Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Transportation Research Part A 126 (2019) 342–359

Contents lists available at ScienceDirect

Transportation Research Part A


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

Lifestyles, residential location, and transport mode use: A


T
hierarchical latent class choice model

Ali Ardeshiria, , Akshay Vijb
a
Research Centre for Integrated Transport Innovation, School of Civil and Environmental Engineering, UNSW Sydney, NSW 2052, Australia
b
Institute for Choice, University of South Australia, WL5-41, City West Campus, Adelaide, SA 5000, Australia

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

Keywords: This study develops a hierarchical latent class choice model that captures the concurrent influ-
Lifestyle ence of lifestyles on household residential neighbourhood location and individual transport mode
Residential location choice use decisions. The model is empirically evaluated using data from the 2010–12 California
Transport mode choice Household Travel Survey. The model identifies six household-level classes that differ in terms of
Hierarchical latent class
their preferences for different neighbourhood attributes when deciding where to live and their
Choice model
household characteristics. Coincidentally, the model also identifies six individual-level classes
that differ in terms of the travel modes that they consider when deciding how to travel, their
sensitivity to different level-of-service attributes, and their individual characteristics. Household
preferences for neighbourhood types and individual preferences for travel modes show expected
patterns of correlation. In general, households that prefer to live in suburban neighbourhoods are
more likely to consist of individuals that are car-dependent, and households that prefer to live in
inner-city neighbourhoods are more likely to consist of individuals that are multimodal.
However, our analysis also reveals interesting patterns of deviation. For example, high-income
migrant households and median-income white households display strong preferences for sub-
urban neighbourhoods, but individuals belonging to these households also have a high likelihood
of being multimodal, with a strong preference for bicycling. We discuss how these patterns of
correlation can be used to inform transport and land use policy in novel ways.

1. Introduction

Issues such as urban sprawl, congestion, oil dependence, climate change and public health, are prompting urban and transpor-
tation planners to turn to land use and urban design to rein in automobile use. As a result, increased interest in jobs-housing balance
(Giuliano, 1991; Cervero and Duncan, 2006), smart growth (Handy, 2005; Cervero and Duncan, 2006), compact cities (Stead et al.,
2000), transit oriented developments (Cervero, 1994; Bernick and Cervero, 1997), and new urbanism (Lund, 2003; Ardeshiri and
Ardeshiri, 2011) have spawned concepts for researchers who diligently probe and dissect the many ways in which urban form,
neighbourhood design, and the overall physical make-up of cities and regions interact with how people travel. The relationship
between urban form and sustainability is currently one of the most hotly debated issues on the international planning agenda. The
way that cities should be developed in the future, and the effect that their form can have on resource depletion and social and
economic sustainability, are central to this debate (Ardeshiri and Ardeshiri, 2011).
One of the implicit beliefs in this effort is that the right land use policies will, in fact, help to reduce automobile use and increase


Corresponding author.
E-mail address: A.Ardeshiri@unsw.edu.au (A. Ardeshiri).

https://doi.org/10.1016/j.tra.2019.06.016
Received 22 September 2018; Received in revised form 3 May 2019; Accepted 27 June 2019
Available online 06 July 2019
0965-8564/ © 2019 Elsevier Ltd. All rights reserved.
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

the use of alternative modes of transportation (Handy, 1996). While the empirical evidence indicates a strong correlation between
neighbourhood location decisions and transport mode choices, closer examination suggests that the relationships are more complex,
and the source of the correlation is not as clear as it may, at first, seem (for an excellent synthesis of the vast literature on the subject,
the reader is referred to Cao et al., 2009). Many studies have concluded that land use and the built environment indicators such as
density, land-use diversity, design, accessibility and distance to transit, etc., influence the travel behaviour in statistically significant
ways. Others have argued that at least part of the observed correlation between household neighbourhood location decisions and
individual travel mode choice behaviours can be explained by self-selection effects, where individuals who prefer particular modes of
transport may deliberately choose to live in neighbourhoods conducive to the use of these same modes.
For example, a number of studies have looked specifically at how per capita vehicle miles travelled (VMT) vary between different
neighbourhood types, and to what extent can these differences be attributed to differences in the built environment itself, and to what
extent are they determined by residential self-selection effects. Zhou and Kockelman (2008) find that 58% to 90% of the difference in
VMT between urban and suburban neighbourhoods in Austin, Texas can be explained by differences in the built environment, the
balance being due to self-selection. Bhat et al. (2009) report that 87% of the difference in VMT between conventional and neo-
urbanist neighbourhoods in the San Francisco Bay Area can be explained by differences in the built environment, and the remainder
to self-selection effects. Similarly, Cao et al. (2010) find that 80% of the difference in VMT between exurban and inner-suburban
neighbourhoods in Raleigh, NC could be ascribed to differences in the built environment, with the balance being due to self-selection
effects.
Findings such as these indicate that there is likely great heterogeneity within any population. Some individuals might be re-
sponsive to changes in the built environment in their travel mode choice behaviours; others may be very set in their ways and
immune to such changes. This study uses the construct of lifestyles to explore how preferences for different neighbourhood types and
transport modes vary across the population in systematic ways, in an attempt to lend structure and meaning to the underlying
covariance between observable patterns of land use and travel behaviour.
In doing so, this study also contributes to the growing body of literature on lifestyle effects in the context of different land use and
travel behaviours. In particular, several studies have examined the relationship between lifestyle and residential location choice (Ben-
Akiva and Lerman, 1985; Bagley and Mokhtarian, 1999; Aeroe, 2001; Waddell, 2001; Walker and Li, 2007); and separately, the
relationship between lifestyles and travel behaviour (Salomon and Ben-Akiva, 1983; Simma and Axhausen, 2001; Diana and
Mokhtarian, 2009; Kitamura, 2009; Ohnmacht et al., 2009; Van Acker et al., 2010; Kuhnimhof et al., 2012; Vij et al., 2013; Xiong
et al., 2015; Prato et al., 2016). However, these studies have examined the influence of lifestyles on different dimensions of trans-
portation and land use behaviours in isolation from each other.
In this study we propose an integrated approach through the development of a hierarchical1 latent class choice model (HLCCM)
that recognizes the simultaneous influence of lifestyles on both household neighbourhood location and individual transport mode
use. The benefits of the proposed framework are three-fold: (1) it provides a basis for explicitly integrating travel demand and land
use analysis through the construct of lifestyles; (2) it offers a deeper understanding of the decision-making process underlying the
complex relationship between these inter-related decisions; and (3) it contributes fresh insights to planners and policy-makers on how
they can intervene in the decision-making process to produce societally beneficial outcomes.
The remainder of the paper is constituted as follows: Section 2 provides an overview of the literature and background studies
related to land use and travel behaviour, lifestyle and modality style and models of group decision-making. Section 3 describe the
data used by this study. Section 4 introduces the econometric framework that we use for our analysis; and Section 5 presents the
estimation results. Finally, Section 6 concludes with a summary of our key findings and a discussion on directions for future research.

2. Literature review

The relationship between the built environment and travel behaviour has been the subject of numerous academic enquiries. We
will not review this extensive body of literature; numerous reviews have been written already, and the interested reader is pointed to,
among others, Litman (2005), Boarnet (2011) and Cao et al. (2009). By and large, there is consensus that built environment attri-
butes, such as density and diversity, do indeed impact travel behaviours; disagreement is usually limited to quantifying the salience of
this effect in relation to other determinants of travel behaviour.
We focus our attention instead on an ongoing and parallel stream of research that has examined the influence of more subjective
determinants on travel behaviour, such as habits, attitudes and lifestyle. For example, it has been argued that congestion is not a
problem in and of itself, but merely a symptom of the true problem, which is the “lifestyle” selection of residents (Kitamura, 2009).
The term ‘‘lifestyle’’ as used in the literature has two meanings: (a) activity and time use patterns and (b) values and behavioural
orientation. These two are interrelated, but a critical difference exists: lifestyle as activity patterns may change as an individual
adapts to a change in the environment, whereas lifestyle as orientation is one that the individual attempts to maintain by modifying
behavioural patterns and adapting to the change. Change in lifestyle as orientation takes place in the long term through changes in
values, attitudes, and preferences.
As mentioned previously, multiple studies have examined the influence of lifestyles on household preferences for residential
location (Walker and Li, 2007; Scheiner, 2010; Pisman et al., 2011; Smith and Olaru, 2013; Ardeshiri, 2014; Zhang, 2014; Liao et al.,
2015; Ardeshiri et al., 2016, 2018). By and large, these studies report strong and statistically significant relationships between

