Content uploaded by Stefano De Luca
Author content
All content in this area was uploaded by Stefano De Luca on Jan 17, 2018
Content may be subject to copyright.
Modelling users’ behaviour in inter-urban carsharing program:
A stated preference approach
Stefano de Luca
⇑
, Roberta Di Pace
Department of Civil Engineering, University of Salerno, Via Giovanni Paolo II 132, 84084 Fisciano (SA), Italy
article info
Article history:
Received 22 June 2013
Received in revised form 2 November 2014
Accepted 3 November 2014
Keywords:
Carsharing
Inter-urban
Mode choice
Random utility theory
Stated preferences
Switching behaviour
abstract
In this paper, the effects of a inter-urban carsharing program on users’ mode choice behav-
iour were investigated and modelled through specification, calibration and validation of
different modelling approaches founded on the behavioural paradigm of the random utility
theory. To this end, switching models conditional on the usually chosen transport mode,
unconditional switching models and holding models were investigated and compared.
The aim was threefold: (i) to analyse the feasibility of a inter-urban carsharing program;
(ii) to investigate the main determinants of the choice behaviour; (iii) to compare different
approaches (switching vs. holding; conditional vs. unconditional); (iv) to investigate differ-
ent modelling solutions within the random utility framework (homoscedastic, heterosced-
astic and cross-correlated closed-form solutions). The set of models was calibrated on a
stated preferences survey carried out on users commuting within the metropolitan area
of Salerno, in particular with regard to the home-to-work trips from/to Salerno (the capital
city of the Salerno province) to/from the three main municipalities belonging to the
metropolitan area of Salerno. All of the involved municipalities significantly interact each
other, the average trip length is about 30 km a day and all are served by public transport.
The proposed carsharing program was a one-way service, working alongside public trans-
port, with the possibility of sharing the same car among different users, with free parking
slots and free access to the existent restricted traffic areas. Results indicated that the inter-
urban carsharing service may be a substitute of the car transport mode, but also it could be
a complementary alternative to the transit system in those time periods in which the ser-
vice is not guaranteed or efficient. Estimation results highlighted that the conditional
switching approach is the most effective one, whereas travel monetary cost, access time
to carsharing parking slots, gender, age, trip frequency, car availability and the type of trip
(home-based) were the most significant attributes. Elasticity results showed that access
time to the parking slots predominantly influences choice probability for bus and carpool
users; change in carsharing travel costs mainly affects carpool users; change in travel costs
of the usually chosen transport mode mainly affects car and carpool users.
Ó2014 Elsevier Ltd. All rights reserved.
1. Introduction and motivations
The carsharing industry has grown significantly over recent years throughout the world and it has shown great potential
as well as becoming a sustainable transport solution, which in turn implies significant business opportunities.
http://dx.doi.org/10.1016/j.tra.2014.11.001
0965-8564/Ó2014 Elsevier Ltd. All rights reserved.
⇑
Corresponding author at: Dept. of Civil Engineering, Transportation System Analysis group, University of Salerno, Via Giovanni Paolo II, 84084 Fisciano
(SA), Italy. Tel.: +39 089 964122, cell: +39 320 7406257.
E-mail address: sdeluca@unisa.it (S. de Luca).
Transportation Research Part A 71 (2015) 59–76
Contents lists available at ScienceDirect
Transportation Research Part A
journal homepage: www.elsevier.com/locate/tra
As shown by the Transportation Sustainability Research Center at the University of California (Shaheen and Cohen,
2013b) carsharing operates in 27 countries and on 5 continents, accounting for an estimated 1,788,000 members sharing
over 43,550 vehicles. North America remains the largest carsharing region, with Europe and North America accounting
for 38.7% and 50.8% of worldwide carsharing membership, respectively. Currently, Europe accounts for the majority of fleets
deployed in 2012: 47.0% in contrast to 36.2% in North America. Furthermore, as highlighted by a report from Navigant
Research
1
(2013), worldwide membership in carsharing programs will grow from 2.3 million in 2013 to more than 12 million
by 2020, and global carsharing services revenue will approach $1 billion in 2013 and grow to $6.2 billion by 2020.
The rapid growth of carsharing derives from two different but complementary reasons.
