Travel Behaviour and Society 32 (2023) 100570

Contents lists available at ScienceDirect

Travel Behaviour and Society

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

Differential impacts of autonomous and connected-autonomous vehicles on

household residential location
Md Mehedi Hasnat a, Eleni Bardaka a, *, M. Shoaib Samandar b
a
Department of Civil, Construction, and Environmental Engineering, North Carolina State University, Fitts-Woolard Hall, 915 Partners Way, Raleigh, NC 27606, USA
b
Institute of Transportation Research and Education, North Carolina State University, Research IV, 909 Capability Dr, Raleigh, NC 27606, USA

A R T I C L E I N F O A B S T R A C T

Keywords: High market penetration of autonomous vehicles (AVs) and connected-autonomous vehicles (CAVs) is expected
Residential location choice to impact transportation network performance, which is an important determinant of residential location de­
Autonomous vehicles cisions, especially for households who commute to work by personal vehicle. This study examines and compares
Connected-autonomous vehicles
the impacts of privately owned AVs and CAVs on the location and commute characteristics as well as the spatial
Electric vehicles
Cluster analysis
distribution of households within the Triangle Region in North Carolina. A Mixed Multinomial Logit model is
Suburbanization developed using recent household survey data to capture household preferences. In addition, the region’s travel
demand model, the Triangle Regional Model, is used to predict the network-level impacts of AV and CAV
adoption, and cluster analysis is conducted to explore how network performance changes vary with trans­
portation demand and supply zone characteristics at a local and regional level. Residential location patterns are
predicted for a number of AV and CAV scenarios for the year 2045 using the outputs of the econometric analysis
and the Triangle Regional Model. We find that extensive adoption of private CAVs improves network conditions
and encourages households to live farther from work, leading up to a 5.6% increase in suburban and rural
households that commute to work by personal vehicles. A high market share of AVs is associated with deteri­
orated transportation network performance and up to a 2.8% increase in urban households. Results vary by
market penetration rate of each technology, mix of AVs, CAVs, and human driven vehicles in the traffic stream,
and fuel type (conventional-fuel versus electric vehicles).

1. Introduction
The adoption of autonomous vehicle (AV) and connected-
autonomous vehicle (CAV) technologies in the transportation sector
will substantially impact the ways people and goods are transferred from
one place to another (Berglund et al., 2020). Although researchers and
policy makers are eagerly waiting to harness the benefits of these
technologies, they are also concerned about potential unforeseen con­
sequences that might arise from wide adoption (Bardaka et al., 2021;
Dubljevic et al., 2021). Numerous studies have reported expected im­
pacts in travel demand (Bardaka and Hasnat, 2021; Truong et al., 2017;
Wadud et al., 2016), highway capacity (Adebisi et al., 2020; Tientrakool
et al., 2011), and safety (Anderson et al., 2014) due to AV and CAV
adoption. In addition to these direct impacts, long-term indirect changes
in land use and development are likely to follow.
Widespread AV and CAV adoption is expected to change trans­
portation network performance, which has been shown to significantly
affect households’ residential location decisions (Guo and Peeta, 2020;
Bruns and Matthes, 2019; Bhat and Guo, 2004). CAVs are anticipated to
substantially improve freeway capacity because of their ability to
communicate with other CAVs and the infrastructure and operate safely
with smaller headways compared to human driven vehicles (Shladover
et al., 2012; Tientrakool et al., 2011). On the other hand, AV de­
ployments are more likely to focus on traffic safety and incorporate
decision algorithms that are more conservative compared to human
driving, leading to capacity reductions and degraded network perfor­
mance (Hasnat et al., 2021; Adebisi et al., 2020; Bierstedt et al., 2014). If
the direct impacts of widespread adoption of AVs and CAVs on capacity
and overall transportation network performance are heterogeneous,
then their long-term effects on metropolitan areas may also be different.
In this study, we simulate the changes in network demand due to
privately owned AVs and CAVs in a US metropolitan area and predict
households’ location decisions to better understand the differential
impacts of AV and CAV adoption on the urban form. Although it is

* Corresponding author.
E-mail addresses: mhasnat@ncsu.edu (M.M. Hasnat), ebardak@ncsu.edu (E. Bardaka), smsamand@ncsu.edu (M.S. Samandar).