1
Throughout the paper, the word ‘hierarchical’ is not used in the Bayesian sense of the word.

343
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

lifestyle-related variables, such as income and household structure, and household preferences for residential location. For example,
most studies find high-income households are likely to prefer suburban low-density neighbourhoods, and low-income households are
likely to prefer urban compact developments. Similarly, young households with children are likely to prioritize larger lot and house
sizes over access to amenities, while the opposite generally holds true for older households without children.
Relatedly, multiple studies have also examined the influence of lifestyles on individual preferences for transport mode use (e.g.
Kuhnimhof and Gringmuth, 2009; Vij et al., 2013; Prato et al., 2016). In general, transport mode use is found to be strongly correlated
with lifecycle stage. For example, car use is found to increase initially with life cycle stage, as individuals enter the labour force and
single-person households transition to two-person households; it peaks when the household has children; and it starts to decline when
grown-up children leave the parental home, the parents retire, and old age sets in.
Within this broader literature, the construct of modality styles has been introduced to denote that aspect of individual and
household lifestyles constructed around the use of one or more travel modes (Vij et al., 2013). Previous studies have typically focused
on examining the influence of attitude and opinions, reflective of deeper differences in modality styles, on one dimension of be-
haviour, such as mode choice (Krueger et al., 2018), car ownership (Banerjee, 2011) or residential location (Walker and Li, 2007).
However, modality styles are hypothesized to influence all dimensions of travel and activity behaviour. For example, where a man
lives is not a very different question to ask from how many cars does a man need (Zhang, 2014). Therefore, we propose a more
integrated approach that explicitly captures the simultaneous influence of modality styles on multiple dimensions of behaviour.
Our work also builds upon past research on group decision-making within the context of individual and household travel and
activity behaviour. Many of the dimensions of travel and activity behaviour mentioned in previous paragraphs involve choices made
at the level of the household. The preferences of the household are the outcome of a process of negotiation between the individuals
that comprise the household and their respective preferences (Corfman, 1991; Lee and Beatty, 2002). In turn, the preferences of the
individuals themselves are shaped by the preferences of other household members, and are therefore some reflection of the pre-
ferences of the household as a whole (Davis, 1973; Menasco and Curry, 1989). A comprehensive travel demand model must recognize
the dialogue between individual and household preferences that underlies observable behaviour. However, unlike previous studies on
group decision-making that have been very detailed in their representation of intra-group dynamics (e.g. Rao and Steckel, 1991;
Arora and Allenby, 1999; Aribarg et al., 2002; Rose and Hensher, 2004; Hensher and Stanley, 2008), we won’t be as explicit. That
being said, we will be relying on findings from these studies to develop a simpler framework that captures the concurrent influence of
individual and household modality styles on different dimensions of travel and activity behaviour.

3. Methodological framework

Existing travel demand modelling practice relies on sequentially nested logit models for a hierarchical representation of different
choice dimensions. Lower dimensions, such as travel mode choice, are conditioned on purportedly higher dimensions, such as
residential location, creating a vertical chain of inter-connected nests that in their totality represent an individual’s travel and activity
behaviour. While such a representation is convenient from the standpoint of estimation, it overlooks the concurrent influence of
modality styles on all dimensions of travel and activity behaviour.
In developing a framework that captures the concurrent influence of lifestyles and modality styles on both household-level
decisions, such as neighbourhood location, and individual-level decisions, such as travel mode choices, we propose using a hier-
archical Latent Class Choice Model (HLCCM), illustrated in Fig. 1. Latent classes at the higher level represent household modality
styles and latent classes at the lower level represent individual modality styles. We argue that households may be decomposed into
discrete segments that differ in their predisposition towards different neighbourhood types and their sensitivity to different neigh-
bourhood attributes. A household’s modality style is hypothesized to be a function of observable sociodemographic variables, such as
income and household structure. Similarly, household members may be decomposed into discrete segments themselves that differ in
their consideration of different travel modes, and their sensitivity to different level-of-service attributes of the transportation system.
Individual modality styles are a function not only of the sociodemographic variables specific to the individual, such as gender, age,
employment, etc., but also of the modality style of the household as a whole. In this way, the model implicitly captures the influence
of intra-household interactions on household members’ individual preferences. Together, these differences in household and in-
dividual preferences are reflective of deeper differences in modality styles that concurrently influence neighbourhood choice and
travel mode choices over time. To the best of our knowledge, hierarchical LCCMs have not been used previously in the literature. In
this regard, this study makes a methodological contribution as well.
The proposed model framework comprises four components: (1) household class membership model; (2) residential neigh-
bourhood location choice model; (3) individual class membership model; and (4) travel mode choice models. Over subsequent
subsections, we describe each of these components in greater detail. We conclude with a description of how each of these sub-models
are brought together through a single likelihood function, and how the parameters corresponding to each of the sub-models are
estimated in practice.

3.1. Household class membership model

The household class membership model predicts the probability that household h belongs to household modality style r , and is
formulated as a multinomial logit model:

344
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Fig. 1. The proposed hierarchical latent class choice model of household residential neighbourhood choice and individual travel mode choices.

exp(z'h α r )
P(qhr = 1) = R
∑r'= 1 exp(z'h α s' ) (1)

where qhr equals one if household h belongs to modality style r , and zero otherwise; zh is a vector of characteristics of household h ,
such as income and household structure; α r is a vector of parameters associated with the household’s characteristics; and R denotes
the number of household modality styles in the sample population. Note that R must be determined exogenously, by estimating
models with different numbers of classes and comparing estimation results in terms of both model fit and behavioural interpretation.

3.2. Residential neighbourhood location choice model

Residential neighbourhood location is subsequently conditioned on the household’s modality style. Let uhj|r be the utility of
neighbourhood j for householdh , given that the household belongs to latent class r :

uhj|r = x 'jβr + εhj|r (2)

where x j is a vector of attributes of neighbourhood j; βr is a vector of class-specific parameters denoting sensitivities to each of these
attributes; and εhj|r is the stochastic component of the utility specification, assumed to be i.i.d. Extreme Value across households,
neighbourhoods and classes with location zero and scale one. We are implicitly assuming that the utilities of different neighbour-
hoods are not spatially correlated. Given the spatial nature of our data, and likely unobserved variables that are shared by neigh-
bourhoods in close proximity to each other, the assumption might be overly restrictive. However, many empirical studies of re-
sidential location choice have overlooked similar patterns of spatial correlation in their analysis as a matter of convenience (e.g.
McFadden, 1978, 1980; Bhat and Guo, 2007; Zhang and Guhathakurta, 2018). For these same reasons, a residential location sub-
model that explicitly accounts for potential sources of spatial correlation is deemed beyond the scope of the present study, and left as
an important direction for future research.
Assuming that all individuals are utility maximizers, the class-specific neighbourhood choice model may be formulated as follows:

exp(x 'jβr )
P(yhj=1|qhr = 1) =
∑ j' ∈ Jr exp(x ' ' βr )
j (3)

where yhj equals one if household h resides in neighbourhood j , and zero otherwise; and Jr denotes the set of all neighbourhood s
considered by households belonging to class r . In our empirical application, the choice set Jr is defined to consist of all 1577 census
tracts in our area of analysis. The estimation of choice models with such a large number of alternatives can prove to be computa-
tionally intractable. We operationalize the model by randomly sampling 49 non-chosen census tracts from the full choice set for each

345
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

household. If the choice sub-model is multinomial logit, as it is in our case, McFadden (1978) shows that the parameter estimates are
asymptotically unbiased for any random subset of the full choice set.
Heterogeneity in the decision-making process is captured by allowing the taste parameters βr . For example, some classes might be
more sensitive to land use variables, such as density and diversity, and others might base their decision on other variables, such as
quality of schooling or crime rate. Eq. (3) may be combined iteratively over all neighbourhoods in the set Jr to yield the following
conditional probability of observing the vector of neighbourhood location choices yh for household h :

f y (yh |qhr = 1) = ∏ [P(yhj=1|qhr = 1)]yhj


j' ∈ Jr (4)

3.3. Individual class membership model

The probability that individual n from household h has modality style s , conditional on the household belonging to modality style
r , is assumed to be multinomial logit, and can be expressed as:

exp(w 'hn γrs )


P(ghns=1|qhr = 1) = S
∑s'= 1 exp(w 'hn γ rs' ) (5)

where ghns equals one if individual n from household h belongs to modality style s , and zero otherwise; whn is a vector of char-
acteristics of the individual, such as age and gender; γrs is a vector of model parameters associated with the individual’s char-
acteristics; and S denotes the number of individual modality styles in the sample population. As was the case with the number of
household modality styles R , note that S too must be determined exogenously, by estimating models with different numbers of classes
and comparing estimation results in terms of both model fit and behavioural interpretation.
The interaction between household members is captured by conditioning the model parameters γrs on the modality style of the
household as a whole. In other words, individuals with the same set of demographic characteristics could potentially belong to very
different modality styles, depending on the household’s modality style. For example, young men belonging to households that are
collectively car-oriented may be more likely to be car-oriented themselves, and similarly, young men belonging to households that
are more multimodal may be more likely to be multimodal themselves. Differences in the probability of the same demographic
subgroup to belong to different individual modality styles, due to differences in household modality styles, can offer insight on the
influence of household interactions on individual lifestyles.

3.4. Travel mode choice model

Individual travel mode choices are conditioned on the individual’s modality style. Let vhndtk|s be the utility of travel mode k over
tour t for tour purpose d and individual n belonging to household h , given that the individual belongs to latent class s :

vhndtk|s = c'hndtk λ ds + ηhndtk|s (6)

where c hndtk is a vector of attributes of travel mode k over tour t for tour purpose d and individual n belonging to household h ; λ ds is a
vector of class-specific parameters denoting sensitivities to each of these attributes for tour purpose d and class s ; and ηhndtk|s is the
stochastic component of the utility specification, assumed to be i.i.d. Extreme Value across households, individuals, tour purposes,
tours, travel modes and classes with location zero and scale one. Assuming that all individuals are utility maximizers, the class-
specific travel mode choice model may be formulated as follows:

exp(c'hndtk λ s )
P(mhndtk=1|ghns = 1) =
∑k'∈ K ds exp(c' 'λ s ) (7)
hndtk

where mhndtk equals one if individual n from household h chooses travel mode k over tour t for tour purpose d , and zero otherwise;
and K ds denotes the set of all travel modes in the consideration set of individuals belonging to class s for tours with purpose d . We
specify separate class-specific mode choice models for mandatory and non-mandatory tours, as denoted by the subscript d . We
capture heterogeneity in the decision-making process by allowing both the taste parameters λ ds and the consideration set K ds to vary
across modality styles. For example, individuals belonging to a particular modality style might always drive, regardless of the level-
of-service of other travel modes, whereas individuals belonging to a different class might consider all travel modes but have a high
value of time that similarly predisposes them towards the car.
Eq. (7) may be combined iteratively over travel modes, tours and tour purposes to yield the following conditional probability of
observing the vector of choices mhn :
D T hnd
fm (mhn |ghns = 1) = ∏∏ ∏ [P(mhndtk=1|ghns = 1)]mhndtk
d = 1 t = 1 k'∈ K ds (8)

where D denotes the number of tour purposes, two in our case; and Thnd denotes the number of observed tours for individual n from
household h for tour purpose d .

