these innovative modes is a lack of direct ownership by the user; in each of these mobility offerings, a resource not owned by the end users’ is shared for fulfilling travel needs. This concept has flourished and is being hailed as a potential option for autonomous vehicle operation moving forward. However, substantial investigation into how new shared modes affect travel behaviors and integrate into existing transportation networks is lacking. This paper explores whether the growth in the adoption and usage of these modes is unbounded, or if there is a limit to their uptake. Recent trends and shifts in Uber demand usage from New York City were investigated to explore the hypothesis. Using publicly available data about Uber trips, temporal trends in the weekly demand for Uber were explored in the borough of Manhattan. A panel-based random effects model accounting for both heteroscedasticity and autocorrelation effects was estimated wherein weekly demand was expressed as a function of a variety of demographic, land use, and environmental factors. It was observed that demand appeared to initially increase after the introduction of Uber, but seemed to have stagnated and waned over time in heavily residential portions of the island, contradicting the observed macroscopic unbounded growth. The implications extend beyond already existing fully shared systems and also affect the planning of future mobility offerings.
Article
This paper describes the application of a state-of-the-art, tour-based travel demand microsimulation model to estimate impact on mobility and accessibility on different populations to support development of a countywide transportation plan. Equity analyses based on traditional travel demand forecast models are compromised by aggregation biases and data availability limitations. Use of the disaggregate (individual person-level) San Francisco, California, tour-based microsimulation model made it possible to estimate benefits to and impact on different communities of concern on the basis of individual characteristics, such as gender, income, automobile availability, and household structure. The concepts and policy context of equity analysis in transportation are presented. Identifying communities of concern and relevant measures of transportation system performance are outlined. The San Francisco model structure is described, and the results of the equity analysis are presented.
Article
The rapid expansion of transportation network companies (TNCs), such as Uber and Lyft, has led policy makers across the country to consider legislation to legalize and regulate TNC operations. In 2012, TNCs introduced a new travel option that used digital technology to provide an on-demand and highly automated private ride service, which has received support from consumers and investors. The emergence of TNCs has generated uncertainty about the legal status of TNC services, criticism from the taxicab industry, and concerns about public safety. In almost every U.S. state, policy makers have considered TNC legislation to address these issues, but there is no comprehensive source of information on the content of these policies. In this study, researchers systematically compiled a database of state TNC legislation and evaluated a set of policy issues addressed in the legislation. Thirty-four states and Washington, D.C., enacted legislation to authorize TNC operations through May 2016. These laws addressed policy areas including permits and fees, insurance and financial responsibility, driver and vehicle requirements, operational requirements, passenger protections, data reporting, and regulatory and rulemaking authority. This database can help policy makers navigate the evolving policy considerations presented by the rising popularity—and accompanying controversy—of TNCs. Key policy questions that emerged from this review of state TNC legislation include whether to regulate TNCs; if so, at what level of government; how to harmonize TNC policies with existing taxi and transportation policies; and how to address public safety without suppressing market competition.