https://doi.org/10.1016/j.tbs.2023.02.007
Received 10 March 2021; Received in revised form 7 February 2023; Accepted 9 February 2023
Available online 2 March 2023
2214-367X/© 2023 Hong Kong Society for Transportation Studies. Published by Elsevier Ltd. All rights reserved.
M.M. Hasnat et al.
uncertain whether AVs and CAVs in the US will be primarily privately
owned (Zhang et al., 2018) or operate as part of shared fleets (Shen
et al., 2017), this research concentrates on a potential future where
private vehicles dominate the market of AVs and CAVs. We also explore
whether the adoption of electric AVs and CAVs will result in substan­
tially different outcomes compared to conventional-fuel AVs and CAVs.
The study focuses on the Raleigh-Durham-Chapel Hill combined
metropolitan statistical (CSA) area, also known as the Triangle Region,
in North Carolina. AV and CAV scenarios for the year 2045 are modeled
using the regional macroscopic travel demand model for the Triangle
Region. Cluster analysis is also used to provide insights on the spatial
distribution of the changes in network performance across the region.
Household residential location in 2045 is predicted and analyzed using
data from the Triangle household survey and the travel demand model
simulation results.
Researchers have envisioned personal self-driving vehicles inducing
suburbanization and dispersed land development (Litman, 2020; Meyer
et al., 2017; Zakharenko, 2016; Anderson et al., 2014), and shared ve­
hicles offering fast and inexpensive service and attracting more people
to live in high-density areas (Durand et al., 2018). However, quantita­
tive research in this subject has been rather limited (Bansal and Kock­
elman, 2018; Carrese et al., 2019; Gelauff et al., 2019; Kim et al., 2020;
Krueger et al., 2019; Moore et al., 2020; Zhang and Guhathakurta,
2018). Our study contributes to this limited literature and advances the
understanding of the varying effects of AVs and CAVs, a topic that
previous research has not focused on. In addition, we deviate from
studies that have assumed excessive changes in travel behavior and
roadway capacity because such changes are not supported by the latest
related research (Singleton, 2019; Krueger et al., 2019; Adebisi et al.,
2020). Our study introduces capacity adjustments based on microscopic
simulation analysis that accounts for the interactions of AVs and CAVs
with traditional vehicles. Furthermore, our study takes a conservative
approach and does not assume a decrease in the value of travel time for
AV and CAV commute trips. Value of travel time reductions have been
based on the hypothesis that passengers of self-driving vehicles will be
able to undertake productive activities while driving, such as working,
reading, or sleeping. Research has shown that such productivity gains,
which can be currently realized during travel on airplanes or fixed
guideway transit systems, are possible for long-distance AV and CAV
trips on limited-access facilities but may not be relevant for commute
trips to work, which are typically shorter and include regular speed and
direction changes (Singleton, 2019). Recently, Krueger et al. (2019)
found no changes in the value of travel time due to self-driving vehicles
based on a stated residential location preference survey of commuters in
Sydney, Australia.
This research can assist transportation agencies and metropolitan
area planning organizations to gain a better understanding of the effects
of AVs and CAVs on transportation system performance and distribution
of households within a metropolitan area in the future. Although the
direction of the changes may be already understood or anticipated,
reliable predictions on their magnitude remain scarce. Furthermore,
transportation engineers and urban planners can use the outcomes of
this study to inform decisions and plans related to future land devel­
opment and infrastructure investments.
The paper is organized into nine sections. The second section dis­
cusses the literature on residential location choice as well as previous
research on the impact of self-driving vehicles on residential location
decisions. In the third section, we describe the study region. The fourth
section includes the study methodology and discusses how we estimate
and predict residential location choice under a number of AV and CAV
scenarios. Section 5 focuses on the analysis of household residential
location preferences for our study region, and Section 6 discusses the
network-level impacts of AV and CAV adoption. The predicted changes
in household residential location for 2045 are presented in Section 7.
Section 8 discusses the sensitivity our results to a number of inputs. The
final section concludes the study.
2
Travel Behaviour and Society 32 (2023) 100570
2. Literature review
2.1. Modeling residential location choice
The determinants of households’ choice of residential location have
been studied extensively (Eliasson, 2010; Gehrke et al., 2019; Kroesen,
2019). The majority of past studies have used household survey data to
develop discrete choice models where an individual or a household
seeks to maximize their utility with respect to residential location
(Akbari et al., 2020; Pinjari et al., 2011). The granularity of the unit of
analysis, such as dwelling unit, traffic analysis zone (TAZ), and census
geography, largely depends on the research question and the available
dataset. Explanatory variables previously used in residential location
choice models include built environment characteristics of home and
work locations (walkability, street density, access to different destina­
tions), residential unit features (size, number of bedrooms, price),
household characteristics (owner or renter, size, age of the members,
income, number of workers, number of children, vehicle ownership),
individual-level information (age, education, income, commute time,
commute cost), and neighborhood attributes (median household in­
come, racial composition, proximity to downtown, parks, and schools)
(Akbari et al., 2020; Pinjari et al., 2011).
Results from previous research suggest that commute time has an
important but heterogeneous effect on households’ residential location
decision (Guo and Bhat, 2007; Zhou and Kockelman, 2009). For
example, Guo and Bhat (2007) found that single-person households and
households with female workers have a stronger preference to reside
closer to their workplaces. Studies have also reported that young and
single-person households prefer central, high-density areas, whereas
families with children and high-income households are drawn towards
low-density suburban areas (Pinjari et al., 2011). After reviewing a
number of studies, Schirmer et al. (2014) concluded that the probability
of relocating closer to urban city centers or moving to suburban areas
changes over the households’ life span and is influenced by different
events like marriage, children, or job status.
In this study, we develop a Mixed multinomial logit (MNL) model to
explain and predict residential location choice for households in the
Triangle Region. We follow the directions of previous research to select
explanatory variables for our analysis. The methodological approach is
presented in Section 4.1, and the data description and analysis are dis­
cussed in Section 5.
2.2. Autonomous versus connected-autonomous vehicles
Autonomous vehicles (AVs) are designed to ensure a more efficient
and safe travel experience. According to the latest definition by the
Society of Automotive Engineers (SAE, 2021), there are 5 levels of
vehicle automation. Level 1 and Level 2 vehicles include minimal driver
assistance features, such as blind spot detection, automatic emergency
braking, and adaptive cruise control. Level 3 vehicles act as traffic jam
chauffeurs and can self-maneuver in a stop-and-go traffic situation. Up
to Level 3 of automation, the driver must constantly engage in the
driving process. Level 4 and Level 5 vehicles include the most advanced
levels of automation and are considered “self-driving”. Level 4 AVs will
be able to operate without human intervention under limited traffic and
roadway conditions, such as inside residential and academic compounds
or in low-speed urban streets. Level 5 vehicles are envisioned to be able
to operate safely and efficiently under any traffic and roadway envi­
ronment without human intervention (SAE, 2021). On the other hand,
Connected vehicles (CVs) are described as vehicles that are capable to
communicate with each other using vehicle-to-vehicle (V2V) technolo­
gies and with the roadway infrastructure using vehicle-to-infrastructure
(V2I) technologies. This communication increases the operating effi­
ciency and enhances safety and mobility. Communication between ve­
hicles in the traffic stream has been shown to increase situation
awareness and help mitigate crashes (Uhlemann, 2015). Connected-
M.M. Hasnat et al.
autonomous vehicles (CAVs) include both automation and connectivity
technologies. Our literature review as well as our study’s scope focuses
on AVs and CAVs with high levels of automation (Levels 4 and 5), also
referred to as self-driving vehicles.
Past studies have followed theoretical approaches or simulation
techniques to investigate the impact of AVs and CAVs on roadway ca­
pacity (Chang and Lai, 1997; Tientrakool et al., 2011; Bierstedt et al.,
2014; Adebisi et al., 2020; Hajbabaie et al., 2021). These studies
modeled the single or combined effect of different factors that may
impact the capacity of a freeway or uninterrupted highway segment
including the time gap, platooning, car following behavior, and desired
vehicle speed. Many of the initial studies had assumed “aggressive”
driving behavior (such as large reductions in the time gap between ve­
hicles) for AVs and had therefore reported moderate to high capacity
improvements (Chang and Lai, 1997; Tientrakool et al., 2011). How­
ever, recently it has become more clear that the original equipment
manufacturers (OEMs) of AVs are placing significant emphasis on traffic
safety and developing vehicles that operate in a more conservative
manner compared to human driving, considering multiple factors,
including the maturity and reliability of the associated technologies,
liability aversion, and eminence of the industry (Adebisi et al., 2020;
Hasnat et al., 2021). Previous research has concluded that this is ex­
pected to negatively impact highway capacity (Adebisi et al., 2020;
Bierstedt et al., 2014). On the contrary, because CAVs will have the
ability to communicate with other CAVs and the infrastructure and
constantly track the environment, they are expected to maintain shorter
time gaps between alike vehicle types and form platoons with short
headways (Shladover et al., 2012; Tientrakool et al., 2011). Prior studies
have reported up to a 270% increase in capacity due to different market
penetration levels of CAVs (Shladover et al., 2012; Tientrakool et al.,
2011). These reported gains in capacity are found to be directly pro­
portional to the share of CAVs in the traffic stream.
2.3. Impacts of emerging vehicle technologies on residential location
choice
A few studies have simulated the changes in travel time and acces­
sibility under private and shared self-driving vehicle scenarios and have
used the results to predict the impacts on households’ residential loca­
tion choices (Zhang and Guhathakurta, 2018; Gelauff et al., 2019).
Focusing on shared AVs, Zhang and Guhathakurta (2018) integrated a
residential location choice model with an agent-based simulation model
to identify household location changes for the Atlanta metropolitan
area. A Mixed MNL residential location choice model was first devel­
oped to capture the current location choice behavior of households in
Atlanta using data from a regional travel survey. The agent-based
simulation assumed that all travel demand would be served by shared
AVs that could only carry a single person at a time. Other assumptions
included a decrease in parking spaces needed, lower cost of travel
compared to traditional vehicles, and multitasking opportunities during
travel which were reflected by a 25% to 100% reduction in the perceived
cost of in-vehicle travel time. Based on these assumptions, a 48.4% to
72.4% decrease in commuting travel cost was estimated, which led to
most households choosing to reside further away from their workplace
and the central business district (CBD) (Zhang and Guhathakurta, 2018).
Gelauff et al. (2019) used a spatial general equilibrium model to simu­
late potential population migration between city and rural areas in the
Netherlands under privately owned and shared AV scenarios for the year
2050. The study assumed that private, fully autonomous vehicles would
provide more productive use of in-vehicle travel time compared to
traditional vehicles and therefore, reduced the perceived in-vehicle
travel time cost by 20%. Shared AVs were assumed to provide demand
response ridesharing, be more efficient than traditional public transport,
but not as efficient as personal vehicles due to detouring for other pas­
sengers; these assumptions were reflected in a 20% increase in the
perceived cost of in-vehicle travel time compared to traditional vehicles.
3
Travel Behaviour and Society 32 (2023) 100570
In addition, Gelauff et al. (2019) assumed zero changes in transportation
network performance due to automation, expecting that improvements
in capacity would be counterbalanced by higher travel demand, leaving
travel time unchanged. The results suggested that adoption of private
AVs would increase population density in non-urban areas by up to 1%,
whereas adoption of shared AVs would increase population density in
urban areas by up to 3%.
Two prior studies used regional or national-level transportation land
use models to simulate the impact of AVs on urban land-use patterns
(Meyer et al., 2017; Cordera et al., 2021). Cordera et al. (2021) modeled
the population distribution for the Bay of Santander (Cantabria, Spain)
under different private and shared AV scenarios. The model used in
Cordera et al. (2021) allocates population on different zones based on
their accessibility measures where zones with higher accessibility at­
tracts more population. For moderate market share of AVs Cordera et al.
(2021) assumed 40% increase in interurban roadway capacity and no
additional trips from induced demand. This improved the accessibility
for the areas outside the city center and increased the probabilities of
population influx in the area surrounding the urban center. For high
market share of AVs Cordera et al. (2021) incorporated 80% increased
capacity for urban roadway which increased the accessibility for the
urban center resulting in an increase of total population. Meyer et al.
(2017) introduced similar assumptions to the Swiss national transport
model and found that increased accessibility derived from improved
roadway capacity due to AVs would favor urban sprawl in the future. In
addition, a number of studies developed analytical frameworks to model
the possible shifts in population and their impact on urban or suburban
sprawl (Liu et al., 2021; Larson and Zhao, 2020; Zakharenko, 2016). Liu
et al. (2021) proposed an analytical framework to find urban land-use
patterns by deriving two equilibrium: urban equilibrium defined by
residential location and consumption choice, and transportation equi­
librium for suburban residents which consists of transportation mode
choice (including AVs) and departure time choice. Liu et al. (2021)
investigated the impact of different levels of automation on land uses
and concluded that with an increase in automation level, the land rent in
the urban core will reduce resulting in an expansion of the suburban
areas (Liu et al., 2021). Larson and Zhao (2020) and Zakharenko (2016)
used a hypothetical monochromatic city to model the impact of AVs on
urban and suburban sprawl. Larson and Zhao (2020) investigated three
distinct scenarios: use of personal AVs for commuting trips, restricting
parking inside the city so that owner has to send the AV home, and
shared AV services. For all the cases, AVs were found to increase the
accessibility to the suburban areas making them more attractive for
households which ultimately led to increased population outside the city
(Larson and Zhao, 2020). Zakharenko (2016) showed that with the
decrease in AV ownership costs, commuters will move further away
from their work location.
Researchers have recently conducted stated preference surveys
related to residential location decisions and self-driving vehicle sce­
narios to better understand whether behavioral changes should be ex­
pected in the future (Bansal and Kockelman, 2018; Carrese et al., 2019;
Kim et al., 2020; Krueger et al., 2019; Moore et al., 2020). Among other
hypotheses, these studies examined whether individuals would consider
moving their residence to a location that would require a longer
commute if they had access to a personal vehicle with self-driving ca­
pabilities, all else constant. It was hypothesized that people would be
willing to commute longer if they felt that they could spend their
commute time more productively; in that case, their perceived cost of in-
vehicle travel time would be lower than that of traditional vehicles. Kim
et al. (2020) found that only 11.0% of Atlanta residents would consider
moving farther from work if they owned fully autonomous vehicles.
Similarly, Bansal and Kockelman (2018), who conducted a survey in
Austin, Texas, documented that 11.1% of the survey respondents would
consider moving farther from the city center assuming travelling in AV
will be much easier than conventional human-driven vehicles. The study
also found that 7.4% of the respondents might choose to live closer to
M.M. Hasnat et al.
city centers to exploit the lower cost of shared AVs (Bansal and Kock­
elman, 2018). Moore et al. (2020) found that a larger percentage of
respondents (25.8%) would be willing to move and increase their
commute time by more than 10 min if they owned self-driving vehicles.
The sample of that study was, however, skewed towards educated, full-
time employed, non-Hispanic White individuals. Following a different
approach, Krueger et al. (2019) asked survey respondents to choose
between combinations of residential locations and commute mode op­
tions, and estimated the value of time for traditional and fully autono­
mous vehicles. Their results suggest no statistically significant
differences between the value of travel time for fully autonomous and
traditional vehicles. Overall, stated preference surveys have not pro­
vided strong evidence of behavioral changes in residential location
choice due to self-driving vehicle adoption for the majority of the pop­
ulation. Based upon these recent results as well as the analysis conducted
by Singleton (2019), no changes are considered in the value of travel
time in the AV and CAV scenarios for this study. Specifically, we assume
that future changes in households’ residential location are mainly an
outcome of changes in commute travel time and accessibility due to
widespread AV and CAV adoption, and not an outcome of more pro­
ductive use of the time spent in a personal AV or CAV.
Given the recent findings of the literature on the varying effects of
AVs and CAVs on highway capacity, it is important to account for the
type of vehicle technology while investigating the future changes in
land-use patterns. As evidenced from this literature review, previous
studies on the subject of land-use impacts have solely focused on self-
driving vehicles assuming that they will improve or not effectively
impact network performance. To the authors’ best knowledge, so far,
none of the analytical, empirical, or simulation studies on emerging
vehicle technologies have investigated the differential impacts of AVs
and CAVs on land use.
3. Study setting
The Raleigh-Durham-Chapel Hill CSA, also known as the Triangle
Region, is the second largest CSA in North Carolina. It covers 11 counties
Fig. 1. Triangle Regional Model region in 2013 and 2045.
4
Travel Behaviour and Society 32 (2023) 100570
with a total area of 5510 square miles and a total population of
2,238,315. The three main population centers are the cities of Raleigh,
Durham, and Chapel Hill, which are approximately a 20 to 30-min drive
apart. Growing numbers of high-skill job opportunities and multiple
educational institutions have contributed to the region’s high popula­
tion growth. Currently, the median household income of the Triangle
Region is 25% higher than the NC median household income and 10%
higher than the US median household income (US Census Bureau,
2019).
The Triangle Regional Model (TRM) is the Triangle Region’s travel
demand model. The TRM covers an area of 3380 square miles, that in­
cludes Wake, Durham, and Orange Counties, and parts of Chatham,
Franklin, Granville, Harnett, Johnston, Lee, and Person Counties of the
Raleigh-Durham-Chapel Hill CSA. The TRM region constitutes the study
area for this research and is composed of 2857 traffic analysis zones
(TAZs) (Fig. 1). TRM scenarios representing current conditions are based
on 2013 socioeconomic data for the region, while future conditions (up
to year 2045) are modeled using socioeconomic predictions from the
Triangle CommunityViz 2.0. The Triangle CommunityViz 2.0 is a GIS-
based planning and decision support tool that combines statistical and
spatial analysis with inputs from stakeholders to predict the magnitude
and location of future developments in the Triangle Region (Commu­