346
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

3.5. The likelihood function and model estimation

Eq. (8) may be marginalized over the distribution of individual modality styles to yield the probability function of observing the
vector of choices mhn , conditional on the household’s modality style. This may subsequently be combined iteratively over all in-
dividuals belonging to the household, multiplied by Eq. (4), marginalized over the distribution of household modality styles, and
iteratively combined over all households in the sample population to yield the following likelihood function:
H R Nh S
L(y , m) = ∏ ∑ P(qhr = 1)f y (yh |qhr = 1) ∏ ∑ fm (mhn |ghns = 1)P(ghns=1|qhr = 1)
h=1 r=1 n=1 s=1 (9)

where H denotes the number of households in the sample population; and Nh denotes the number of individuals that belong to
household h .
The unknown model parameters α , β , γ and λ may be estimated by maximizing the likelihood function given by Eq. (9). However,
simultaneous estimation of the full model proved to be computationally challenging. The model reported in this paper was ultimately
estimated sequentially through the following four-step procedure: (1) estimate an HLCCM of household neighbourhood choice be-
haviour to determine R , the number of household modality styles, β , the vector of class-specific residential neighbourhood location
choice model parameters, and α , the vector of household modality style class membership model parameters; (2) estimate an HLCCM
of individual travel mode choice behaviour to determine S , the number of individual modality styles, and λ , the vector of class-
specific model choice model parameters; (3) use the estimates for α and β from the first step and λ from the second, and maximize Eq.
(9) for γ ; and (4) use the estimates for β from the first step, λ from the second step, and γ from he third step, and maximize Eq. (9) to
re-estimate α .
In our case, Step 1 was implemented in Python Biogeme (Bierlaire, 2016) using maximum likelihood estimation methods. Step 2
was implemented in Python using the Expectation-Maximization algorithm, which itself used an implementation of the BFGS al-
gorithm contained in the SciPy library (Jones et al., 2001) for the M-step of the algorithm. Steps 3 and 4 were implemented in Python
using the same BFGS algorithm directly.
The modelling approach is exploratory: the appropriate number of household and individual modality styles and the behaviour of
each modality style is determined based on a comparison between different model specifications across goodness-of-fit and beha-
vioural interpretation. As we’ll demonstrate in the subsequent case study, differences in taste parameters across the neighbourhood
and travel mode choice class-specific models, and how these differences are correlated through the individual and household
modality styles constructs, can offer insight on the extent to which households self-select into different built environments and the
extent to which the built environment influences travel behaviour.

4. Data

The data used for investigating the proposed hierarchical latent class choice model of household residential neighbourhood choice
and individual travel mode choices is drawn from the 2010–2012 California Household Travel Survey (CHTS). CHTS is a multi-modal
study of the demographic and travel behaviour characteristics of residents across the entire state of California. For more details about
the survey, the reader is referred to NuStats, LLC (2013). For the purposes of this study, we only looked at households and individuals
residing in the San Francisco Bay area (henceforward the Bay Area), which includes nine counties: Alameda, Contra Costa, Marin,
Napa, San Francisco, San Mateo, Santa Clara, Solano, and Sonoma.
In all, we have data from 8208 households. For each household, we know the census tract location of their current residence.
Secondary data related to the demographic characteristics of each of the 1577 census tracts in the Bay Area was extracted from the
2013 TIGER/Line Shapefiles and the 2009–2013 American Community Survey (ACS) 5-year estimates data. Supplemental in-
formation related to tract school quality and crime rate were also added to better understand residential location decisions. We used
the 2012 Academic Performance Index (API) score available from the California Department of Education website. The API is a single
number, ranging from a low of 200 to a high of 1000, which reflects a school’s, local educational agencies, or a student group’s
performance level, based on the results of state-wide assessments. Each school with its relevant API scores were georeferenced on the
2013 TIGER/Line Shapefiles, and for each 1577 census tracts an average API score from all available schools on the tract was
calculated. Crime data for each county were sourced from the county Sheriff crime office 2012 reports and maps. These crime data
were then normalised to a 10-point scale (1 representing a safe place and 10 representing an unsafe place) to represents safety scale
across all census tract. Table 1, provides an exhaustive list of variables used in the model with a description of how they have been
used in the model.
Each household member was requested to keep a travel diary over a one-day period, recording which activities were conducted
where, when, for how long, with whom and using what mode of travel. We processed this information into home-based tours. Tours
that included a work or school stop were subsequently classified as mandatory tours, and all other tours were classified as non-
mandatory. In all, we have data for 9762 mandatory tours and 17,292 non-mandatory tours made by 17,680 individuals from the
8208 households. For each tour, individuals are assumed to have access to the following five feasible travel modes: private vehicle,
private transit, public transit, bike, and walk. Private vehicle denotes cases where the individual used a motorized vehicle, owned
either by themselves or someone they know, as a driver or a passenger. Private transit refers to travel modes such as taxis, car share,
ride share, rental cars and private shuttles. Public transit includes modes such as buses, trains, ferries, etc. Each of these modes is
defined in terms of two attributes, namely travel times and costs, using network skims from SF MTC. Unfortunately, we are not able to

347
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Table 1
Variables used to model our residential neighbourhood choice and individual travel mode choices.
Tract profile Description

Density Number of residents and employees per square mile.


Diversity An entropy measure calculated using the proportions of the population that are residents or employees
belonging to one of fourteen different industries
Design Number of census blocks per square mile
Owen dwelling Proportion of owned dwelling
API Score The Academic Performance Index (API) is a single number on a scale of 200 (performing very bad) to 1000
(performing very well) that indicates how well students in a school or district performed on the previous
spring’s tests.
Crime score A scale representing crime ranging from 1 (safe place) to 10 (unsafe place).
Gini Score A measure of statistical used to measure inequality. A Gini coefficient of zero expresses perfect equality,
where all values are the same (for example, where everyone has the same income) and A Gini coefficient of
1 expresses maximal inequality among values.
Tract Population Average tract population.
Median Rent Median rent of a residential unit in the census tract.
Median Value Median value of a residential unit in the census tract.
Blacks Proportion of black residents in the census tract.
Immigrants (Asians) Proportion of Immigrant residents in the census tract.
Whites Proportion of white residents in the census tract.
Hispanics Proportion of Hispanic residents in the census tract.
Upper Class Proportion of upper-class residents in the census tract based on income and family size, as defined by the US
Census.
Upper Middle Class Proportion of upper-middle residents in the census tract based on income and family size, as defined by the
US Census.
Lower Middle Class Proportion of lower-middle class residents in the census tract based on income and family size, as defined
by the US Census.
Lower Class Proportion of lower-class residents in the census tract based on income and family size, as defined by the US
Census.
Below Poverty Line Proportion of the tract population living with income at or below the Federal poverty thresholds by family
size
Low Level of Education Proportion of the population with a degree equivalent to high school or below
Household variables Description
Number of vehicles Number of vehicle/s owned in the household.
Driver’s license holding An indicator of number of households that hold a driver’s license.
Number of bikes Number of bike/s owned in the household.
Single Family House A 0–1 dummy variable representing if the household residential type is a single-family house.
Residence Type: Building with 2–4 Apartments A 0–1 dummy variable representing if the household residential type is a building with 2–4 apartments.
Residence Type: Building with 5–19 Apartments A 0–1 dummy variable representing if the household residential type is a building with 5–19 apartments.
Residence Type: Building with 20 or more A 0–1 dummy variable representing if the household residential type is a building with more than 20
Apartments apartments.
Home Owner A 0–1 dummy variable indicating if the dwelling is owed by the household.
Annual household income The total annual household income.
Household size A measure representing the household size.
Residence tenure Representing residence tenure in year.
Number of workers An indicator of number of individual contributing to the total household income.

Household variables Description

Number of students An indicator of number of students in the household.


Unrelated household A 0–1 dummy variable indicating if household members are related to each other.
Pre-school kids A 0–1 dummy variable representing the presence of pre-school kids in the household.
School kids A 0–1 dummy variable representing the presence of school kids in the household.
Hispanic household A 0–1 dummy variable indicating if the household is Hispanic.
White house hold A 0–1 dummy variable indicating if the household is white.
Black household A 0–1 dummy variable indicating if the household is black.
American household A 0–1 dummy variable indicating if the household is native American.
Travel mode variables Description
Travel mode The travel modes examined for this study are car, private transit, public transit, bike and walk.
Travel time The travel time base on mode specific.
Travel cost The travel cost base on mode specific.
Individual variables Description
Male A 0–1 dummy variable indicating if the individual is male.
Married A 0–1 dummy variable indicating if the individual is married.
Parent A 0–1 dummy variable indicating if the individual is a parent.
Employed A 0–1 dummy variable indicating if the individual is employed.
Student A 0–1 dummy variable indicating if the individual is student.
Age An indicator representing the individual’s age.