Firstly, carsharing has become an alternative mode of urban transportation and makes it possible to accomplish several
transportation planning goals while contributing to sustainable urban development. Overall, carsharing represents a
significant potential for reducing car ownership as well as the total amount of car trips made in urban areas. Moreover, it
allows cars to be used properly, it makes it possible to use the appropriate mode of transport for each journey, it favours
trip-chaining and reduces impulsive trips.
Secondly, as stated by Huwer (2004), carsharing is a sort of ‘carrot’ given to the car users, unlike most of the transport
policies that aim to offer only ‘sticks’ against cars. Indeed, transportation users can benefit from the car’s flexibility without
having to bear all its inherent costs.
Currently, carsharing has been mainly implemented in urban contexts, and in particular in highly populated cities, with
significant congestion and parking problems. The present trend indicates that the carsharing business is battling to obtain
profitability even as membership increases over time. In fact, many of the existing carsharing programs have been (and
are) financially sustained by communities and governments through tax incentives, starting investments, free parking
spaces, marketing, etc. At the same time alternative approaches to make carsharing more profitable are focusing on the
increase the number of members (with fee) but decrease the frequency of use.
In this context, the most critical issues depend on the fixed costs of the service, but also on the overestimated (or
underestimated) revenues, mainly determined by the misunderstanding of users’ behaviour (Wagner and Shaheen, 1998).
Indeed, although carsharing may rely on several case studies and it is already relatively established in many cities, it should
be noted that not many attempts to model choice behaviours exist in literature.
The majority of studies are mainly concentrated in North America, in urban contexts and they have most frequently
been addressed through focus groups and/or analyses of real data or through revealed preferences (ex-post). Among mod-
elling approaches, different solutions have been applied in order to predict changes in individual car ownership, mode
choice and carsharing usage. Most of them are based on ex-post revealed preferences surveys and have usually interpreted
carsharing as an alternative to the car transportation mode. Those which are less investigated are carsharing programs on
an inter-urban scale where different municipalities significantly interact each other (e.g. bidirectional commuting flows)
and carsharing may be an alternative to car, carpool and might become a potentially complementary alternative to public
transport.
Indeed, as also highlighted by Shaheen et al. (2006), there are five major demographic markets for which carsharing may
be a valid transport alternative: neighbourhood, business, college, low-income and commuter. Among them, it is the authors’
opinion that the inter-urban travel demand may be an interesting potential market, especially between municipalities that
significantly interact each other.
First of all, the household’s car ownership rate for inter-urban systematic travellers is usually greater than the rate of
urban systematic users, thus the fixed travel costs with which they have to deal are much greater than the costs perceived
by urban travellers. In this context, car users may be more inclined to reduce the number of owned cars.
At the same time, inter-urban public transport users may be significantly attracted by a carsharing service. Indeed, if the
involved municipalities are not served by a frequent and continuous (spatially and temporally) transit system (e.g. schedule
based with few stops/stations within each municipality), carsharing could be a complementary alternative to the transit
system in those time periods in which the service is not guaranteed, efficient or not accessible. Moreover, carpooling users
may be interested (captive) in (to) the service, since they already take part in a sort of ‘‘self-organised’’ carsharing. In this
case, carsharing would allow a more flexible solution.
From an operational point of view, if some doubts may arise on the effective feasibility of an inter-urban carsharing
program, on the other hand it can be pointed out that the travel distances between municipalities daily interacting each
other are comparable to those occurring in most typical contexts (intra-urban scale) in which carsharing has been already
developed. Moreover, if the involved municipalities significantly interact with each other and the travel demand is
bidirectional, the demand for shared cars will redistribute the cars during the day. Furthermore, users might share the
car with other people, thus facing travel costs which are comparable to public transport and smaller than those faced
travelling by car.
Based on said motivations, in this paper an ex-ante analysis of the acceptability of an inter-urban (short-distance)
carsharing program, and its effects on mode choice behaviour were investigated and modelled.
The analysis was carried out through the specification, calibration and validation of different modelling solutions which
were founded on the behavioural paradigm of the random utility theory. The aim was fourfold.
1
http://www.navigantresearch.com/research/carsharing-programs.