nityViz, 2018; Stantec, 2015). At first, CommunityViz identifies the
development constraints for individual parcels. Then the parcels are
assigned a development status (developed, committed or asserted, un­
developed, underdeveloped, and redeveloped). Regions are classified as
developed if there are no plans for future development. New growth is
assigned to the committed or asserted regions based on under-
construction buildings and anchor institution plans (CommunityViz,
2018). Undeveloped, underdeveloped, and redeveloped regions undergo
development according to the CommuntiyViz algorithm. CommunityViz
combines current growth patterns and inputs from stakeholders to
perform statistical and geospatial analysis to estimate the future socio­
economic characteristics (Stantec, 2015). The model uses land suit­
ability analysis (measure of how “attractive” a parcel is for new growth)
to assign new growth to parcels. This is based on a regression analysis
M.M. Hasnat et al.
that utilizes 14 different suitability factors including proximity to
transportation investments and major activity centers, availability of
sewer services, location within local government planned growth, and
other criteria which are associated with suitability for future develop­
ment (CommunityViz, 2018). Some of the suitability factors are
weighted based on focus group meetings and workshops with stake­
holders (business owners, developers, local government officials, and
others). Features of the suitability measures are layered and modeled for
grid cells which are combined to parcels in the final output (Stantec,
2015). The software categorizes parcels into 37 place (land-use) types
within five main development patterns: rural, suburban, city and town,
industrial, and special place types (airport, university campus, health
care campus, civic and institutional) (CommunityViz, 2018). Compared
to other regional land use forecasting tools, CommunityViz provides
dynamic and annual estimates like the California Urban Futures (CUF)
and UrbanSim, and models residential and employment growth like the
Disaggregated Residential Allocation Model (DRAM) and Employment
Allocation Model (EMPAL) (Waddell, 2002). Additional information on
the modeling assumptions and parameters used in the CommunityViz
model can be found in CommunityViz (2018) and Stantec (2015).
It is estimated that between 2013 and 2045, the Triangle Region will
experience a 76% increase in population and a 60.4% increase in
employment (CommunityViz, 2018). The region is classified into four
area types (CBD, urban, suburban, rural) based on employment and
land-use density. Substantial growth is anticipated for the CBD, urban,
and suburban areas in the next decades. Specifically, between 2013 and
2045, the CBD, urban, and suburban land areas are expected to increase
by 226%, 65%, and 55%, respectively (Fig. 1). Additional information
on the region’s socioeconomic characteristics is provided in Section 5.2.
4. Methodology
The methodological framework of our study is presented in Fig. 2.
We begin by modeling the current residential location choice behavior
of households residing in the Triangle Region using household survey
data as well as the latest information on the socioeconomic and travel
characteristics of the region. To predict the future location choice of
households in an AV and CAV era, we start by developing scenarios
related to the adoption and use of these technologies in the future. We
Fig. 2. Methodological framework.
5
Travel Behaviour and Society 32 (2023) 100570
assume that by 2045, AVs and CAVs could be owned by the high-income
and in some cases, the medium-income households of our study region.
We also assume that AVs and CAVs will improve access to opportunities
and destinations for population groups with limited driving abilities,
such as young and senior household members, and therefore induce
more trips from these groups. Additional vehicle miles may also be
added to the network due to empty AVs and CAVs searching for parking
or returning home to serve other household members. We hypothesize
that AVs and CAVs will have varying impacts on land uses because the
interaction among AVs, CAVs, and traditional human-driven vehicles is
expected to impact highway capacity in different ways depending on the
market penetration rate and mix of these vehicles in the traffic stream.
We modify the region’s travel demand model (Triangle Regional Model)
to estimate commuting and accessibility measures for 2045 based on the
aforementioned assumptions. These measures, along with socioeco­
nomic characteristics for 2045 and vehicle operating cost based on fuel
type (electric or conventional AVs and CAVs), serve as inputs to the
residential location choice model to predict the households’ future res­
idential location given different AV and CAV adoption rates. Cluster
analysis is also conducted to help us gain a better understanding of the
spatial variation of network performance changes in the Triangle Region
due to AVs and CAVs and their relationship with the built environment
characteristics of each zone. The study scenarios and assumptions are
summarized in Table 1. Additional information, theoretical background,
and the literature that informed these scenarios are provided in the
following sections.
4.1. Modeling households’ residential location choice behavior
Household preferences are modeled on the basis of random utility
theory (McFadden, 1978). A location (in our case, a TAZ) is chosen by a
household from a set of mutually exclusive alternatives if it provides the
highest utility compared to all other alternatives. We use a Mixed MNL
model that accounts for correlation across alternatives and unobserved
heterogeneity by allowing coefficients to vary across individuals
(McFadden and Train, 2000). Several past studies on residential location
choice have applied the Mixed MNL model to capture household or in­
dividual preferences (Zhang and Guhathakurta, 2018; Krueger et al.,
2019).
M.M. Hasnat et al.
Table 1
Scenarios and assumptions for AVs and CAVs in 2045.
Predicting transportation network performance
(Triangle Regional Model)
Market AV or CAV owners Induced demand
Penetration
(A1) 30% AV High-income households 30% increase in
(A2) 30% CAV with personal vehicles home-based other,
(A3) 15% AV, own AVs and CAVs non-home non-work,
15% CAV and home-based
(B1) 75% AV High-income households with shopping trips for
(B2) 75% CAV personal vehicles and medium- household members
(B3) 37.55% AV, income households with vehicles with age 12–17 and
37.5% CAV ⩾ workers own AVs or CAVs age 65 and over
We analyze data from 2356 households in the Triangle Region who
participated in a household survey in 2016, assuming that each house­
hold can choose to reside in one of the TAZs in our study area. The utility
that a household h ∈ (1, 2, …, H) derives from an alternative TAZ j ∈ (1,
2, …, J) is given by (Train, 2009):
Uhj = β′h xhj + εhj
where xhj is a row vector of explanatory variables of household char­
acteristics and TAZ attributes, βh is a column vector of coefficients for
household h, and εhj is an independently and identically distributed
random error term that follows an extreme value distribution. The co­
efficients β are assumed to vary across households (due to variation of
preferences) with density function f(β|ϕ), where parameters ϕ represent
the density function’s attributes (for example, mean and covariance of
the β). Under these assumptions, the probability of household h choosing
TAZ i is given as follows (Train, 2009):
∫ ∫ ′
eβh xhi
Phi = Lhi (β)f (β|ϕ)dβ = ∑ β′ x f (β|ϕ)dβ (2)
e h hj
j to be households with high income and multiple vehicles (Hjorthol,
This probability represents a weighted average of the conditional logit
probability Lhi (β) for different values of β and weights provided by the
density function f(β|ϕ) (Train, 2009). This study uses a continuous form
of the density function. For the model estimation, we specifically assume
that coefficients follow a normal distribution, β ∼ N(b, W), with pa­
rameters b and W to be estimated. The Mixed MNL is estimated using
simulated log likelihood. Exact maximum likelihood estimation is not
feasible because Eq. 2 cannot be evaluated analytically (Train, 2009).
In addition, assessing the full set of alternative choices for each
household in model estimation is not attainable, especially for large
study areas with thousands of alternatives. Typically, a sub-sample of
the universal choice set is randomly selected to serve as the choice set for
an individual household (Guo and Bhat, 2007; Zhou and Kockelman,
2009). Based on this method, n alternatives are selected at random from
the available J alternatives, resulting in a choice set with n + 1 alter­
natives (including the 200 observed choice). Random selection of al­
ternatives has been shown to yield consistent parameter estimation
(McFadden, 1978). The value of n chosen by previous studies varies by
unit of analysis, total number of alternatives in the universal choice set,
household sample size, and modeling approach (Zhang and Guhatha­
kurta, 2018; Yan, 2020). Following the direction of previous research,
we randomly select n = 59 alternatives for each household from a
universal choice set of J = 2819 TAZ alternatives.
4.2. Simulating private AV and CAV adoption in the Triangle Region
Scenarios of adoption of privately owned AVs and CAVs are simu­
lated in the Triangle Regional Model (TRM) for the year 2045. The TRM
is an aggregated trip-based model for the Triangle Region with four
Travel Behaviour and Society 32 (2023) 100570
Predicting residential location choice
(Mixed Logit model)
Empty trips %Change in Conventional-fuel Electric
highway capacity AVs or CAVs AVs or CAVs
10% - 4.90% Operating cost Operating cost
additional 4.70% reduced by 5% reduced by 33%
trips for 1.70% compared to compared to
all trip human-driven human-driven
purposes − 7.80% vehicles vehicles
36.10%
3.30%
major steps: trip generation, trip distribution, mode choice, and trip
assignment. Increased trip rates for some AV or CAV user groups and
capacity changes on certain highway segments are implemented to TRM
v6 to simulate AV and CAV scenarios. Assumptions and analysis related
to vehicle automation and connectivity, market penetration rate,
induced demand, empty trips, and changes in highway capacity are
(1)
discussed in the following sections.
4.2.1. Market penetration of private AVs and CAVs
With respect to automation, this study focuses on self-driving vehi­
cles (SAE Levels 4 and 5) (SAE, 2021). In addition to self-driving capa­
bilities, CAVs are assumed to have the ability to communicate with other
CAVs and the infrastructure to collect traffic and other network infor­
mation and to form platoons. This communication enables CAVs to
operate more safely and efficiently compared to AVs. The market share
of self-driving vehicles is predicted to reach its saturation point by 2070
(Litman, 2020). By 2045, up to 50% of the new vehicle sales and 40% of
the total VMT could be generated by self-driving vehicles (Litman,
2020). In addition, studies have shown that early adopters are expected
2013; Petersen et al., 2006). Based on these findings from the literature,
we consider two main market penetration scenarios: (i) a conservative
scenario (Scenario A), where high-income households with personal
vehicles in the Triangle Region are assumed to own personal AVs or
CAVs, and (ii) an optimistic scenario (Scenario B), where high income
households with personal vehicles and medium-income households with
as many or more vehicles than employed members own personal AVs or
CAVs. Based on the Triangle Region household socioeconomic charac­
teristics predicted for 2045, Scenario A translates into a 30% market
penetration rate (MPR) of AVs or CAVs, while Scenario B translates into
a 75% MPR of AVs or CAVs. To differentiate between the impacts of AVs
and CAVs, additional sub-scenarios are considered for Scenario A (30%
AVs; 30% CAVs; 15% AVs and 15% CAVs) and Scenario B (75% AVs;
75% CAVs; 37.5% AVs and 37.5% CAVs).
4.2.2. Induced travel
Self-driving vehicles are expected to improve mobility for current
non-drivers, seniors, and individuals with disabilities. Truong et al.
(2017) emphasized the potential of AVs and CAVs to cover the travel
needs of 12–17 year old individuals who are currently dependent on
public transport or their parents. A recently passed House Bill (HB 469)
in North Carolina exempts operators of self-driving vehicles from the
requirement to hold a driver’s license and states that an adult is required
only if a person under 12 years old is in the vehicle (NC General As­
sembly, 2017). Therefore, given the current NC legislation, new trips
could be generated in the future by people as young as 12 years old. In
addition, Wadud et al. (2016) suggested that there is a higher trip rate
decrease after the age of 62, compared to the 44–62 age group, which
may be due to impaired driving abilities. Based on these findings, we
introduce a 30% increase in trip rates for the individuals between 12 and
6
M.M. Hasnat et al.
17 years old and individuals who are 65 years old or older living in
households assumed to own AVs or CAVs in Scenarios A and B. These
changes are implemented for home-based other, non-home non-work,
and home-based shopping trips for the aforementioned age groups. For
individuals between 18 to 64 years old who have access to personal
vehicles, it is assumed that their travel needs are already met (Wadud
et al., 2016; Truong et al., 2017) and therefore, their trip rates remain
unchanged. The TRM does not provide any information on individuals
with disabilities; therefore, additional trips for this population group
could not be incorporated into this analysis.
4.2.3. Empty trips
Vehicles with the highest levels of automation are expected to be
able to operate without the presence of a driver in the future. Individuals
may take advantage of such capabilities and choose to re-position their
vehicles to be used by other household members or to park in free or
lower cost locations (Meyer et al., 2017). Several studies have explored
the potential of empty vehicle-miles travelled (VMT) from privately
owned self-driving vehicles by either considering reallocating vehicles
to serve other members of the household (Vyas et al., 2019), allowing
empty vehicles to park in lower cost areas (Zhao and Kockelman, 2018;
Childress et al., 2015) or exclusively incorporating empty trips in the
trip generation module of regional travel demand tools (Meyer et al.,
2017).
The proportion of the population that would be willing to re-position
their vehicles is still not well-understood. Meyer et al. (2017) included
0.5, 1, and 1.5 additional trips per person per day in the Swiss national
transport model to account for empty trips by self-driving vehicles; this
led to an increase in the daily person-trips by 17.5%, 35%, and 52.4%,
respectively. Other studies reduced the parking cost for self-driving
vehicles (Zhao and Kockelman, 2018; Childress et al., 2015) which led
to additional vehicle miles in the network. Most studies, however, did
not explicitly report the amount of trips or vehicle miles generated from
empty relocation trips. Some information is provided by research on
shared self-driving vehicles. For example, Fagnant and Kockelman
(2014) simulated a 100 square mile grid-based hypothetical city where
all travel needs are met by a fleet of shared AVs and reported that
relocation of shared AVs (to park or to serve the next customer) might
add 10% empty VMTs in the network. (Meyer et al., 2017) assumed 15%
empty trips while simulating shared-AVs in Swiss national transport
model. Also, based on an online survey, Oh et al. (2020) modeled the
share of mobility on demand services in Singapore that would be served
by self-driving vehicles under different pricing scenarios for 2030. The
agent-based simulation suggested that around 20% of the vehicle kilo­
meters travelled generated by self-driving vehicles will be empty. In this
study, we account for the potential of empty trips by introducing 10%
additional trips for all trip purposes for households with access to per­
sonal AVs and CAVs. Given the uncertainty in the amount of empty trips
that will be generated in the future, we also explore how the empty trips
assumed affect our results as part of the sensitivity analysis presented in
Section 8.
4.2.4. Highway capacity changes
The impact of AVs and CAVs on the capacity of freeway and highway
segments operating without interruptions was assessed using state-of-
the-art longitudinal and lateral movement models (Xiao et al., 2017;
Xiao et al., 2018) introduced into SUMO, which is an open source
simulation platform. Several factors including maturity and reliability of
associated technologies, liability aversion, and eminence are anticipated
to result in AV decision algorithms that, on average, are more conser­
vative compared to human driving, with negative implications on
roadway capacity (Adebisi et al., 2020; Hasnat et al., 2021). On the
other hand, CAVs will have the ability to communicate with other CAVs
and the infrastructure and constantly track the environment. CAVs are
therefore expected to maintain shorter time gaps between alike vehicle
types and form platoons with short headways (Shladover et al., 2012;
7
Travel Behaviour and Society 32 (2023) 100570
Tientrakool et al., 2011). Regarding the interactions of AVs, CAVs, and
traditional vehicles in mixed-traffic conditions, they are modeled using a
dynamic car-following algorithm that implements CAV car following (e.
g., short time gaps and platooning) if a CAV follows another CAV, and
AV car following (conservative time gaps and no platooning) if a CAV
follows an AV or a traditional vehicle. The microsimulation results
indicate that CAVs improve capacity at all levels of market penetration
(Hajbabaie et al., 2021). Specifically, a 4.7% increase in capacity is
estimated for 30% MPR and a 36.1% increase for 75% MPR. However,
30% and 75% MPRs of AVs result in reduction of capacity by 4.9% and
7.8%, respectively (Hajbabaie et al., 2021). These results are in line with
previous research on CAVs (Shladover et al., 2012; Tientrakool et al.,
2011) and recent studies on AVs (Adebisi et al., 2020; Bierstedt et al.,
2014). For scenarios with a 50–50 mix of AVs and CAVs, the interactions
between the different vehicle types in the traffic stream is found to result
in small improvements in capacity even for high MPR (1.3% increase for
30% MPR and 3.3% increase for 75% MPR). These capacity changes are
implemented on freeways and other uninterrupted highway segments in
the TRM network.
4.2.5. Cluster analysis
The results of the AV and CAV scenario analysis in the TRM are
further evaluated using cluster analysis. This allows us to gain a better
understanding of the spatial variation of network performance changes
in the Triangle Region due to AVs and CAVs and their relationship with
the built environment characteristics of each zone. We perform clus­
tering for two types of travel time changes: (i) regional, based on the
average travel time from each TAZ to all other TAZs, and (ii) local, based
on the average travel time from each TAZ to the nearest ten TAZs. Travel
time changes are estimated as the percentage change in travel time
between an AV or CAV scenario and the base scenario for 2045. Other
attributes included in the clustering are TAZ roadway density, freeway
density, average travel distance to the three CBD areas in the Triangle
Region, and percentage change in trips generated compared to the 2045
base scenario. These features represent transportation supply and
increased travel demand due to AVs or CAVs for each TAZ. We apply K-
means clustering to partition TAZs into an optimal number of clusters by
minimizing the clustering error (Wu, 2012; Lloyd, 1982). The clustering
error is the sum of squared Euclidean distances between an observed
TAZ attribute and the cluster mean for that attribute over all TAZs and
attributes under consideration (Likas et al., 2003). The optimum number
of clusters is determined based on the elbow method, which calculates
the percentage of within cluster variance (distortion) for different
numbers of clusters (Bholowalia and Kumar, 2014). The distortion
scores in this study are calculated as the sum of squared distances from
the centroid of each TAZ to the center of the cluster where the TAZ is
assigned. Before applying the clustering algorithm, the TAZ attributes
are rescaled so that their values are between zero and one; this is done to
avoid estimation bias due to attributes with high variance (Choudhary
et al., 2016; Han et al., 2012; Greenacre and Primicerio, 2013; Rokach
and Maimon, 2008).
4.3. Predicting households’ future residential location choice
We are interested in understanding how changes in transportation
network performance due to AVs and CAVs are going to impact house­
holds’ residential location in the future. To this end, the explanatory
variables in the residential location choice model are updated with the
socioeconomic data for 2045 and the transportation-related outcomes
from the TRM scenario analysis to produce residential choice pre­
dictions for 2045. Using Halton sequencing, a vector of estimated co­
efficients ̂
β h is drawn from the distributions of the individual coefficients
of the Mixed MNL model. The coefficient vector ̂ β h and the updated
explanatory variables (zhi ) are used to calculate the probability of
household h choosing alternative TAZ i:
M.M. Hasnat et al. Travel Behaviour and Society 32 (2023) 100570