348
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Table 2
Summary statistics for different specifications for the sub-model of individual modality styles and travel mode choices.
Model Parameters Log-likelihood AIC BIC

Two classes 37 −9784 19,641 19,941


Three classes 59 −9682 19,482 19,960
Four classes 75 −9619 19,388 19,996
Five classes 99 −9527 19,252 20,054
Six classes 127 −9477 19,207 20,236
Seven classes 155 −9460 19,229 20,485
Eight classes 173 −9446 19,239 20,641
Nine classes 197 −9442 19,279 20,875
Ten classes 218 −9441 19,318 21,084
Eleven classes 241 −9430 19,342 21,295

separate travel time into its constituent elements, such as in-vehicle time and access time, as this information was not available to us.
Note that a different version of the mode choice dataset has been used previously by Vij et al. (2017).

5. Estimation results

As described in Section 3, we used a four-step procedure to determine the final model specification. In determining this speci-
fication, we estimated numerous models and sub-models where we varied the utility specification, number of classes and choice set
assumptions. Coincidentally, based on a comparison across both statistical measures of fit, such as the Bayesian Information Criterion
(BIC) and the Akaike Information Criterion (AIC) listed in Tables 2 and 3, and behavioural interpretation, a six-class model was
selected as the preferred model specification for both the household-level residential neighbourhood choice model and the in-
dividual-level travel mode choice model. The case for six classes at the individual-level is easily defended, as this model has the
lowest AIC and second lowest BIC. We terminated the estimation algorithm at six classes at the household level due to numerical
difficulties estimating models with higher numbers of household classes. Of the models that we were able to successfully estimate, the
model with six household classes has the lowest AIC and BIC.
The final model has a large number of model parameters. Table 4 reports estimation results for the household-level class
membership model; Table 5 reports estimation results for the class-specific residential neighbourhood choice model; Table 6 reports
estimation results for the individual-level class membership model, which the reader should recall are conditioned on household class
membership; and Tables 7 and 8 report estimation results for the class-specific travel mode choice models for mandatory and non-
mandatory tours.
Section 5.1 describes the six household-level classes identified by the final model, and how they correlate with residential
neighbourhood preferences and behaviours. Similarly, Section 5.2 describes the six individual-level classes identified by the final
model, and how they correlate with travel mode choice preferences and behaviours. Finally, Section 5.3 discusses how these two
dimensions of behaviour correlate with each other, and what they mean for transport and land use policy.

5.1. Household-level modality styles and residential neighbourhood choice patterns

As mentioned previously, the final specification for the household residential neighbourhood choice behaviour sub-model
identified six classes in the sample population. These classes differ from each other in terms of their sensitivity to different neigh-
bourhood attributes, and their demographic characteristics. Over subsequent paragraphs, we summarize some of the key attributes of
each of these six classes. To ease comprehension, we’ve ordered the classes in terms of decreasing preference for urban environments
and increasing preference for suburban and exurban environments. To emphasize behavioural differences between classes, we
conducted a sample enumeration exercise to calculate the expected neighbourhood location probabilities and the expected socio-
economic composition for each class. We’ve included results from the exercise in our description of the classes.
Finally, as an illustration of how different preferences result in different patterns of neighbourhood location, we plot the expected
probabilities of residing in different census tracts across the cities of San Francisco and San Jose for the six classes in Figs. 2 and 3.
These results have been incorporated in our description of the classes. As the reader will notice from the names that we’ve given to
each of the six classes, income, race and ethnicity are inextricably linked with neighbourhood preferences, raising issues of equity

Table 3
Summary statistics for different specifications for the sub-model of household modality styles and residential neighbourhood choices.
Model Parameters Log-likelihood AIC BIC

Two classes 79 −35,462 71,082 71,082


Three classes 126 −34,728 69,707 69,706
Four classes 176 −34,297 68,947 68,945
Five classes 224 −34,060 68,568 68,567
Six classes 274 −33,908 68,364 68,362

349
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Table 4
Parameter estimates (and t-statistics) for the household class membership model.
Parameters DINK Low-income Black High-Income White High-Income Median-Income Low-Income
households Households Households Migrant White Households Hispanic
Households Households

Constant 0.07* 0.00 −1.76 −0.22* −1.10 1.98


(0.24) (–) (−2.42) (−0.52) (−2.80) (4.17)
Number of vehicles −1.06 0.00 2.23 2.70 2.85 2.72
(−4.55) (–) (9.75) (10.38) (14.00) (10.04)
Driver’s license holding −2.93 0.00 −3.12 −3.22 −3.18 −4.76
(−11.01) (–) (−8.78) (−11.68) (−13.78) (−11.30)
Number of bikes 0.56 0.00 0.60 −0.07* 0.31* −0.23*
(2.95) (–) (2.80) (−0.29) (1.48) (−1.46)
Residence Typea: Single −2.15 0.00 −6.56 −3.55 −6.80 −4.90
Family House (−5.12) (–) (−23.20) (−9.45) (−25.19) (−13.20)
Residence Type: Building with 3.36 0.00 −4.72 −1.78 −3.31 −1.47
2–4 Apartments (10.54) (–) (−9.79) (−1.96) (−7.82) (−3.60)
Residence Type: Building with 2.38 0.00 −13.18 −0.70* −3.90 −3.85
5–19 Apartments (7.83) (–) (−17.66) (−1.16) (−7.08) (−10.89)
Residence Type: Building with 6.66 0.00 −7.93 1.81 −2.81 0.08*
20 or more Apartments (14.68) (–) (−20.14) (5.21) (−3.12) (0.25)
Residence Owned 0.32* 0.00 1.36 2.81 2.77 1.84
(1.02) (–) (3.28) (5.98) (7.27) (6.78)
Annual household income 7.76 0.00 8.38 7.05 5.00 2.41
(32.08) (–) (34.36) (27.13) (23.70) (6.86)
Household size −1.24 0.00 −1.52 −0.23* −1.71 0.97
(−5.25) (–) (−5.53) (−1.09) (−11.20) (4.71)
Residence tenure (years) 2.87 0.00 3.30 −2.72 −6.66 −4.87
(5.61) (–) (7.06) (−5.07) (−15.89) (−12.33)
Number of workers 4.99 0.00 3.45 3.31 3.95 3.95
(14.67) (–) (11.00) (11.71) (13.92) (17.31)
Number of students −0.55* 0.00 0.85* −0.82 0.51* −0.58
(−1.38) (–) (1.79) (−2.20) (1.28) (−1.99)
Unrelated household 0.01* 0.00 −1.02 −4.47 −1.22 −2.27
(0.01) (–) (−2.58) (−5.64) (−2.79) (−3.62)
Presence of pre-school kids 0.74 0.00 −0.50* 0.71 −0.13* 0.43*
(1.95) (–) (−1.10) (2.43) (−0.44) (0.81)
Presence of school kids 1.82 0.00 1.60 4.29 1.50 1.67
(2.41) (–) (4.95) (8.90) (4.11) (2.38)
Raceb: Hispanic household 1.01 0.00 −0.67* 0.58* 0.80* 5.55
(2.71) (–) (−1.10) (1.80) (1.49) (18.60)
Race: White house hold 7.17 0.00 9.17 3.23 9.53 3.77
(27.07) (–) (22.48) (10.80) (24.62) (7.83)
Race: Black household −12.07 0.00 −15.37 −13.14 −14.49 −12.19
(−29.97) (–) (−40.67) (−41.48) (−25.95) (−37.59)
Race: American household 2.43 0.00 2.76 0.60* 5.18 0.99
(8.03) (–) (4.28) (1.76) (12.2) (3.32)

* Not Significant at 95% level.


a
The base reference group for residential type is “mobile home such as boat, RV, Van, etc.”
b
The base reference group for race is “Native Hawaiian and other Pacific Islander alone”.

with regards to mobility and access. We visit these issues in more detail in the following subsection.

5.1.1. Household class 1 – double income no kids (DINK) households


This class constitutes 12% of the sample population and consists largely of double income households with no kids (DINKs): they
tend to have small households (average household size of 1.78) that are unlikely to include children (15% of these households tend to
have children), and high household incomes ($93,400 per year). Households belonging to this class tend to have low levels of both
home ownership (36%) and car ownership (0.91 cars per household). These households usually have the lowest average residency
tenure (11.94 years) among all classes. They are less likely to reside in a single-family detached house (6.4%) and more likely to live
in apartments (78%). Households belonging to this class incline to prefer neighbourhoods that are diverse and walkable, and are
highly concentrated in the San Francisco inner city neighbourhoods.

5.1.2. Household class 2 – low-income black households


This class constitutes 2.1% of the sample population and consists largely of black households (94%) with low average household
incomes ($50,815), low levels of home ownership (49%) and low levels of car ownership (1.25 cars per household). Compared to
other classes, this class has the second lowest (Class1 has the lowest) tendency of living in a single-family detached house (48%) and
the highest propensity of living in a single-family attached house such as townhouse, duplex or triplex (19%). In terms of the built

350
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Table 5
Parameter estimates (and t-statistics) for the class-specific neighbourhood location choice model.
Parameters DINK Low-income Black High-Income White High-Income Median-Income Low-Income Hispanic
households Households Households Migrant Households White Households Households