̂
β’ h zhi work trip; 97% of these households used personal vehicles to
̂hi = e
P (3) commute to work. The final dataset that we use for our analysis includes
∑ ̂β’ z
e h hj responses from 2356 households who live and work within the TRM
j
region, and commute to work by personal vehicle. The households’
This process is repeated 2000 times for each household. In other words, home and work locations are geocoded in ArcGIS and spatially joined
with the TAZ layer.
for each household h, 2000 ̂ β vectors are drawn using Halton
A few zones (38 TAZs out of the 2857 TAZs) are excluded from the
sequencing; the individual probabilities are estimated using Eq. 3 and
analysis because they do not contain any residential areas. These zones
averaged to provide an estimate of the average probability of choosing
include recreational parks and protected green spaces, airports, uni­
alternative i. For every household, the process is repeated for each of the
versity campuses, and shopping complexes. For the rest of the 2819
2857 TAZs in the study region. The TAZ with the maximum average
TAZs, land-use and built environment information such as number of
probability value is reported as the chosen alternative for that
commercial establishments, educational institutions, medical in­
household.
stitutions, parks, tourist spots, and recreational establishments are
collected from various NC GIS data libraries. Median house value for
4.3.1. Future vehicle operating costs
census block groups is available from the 2018 5-year estimates of the
By 2040, the largest automotive markets (including Europe, US, and
American Community Survey. This information is spatially joined with
China) are likely to be dominated by electric vehicles (McKinsey &
the TAZ layer in ArcGIS. Transportation system information (total and
Company, 2021; General Motors, 2022). The major industries working
non-motorized roadway length), mean household income in each TAZ,
on self-driving technologies are investing heavily on battery electric
and employment opportunities (retail jobs, industrial jobs, employment
vehicles as well (Mullaney, 2020; General Motors, 2022). Previous
density) are available for 2013 through the Triangle CommunityViz 2.0
studies on cost-based analysis of autonomous mobility services have
(CommunityViz, 2018).
considered and compared conventional-fuel and electric self-driving
vehicles (Millard-Ball, 2019; Bösch et al., 2018; Stephens et al., 2016).
5.2. Variable description
These studies have suggested that self-driving vehicles will have a lower
operating cost than human-driven vehicles, due to more balanced
Table 2 presents descriptive statistics for the variables considered in
driving and lower insurance rates (Millard-Ball, 2019; Bösch et al., 2018;
Stephens et al., 2016). The reduction in the operating cost is expected to the analysis of residential location choice. The first part of the table
focuses on built environment, land-use, and socioeconomic TAZ-level
vary by roadway facility type and fuel type (Stephens et al., 2016). Es­
timates of vehicle operating cost reduction differ by study, but overall, characteristics, while the second part of the table focuses on house­
hold information retrieved from the 2016 Triangle Region Household
previous research has suggested an average 5% decrease in the oper­
ating cost of conventional-fuel self-driving vehicles and 33% decrease in Travel Survey. The majority of the households in the survey sample have
at least two working members, two or more vehicles, and live in owner-
the operating cost of electric self-driving vehicles, compared to
conventional-fuel human driven vehicles (Millard-Ball, 2019; Bösch occupied single-family houses. Households’ annual incomes are recor­
et al., 2018; Stephens et al., 2016). Because commute cost is a significant ded within ten income brackets. The average (unweighted) household
factor in households’ residential location choice, we account for these income in our sample is $50,970.
differences in vehicle operating cost in our prediction of household We estimate employment accessibility for TAZ i, Ai , to be propor­
decisions and make comparisons between electric and conventional-fuel tional to the employment opportunities in all TAZs in the study region
AVs and CAVs. Operating cost differences between AVs and CAVs have and inversely proportional to the travel time between TAZ i and all other
not been reported in the literature yet. TAZs:

1∑ J
Total Employmentj
5. Analysis of household residential location preferences in the Ai = (4)
J j=1 Travel Timeij
Triangle Region
In addition, commute time and commute cost are estimated for each
5.1. Data preparation
household. Commute time is the network travel time between TAZ
centroids of a household’s home and work location. A household’s
To model household residential location preferences, we use data
commute cost is estimated as the average vehicle operating cost for the
from the 2016 Triangle Region Household Travel Survey. The survey
distance travelled to work. Vehicle type information from the household
was conducted between February and April 2016. A stratified random
responses is used to calculate the average operating cost per mile based
sample was used based on USPS delivery addresses with compensatory
on the American Automobile Association (AAA, 2018). For each
oversampling of low-income households, households with public transit
household with multiple workers, commute time and distance to work
users, zero-vehicle households, and households with college or univer­
are calculated as the total commute time and distance for all the workers
sity students, to ensure that the final responses would be representative
in the household. As shown in Table 2, the sample’s mean household
of the household distribution in the region (RSG, 2016). A total of
commute time and cost is 27.89 min and $3.26, respectively.
76,097 households were invited through mail to participate in the sur­
vey; 4,194 households participated by completing a one-day travel diary
5.3. Mixed multinomial logit results and discussion
for every member of the household.
The survey included questions about household characteristics
The results of the Mixed MNL model are presented in Table 3. The
(number of adults, number of children, number of licensed drivers, in­
model was estimated in STATA 15 using simulated log likelihood and
come), vehicles (number, type), and occupation (location of primary
2000 Halton draws (Hole, 2007). Statistically insignificant parameters
and secondary work for workers, school location for students). The
have been removed to reduce computational burden in the model esti­
survey also asked for a one-day travel diary and contained other ques­
mation. Overall, the results are intuitive and consistent with our original
tions about travel characteristics. Out of the 4194 survey responses from
hypotheses. We find that TAZs with higher population density, more
households, 471 responses were removed because the home or work
medical and recreational facilities, lower road density, and lower crime
location was outside of the TRM region, no work location was provided,
index have a higher probability of being chosen by a household, on
or multiple home locations were reported. From the remaining 3723
average. In addition, our results suggest that household preferences are
household responses, only 2435 households reported a home-based
heterogeneous with respect to TAZ employment density. The estimated

8
M.M. Hasnat et al. Travel Behaviour and Society 32 (2023) 100570

Table 2
Summary statistics of TAZ and household characteristics.
TAZ-level characteristics

Variable Mean Standard Minimum Maximum

deviation

TAZ area (square miles) 1.19 2.07 0.01 24.16

Population density, 2013 (1000 per square mile) 1.84 2.50 0.00 34.79
Percentage of White population, 2017 0.67 0.19 0.00 0.98
Percentage of African American population, 2017 0.22 0.18 0.00 0.85
Employment density, 2013 (1000 jobs per square mile) 2.55 14.65 0.00 499.32
Employment accessibility, 2013 15.24 3.95 6.13 23.39
Roadway density, 2013 (roadway miles per sq. miles of area) 13.21 9.80 0.00 72.98
Non-motorized path length, 2013 (miles) 1.21 2.09 0.00 37.16
Number of educational institutions, 2018 0.79 1.27 0.00 11.00
Number of medical facilities, 2016 0.17 0.59 0.00 7.00
Number of recreational centers, 2016 0.12 0.39 0.00 7.00
Number of religious institutions, 2016 0.78 1.20 0.00 9.00
Number of government offices, 2016 0.39 1.46 0.00 25.00
Number of libraries and museums, 2016 0.13 0.50 0.00 7.00
Crime index, 2017 (number of violent crimes per 100,000 population) 267.24 99.95 175.89 461.86
Mean household income, 2013 (1000 USD) 75.04 36.24 32.80 302.94
Median house value, 2018 (in 1000 USD) 251.01 103.30 63.90 699.31

Household characteristics (unweighted), 2016 Triangle Region Household Travel Survey

0 1 2 3 or more
Household size – 21% 39.5% 39.5%
Number of adults – 24.2% 66.6% 9.2%
Number of children 64.7% 16.1% 14.4% 4.8%
Number of students 57.9% 20.3% 15.9% 5.9%
Number of workers – 45.7% 50.3% 4.0%
Number of vehicles 0.0% 27.5% 52.0% 20.5%
Number of bikes 41.0% 18.3% 18.6% 22.1%
Household members with a driver’s license 0.0% 26.4% 64.4% 9.2%

Home ownership owned rented other unspecified

77.2% 21.1% 0.4% 1.3%

Life cycle working retirees retirees working

with children without children with children without children
34.3% 8.1% 1.0% 56.6%

Resident type single-family house town house building with ⩾ 2 units other
74.0% 10.5% 15.0% 0.5%

Residence area type CBD Urban Suburban Rural

0.3% 44.5% 38.6% 16.6%

Mean Standard deviation Minimum Maximum

Household annual income, 2016 (1000 USD) 101.98 59.32 7.5 300.00
Household commute time, 2016 (minutes) 26.57 16.65 3.00 116.06
Household commute cost, 2016 (USD) 3.12 2.61 0.02 17.47

coefficient of employment density is normally distributed with a mean of
− 0.2108 and standard deviation of 0.0812, indicating that for 99.5% of
cases, higher employment density reduces the probability of choosing a
TAZ for residential location. Furthermore, the results indicate that
households of similar income and racial composition tend to be spatially
concentrated. This outcome, which has also been reported by previous
studies (Pinjari et al., 2011; Waddell, 1992), is associated with housing
affordability considerations as well as the residential segregation that
persists in many US regions. Households that include students are more
likely to choose TAZs with higher number of educational institutions, all
else being equal. Additionally, we find that households with a higher
number of vehicles tend to live farther from CBD areas, and households
that own at least two bicycles tend to reside in areas with more non-
motorized paths. The positive coefficients for the indicator variables
representing urban and suburban TAZs provide evidence of stronger
preference for urban and suburban living.
Our results confirm that the travel time between home and work
plays an important role in a household’s residential location choice.
Higher commute time is associated with lower probability of a TAZ
being selected as home location. In line with previous research (Guo and
Bhat, 2007; Pinjari et al., 2011), we also find that the effect of commute
time varies among households. Household commute time and cost are
highly correlated because commute cost is directly proportional to
commute distance. To avoid a multicollinearity problem and to account
for variability in cost sensitivity across households, we divide the
household commute cost by the household relative income (household
income divided by the sample average household income). As expected,
we find that on average, household commute cost is negatively associ­
ated with a location’s utility. Last, our results indicate that higher
employment accessibility reduces the probability of a TAZ being
selected, restating the common preference for residential areas farther
from employment centers for households in the Triangle Region.

9
M.M. Hasnat et al. Travel Behaviour and Society 32 (2023) 100570

Table 3
Mixed multinomial logit estimation results for household residential location choice (random parameters are normally distributed; standard errors are in parentheses).
Variable Estimated Standard Deviation of
Parameter Parameter Distribution

TAZ-level characteristics

Population density 0.1052 (0.0095)*** –

Employment density − 0.2108 (0.0185)*** 0.0812 (0.0146)***
Roadway density − 0.0363 (0.0046)*** –
TAZ area 0.1410 (0.0136)*** –
Number of medical facilities 0.0868 (0.0359)* –
Number of recreational facilities 0.1985 (0.0461)** –
Crime index − 0.0008 (0.0003)* –
Urban TAZ indicator 1.3207 (0.1109)*** –
Suburban TAZ indicator 1.0892 (0.0864)*** –

TAZ-level characteristics interacted with household demographics

Absolute difference in household income and TAZ-level mean household income − 0.0121 (0.0008)*** –
Median house value/ Household Income − 0.0402 (0.0144)** –
White household × Percentage of White population in TAZ 2.8543 (0.1917)*** –
African American household × Percentage of African American population in TAZ 6.5548 (0.4413)*** –
Household with student member × Number of educational institutions in TAZ 0.1795 (0.0282)*** –
Household with more vehicles than adults × Distance to closest CBD 0.0546 (0.0081)*** –
Household with at least 2 bicycles × TAZ non-motorized path length 0.0881 (0.0096)*** –

Commuting and accessibility

Household commute time − 0.1000 (0.0047)*** 0.0431 (0.0037)***

Household commute cost/ (Household income/ Average household income) − 0.0789 (0.0164)*** 0.0547 (0.0153)***
Employment accessibility − 0.1271 (0.0110)*** –

Number of households 2356

Log-likelihood at zero − 7363.09
Log-likelihood at convergence − 7106.57

***ρ < 0.001, **ρ < 0.005

5.4. Predictive performance of the estimated model
We estimate the Brier score (Brier, 1950) and the average of the
predicted probabilities for the chosen alternative (choice actually made
by each household) to evaluate the predictive accuracy of the model
(Danaf et al., 2019; Xie et al., 2020; Haque et al., 2019). The Brier score
is defined as follows:
1 ∑ H ∑ J 2
BS = ̂
(1{yn = j} − P hj )
HJ h=1 j=1
where H is the number of households in the test sample, J is the total
number of alternatives, P̂ hj is the predicted probability of household h
{ }
choosing alternative j, and indicator 1 yn = j equals to one if the
alternative is the actual chosen alternative. The Brier score accounts for
the predicted choice probability of the entire choice set. True predictive
probabilities minimize the value of the Brier score. Therefore, a Brier
score close to zero reflects better predictive capability of a model
(Krueger et al., 2021; Gneiting and Raftery, 2007). We randomly split
the data into a training sample (80%) for model estimation and a testing
sample (20%) for model validation. We then estimate the model using
the training sample and predict the location choice probability on the
test sample. This process is repeated five times. All the estimated models
provided consistent parameter estimates of the exogenous variables.
Table 4 shows the average predicted probability for the chosen alter­
natives pchosen and the Brier score for the five subsets of the data.
It can be seen from Table 4 that the Brier scores for all the subsets are
very close to zero which indicates high predictive ability of the esti­
mated models. The average of the predicted probabilities for the actual
choices varies from 0.0076 to 0.0045. The equal choice probability of
the households choosing from 2819 alternative TAZs is 0.00035 (1/
2819). Therefore, the estimated models improve the prediction capa­
bility by 13–22 times compared to randomly choosing from 2819
alternatives.
6. Network-level impacts of AV and CAV adoption
The TRM model was run for (i) a base scenario, which provides the
baseline results on system performance considering zero AV or CAV
adoption for the year 2045, (ii) three scenarios (A1, A2, A3) that assume
(5)
30% MPR of AVs or CAVs, and (iii) three scenarios (B1, B2, B3) that
assume 75% MPR of AVs or CAVs. All AV and CAV scenarios incorporate
induced travel demand (discussed in Section 4.2.2), empty trips (dis­
cussed in Section 4.2.3), and changes in multi-lane highway and freeway
segments’ capacity (discussed in Section 4.2.4). The daily vehicle-miles
traveled (VMT), average freeway speed, daily hours of delay, and
average travel time for trips to work for the TRM network are presented
by scenario and compared with the base 2045 scenario in Table 5.
Overall, the average changes estimated for daily VMT and average
commute time compared to the base scenario are relatively small and
less than 7%. We find moderate changes in the average freeway speed
which vary from a 3.4% decrease in the 75% AV scenario to a 5.8%
increase in the 75% CAV scenario. The estimated changes are more
substantial for daily hours of delay, but still below 30% of the base
scenario. We find an increase in VMT for all scenarios. This change in
VMT is greater for higher market penetration of privately-owned AVs
and CAVs. In terms of network performance, results indicate substantial
differences between adoption of AVs and CAVs. Scenarios representing
AV-only adoption (A1 and B1) lead to lower speed and higher delays and
travel time, compared to the base scenario. CAV-only scenarios (A2 and
B2) lead to improvements in freeway speed and a minor increase in the
daily delay. For scenarios that simulate mixed MPR of AVs and CAVs (A3