Density −0.02* −0.27* −2.77 −1.08* −3.22 −1.35


(100,000 per square (−0.19) (−0.47) (−6.17) (−1.15) (−4.95) (−2.41)
mile)
Diversity 0.28 −0.07* 0.08* 0.28* 0.11* 0.26
(Land use entropy (3.78) (−0.27) (1.17) (1.78) (1.32) (2.05)
measure)
Design 0.26 0.31 0.40 −0.17* −0.01* −0.03*
(Census blocks per sq. (5.62) (2.88) (7.95) (−1.12) (−0.18) (−0.31)
mile)
Own dwelling −3.85 0.86* 1.05 0.54* 1.34 0.18*
(−15.01) (1.25) (4.89) (1.00) (5.87) (0.40)
API Score −0.16 −0.02* 0.02* 0.13* −0.05* −0.10*
(−5.34) (−0.24) (0.91) (1.82) (−1.82) (−1.81)
Crime 0.01* 0.05 −0.07 0.07 −0.02* 0.01*
(0.92) (1.38) (−4.57) (2.15) (−1.03) (0.49)
Gini Score 4.90 3.74* 0.40* 0.60* −0.13* 1.28*
(7.04) (1.84) (0.74) (0.41) (−0.21) (1.06)
Median Rent −0.24* −0.20* −0.31 0.09* −0.61 −0.28*
(−1.60) (−0.51) (−3.31) (0.40) (−5.67) (−1.22)
Median Value 0.02* −0.31 0.23 0.01* −0.08 −0.05*
(0.83) (−4.02) (10.42) (0.23) (−3.08) (−1.18)
Tract population −0.03* −0.34* −0.01* −0.37 −0.35 −0.07*
(−0.61) (−0.75) (−0.10) (−2.41) (−2.20) (−0.46)
Vacant Units 0.07* −2.34* −2.46 −3.18* 0.30* 1.36*
(0.12) (−1.40) (−4.44) (−1.94) (0.73) (1.40)
Racea: Blacks −4.30 8.83 −4.76 −0.41* −1.42* −2.75*
(−2.52) (3.15) (−2.93) (−0.12) (−1.19) (−1.66)
Race: Immigrants −7.35 −6.43* −2.73 −2.30* −1.30* 2.41*
(−4.57) (−1.71) (−2.14) (−0.79) (−0.99) (1.17)
Race: Native Americans −8.40* −12.30* −6.08* −13.30* 2.79* −18.50
(−1.08) (−0.69) (−0.96) (−0.68) (0.65) (−2.31)
Race: Pacific Islanders −10.20* 10.10* −9.75 7.05* 1.76* 1.84*
(−1.80) (1.51) (−2.11) (0.80) (0.51) (0.50)
Race: Whites −1.79* −2.59* 0.27* −4.70* 1.81 0.32*
(−1.35) (−1.04) (0.25) (−1.79) (1.97) (0.29)
Classb: Upper Middle 3.13 3.27* 1.19 1.01* 3.00 2.03*
(4.07) (1.44) (2.39) (0.86) (4.98) (1.56)
Class: Lower Middle 2.80 2.12* 0.94 1.84* 2.58 1.35*
(4.80) (1.21) (2.10) (1.83) (4.75) (1.28)
Class: Lower 0.74* 0.89* −0.61* −2.57* 2.86 1.56*
(1.24) (0.47) (−1.23) (−1.93) (5.37) (1.44)
Below Poverty Line −3.34 2.67 −1.42 −0.50* −2.76 −0.22*
(−5.41) (2.03) (−2.08) (−0.29) (−4.77) (−0.25)
Low Level of Education −4.37 −1.36* −0.80* 1.54* −1.65 2.80
(−6.68) (−0.94) (−1.43) (1.17) (−3.22) (3.40)

* Not Significant at 95% level.


a
The base reference group for race is “other”.
b
The base reference group for class is “upper”.

environment, households belonging to this class prefer neighbourhoods that are walkable. In terms of other attributes, households
belonging to this class are most likely to live in poorly performing school districts and neighbourhoods with high crime rates. These
households are highly concentrated in urban neighbourhoods in Oakland, Fruitvale and Richmond.

5.1.3. Household class 3 – high-income white households


This class constitutes 40% of the sample population and consists largely of white households (95%) with high household incomes
($152,500 per year) and high rates of car ownership (2.08 cars per household). On average, 37% of the households belonging to this
class tend to have children, with an average household size of 2.75. These households tend to have the highest average residency
tenure (21 years) among all classes, are highly likely to be living in single family houses (99%) and incline to have high levels of home
ownership (95%). They prefer neighbourhoods that have low densities and are walkable. These households tend to live in high
income neighbourhoods with good schooling and low crime rates. They tend to be concentrated in Marin County, the Berkeley-
Oakland hills, and along the corridor connecting south San Francisco to San Jose.

351
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Table 6
Parameter estimates (and t-statistics) for the individual class membership model, conditional on the household class.
Household-level modality style Variable Individual-level modality style

Complete car Near complete car dependents Partial car Multimodals


dependents dependents
Class 1 Class 2 Class 3 Class 4 Class 5 Class 6

Class 1: DINK Households Class-specific 0.00 −0.06 *


−1.95 −271.62 −0.43 *
−103.35
constant (–) (−0.05) (−2.02) (−278.1) (−0.43) (−111.40)
Male 0.00 −2.20 −0.72* 0.12* −1.04* −0.53*
(–) (−2.18) (−1.23) (0.13) (−1.31) (−0.75)
Married 0.00 −27.63 −0.79* −0.60* −1.28* −0.54*
(–) (−27.63) (−0.96) (−0.69) (−1.38) (−0.79)
Parent 0.00 −30.85 −0.79* 0.33* −0.60* 0.82*
(–) (−30.85) (−0.78) (0.37) (−0.59) (0.94)
Employed 0.00 1.02* 2.11 273.47 1.63* 104.3
(–) (1.04) (3.45) (280.00) (1.89) (114.8)
Student 0.00 −30.85 169.85 272.06 −30.98 272.30
(–) (−30.84) (183.40) (286.70) (−30.90) (320.60)
Age (years) 0.00 −0.01* 0.05 −0.03* 0.01* 0.02*
(–) (−0.50) (2.80) (−0.70) (0.60) (1.10)

Class 2: Low-Income Black Class-specific 0.00 55.63 56.54 −358.96 −322.74 −56.94
Households constant (–) (49.48) (50.65) (−358.59) (−287.60) (−30.55)
Male 0.00 −148.88 −147.05 −208.95 −352.05 48.60
(–) (−110.47) (−99.38) (−205.65) (−352.05) (26.50)
Married 0.00 14.87 129.40 8.13 15.53 −37.03
(–) (14.87) (127.97) (8.01) (15.53) (−37.03)
Parent 0.00 −30.85 104.92 149.45 342.75 63.53
(–) (−30.85) (55.80) (83.01) (305.43) (51.94)
Employed 0.00 92.93 89.81 −43.96 −67.71 −23.71
(–) (68.37) (67.06) (−43.96) (−65.72) (−23.31)
Student 0.00 43.81 2.33* −55.70 −33.06 266.03
(–) (33.10) (1.91) (−55.70) (−33.06) (261.80)
Age (years) 0.00 0.01* 0.03* 6.48 2.90 0.14
(–) (0.30) (0.70) (6.32) (14.50) (2.20)

Class 3: High-Income White Class-specific 0.00 −297.15 3.29 −85.92 2.42* 4.67
Households constant (–) (−294.35) (2.15) (−73.63) (1.57) (3.13)
Male 0.00 1.04* 0.92* 3.43 0.42* 0.99*
(–) (1.02) (1.16) (3.23) (0.44) (1.23)
Married 0.00 239.91 118.18 210.06 118.30 119.12
(–) (237.65) (154.63) (180.02) (133.84) (156.86)
Parent 0.00 −4.79 −4.21 −59.55 −145.13 −6.27
(–) (−5.27) (−5.19) (−59.55) (−145.13) (−7.72)
Employed 0.00 179.91 2.09 1.14* 2.63 2.28
(–) (178.22) (2.54) (1.05) (2.86) (2.60)
Student 0.00 67.88 −2.02* −59.55 −59.54 −1.50*
(–) (67.80) (1.50) (−59.55) (−59.54) (−1.10)
Age (years) 0.00 0.04* 0.02* −0.07 −0.00* −0.01*
(–) (1.10) (0.60) (−1.96) (−0.07) (−0.50)
Class 4: High-Income Migrant Class-specific 0.00 −102.37 216.95 −468.19 79.06 215.78
Households constant (–) (−93.87) (216.36) (−403.23) (79.07) (211.25)
Male 0.00 1.86* 108.87 202.36 −293.14 −9.59
(–) (0.81) (61.34) (174.29) (−293.14) (−9.01)
Married 0.00 108.13 −250.00 114.89 −45.99 64.26
(–) (85.06) (−144.59) (98.95) (−39.18) (55.13)
Parent 0.00 172.31 −190.13 106.72 −356.69 139.13
(–) (111.24) (−182.10) (91.53) (−356.69) (134.75)
Employed 0.00 84.80 92.35 129.29 −271.50 −50.00
(–) (62.30) (89.99) (87.00) (−271.50) (−49.67)
Student 0.00 342.56 −110.63 15.53 −159.34 4.93
(–) (286.50) (−91.9) (15.53) (−159.34) (4.93)
Age (years) 0.00 0.45* 0.83* 3.39 3.34 −2.26
(–) (1.10) (1.80) (7.50) (7.20) (−4.90)

(continued on next page)

352
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Table 6 (continued)