10
M.M. Hasnat et al. Travel Behaviour and Society 32 (2023) 100570

Table 4
Validation of Mixed Logit Model.
Subset Households in Initial Final Households in pchosen Brier score
estimation set log-likelihood log-likelihood testing set

1 1885 − 5878.41 − 5672.99 471 0.0054 0.0004

2 1885 − 5885.64 − 5687.77 471 0.0051 0.0004
3 1885 − 5883.08 − 5681.82 471 0.0045 0.0004
4 1885 − 5901.93 − 5703.35 471 0.0063 0.0004
5 1885 − 5928.76 − 5719.51 471 0.0076 0.0004

Table 5
Network-level impacts of AV and CAV adoption for the Triangle Regional Model region in 2045.
Market Total daily vehicle-miles Average freeway speed Total daily hours of Average travel time: trips to Average employment
traveled delay work accessibility
Penetration Value (million) %Changeb Value (mph) %Change b
Value %Changeb Value (min) %Changeb Value %Changeb

Base Scenarioa 90.79 – 55.40 451,335 – 25.84 – 19.62 –

(A1) 30% AV 92.47 1.85 54.60 − 1.44% 493,824 9.41 26.47 2.44 19.34 − 1.44%
(A2) 30% CAV 92.71 2.11 56.10 1.26% 456,778 1.21 25.97 0.50 19.69 0.35%
(A3) 15% AV, 92.62 2.01 55.60 0.36% 469,083 3.93 26.13 1.12 19.56 − 0.27%
15% CAV
(B1) 75% AV 95.61 5.30 53.50 − 3.43% 582,580 29.08 27.4 6.04 19.05 − 2.92%
(B2) 75% CAV 96.49 6.27 58.60 5.78% 451,608 0.06 25.63 − 0.81 20.21 3.03%
(B3) 37.55% AV, 95.91 5.63 55.30 − 0.18% 531,755 17.82 26.74 3.48 19.55 − 0.35%
37.5% CAV

a The base scenario reflects the network performance for zero market penetration of AVs or CAVs in 2045.
b % Change is the percent change between a performance measure for an AV or CAV market penetration scenario and the base scenario.

and B3), the daily hours of delay increase up to 17.8% as a result of the
additional demand, empty trips, and only small improvements in high­
way capacity due to interactions of different vehicle types in the traffic
stream. Table 5 also reports the average employment accessibility for
different scenarios. Except for 30% CAV scenario, a reduction in travel
time increases employment accessibility and vice versa as it is inversely
proportional to the travel time (Eq. 4). Focusing on the most impactful
scenarios, a 75% MPR of AVs is expected to lead to deteriorated network
performance (29.1% increase in daily hours of delay and 3.4% decrease
in average freeway speed), while a 75% MPR of CAVs is expected to
increase the average freeway speed by 5.8% with very little impact on
daily hours of delay.
We use K-means clustering to explore how the changes in network
performance vary with transportation supply, demand, and other zone
attributes. Network performance changes are captured by TAZ-to-TAZ
travel time. We calculate the percentage change in average travel time
from each TAZ to all other TAZs compared to the base scenario to cap­
ture average regional network performance changes. We also calculate
the percentage change in average travel time from each TAZ to the
nearest ten TAZs compared to the base scenario to capture the changes
in local network performance. The average regional travel time de­
creases (compared to the base scenario) only for the scenarios of CAV
adoption (A2 and B2). Average local travel time increases for all sce­
narios, with a small decrease for the scenario of 75% MPR of CAVs (B2).
The cluster analysis includes four additional TAZ attributes for 2045:
roadway density, freeway density, average distance to the three CBD
areas in the Triangle Region, and percentage change in total trips
generated compared to the base scenario. For both regional and local
analyses, the optimum number of clusters is found to be four.
Table 6 and Fig. 3 present the results of the regional cluster analysis
for 75% MPR of AVs and 75% MPR of CAVs (scenarios B1 and B2). The
results are consistent for the rest of the scenarios, but the magnitude of
the network performance changes is smaller. The clusters are named
based on the level of travel time changes for the TAZs within each
cluster. For 75% MPR of AVs, TAZs in the “High increase” cluster
experience an average regional travel time increase of 4.5%. These TAZs
are mainly located in CBD areas or other areas close to the CBD with
high roadway and freeway density. As we move further away from the
CBD areas, the regional travel time impacts decrease. TAZs within the
“Low-medium increase” cluster have an average regional travel time
increase of 2.8% and are located along the peripheral zones, mainly in
rural areas with low roadway and freeway density. For 75% MPR of
CAVs, all four clusters experience a reduction in average travel time
compared to the base scenario. TAZs in the “High reduction” cluster
experience an average regional travel time decrease of 4.4%. These TAZs
are located in CBD areas and other urban or suburban areas with high
roadway and freeway density. Areas with medium and small travel time
decrease include TAZs with very low freeway density and higher dis­
tance from the CBD compared to other clusters.
Different results are found for the cluster analysis related to local
network performance changes, as shown in Table 7 and Fig. 4. For 75%
MPR of AVs, TAZs within the “High increase” cluster experience a 2.7%
average increase in local travel time and are mainly located along the
urban fringe, whereas “Medium increase” TAZs are primarily in urban
areas closer to the CBD and have the highest average roadway density
compared to other clusters. Smaller local travel time changes are found
for the majority of suburban and rural TAZs with lower road and
freeway densities. For 75% MPR of CAVs, the majority of TAZs experi­
ence a moderate to low increase in local travel time. The highest in­
crease is observed in the TAZs with the highest density of roadways and
freeways. Even though this result seems counter-intuitive, it can be
explained through the consideration of network effects in traffic
assignment: Because of the increase in capacity of freeway and certain
highway segments due to CAV adoption, more traffic is diverted to these
facilities creating local traffic nearby and reducing traffic elsewhere.
Lastly, a small reduction in local travel time is observed for some TAZs
located mainly in rural and outer suburban areas of the Triangle Region.
7. Changes in household residential location in 2045 under AV
and CAV adoption scenarios
Future residential location choices are predicted for the 2045 base
scenario as well as for the six AV and CAV scenarios for households that
commute to work by personal vehicle. These predictions are carried out
by applying the estimated Mixed MNL model using updated variable
values that reflect 2045 conditions based on the methodology described

11
M.M. Hasnat et al. Travel Behaviour and Society 32 (2023) 100570

Table 6
K-means clustering results – Regional travel time changes in 2045.
Cluster name % Change in Roadway density Freeway density Average distance % Change in
(number of TAZs) travel time* (miles/sq. mile) (miles/sq. mile) from CBDs (miles) trips generated

75% MPR of AVs

High increase (135) 4.49 73.91 25.95 16.56 6.00

High-medium increase (1592) 4.26 15.49 2.10 18.46 6.74
Medium increase (701) 2.89 8.25 0.71 28.45 7.32
Low-medium increase (429) 2.75 6.48 0.26 44.91 6.97

75% MPR of CAVs

High reduction (127) − 4.37 75.2 27.88 16.71 6.07

High-medium reduction (1271) − 4.04 15.15 2.6 19.53 6.87
Medium reduction (574) − 3.12 6.63 0.32 42.32 7.03
Low reduction (885) − 2.76 11.97 0.48 22.14 7.02

* Percentage change in average travel time from each TAZ to all other TAZs compared to the base scenario in 2045.

Fig. 3. K-means clustering results – Regional travel time changes in 2045.

Table 7
K-means clustering results – Local travel time changes in 2045.
Cluster name % Change in Roadway density Freeway density Average distance % Change in
(number of TAZs) travel time* (miles/sq. mile) (miles/sq. mile) from CBDs (miles) trips generated

75% MPR of AVs

High increase (511) 2.69 13.86 3.94 21.52 6.83

Medium increase (1316) 0.79 21.71 3.72 17.29 6.67
Low-medium increase (640) 0.67 7.77 0.69 30.1 7.34
Low increase (390) 0.3 6.62 0.26 45.67 6.9

75% MPR of CAVs

Medium increase (129) 0.35 75.09 27.31 16.78 6.04

Low-medium increase (1469) 0.31 16.3 2.16 17.66 6.74
Low increase (448) 0.2 6.62 0.3 44.56 6.95
Low reduction (811) − 0.02 8.14 0.78 28.06 7.35

* Percentage change in average travel time from each TAZ to the nearest 10 TAZs compared to the base scenario in 2045.

12
M.M. Hasnat et al.
Fig. 4. K-means clustering results – Local travel time changes in 2045.
in Section 4.3. TAZ-level characteristics, including population density,
employment density, mean household income, roadway density, and
non-motorized path length are updated for 2045 using the information
available from the Triangle CommunityViz 2.0. Employment accessi­
bility and household commute time and cost are estimated for 2045
based on the TAZ-to-TAZ TRM results for each scenario. We assume that
all households in the survey sample, irrespective of AV/CAV ownership,
experience the same travel time for a given origin–destination pair and a
given scenario because the travel times reflect network-level conditions.
We note that the analysis discussed in Section 6 includes all zones
and households in our study area to forecast network conditions, while
the household survey sample used here (as well as in Section 5) contains
only households who commute to work by personal vehicle. Although
our analysis is restricted to this group of households, we use the survey
weights so that our results are representative of the Triangle Region’s
working households with personal vehicles. The weight of each survey
record is based on household size, number of workers, number of ve­
hicles, age of the head of the household, income, and number of children
in the household to match the demographic data targets from the
2010–2014 five-year estimates of the American Community Survey RSG,
2016). In addition, out of the 2356 household survey records whose
behavior was analyzed in Section 5, 1452 (unweighted) households
qualify for AV or CAV ownership under Scenario A (30% market pene­
tration) and 2210 (unweighted) households qualify under Scenario B
(75% market penetration). We first focus on the changes of commute
and location characteristics of these households, and then we present the
overall results of residential location for all working households who
drive to work.
Fig. 5 presents the changes in location and commute characteristics
by scenario for households with conventional-fuel and electric AVs and
CAVs. These results constitute percentage changes of weighted averages
between an AV/CAV scenario and the respective 2045 base scenario. We
find that extensive adoption of conventional-fuel AVs leads to house­
holds choosing home locations closer to their work (1.2% and 5.2%
decrease in average commute distance for 30% and 75% MPR of AVs,
respectively, compared to the base scenario) in order to counteract the
deteriorated transportation network conditions. Changes in the
commute time are not proportionate to the changes in commute
13
Travel Behaviour and Society 32 (2023) 100570
distance, reflecting the negative impact of AVs’ market penetration on
network speeds and delays. In the case of a 30% MPR of electric AVs, the
effect of lower vehicle operating cost on households’ utility outweighs
the impact of higher network travel times resulting in households
moving farther from work and a 0.9% increase in commute time, on
average. The same does not hold for a 75% MPR of electric AVs, where
the lower network performance leads to a decrease in commute distance,
despite the lower cost of travel. Scenarios of CAV adoption lead to
households, on average, choosing to reside further away from work and
CBD areas compared to the 2045 base scenario. The increase in the
average distance between home and work is partially motivated by the
network performance improvements due to widespread CAV adoption.
The average household commute time does not change in the case of
conventional-fuel CAVs despite the increase in commute distance due to
the improved regional and local network conditions. Regarding electric
CAVs, the decrease in vehicle operating cost reduces the disutility of
longer commutes even further and leads to more substantial changes in
average commute distance and the households’ distance from the CBD
areas. Overall, the results indicate that the market penetration of CAVs
enables the average household to choose a more attractive residential
location in terms of amenities and neighborhood characteristics without
significant changes in commute time despite the greater commute dis­
tance compared to the base scenario. As shown in Section 6, when there
is a mixture of AVs, CAVs, and traditional vehicles in the traffic stream
the positive impacts of CAVs on the transportation network are rela­
tively counterbalanced by the negative AV effects, leading to small
overall changes. Moderate impacts in household location are found for
these scenarios mainly for the case of electric vehicles where there is a
large decrease in vehicle operating costs.
The choices of households with traditional vehicles may also be
affected given the predicted impacts of vehicle automation and con­
nectivity on the transportation system performance. It is therefore
important to study the overall impacts of these technologies on the
spatial distribution of households within the study region. The CBD,
urban, and suburban areas within the Triangle Region are predicted to
grow substantially by 2045 (Fig. 1). Fig. 6 presents the predicted loca­
tion of households who commute to work by personal vehicle (weighted
proportion of households) for the 2045 base scenario and the AV/CAV
M.M. Hasnat et al. Travel Behaviour and Society 32 (2023) 100570