Household-level modality style Variable Individual-level modality style

Complete car Near complete car dependents Partial car Multimodals


dependents dependents
Class 1 Class 2 Class 3 Class 4 Class 5 Class 6

Class 5: Median-Income White Class-specific 0.00 −37.63 −106.91 −41.81 −42.51 −39.86
Households constant (–) (−39.01) (−48.77) (−36.57) (−32.54) (−40.12)
Male 0.00 −20.54 −3.14* −18.90 −19.60 −20.45
(–) (−21.89) (−0.86) (−16.72) (−19.07) (−20.70)
Married 0.00 −11.72 −59.56 −59.54 −12.90 −12.19
(–) (−16.42) (−59.56) (−59.55) (−13.80) (−16.25)
Parent 0.00 −2.60 15.10 −30.85 −3.61 −3.11
(–) (−2.75) (8.42) (−30.84) (−3.42) (−3.31)
Employed 0.00 8.38 0.67* 8.49 8.50 8.89
(–) (13.15) (0.61) (8.72) (9.81) (12.78)
Student 0.00 48.39 −35.76 −59.55 49.20 48.29
(–) (49.10) (−35.76) (−59.55) (42.80) (45.80)
Age (years) 0.00 6.53 7.29 6.56 6.57 6.56
(–) (3.90) (4.10) (3.90) (3.90) (3.90)
Class 6: Low-Income Hispanic Class-specific 0.00 −31.56 133.37 −182.84 −503.91 661.50
Households constant (–) (−31.56) (112.12) (−182.02) (−503.91) (440.80)
Male 0.00 13.41 512.41 227.81 −35.81 −80.49
(–) (13.41) (362.25) (227.31) (−35.81) (−71.92)
Married 0.00 −15.02 −32.29 117.97 −107.95 −105.22
(–) (−15.02) (−32.25) (117.43) (−107.95) (−101.70)
Parent 0.00 12.40 −77.45 193.09 −30.89 239.07
(–) (12.40) (−61.62) (193.11) (−30.89) (108.90)
Employed 0.00 −59.55 −111.16 −123.76 −49.01 −252.62
(–) (−59.55) (−61.17) (−123.58) (−49.01) (−241.65)
Student 0.00 −59.55 −539.54 −382.18 243.01 −205.46
(–) (−59.55) (−526.20) (−382.00) (243.01) (−139.10)
Age (years) 0.00 −30.85 0.19 2.34 4.80 −5.96
(–) (−30.85) (3.70) (14.50) (4.90) (−89.20)

* Not Significant at 95% level.

Table 7
Parameter estimates (and t-statistics) for the class-specific travel mode choice model for mandatory tours.
Model Complete car dependents Near complete car dependents Partial car dependents Multimodals

Class 1 Class 2 Class 3 Class 4 Class 5 Class 6

Alternative specific constants


Private vehicle 0.000 0.000 0.000 0.000 0.000 0.000
(–) (–) (–) (–) (–) (–)
Private transit – – – −2.80 −1.67 –
(−13.99) (−2.48)
Public transit – 2.84 0.47* −1.44 4.98 10.91
(5.46) (1.50) (−5.38) (8.20) (19.66)
Bike – – −1.42 −5.46 3.68 –
(−9.35) (−2.66) (9.92)
Walk – 0.45* 1.25 −2.59 6.57 11.33
(1.35) (5.61) (−5.52) (12.63) (16.46)

Level-of-service
Travel time (min) – −0.10 −0.07 −0.00* −0.09 −0.06
(−15.70) (−17.70) (−1.41) (−13.65) (−13.66)
Travel cost ($) – −1.00 – – −0.06* −0.53
(−16.83) (−1.04) (−6.72)

Choice probability elasticities


Travel time (min) 0.00 −18.44 −11.24 −0.17 −10.55 −8.46
Travel cost ($) 0.00 −1.09 0.00 0.00 −0.19 −0.57

Marginal rates of substitution


Value of time ($/h) – 5.73 ∞ ∞ 81.84 6.83

* Not Significant at 95% level.

353
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Table 8
Parameter estimates (and t-statistics) for the class-specific travel mode choice model for non-mandatory tours.
Model Complete car dependents Near complete car dependents Partial car dependents Multimodals

Class 1 Class 2 Class 3 Class 4 Class 5 Class 6

Alternative specific constants


Private vehicle 0.00 0.00 0.00 0.00 0.00 0.00
(–) (–) (–) (–) (–) (–)
Private transit – – – −3.32 −4.66 –
(−10.32) (−5.02)
Public transit – −1.62 0.47* −1.32 1.35 2.78
(−2.66) (0.98) (−4.99) (2.69) (5.21)
Bike – – −2.96 −1.68 0.63 −3.80
(−13.60) (−8.99) (3.63) (−2.18)
Walk – 1.82 1.26 0.36 2.53 3.62
(12.40) (5.25) (2.27) (8.73) (6.31)

Level-of-service
Travel time (min) – −0.05 −0.14 −0.01 −0.07 −0.06
(−22.39) (−18.42) (−9.38) (−14.24) (−8.70)
Travel cost ($) – −0.15* – −0.01* – −0.48
(−1.45) (−0.52) (−4.30)

Choice probability elasticities


Travel time (min) 0.00 −5.15 −13.64 −0.56 −5.20 −5.08
Travel cost ($) 0.00 −0.12 0.00 −0.04 0.00 −0.31

Marginal rates of substitution


Value of time ($/h) – 19.32 ∞ 34.10 ∞ 7.38

* Not Significant at 95% level.

Fig. 2. Expected neighbourhood choices in the city of San Francisco across the six household-level classes.

5.1.4. Household class 4 – high-income migrant households


This class constitutes 7.3% of the sample population and comprises a large share of households from minority races and migrant
backgrounds (27%) with high household incomes ($130,150 per year) and high rates of car ownership (2.13 cars per household). On
average, 60% of the households belonging to this class tend to have children, with an average household size of 3.10. Like Class 3,
they are highly likely to be living in single family houses (84%) in high income neighbourhoods with good schooling and low crime
rates and tend to have high levels of home ownership (89%). However, compared to Class 3, they incline to have a lower average
residency tenure of 13 years, and perhaps more interestingly, they tend to not be sensitive to any of the built environment variables.
These households tend to be concentrated in San Jose, Fremont and Hayward.

354
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Fig. 3. Expected neighbourhood choices in and around the city of San Jose across the six household-level classes.

5.1.5. Household class 5 – median-income white households


This class constitutes 27.7% of the sample population and consists largely of white households (97%) with median household
incomes ($85,044 per year) and high rates of car ownership (2.07 cars per household). Like Classes 3 and 4, they are highly likely to
be living in single family homes (92.7%) in neighbourhoods with good schooling and low crime rates, and tend to have high levels of
home ownership (91%). However, unlike Classes 3 and 4, these neighbourhoods tend to have lower average incomes, and households
belonging to this class incline to have a distinct preference for low density car-oriented neighbourhoods that aren’t very walkable.
They tend to be concentrated in suburban neighbourhoods in and around Concord and Santa Rosa.

5.1.6. Household class 6 – low-income hispanic households


This class constitutes 10.2% of the sample population and consists largely of Hispanic households (86%) with low household
incomes ($49,452 per year), low levels of home ownership (54%) and median levels of car ownership (1.75 cars per household). On
average, 76% of the households belonging to this class tend to have children, with an average household size of 3.77. In terms of the
built environment, households belonging to this class incline to prefer low-density mixed-use neighbourhoods. They are usually
concentrated in small pockets in and around South San Francisco, Palo Alto, San Jose, Richmond, Antioch and Santa Rosa.
In Fig. 4, we illustrate how our model predictions for the city of San Francisco compare with available observed data. The latter
has been sourced from the StatisticalAtlas website, which itself uses a mix of data from the 2010 US Census and the 2009–2013
American Community Survey, both of which are in the same relative timeframe as the CHTS data that was used to calibrate our
model. We’ve relied on the most dominant traits underlying each of the six household classes to select the most appropriate sub-
population for comparison. For example, Fig. 4-1 compares our predicted distribution of Class 1 with the observed distribution of
two-member households in the region, and similarly, Fig. 4-2 compares our predicted distribution of Class 2 with the observed
distribution of black households. The close correspondence within different pairs of images attests to our model’s ability to reproduce
patterns of residential location at the sub-population level to a high degree of accuracy.

5.2. Individual-level modality styles and travel mode choice patterns

As mentioned previously, the final specification for the individual travel mode choice behaviour sub-model identified six classes
in the sample population. These classes differ from each other in terms of the travel modes that they consider when deciding how to
travel, their relative sensitivity to travel times and travel costs, and their demographic characteristics.
The reader should note from Tables 7 and 8 that we did not find travel costs to be statistically significant across most classes (in
cases where estimates are missing, this is because our estimation routine constrained parameters corresponding to travel times and
costs to be negative, and the estimation procedure reported a final estimate of zero, i.e. the boundary condition). We speculate that
the lack of statistical significance might have to do with the revealed preference nature of our data, where we simply do not have
enough variability in travel costs within our data to detect sensitivity to the same. In other words, our analysis indicates the most
consumers are insensitive to changes in travel costs within the variability contained within our data.
Over subsequent paragraphs, we summarize some of the key attributes of each of these six classes. To ease comprehension, we’ve

355
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

Fig. 4. A comparison between expected neighbourhood choices for each household modality style and with the available statistical data for the San
Francisco Area (Map image source: https://statisticalatlas.com/place/California/San-Francisco/tract/).

ordered the classes in terms of decreasing car dependence. To underscore behavioural differences between classes, a sample enu-
meration similar to that described in Section 5.1 was carried out to characterize the kinds of individuals that are most likely to belong
to different classes, and the travel modes that they are most likely to use. Note that due to the similarity of the mode choice data used
by the present study and Vij et al. (2017), some of the results between the two studies expectedly overlap.

5.2.1. Individual class 1 – complete car dependents


This class constitutes 28.4% of the sample population. Individuals belonging to this class deterministically choose private vehicle
for both mandatory and non-mandatory tours. In other words, they are completely dependent on their cars to fulfil their mobility
needs. Married adults with children are most likely to be in this class.

5.2.2. Individual class 2 – near complete car dependents


This class constitutes 40.2% of the sample population. Individuals belonging to this class consider private vehicle, public transit
and walk for both mandatory and non-mandatory tours. However, in terms of actual use, the class is almost equally dependent on the
car, with mode shares of 94% and 79% for mandatory and non-mandatory tours, respectively. This class is sensitive to both travel
times and travel costs, and the value of time was estimated to be 5.7$/h for mandatory tours and 19.3$/h for non-mandatory tours.
Individuals belonging to this class have an average age of 40 years, with 46% being employed.