Fig. 5. Average changes in location and commute characteristics of households with AVs/CAVs compared to the respective 2045 base scenario. 2045 base scenario
weighted averages for 30% market penetration (1452 survey records): distance to work = 10.35 miles; commute time = 23.61 min; commute cost = $1.92; distance
to closest CBD = 11.78 miles. 2045 base scenario weighted averages for 75% market penetration (2210 survey records): distance to work = 9.02 miles; commute
time = 21.03 min; commute cost = $1.68; distance to closest CBD = 11.43 miles.

scenarios. The results show that the adoption of conventional-fuel or
electric AVs leads to a higher proportion of households residing in urban
areas (up to 1.6 percentage points higher compared to the base sce­
nario). On the other hand, conventional-fuel CAVs lead to a higher
proportion of households in suburban and rural areas (up to 0.6 per­
centage points higher compared to the base scenario). In the case of
electric CAVs, the reduced vehicle operating cost reinforces the subur­
banization trends and results in a larger proportion of households
choosing to reside in suburban and rural areas. The weighted percentage
of households living inside suburban and rural areas increases by 2.4
percentage points (a 5.6% increase) for a 75% MPR of electric CAVs.
We note that the residential location choice of each household is a
complex decision arising from the consideration of multiple factors as
well as random variation (Table 3). The widespread adoption of AVs and
CAVs translates into distinctive regional and local network impacts,
which may affect households differently depending on their job location

14
M.M. Hasnat et al. Travel Behaviour and Society 32 (2023) 100570

Fig. 6. Predicted residential location of households that commute to work by personal vehicle.

and socioeconomic and other attributes. This is demonstrated in Fig. 7. It
is shown that, for example, 1.5% of households that were predicted to
reside in CBD or urban areas in 2045 chose a suburban or rural area
when a 30% MPR of electric CAVs was assumed; at the same time, 0.5%
of households that were predicted to reside in suburban or rural areas in
2045 chose a CBD or urban area when a 30% MPR of electric CAVs was
assumed. This result indicates that there is variation among household
decisions but the net difference, which is reflected in Fig. 6, suggests a
suburbanization trend for CAVs. Fig. 7 also reveals that only a small
number of households (up to 3.2%) is predicted to choose a different
area type between the 2045 base scenario and any AV or CAV scenario.

Fig. 7. Predicted residential location shifts between the 2045 base scenario and AV/CAV scenarios for households that commute to work by personal vehicle.

15
M.M. Hasnat et al. Travel Behaviour and Society 32 (2023) 100570

8. Sensitivity analysis
We conduct additional analysis to understand the sensitivity of our
results to the main study assumptions, including the level of induced
demand, empty trips, and highway capacity changes due to the adoption
of AVs and CAVs. The sensitivity analysis focuses on the optimistic
scenario of 75% market penetration of AVs and CAVs. The first five
columns of Table 8 include the assumptions of each scenario run and the
remaining table presents the resulting distribution of households in
urban/CBD and suburban/rural areas. Fig. 8 presents the changes in the
percentages of weighted households in urban/CBD areas and suburban/
rural areas compared to the base 2045 scenario.
Figs. 8(a) and (b) present the changes in the proportion of house­
holds inside suburban/rural areas and urban/CBD areas, respectively,
for three different rates of induced demand (10%, 30%, and 50%). Our
results are not particularly sensitive to the amount of induced trips by
young and senior household members, as these trips represent a smaller
portion of the total trips in our region. We also test the influence of the
amount of empty trips in the study outcomes (Figs. 8(c) and (d)). A large
number of empty trips increases the total VMT and delays in the trans­
portation network and reverses the suburbanization trends for the case
of conventional-fuel CAVs; electric CAVs would still lead to more
households living in suburban and rural areas even with 20% additional
empty trips. Finally, we explore how large changes in highway capacity
due to AVs or CAVs may affect our results (Figs. 8(e) and (f)). As ex­
pected, if highway capacity further deteriorates due to AVs, more
households would choose to live closer to work and therefore the pop­
ulation of urban areas would increase; the opposite holds for the case of
highway capacity improvements due to CAVs. Overall, the direction of
the changes estimated as part of this sensitivity analysis is as expected.
For example, additional demand and empty trips create more adverse
network conditions and discourage more households from residing
farther from work. In terms of the magnitude of the estimated changes,
the results are more sensitive to the amount of empty trips added to the
network compared to adjustments in the capacity of uninterrupted
highway and freeway segments and induced demand from young and
senior household members.
9. Conclusion
Mass adoption of self-driving vehicle technologies is expected to
significantly impact transportation system performance and mobility,
which are vital factors of residential location decisions for households.
This study investigates and compares the long-term effects of moderate
to high market penetration rates of personal AVs and CAVs on the dis­
tribution of households within a metropolitan area. First, this study
estimates a Mixed Multinomial Logit model to capture the existing res­
idential location choice preferences of households living in the Triangle
Region of North Carolina and commuting to work by personal vehicle.
Then, the region’s transportation network performance for several AV
and CAV-related scenarios for the year 2045 is simulated using the
Triangle Region four-step travel demand model. The outputs from the
travel demand model along with predicted sociodemographic variables
for 2045 are used to forecast the future residential location of the studied
household population by AV and CAV scenario. The analysis encom­
passes a wide range of scenarios, including conservative and optimistic
levels of market penetration, self-driving vehicles with and without
vehicle connectivity components, and fuel types associated with
different operating costs, providing a broader spectrum of the potential
effects of self-driving vehicle technologies.

Table 8
Sensitivity analysis assumptions and results.
MPR Fuel type Induced Capacity Empty Weighted % of % Change in weighted Weighted % of % Change in weighted
demand adjustment trips households in households inside households in households inside
Urban/CBD areas Urban/CBD areas Suburban/Rural Suburban/Rural areas
areas

0 – – – – 57.5% – 42.5% –

75% AV Conventional 10% − 7.8% 0.0% 58.1% 1.0% 41.9% − 1.4%

fuel
30% − 7.8% 0.0% 58.4% 1.6% 41.6% − 2.2%
50% − 7.8% 0.0% 58.6% 1.8% 41.4% − 2.5%
30% − 7.8% 10.0% 59.1% 2.7% 40.9% − 3.7%
30% − 7.8% 20.0% 59.8% 4.0% 40.2% − 5.4%
30% − 11.7% 0.0% 58.5% 1.7% 41.5% − 2.3%
30% − 15.6% 0.0% 58.8% 2.2% 41.2% − 3.0%
Electric 10% − 7.8% 0.0% 56.8% − 1.2% 43.2% 1.6%
30% − 7.8% 0.0% 57.5% 0.0% 42.5% 0.0%
50% − 7.8% 0.0% 57.7% 0.4% 42.3% − 0.5%
30% − 7.8% 10.0% 58.1% 1.1% 41.9% − 1.5%
30% − 7.8% 20.0% 58.6% 1.9% 41.4% − 2.6%
30% − 11.7% 0.0% 57.8% 0.4% 42.2% − 0.6%
30% − 15.6% 0.0% 58.1% 1.1% 41.9% − 1.5%

75% CAV Conventional 10% 36.1% 0.0% 55.2% − 4.0% 44.8% 5.5%
fuel
30% 36.1% 0.0% 55.9% − 2.8% 44.1% 3.8%
50% 36.1% 0.0% 55.9% − 2.8% 44.1% 3.8%
30% 36.1% 10.0% 56.9% − 1.1% 43.1% 1.5%
30% 36.1% 20.0% 57.7% 0.4% 42.3% − 0.5%
30% 54.2% 0.0% 55.0% − 4.3% 45.0% 5.9%
30% 72.2% 0.0% 54.6% − 5.1% 45.4% 6.9%
Electric 10% 36.1% 0.0% 54.1% − 6.0% 45.9% 8.1%
30% 36.1% 0.0% 54.3% − 5.6% 45.7% 7.5%
50% 36.1% 0.0% 54.3% − 5.5% 45.7% 7.4%
30% 36.1% 10.0% 55.1% − 4.1% 44.9% 5.6%
30% 36.1% 20.0% 56.3% − 2.1% 43.7% 2.8%
30% 54.2% 0.0% 53.7% − 6.7% 46.3% 9.0%
30% 72.2% 0.0% 53.4% − 7.2% 46.6% 9.7%