5.2.3. Individual class 3 – near complete car dependents


This class constitutes 23.0% of the sample population. Individuals belonging to this class consider all travel modes except private
transit for both mandatory and non-mandatory tours. In addition to the difference in availability of transportation mode between
class 2 and 3 consideration sets, for class 3 mandatory trips, walking is the second preferred option whereas public transit is the
second preferred option for class 2 mandatory trips. In terms of actual use though, like Class 2, the class is highly dependent on the
car, with mode shares of 88% and 93% for mandatory and non-mandatory tours, respectively. However, unlike Class 2, individuals
belonging to the class are insensitive to travel costs but have high mean elasticities to travel times. Similar to class 1, married adults
with children are most likely to be in this class. Individuals belonging to this class are older (average age of 48 years, compared to
40 years for Class 2), have higher employment rates (65%, compared to 46% for Class 2), and higher school enrolment rates (10%,
compared to 34% for Class 2).

5.2.4. Individual class 4 – partial car dependents


This class constitutes 3.6% of the sample population. Individuals belonging to this class consider all travel modes for both

356
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

mandatory and non-mandatory tours, though 75% of mandatory tours and 54% of non-mandatory tours are still made by private
vehicle. Of the other modes, public transit is most popular for mandatory tours, with a mode share of 13%, and walk is most popular
for non-mandatory tours, with a mode share of 34%. The class has low mean elasticities towards both travel times and costs, for both
mandatory and non-mandatory tours, indicating that travel mode decisions are determined by factors other than these level-of-
service attributes. The class has the highest average age (50 years) and the lowest rate of parenthood (11%), indicating that in-
dividuals belonging to this class likely lie later in their life cycles than other classes.

5.2.5. Individual class 5 – multimodals


This class constitutes 2.3% of the sample population. Like Class 4, individuals belonging to this class consider all travel modes for
both mandatory and non-mandatory tours. However, unlike Class 4, travel mode shares for private vehicle are significantly lower,
with 38% of mandatory tours and 56% of non-mandatory tours being made by that mode. The class has high mean elasticities
towards travel time and high values of time for both tour types. Bike is by far the most popular alternative mode of transport,
accounting for 37% of mandatory tours and 23% of non-mandatory tours. This is the only class with a significant skew towards one
gender: 78% of individuals belonging to this class are male.

5.2.6. Individual class 6 – multimodals


This class constitutes 2.5% of the sample population. Individuals belonging to this class consider all travel modes except private
transit and bike for mandatory tours, and all travel modes except private transit for non-mandatory tours. Like Class 5, travel mode
shares for private vehicle are low, with 23% of mandatory tours and 57% of non-mandatory tours being made by that mode. Unlike
Class 5, the class has a low value of time of 6.8$/h and 7.4$/h for mandatory and non-mandatory tours, respectively. Walk is by far
the most popular alternative mode of transport, accounting for 40% of mandatory tours and 35% of non-mandatory tours. Public
transit usage for mandatory tours is quite high as well, at 38%, but the corresponding figure for non-mandatory tours is much lower at
7%. The class comprises a mix of adult parents and children still in school.

5.3. Lifestyles, residential location and transport mode use

In this section, we explore patterns of correlation between preferences for different neighbourhood types and preferences for
different transport modes, as arising from the shared influence of common lifestyle variables. Table 9 presents the percentage of the
total sampled population belonging to one of the six household and individual modality styles. Each column represents one of the six
household modality styles, ordered in terms of their decreasing likelihood to live in inner city neighbourhoods and increasing
likelihood to live in suburban and exurban neighbourhoods in the San Francisco Bay Area. Each row represents one of the six
individual modality styles and moving, ordered in terms of decreasing car dependency. Each cell represents the proportion of the total
sampled population that belongs to one of the 36 combinations of household and individual modality styles.
There is a clear pattern of correlation between preferences for suburban and exurban neighbourhoods on one hand, and car
dependency on the other, as indicated by the large proportion of the sample that belongs to cells along the diagonal from the bottom
left to the top right. For example, DINK households and low-income black households who are most likely to live in inner city
neighbourhoods are also most likely to have individuals who are either multimodal or partial car dependents. Similarly, median-
income white households and low-income Hispanic households who are most likely to live in outer suburban and exurban neigh-
bourhoods are also most likely to have individuals who are either entirely or nearly entirely dependent on the car to fulfil their
mobility needs.
There are some interesting deviations from this pattern as well that are worth pointing out. For example, low-income Hispanic
households are highly likely to reside in outer suburban and exurban neighbourhoods, but individuals belonging to these households

Table 9
Total sampled proportion of class membership model normalised on both individual modality styles and the household modality style.
Moving away from the CBD and towards suburban and exurban neighbourhoods

Household Modality Styles


Median-
Low-Income High-Income High-Income Low-Income
DINK Income
Black White Migrant Hispanic
households White
Households Households Households Households
Households
Complete car
Class 1 0.07% 0.74% 2.34% 7.24% 6.34% 0.00%
dependents
Towards less car dependency

Individual Modality Styles

Near Class 2 0.88% 1.36% 3.28% 1.36% 2.26% 7.55%


complete car
dependents Class 3 1.22% 3.19% 3.91% 2.55% 0.17% 5.64%
Partial car
Class 4 3.83% 5.61% 2.72% 1.39% 1.92% 1.16%
dependents
Class 5 4.32% 0.90% 2.48% 4.12% 4.64% 0.20%
Multimodals
Class 6 6.35% 4.86% 1.94% 0.00% 1.33% 2.12%

357
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

are least likely to be complete car dependents. Financial hardship could be one explanation for this pattern. Analogously, high-
income migrant households and median-income white households display strong preferences for suburban and exurban neigh-
bourhoods, but individuals belonging to these households also have a high likelihood of being multimodal, with a strong preference
for bicycling.
These patterns of correlation can be explored in further detail through a closer examination of the estimation results reported for
the individual-level class membership models, conditional on the household-level class, reported in Table 6. These estimates reveal
some of the tension between the individual and their role in the household. For example, men that are part of low-income Hispanic
households are most likely to be near complete car dependents, whereas women that are part of the same household-level class are
most likely to be multimodal, with a strong preference for public transit. Similar patterns can be discerned for other demographic
groups across different household-level classes.2
These patterns of correlation can also be used to inform transport and land use policy in novel ways. For example, high-income
migrant households, median-income white households and low-income Hispanic households have the greatest likelihood of living in
suburban and exurban neighbourhoods. However, individuals belonging to these different households differ in significant ways in
terms of their preferences for different travel modes, and policies discouraging the use of car and encouraging the use of alternative
travel modes can be tailored accordingly to increase their effectiveness.
In the case of high-income migrant households and median-income white households, individuals belonging to these households
are most likely to be either completely dependent on the car, or to be multimodal with a strong preference for bicycling. Therefore,
policies that encourage greater public transport use through, say, improvements in the level-of-service of the local public transport
network are likely not to be effective in neighbourhoods with high concentrations of these households, as a high proportion of
individuals belonging to these households do not consider public transport when deciding how to travel. Instead, policies that support
the provision of improved bicycling infrastructure in these neighbourhoods are much more likely to succeed in discouraging car use.
In contrast, individuals belonging to low-income Hispanic households are most likely to be ‘near complete car dependents’, i.e.
they are willing to consider other modes of transport, even though they are highly dependent on the car to fulfil their mobility needs.
Our analysis also indicates that these households have low incomes and high levels of car ownership, and are likely to live in low-cost
outer suburban and exurban neighbourhoods that have poor public transport connectivity. Therefore, high car use appears to be
borne from necessity for this segment, and policies intending to increase public transport patronage are likely to be more effective if
they target individuals belonging to these households than, say, high-income migrant households or median-income white house-
holds. These policies themselves can take one of two broad approaches. They can either increase public transport connectivity in
neighbourhoods with high concentrations of low-income Hispanic households. Or they can provide low-cost housing that caters to the
preferences of these households in existing neighbourhoods with good public transport connectivity.

6. Conclusions

This study developed a methodological framework to capture the concurrent influence of lifestyles and modality styles on
household neighbourhood location and individual travel mode choice behaviours. The framework was empirically tested using travel
diary data collected from households residing in the San Francisco Bay Area, United States, as part of the 2010–12 California
Household Travel Survey (CHTS). The model identified six household-level classes that differ in terms of their preferences for dif-
ferent neighbourhood attributes when deciding where to live, and their household characteristics. Coincidentally, the model also
identified six individual-level classes that differ in terms of the travel modes that they consider when deciding how to travel, their
sensitivity to different level-of-service attributes, and their individual characteristics.
Household preferences for neighbourhood types and individual preferences for travel modes showed expected patterns of cor-
relation. In general, households that preferred to live in suburban neighbourhoods were found to more likely consist of individuals
that are car dependent, and households that preferred to live in inner city neighbourhoods were found to more likely consist of
individuals that are multimodal. However, our analysis also revealed interesting patterns of deviation. For example, high-income
migrant households and median-income white households displayed strong preferences for suburban neighbourhoods, but in-
dividuals belonging to these households were also found to have a high likelihood of being multimodal, with a strong preference for
bicycling.
There are a number of different directions in which future research might build on the work presented here. From a methodo-
logical standpoint, the framework can be extended to include the influence of lifestyles and modality styles on other dimensions of
travel and activity behaviour, such as workplace location and car ownership. From an empirical standpoint, the framework can be

2
The reader should note that many of the parameters in Table 6 have very large positive or negative values. Our analysis suggests that these
parameter values are indicative of incompatibilities between certain sets of household-level and individual-level classes. Consider, for example, the
individual class membership model parameters for Individual Class 4 (i.e. partial car dependents), conditional on the household belonging to
Household Class 1 (i.e. DINK households). The class-specific constant is −271.6, but the parameter corresponding to the individual being employed
is 273.5, and the parameter corresponding to the individual being a student is 272.1. This implies that if a household belongs to Household Class 1,
only if the individual is either employed or a student do they have a non-zero probability of being in Individual Class 4. Similar patterns can be
gleaned for other combinations as well. These patterns are borne out by our sample enumerations as well, which are what we rely on primarily to
help characterize the different classes. In general, large (positive or negative) parameter values indicate that the process of individual class as-
signment is more deterministic than it is probabilistic. There is a lot of detail contained within these tables. For the sake of readability, we’ve tried to
be as brief as possible in our discussion here.

358
A. Ardeshiri and A. Vij Transportation Research Part A 126 (2019) 342–359

applied to data from other contexts, to see how patterns of correlation between household preferences for neighbourhood types and
individual preferences for travel modes vary across geographies, cultures and time periods.

References

Aeroe, T., 2001. Residential preferences, choice of housing, and lifestyle. PhD dissertation (English summary). Aalborg University.
Ardeshiri, A., 2014. Evaluating urban services using economic valuation techniques: towards better urban environmental quality and promotion of sustainable
development.
Ardeshiri, A., et al., 2016. The values and benefits of environmental elements on housing rents. Habitat Int. 55 (Suppl. C), 67–78.
Ardeshiri, A., et al., 2018. Exploring preference homogeneity and heterogeneity for proximity to urban public services. Cities 81, 190–202.
Ardeshiri, M., Ardeshiri, A., 2011. The issue of sprawl vs compact city towards sustainability in developing countries. Built Nat. Environ. Res. Papers 20.
Aribarg, A., et al., 2002. Understanding the role of preference revision and concession in group decisions. J. Market. Res. 39 (3), 336–349.
Arora, N., Allenby, G.M., 1999. Measuring the influence of individual preference structures in group decision making. J. Market. Res. 476–487.
Bagley, M.N., Mokhtarian, P.L., 1999. The role of lifestyle and attitudinal characteristics in residential neighborhood choice. University of California Transportation
Center.
Banerjee, I., 2011. Automobility in India: A Study of Car Acquisition and Ownership Trends in the City of Surat. UC Berkeley.
Ben-Akiva, M.E., Lerman, S.R., 1985. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT press.
Bernick, M., Cervero, R., 1997. Transit villages in the 21st century.
Bhat, C.R., Guo, J.Y., 2007. A comprehensive analysis of built environment characteristics on household residential choice and auto ownership levels. Transp. Res. Part
B: Methodol. 41 (5), 506–526.
Bhat, C.R., et al., 2009. The impact of demographics, built environment attributes, vehicle characteristics, and gasoline prices on household vehicle holdings and use.
Transp. Res. Part B: Methodol. 43 (1), 1–18.
Bierlaire, M., 2016. PythonBiogeme: a short introduction. Report TRANSP-OR 160706, Series on Biogeme. Transport and Mobility Laboratory, School of Architecture,
Civil and Environmental Engineering, Ecole Polytechnique Fédérale de Lausanne, Switzerland.
Boarnet, M.G., 2011. A broader context for land use and travel behavior, and a research agenda. J. Am. Plan. Assoc. 77 (3), 197–213.
Cao, X., et al., 2009. Examining the impacts of residential self-selection on travel behaviour: a focus on empirical findings. Transp. Rev. 29 (3), 359–395.
Cao, X.J., et al., 2010. Exploring the connections among residential location, self-selection, and driving: propensity score matching with multiple treatments. Transp.
Res. Part A: Policy Practice 44 (10), 797–805.
Cervero, R., 1994. Transit villages: from idea to implementation. ACCESS Mag. 1 (5).
Cervero, R., Duncan, M., 2006. Which reduces vehicle travel more: jobs-housing balance or retail-housing mixing? J. Am. Plan. Assoc. 72 (4), 475–490.
Corfman, K.P., 1991. Perceptions of relative influence: formation and measurement. J. Market. Res. 125–136.
Davis, J.H., 1973. Group decision and social interaction: a theory of social decision schemes.
Diana, M., Mokhtarian, P.L., 2009. Grouping travelers on the basis of their different car and transit levels of use. Transportation 36 (4), 455–467.
Giuliano, G., 1991. Is jobs-housing balance a transportation issue? University of California Transportation Center.
Handy, S., 1996. Methodologies for exploring the link between urban form and travel behavior. Transp. Res. Part D: Transp. Environ. 1 (2), 151–165.
Handy, S., 2005. Smart growth and the transportation-land use connection: what does the research tell us? Int. Reg. Sci. Rev. 28 (2), 146–167.
Hensher, D.A., Stanley, J., 2008. Transacting under a performance-based contract: the role of negotiation and competitive tendering. Transp. Res. Part A: Policy
Practice 42 (9), 1143–1151.
Jones, E., et al., 2001. Open source scientific tools for Python, Scipy.
Kitamura, R., 2009. Life-style and travel demand. Transportation 36 (6), 679–710.
Krueger, R., et al., 2018. Normative beliefs and modality styles: a latent class and latent variable model of travel behaviour. Transportation 45 (3), 789–825.
Kuhnimhof, T., et al., 2012. Travel trends among young adults in Germany: increasing multimodality and declining car use for men. J. Transp. Geogr. 24, 443–450.
Kuhnimhof, T., Gringmuth, C., 2009. Multiday multiagent model of travel behavior with activity scheduling. Transp. Res. Rec. 2134 (1), 178–185.
Lee, C.K., Beatty, S.E., 2002. Family structure and influence in family decision making. J. Consumer Market. 19 (1), 24–41.
Liao, F.H., et al., 2015. Compact development and preference heterogeneity in residential location choice behaviour: a latent class analysis. Urban Stud. 52 (2),
314–337.
Litman, T.A., 2005. Land Use Impacts on Transport: How Land Use Factors Affect Travel Behavior. Victoria Transport Institute.
Lund, H., 2003. Testing the claims of new urbanism: local access, pedestrian travel, and neighboring behaviors. J. Am. Plan. Assoc. 69 (4), 414–429.
McFadden, D., 1978. Modeling the choice of residential location. Transp. Res. Rec. 673.
McFadden, D., 1980. Econometric models for probabilistic choice among products. J. Bus. S13–S29.
Menasco, M.B., Curry, D.J., 1989. Utility and choice: an empirical study of wife/husband decision making. J. Consumer Res. 16 (1), 87–97.
Ohnmacht, T., et al., 2009. Leisure mobility styles in Swiss conurbations: construction and empirical analysis. Transportation 36 (2), 243–265.
Pisman, A., et al., 2011. Urban and suburban lifestyles and residential preferences in a highly urbanized society. Experiences from a case study in Ghent (Flanders,
Belgium), Belgeo. Revue belge de géographie 1–2, 89–104.
Prato, C.G., et al., 2016. Latent lifestyle and mode choice decisions when travelling short distances. Transportation 1–21.
Rao, V.R., Steckel, J.H., 1991. A polarization model for describing group preferences. J. Consumer Res. 18 (1), 108–118.
Rose, J., Hensher, D.A., 2004. Modelling agent interdependency in group decision making. Transp. Res. Part E: Logistics Transp. Rev. 40 (1), 63–79.
Salomon, I., Ben-Akiva, M., 1983. The use of the life-style concept in travel demand models. Environ. Plan. A 15 (5), 623–638.
Scheiner, J., 2010. Social inequalities in travel behaviour: trip distances in the context of residential self-selection and lifestyles. J. Transp. Geogr. 18 (6), 679–690.
Simma, A., Axhausen, K.W., 2001. Structures of commitment in mode use: a comparison of Switzerland, Germany and Great Britain. Transp. Policy 8 (4), 279–288.
Smith, B., Olaru, D., 2013. Lifecycle stages and residential location choice in the presence of latent preference heterogeneity. Environ. Plan. A 45 (10), 2495–2514.
Stead, D., et al., 2000. Land use, transport and people: identifying the connections. Ach. Sustain. Urban Form 174–186.
Van Acker, V., et al., 2010. When transport geography meets social psychology: toward a conceptual model of travel behaviour. Transp. Rev. 30 (2), 219–240.
Vij, A., et al., 2013. Incorporating the influence of latent modal preferences on travel mode choice behavior. Transp. Res. Part A: Policy Practice 54, 164–178.
Vij, A., et al., 2017. From trend spotting to trend’splaining: understanding modal preference shifts in the San Francisco Bay Area. Transp. Res. Part A: Policy Practice
95, 238–258.
Waddell, P., 2001. Towards a Behavioral Integration of Land Use and Transportation Modeling.
Walker, J.L., Li, J., 2007. Latent lifestyle preferences and household location decisions. J. Geogr. Syst. 9 (1), 77–101.
Xiong, C., et al., 2015. The analysis of dynamic travel mode choice: a heterogeneous hidden Markov approach. Transportation 42 (6), 985–1002.
Zhang, J., 2014. Revisiting residential self-selection issues: a life-oriented approach. J. Transp. Land use 7 (3), 29–45.
Zhang, W., Guhathakurta, S., 2018. Residential location choice in the era of shared autonomous vehicles. J. Plan. Educat. Res 0739456X18776062.
Zhou, B., Kockelman, K., 2008. Self-selection in home choice: use of treatment effects in evaluating relationship between built environment and travel behavior.
Transp. Res. Rec.: J. Transp. Res. Board 2077, 54–61.

359

You might also like