16
M.M. Hasnat et al.
Fig. 8. Sensitivity analysis.
High market penetration of AVs is characterized by reduced highway
capacity, which adversely impacts transportation network speeds, travel
time, and delays. Specifically, a 75% MPR of AVs is associated with a 29%
increase in daily hours of delay and 3.4% decrease in average freeway
speed. In addition, a higher rise in average travel time from each TAZ to
all other TAZs is experienced in CBD and urban areas with higher road and
freeway density. Locally, the travel times are higher around the urban
fringe, while a medium increase is found within urban areas. For high
market penetration of conventional-fuel and electric AVs, households
tend to reside closer to work to partially offset the degraded trans­
portation network conditions. Our results indicate that extensive adop­
tion of AVs is associated with up to a 1.6 percentage point increase in the
share of households residing in urban areas compared to the 2045 base
17
Travel Behaviour and Society 32 (2023) 100570
scenario. This translates to a 2.8% increase in urban households that
commute to work by personal vehicle. This impact for personal self-
driving vehicles has not been reported in most previous studies as those
have assumed improved network performance due to self-driving vehi­
cles. Only Cordera et al. (2021) reported an increase in urban population
due to a high market share of personal AVs. Studies that have explored
high market shares of shared-AVs or a combination of shared and private
AVs have also reported a shift in population from non-urban to urban
areas (Gelauff et al., 2019; Bansal and Kockelman, 2018). For example,
Gelauff et al. (2019) simulated a combined “full automation” scenario
where private AVs and shared-AVs constitute 60% and 14% of the total
modal share, respectively, and resulted up to 1.5% increase in the total
population inside the four largest cities of the Netherlands.
M.M. Hasnat et al.
Adoption of CAVs is expected to enhance highway throughput
compared to AVs and human driven vehicles. We show that a moderate
to high market penetration of CAVs will improve the overall trans­
portation network performance. The cluster analysis for regional
network performance changes suggests higher reductions in average
travel time from each TAZ to all other TAZs in CBD and urban areas that
have higher roadway and freeway density. Local travel times also
decrease in some cases, with the highest reductions in suburban and
rural areas with low roadway and freeway density. These conditions
motivate households to reside further away from their work location in
search of more preferable neighborhood amenities and other charac­
teristics without increasing their commute time compared to the 2045
base scenario. For a 75% MPR of electric CAVs, the average commute
distance of households with personal CAVs increases by 5.8% compared
to the 2045 base scenario. This leads to a 2.4 percentage point increase
in the share of households residing in suburban or rural areas within the
Triangle Region. It reflects a 5.6% increase in this region’s suburban and
rural population who commutes by personal vehicle through shifts from
urban zones. Similar trends have also been reported in previous related
studies that have assumed improved network performance or better
travel experience due to the adoption of personal self-driving vehicles
(Zhang and Guhathakurta, 2018; Gelauff et al., 2019; Cordera et al.,
2021). For instance, Gelauff et al. (2019) reported that the share of non-
urban population would increase by 1% due to mass adoption of private
self-driving vehicles in the Netherlands.
The sensitivity analysis provides a broader spectrum on the direction
and magnitude of the households’ movement as response to changes in
highway capacity, induced demand, and empty trips due to AVs and
CAVs. We find that our results are more sensitive to the amount of empty
trips added compared to changes in other parameters. We also find that
if CAVs induce a large number of empty trips, the travel delays may
outweigh the capacity-related benefits of connectivity and result in
population shifting to urban areas. Lastly, our results indicate that if
CAVs have a large positive impact on highway capacity, the proportion
of households inside suburban and rural areas will increase by 9.5%
compared to the base 2045 scenario. Although not directly comparable,
this outcome is close to the findings of several stated preference surveys
(Kim et al., 2020; Bansal and Kockelman, 2018) which have reported
that up to 11% of the households might choose to reside further away
from the city centers if there is a high market share of self-driving ve­
hicles. Overall, the estimated changes in the spatial distribution of
households remain relatively small for all scenarios (less than 10%) but
not trivial. To provide a relative comparison, the urban and suburban
areas in the Triangle Region are expected to grow by 55–65% by 2045
(CommunityViz, 2018).
Some additional conclusions that may be relevant to practitioners
are discussed herein. First, it is important to note that given a trans­
portation network where there are no dedicated lanes for AVs or CAVs in
the majority of the roadway segments, the improved or deteriorated
network conditions due to mixed traffic will be experienced by all
commuters and may lead to different location decisions even for
households that do not own AVs or CAVs. Second, our study suggests
substantial impacts mainly for high market penetration of self-driving
technologies in combination with vehicle electrification. Therefore, it
is likely that it will take more than three decades to realize such impacts.
Simultaneously, public agencies should carefully consider the forecasted
suburbanization trends and promptly explore policies, programs, and
investments that discourage private vehicle ownership and provide
opportunity for shared vehicle programs to be tested and explored.
Policies that include congestion charges and peak-hour fees on empty
rides could be considered to reduce the negative impacts of induced trips
on the transportation network, although their impact on residential
location choice is not clear. If such policies reduce traffic congestion,
they may encourage some households to move farther from employment
centers. At the same time, some self-driving vehicle owners may be
incentivized to move closer to their work to avoid large charges
18
Travel Behaviour and Society 32 (2023) 100570
(depending on their sensitivity to such costs).
This research makes an essential contribution by predicting a range
of long-term changes in the distribution of a large population group
(adults who commute to work by personal vehicle) within a US metro­
politan area given various scenarios. This work also contributes to the so
far limited understanding of the differential impacts of AVs and CAVs.
This study’s results can be used by transportation and planning agencies
to update local or regional plans, or as scenarios in their own analyses.
Other regions could also adopt our methodology to produce region-
specific estimates. We note that comparisons with studies conducted
for other regions are not straightforward due to differences in studies’
assumptions and scope.
There are several limitations related to this research. First, our
analysis focuses only on households with at least one working member
who use their personal vehicles to commute to work and both their home
and work location are within the Triangle Regional Model region
(Fig. 1). This population represents approximately 60% of the total
households in the Triangle Region. However, we note that this consti­
tutes a good representation of the working households in the region
because most working households not included in the analysis were
removed due to their work or home location being outside the TRM
region. In addition, although the literature indicates that there is
geographical and individual-level preference heterogeneity in AV and
CAV adoption (Wali et al., 2021; Hossain and Fatmi, 2022), we were not
able to incorporate such scenarios into our study due to limitations
posed by the available travel demand model. Still, our study makes
relevant assumptions of AV and CAV adoption by a certain socioeco­
nomic strata in line with previous research findings suggesting that the
early adopters of these technologies are expected to be households with
high income and multiple vehicles (Hjorthol, 2013; Petersen et al.,
2006). Further, our work concentrates on privately-owned AVs and
CAVs and does not consider the possible emergence of autonomous
transit and ridesharing services which could attract more households to
live closer to urban centers and take advantage of such amenities.
Moreover, our analysis is limited to commute trips and it does not ac­
count for the impact of any changes in the perceived value of time due to
increased comfort and productivity inside a self-driving vehicle. We note
though that recent studies highlight that such impacts are going to be
mainly realized during longer trips and not during daily commute to
work (Singleton, 2019). The mixed MNL predictions for 2045 (Section 7
and Section 8) ignore some supply-side considerations, including the
availability of residential units in different TAZs. This does not hold for
the 2045 network-level analysis results (Section 6) because those are
based on population and land-use characteristics for 2045 incorporated
in the TRM from the Triangle CommunityViz 2.0. Differences in vehicle
operating cost are not affecting network performance due to the current
setting of the Triangle Regional Model that does not allow the user to set
a different vehicle operating cost by socioeconomic strata. This limita­
tion poses a minimal impact to our results for the following reasons. The
vehicle operating cost is a parameter only in the mode choice step of the
Triangle Regional Model, and 97% of the households in our sample
(households with commute trips within the Triangle Region) commute
to work by car. This is representative of the population because about
2% of the people in the Triangle Region commute to work by transit (Liu
and Bardaka, 2021). The proportion is even smaller for higher income
households who would have access to electric AVs or CAVs (and
therefore lower operating costs) based on our scenarios. Although a
lower vehicle operating cost for higher income households will not
essentially affect mode choice, it does play an important role in resi­
dential location choice where the relevant scenarios on fuel type are
presented and discussed. Additionally, our study does not consider any
further changes in the transportation network performance after the
estimated population shifts take place. Accounting for such effects until
an equilibrium is met could have revealed more moderate population
shifts (for example, an increase in the suburban population could in­
crease traffic in the region and lead to households moving closer to
M.M. Hasnat et al.
work). However, we would not expect substantially different results
from such an analysis because the estimated population shifts are rela­
tively small. Finally, we only focus on residential land uses and we do
not explore changes in commercial space, parking, and other land uses
that are expected to be affected by the widespread adoption of AVs and
CAVs. It is possible that some of these potential changes, such as a lower
need of parking space in downtown areas, can open up space for new
residential areas and may counterbalance some of the predicted subur­
banization trends.
Funding
This work was supported by the North Carolina Department of
Transportation. The contents of this article reflect the views of the au­
thors and do not necessarily reflect the official views or policies of either
the North Carolina Department of Transportation or the Federal High­
way Administration at the time of publication.
CRediT authorship contribution statement
Md Mehedi Hasnat: Conceptualization, Methodology, Formal
analysis, Writing - original draft. Eleni Bardaka: Supervision, Concep­
tualization, Methodology, Writing - original draft, Writing - review &
editing, Funding acquisition, Project administration. M. Shoaib
Samandar: Methodology, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Acknowledgement
The authors are thankful to Jamal Alavi, Director of the Trans­
portation Planning Division at the North Carolina Department of
Transportation (NCDOT), for his valuable insights into this research. The
authors are also grateful to Joseph Huegy, Travel Behavior Modeling
Group, Institute for Transportation Research and Education at North
Carolina State University (NCSU) for his significant assistance with the
Triangle Regional Model. The authors would like to thank Professors
Nagui Rouphail (NCSU), George List (NCSU), Billy Williams (NCSU),
and Missy Cummings (Duke University) for providing valuable input
and region-specific insights to this study.
References
AAA, 2018. Your Driving Costs: How Much Are You Really Paying to Drive? Technical
Report American Automobile Association. URL: https://exchange.aaa.com/wp-
content/uploads/2018/09/18-0090_2018-Your-Driving-Costs-Brochure_FNL-Lo-5-2.
pdf.
Adebisi, A., Liu, Y., Schroeder, B., Ma, J., Cesme, B., Jia, A., Morgan, A., 2020.
Developing Highway Capacity Manual (HCM) Capacity Adjustment Factors (CAF)
for Connected and Automated Traffic on Freeway Segments. Washington D.C, In
Transportation Research Board.
Akbari, S., Hasnine, M.S., Papaioannou, E., Bernardino, A., Habib, K.N., 2020. Home
relocation and mobility tool ownership: Econometric investigations in the context of
rising fuel prices in the Greater Toronto Area. Travel Behav. Soc. 19, 8–19. https://
doi.org/10.1016/j.tbs.2019.10.005.
Anderson, J.M., Nidhi, K., Stanley, K.D., Sorensen, P., Samaras, C., Oluwatola, O.A.,
2014. Autonomous vehicle technology: A guide for policymakers. Rand Corporation.
Bansal, P., Kockelman, K.M., 2018. Are we ready to embrace connected and self-driving
vehicles? A case study of Texans. Transportation 45. https://doi.org/10.1007/
s11116-016-9745-z.
Bardaka, E., Frey, C., Hajbabaie, A., List, G., Rouphail, N., Williams, B., & Cummings, M.
(2021). Impacts of autonomous vehicle technology on transportation systems. URL:
https://connect.ncdot.gov/projects/research/RNAProjDocs/RP2019-11.
Bardaka, E., Hasnat, M., 2021. Impacts of connected and autonomous vehicles on
transportation demand and land use vol. 1. URL: https://connect.ncdot.gov/
projects/research/RNAProjDocs/Volume.
Berglund, E., Monroe, J., Ahmed, I., Noghabaei, M., Do, J., Pesantez, J., Fasaee, M.K.,
Bardaka, E., Han, K., Proestos, G., Levis, J., 2020. Smart infrastructure: A vision for
19
Travel Behaviour and Society 32 (2023) 100570
the role of the civil engineering profession in smart cities. J. Infrastruct. Syst. 26
https://doi.org/10.1061/(ASCE)IS.1943-555X.0000549.
Bhat, C., Guo, J., 2004. A mixed spatially correlated logit model: Formulation and
application to residential choice modeling. Transp. Res. Part B: Methodol. 38,
147–168. https://doi.org/10.1016/S0191-2615(03)00005-5.
Bholowalia, P., & Kumar, A. (2014). EBK-Means: A Clustering Technique based on Elbow
Method and K-Means in WSN. Technical Report 9.
Bierstedt, J., Gooze, A., Gray, C., Peterman, J., Raykin, L., & Walters, J. (2014). Effects of
Next-Generation Vehicles on Travel Demand and Highway Capacity. In FP Think
Working Group 8 (pp. 1–10).
Bösch, P.M., Becker, F., Becker, H., Axhausen, K.W., 2018. Cost-based analysis of
autonomous mobility services. Transp. Policy 64, 76–91. https://doi.org/10.1016/J.
TRANPOL.2017.09.005. URL: https://www.sciencedirect.com/science/article/pii/
S0967070X17300811.
Brier, G.W., 1950. Verification of forecasts expressed in terms of probability. Mon.
Weather Rev. 78, 1–3. URL: https://books.google.com/books?
hl=en&lr=&id=jnbpAAAAMAAJ&oi=fnd&pg=RA1-PA1&dq=Verification+of+
forecasts+expressed+in+terms+of+
probability&ots=0X0X2IqNzK&sig=bX5k1L7ZRQ1PG_
6Mx7D5dmfi2x8#v=onepage&q=Verification.
Bruns, A., Matthes, G., 2019. Moving into and within cities – Interactions of residential
change and the travel behavior and implications for integrated land use and
transport planning strategies. Travel Behav. Soc. 17, 46–61. https://doi.org/
10.1016/j.tbs.2019.06.002.
Carrese, S., Nigro, M., Patella, S.M., & Toniolo, E. (2019). A preliminary study of the
potential impact of autonomous vehicles on residential location in Rome. Res.
Transp. Econ., (pp. 1–7). URL: doi: 10.1016/j.retrec.2019.02.005. DOI: 10.1016/j.
retrec.2019.02.005.
Chang, T.-H., Lai, I.-S., 1997. Analysis of characteristics of mixed traffic flow of autopilot
vehicles and manual vehicles. Transp. Res. Part C: Emerg. Technol. 5, 333–348.
https://doi.org/10.1016/S0968-090X(97)00020-X.
Childress, S., Nichols, B., Charlton, B., Coe, S., 2015. Using an Activity-based Model to
Explore Possible Impacts of Automated Vehicles. Transp. Res. Rec.: J. Transp. Res.
Board 99–106. URL: http://psrc.github.io/attachments/2014/TRB-2015-
Automated-Vehicles-Rev2.pdf.
Choudhary, A., Sharma, P., Singh, M., 2016. Improving K-means through better
initialization and normalization. In: International Conference on Advances in
Computing, Communications and Informatics (ICACCI). IEEE, pp. 2415–2419.
CommunityViz (2018). Triangle CommunityViz 2.0 Technical Overview. Technical
Report.
Cordera, R., Nogués, S., González-González, E., & Moura, J.L. (2021). Modeling the
Impacts of Autonomous Vehicles on Land Use Using a LUTI Model. Sustainability,
13, 1608. URL:https://www.mdpi.com/2071-1050/13/4/1608/htmhttps://www.
mdpi.com/2071-1050/13/4/1608. DOI: 10.3390/SU13041608.
Danaf, M., Becker, F., Song, X., Atasoy, B., Ben-Akiva, M., 2019. Online discrete choice
models: Applications in personalized recommendations. Decis. Support Syst. 119,
35–45. https://doi.org/10.1016/j.dss.2019.02.003. URL: 10.1016/j.
dss.2019.02.003.
Dubljevic, V., List, G., Milojevich, J., Ajmeri, N., Bauer, W.A., Singh, M.P., Bardaka, E.,
Birkland, T.A., Edwards, C.H., Mayer, R.C., Muntean, I., Powers, T.M., Rakha, H.A.,
Ricks, V.A., Samandar, M.S., 2021. Toward a rational and ethical sociotechnical
system of autonomous vehicles: A novel application of multi-criteria decision
analysis. PLoS ONE 16. https://doi.org/10.1371/journal.pone.0256224.
Durand, A., Harms, L., Hoogendoorn-Lanser, S., Zijlstra, T., 2018. Mobility-as-a-Service
and changes in travel preferences and travel behaviour: a literature review.
Netherlands Institute for Transport Policy Analysis. Technical Report KiM.
Eliasson, J., 2010. The influence of accessibility on residential location. In: Residential
Location Choice. Springer International Publishing, pp. 137–164, 10.1007/978-3-
642-12788-5_7.
Fagnant, D.J., Kockelman, K.M., 2014. The travel and environmental implications of
shared autonomous vehicles, using agent-based model scenarios. Transp. Res. Part C:
Emerg. Technol. 40, 1–13. https://doi.org/10.1016/J.TRC.2013.12.001. URL:
https://www.sciencedirect.com/science/article/pii/S0968090X13002581.
Gehrke, S.R., Singleton, P.A., Clifton, K.J., 2019. Understanding stated neighborhood
preferences: The roles of lifecycle stage, mobility style, and lifestyle aspirations.
Travel Behav. Soc. 17, 62–71. https://doi.org/10.1016/j.tbs.2019.07.001.
Gelauff, G., Ossokina, I., Teulings, C., 2019. Spatial and welfare effects of automated
driving: Will cities grow, decline or both? Transp. Res. Part A: Policy Practice 121,
277–294. https://doi.org/10.1016/j.tra.2019.01.013.
General Motors (2022). Why All AVs Should be EVs. URL:https://www.gm.com/stories/
all-avs-should-be-evs.
Gneiting, T., Raftery, A.E., 2007. Strictly Proper Scoring Rules, Prediction, and
Estimation. J. Am. Stat. Assoc. 102, 359–378. https://doi.org/10.1198/
016214506000001437. URL: https://www.tandfonline.com/action/
journalInformation?journalCode=uasa20.
Greenacre, M., Primicerio, R., 2013. Multivariate Analysis of Ecological Data. BBVA
Foundation.
Guo, J.Y., Bhat, C.R., 2007. Operationalizing the concept of neighborhood: Application
to residential location choice analysis. J. Transp. Geogr. 15, 31–45. https://doi.org/
10.1016/j.jtrangeo.2005.11.001.
Guo, Y., Peeta, S., 2020. Impacts of personalized accessibility information on residential
location choice and travel behavior. Travel Behav. Soc. 19, 99–111. https://doi.org/
10.1016/j.tbs.2019.12.007.
Hajbabaie, A., Rouphail, N., Williams, B., Samandar, S., Das, T., Tajalli, M., 2021.
Impacts of connected and autonomous vehicles on transportation capacity - vol. 2.
URL: https://connect.ncdot.gov/projects/research/RNAProjDocs/Volume.
M.M. Hasnat et al.
Han, J., Kamber, M., Pei, J., 2012. Data Mining. In: Data Mining, third ed. Elsevier,
pp. 1–38. 10.1016/b978-0-12-381479-1.00001-0.
Haque, M.B., Choudhury, C., Hess, S., 2019. Modelling residential location choices with
implicit availability of alternatives. J. Transp. Land Use 12, 597–618. https://doi.
org/10.5198/JTLU.2019.1450. URL:https://www.jtlu.org/index.php/jtlu/article/
view/1450.
Hasnat, M.M., Bardaka, E., Samandar, M.S., Rouphail, N.M., List, G.F., Williams, B.,
2021. Impacts of Private Autonomous and Connected Vehicles on Transportation
Network Demand in the Triangle Region, North Carolina. J. Urban Plann. Develop.
147, 13–21. https://doi.org/10.1061/(ASCE)UP.1943-5444.0000649.
Hjorthol, R. (2013). Attitudes, ownership and use of Electric Vehicles-a review of
literature. URL: www.toi.no.
Hossain, M.S., Fatmi, M.R., 2022. Modelling the adoption of autonomous vehicle: How
historical experience inform the future preference. Travel Behav. Soc. 26, 57–66.
https://doi.org/10.1016/J.TBS.2021.09.003.
Kim, S.H., Mokhtarian, P.L., Circella, G., 2020. Will autonomous vehicles change
residential location and vehicle ownership? Glimpses from Georgia. Transp. Res.
Part D: Transp. Environ. 82, 102291 https://doi.org/10.1016/j.trd.2020.102291.
Kroesen, M., 2019. Residential self-selection and the reverse causation hypothesis:
Assessing the endogeneity of stated reasons for residential choice. Travel Behav. Soc.
16, 108–117. https://doi.org/10.1016/j.tbs.2019.05.002. URL: 10.1016/j.
tbs.2019.05.002.
Krueger, R., Bierlaire, M., Daziano, R.A., Rashidi, T.H., Bansal, P., 2021. Evaluating the
predictive abilities of mixed logit models with unobserved inter- and intra-individual
heterogeneity. J. Choice Modell. 41 https://doi.org/10.1016/j.jocm.2021.100323.
Krueger, R., Rashidi, T.H., Dixit, V.V., 2019. Autonomous driving and residential
location preferences: Evidence from a stated choice survey. Transp. Res. Part C:
Emerg. Technol. 108, 255–268. https://doi.org/10.1016/j.trc.2019.09.018.
Larson, W., Zhao, W., 2020. Self-driving cars and the city: Effects on sprawl, energy
consumption, and housing affordability. Regional Sci. Urban Econ. 81, 103484
https://doi.org/10.1016/j.regsciurbeco.2019.103484.
Likas, A., Vlassis, N., Verbeek, J., 2003. The global k-means clustering algorithm. Pattern
Recogn. 32, 451–461.
Litman, T. (2020). Autonomous Vehicle Implementation Predictions. Implications for
Transport Planning. Technical Report Victoria Transport Policy Institute. URL: www.
vtpi.org.
Liu, C., Bardaka, E., 2021. The suburbanization of poverty and changes in access to
public transportation in the Triangle Region, nc. J. Transp. Geogr. 90 https://doi.
org/10.1016/j.jtrangeo.2020.102930.
Liu, P., Xu, S.X., Ong, G.P., Tian, Q., Ma, S., 2021. Effect of autonomous vehicles on
travel and urban characteristics. Transp. Res. Part B: Methodol. 153, 128–148.
https://doi.org/10.1016/J.TRB.2021.08.014.
Lloyd, S.P., 1982. Least Squares Quantization in PCM. IEEE Trans. Inform. Theory 28,
129–137.
McFadden, D. (1978). Modeling the Choice of Residential Location. Transp. Res. Rec.,
(pp. 72–77).
McFadden, D., Train, K., 2000. Mixed MNL Models for Discrete Response. J. Appl.
Econometr. 15, 447–470.
McKinsey & Company (2021). Why the future automotive future is electric. URL:https
://www.mckinsey.com/industries/automotive-and-assembly/our-insights/why-the-
automotive-future-is-electric.
Meyer, J., Becker, H., Osch, P.M.B., Axhausen, K.W., 2017. Autonomous vehicles: The
next jump in accessibilities? Res. Transp. Econ. 62, 80–91. https://doi.org/10.1016/
j.retrec.2017.03.005. URL: 10.1016/j.retrec.2017.03.005.
Millard-Ball, A., 2019. The autonomous vehicle parking problem. Transp. Policy 75,
99–108. https://doi.org/10.1016/j.tranpol.2019.01.003.
Moore, M.A., Lavieri, P.S., Dias, F.F., Bhat, C.R., 2020. On investigating the potential
effects of private autonomous vehicle use on home/work relocations and commute
times. Transp. Res. Part C: Emerg. Technol. 110, 166–185. https://doi.org/10.1016/
j.trc.2019.11.013. URL: 10.1016/j.trc.2019.11.013.
Mullaney, T. (2020). Tesla and the science behind low-cost, next-gen million-mile EV
battery. URL:https://www.cnbc.com/2020/06/30/tesla-and-the-science-of-low-
cost-next-gen-ev-million-mile-battery.html.
NC General Assembly (2017). North Carolina House Bill 469 Session Law 2017–166.
Technical Report General Assembly of North Carolina. URL:https://www.ncleg.
gov/BillLookup/2017/h469.
Oh, S., Seshadri, R., Lima Azevedo, C., Kumar, N., Basak, K., Ben-Akiva, M., 2020.
Assessing the impacts of automated mobility-on-demand through agent-based
simulation: A study of Singapore. Transp. Res. Part A: Policy Practice 138, 367–388.
https://doi.org/10.1016/j.tra.2020.06.004. URL: 10.1016/j.tra.2020.06.004.
Petersen, E., Vovsha, P., Petersen, E., Europe, R., 2006. Intrahousehold Car-Type Choice
for Different Travel Needs. Transp. Res. Rec.: J. Transp. Res. Board 1985, 207–219.
https://doi.org/10.1177/0361198106198500123.
Pinjari, A.R., Pendyala, R.M., Bhat, C.R., Waddell, P.A., 2011. Modeling the choice
continuum: An integrated model of residential location, auto ownership, bicycle
ownership, and commute tour mode choice decisions. Transportation 38, 933–958.
https://doi.org/10.1007/s11116-011-9360-y.
Rokach, L., Maimon, O., 2008. Data Mining with Decision Trees: Theory and
Applications, 2nd ed. World Scientific Publishing, Singapore.
RSG (2016). 2016 Triangle Region Household Travel Survey. Technical Report.
SAE (2021). SAE Levels of Driving Automation - Refined for Clarity and International
Audience. URL:https://www.sae.org/blog/sae-j3016-update.
20
Travel Behaviour and Society 32 (2023) 100570
Schirmer, P.M., Van Eggermond, M.A., & Axhausen, K.W. (2014). The role of location in
residential location choice models: a review of literature. Journal of Transport and
Land Use, 7, 3. URL:https://www.jtlu.org/index.php/jtlu/article/view/740.
10.5198/jtlu.v7i2.740.
Shen, Y., Zhang, H., & Zhao, J. (2017). Embedding Autonomous Vehicle Sharing in
Public Transit System: An Example of Last-Mile Problem. In 96th Annual Meeting of
Transportation Research Board.
Shladover, S.E., Su, D., Lu, X.-Y., 2012. Impacts of Cooperative Adaptive Cruise Control
on Freeway Traffic Flow. Transp. Res. Rec.: J. Transp. Res. Board 2324, 63–70.
https://doi.org/10.3141/2324-08.
Singleton, P.A., 2019. Discussing the Positive utilities of autonomous vehicles: will
travelers really use their time productively? Transport Reviews 39, 50–65. https://
doi.org/10.1080/01441647.2018.1470584.
Stantec (2015). CommunityViz Land Suitability Factor Statistical Analysis for the
Triangle Regional Model. Technical Report DCHC MPO, CAMPO, TJCOG.
Stephens, T.S., Gonder, J., Chen, Y., Lin, Z., Liu, C., Gohlke, D., 2016. Estimated Bounds
and Important Factors for Fuel Use and Consumer Costs of Connected and
Automated Vehicles. Technical Report National Renewable Energy Laboratory. URL:
www.nrel.gov/publications.
Tientrakool, P., Ho, Y.-C., Maxemchuk, N.F., 2011. Highway Capacity Benefits from
Using Vehicle-to-Vehicle Communication and Sensors for Collision Avoidance. In:
2011 IEEE Vehicular Technology Conference (VTC Fall). IEEE, pp. 1–5. https://doi.
org/10.1109/VETECF.2011.6093130.
Train, K., 2009. Discrete Choice Method With Simulation, (2nd ed.). Cambridge
University Press. 10.1017/CBO9780511753930.
Truong, L.T., De Gruyter, C., Currie, G., Delbosc, A., 2017. Estimating the trip generation
impacts of autonomous vehicles on car travel in Victoria, Australia. Transportation
44, 1279–1292. https://doi.org/10.1007/s11116-017-9802-2.
Uhlemann, E., 2015. Introducing Connected Vehicles [Connected Vehicles]. IEEE Veh.
Technol. Mag. URL: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=
7047294&casa_token=HgWMPEOL2JwAAAAA:oZrA0ujo4gVpor0xojZrit5N66B
kdweBdOJyUGE4ZW9D4nzVIQYJHLlDOb8lVkuy9xRGl7UxRA&tag=1.
US Census Bureau (2019). American Community Survey 2014–2018 5-Year Estimates.
URL:https://www.census.gov/newsroom/press-releases/2019/acs-5-year.html.
Vyas, G., Famili, P., Peter Vovsha, Â., Fay, D., Kulshrestha, A., Giaimo, G., Anderson, R.,
2019. Incorporating features of autonomous vehicles in activity-based travel demand
model for Columbus, OH. Transportation 46, 2081–2102. https://doi.org/10.1007/
s11116-019-10030-w. URL: 10.1007/s11116-019-10030-w.
Waddell, P., 1992. A multinomial logit model of race and urban structures. Urban
Geography 13, 127–141. https://doi.org/10.2747/0272-3638.13.2.127. URL:
https://www.tandfonline.com/doi/abs/10.2747/0272-3638.13.2.127.
Waddell, P. (2002). UrbanSim: Modeling Urban Development for Land Use,
Transportation and Environmental Planning. Journal of the American planning
association, 68(3), pp. 297–314, 68, 297–314. URL: www.urbansim.org.
Wadud, Z., Mackenzie, D., Leiby, P., 2016. Help or hindrance? The travel, energy and
carbon impacts of highly automated vehicles. Transp. Res. Part A: Policy Practice 86,
1–18. https://doi.org/10.1016/j.tra.2015.12.001.
Wali, B., Santi, P., Ratti, C., 2021. Modeling consumer affinity towards adopting partially
and fully automated vehicles-The role of preference heterogeneity at different
geographic levels. Transp. Res. Part C 129, 103276. https://doi.org/10.1016/j.
trc.2021.103276. URL: 10.1016/j.trc.2021.103276.
Wu, J., 2012. Advances in K-means clustering: a data mining thinking. Springer Science
& Business Media.
Xiao, L., Wang, M., & Arem, B. v. (2017). Realistic Car-Following Models for Microscopic
Simulation of Adaptive and Cooperative Adaptive Cruise Control Vehicles: doi:
10.3141/2623-01, 2623, 1–9. URL:https://journals.sagepub.com/doi/abs/10.3141/
2623-01. DOI: 10.3141/2623-01.
Xiao, L., Wang, M., Schakel, W., van Arem, B., 2018. Unravelling effects of cooperative
adaptive cruise control deactivation on traffic flow characteristics at merging
bottlenecks. Transp. Res. Part C: Emerging Technol. 96, 380–397. https://doi.org/
10.1016/J.TRC.2018.10.008.
Xie, Y., Zhang, Y., Prakash Akkinepally, A., Ben-Akiva, M., 2020. TRR Personalized
Choice Model for Managed Lane Travel Behavior. Transp. Res. Rec. 2020, 442–455.
https://doi.org/10.1177/0361198120923355.
Yan, X., 2020. Evaluating household residential preferences for walkability and
accessibility across three U.S. regions. Transp. Res. Part D: Transp. Environ. 80,
102255 https://doi.org/10.1016/J.TRD.2020.102255.
Zakharenko, R., 2016. Self-driving cars will change cities. Regional Sci. Urban Econ. 61,
26–37. https://doi.org/10.1016/j.regsciurbeco.2016.09.003.
Zhang, W., Guhathakurta, S., 2018. Residential Location Choice in the Era of Shared
Autonomous Vehicles. J. Plann. Educ. Res. https://doi.org/10.1177/
0739456X18776062.
Zhang, W., Guhathakurta, S., Khalil, E.B., 2018. The impact of private autonomous
vehicles on vehicle ownership and unoccupied VMT generation. Transp. Res. Part C:
Emerg. Technol. 90, 156–165. https://doi.org/10.1016/J.TRC.2018.03.005.
Zhao, Y., Kockelman, K.M., 2018. Anticipating the Regional Impacts of Connected and
Automated Vehicle Travel in Austin, Texas. J. Urban Plann. Develop. 144. URL:
https://ascelibrary.org/doi/pdf/10.1061/ 10.1061/(ASCE)UP.1943.
Zhou, B.B., Kockelman, K.M., 2009. Microsimulation of Residential Land Development
and Household Location Choices. Transp. Res. Rec.: J. Transp. Res. Board 2077,
106–112. https://doi.org/10.3141/2077-14.