60 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
(i) To understand if an inter-urban carsharing program would have been perceived and chosen.
(ii) To compare different modelling approaches and to establish the most effective one. In particular, switching models
conditional on the transport mode which is usually chosen (public transport, car and carpool), unconditional switch-
ing models and holding models were investigated and compared.
(iii) To investigate different modelling solutions within each modelling approach: homoscedastic, heteroscedastic and
cross-correlated closed-form solution.
(iv) To investigate the main determinants of the choice process.
The considered case study was the metropolitan area of Salerno, in particular with regard to the home-to-work trips
from/to Salerno (the capital city of the Salerno province) to/from the three main municipalities belonging to the metropol-
itan area of Salerno. All of the involved municipalities significantly interact each other and the average trip length is about
30 km a day. The study area is served by the transit system, is connected by highways, and commuting flows among each
origin–destination pair may choose among three transport alternatives: car, carpool or bus/train. The proposed carsharing
program was a one-way service, working alongside public transport, with the possibility of sharing the same car among
different users, with free and/or dedicated parking slots and free access to the existent restricted traffic areas.
The survey data were collected from a sample of 500 individuals who were asked to state their switching behaviour
depending on the transport mode used to reach the usual destination. The carsharing program was common to all of the respon-
dents, it was introduced in terms of the main features (the access time to parking sites, the travel time to the destination, the
service fare) without details on the specific parking locations and/or on the type of fare (distance-based or time-based).
The paper is organised as follows: the state of play is discussed in Section 2; the methodological framework is proposed in
Section 3; the case study, the survey and some descriptive results are described in Section 4; estimation results and cross
comparison are discussed in Sections 5 and 6. In Section 7the main conclusions are drawn up.
2. State of play
Carsharing has been investigated since the 70s, but only towards the end of the 80s did it begin to be a viable solution in
urban contexts.
Currently, Carsharing attracts new users by presenting a less expensive option than private car ownership in that a driver
only pays for vehicle use as needed, he/she does not need to pay for or worry about parking, and he/she may be somewhat
protected from rising operating costs. For these reasons, carsharing services have predominantly been implemented on
urban contexts, in the presence of good public transport, cycling and other mobility organizations. Moreover, the market
segment has consisted mainly of non-systematic users.
Against this background, the literature proposes a wide variety of analyses that may be classified according to the
pursued analysis approaches (descriptive ex-post or modelling), the investigated impacts (car ownership and vehicle usage),
the geographical contexts (North America or Europe) and the type of available data (Stated preferences – SP vs. Revealed
Preferences – RP).
Studies on carsharing are mainly concentrated in North America, and are focussed primarily on the feasibility of carshar-
ing programs and on the impact of carsharing on car ownership and vehicle usage. Most of them rely on RP data and mainly
develop descriptive analyses. Interesting overviews are proposed by Meijkamp (1998), Katzev (1999), Litman (2000), Haefeli
et al. (2006), Shaheen et al. (2006, 2009), Barth et al. (2006), Shaheen and Cohen (2007) and Shaheen and Cohen (2013a).
One of the first contributions on carsharing is by Walb and Loudon (1986) and Doherty et al. (1987). The former
investigated the influence of a short-term car rental project on reducing car ownership and increasing transit usage in
San Francisco (data analysis) and the latter investigated a combination of carsharing and carpooling services. In 1996,
Steininger et al. analysed a controlled experiment of voluntary members and carried out a descriptive analysis investigating
pre-membership and membership trip structure and modal split. Shaheen (1999) investigated the impact of information and
communication technology in making carsharing popular in US cities and influencing user behaviour. In the same year,
Shaheen et al. (1999) presented a systematic investigation of commuters’ attitudes towards a carsharing concept over time.
Findings of a pilot study on a commuter-based carsharing program are investigated in Shaheen and Wright (2001) and
carsharing user behaviour based on a survey conducted among carsharing users is dealt in Katzev (2003) and Lane
(2005).Huwer (2004) investigated the benefits of the cooperation of public transport and carsharing and the study shows
that new customer groups for public transport can be reached. Shaheen and Rodier (2005) assert that carsharing may have
significant effects on transit modal share, reducing drive-alone modal share and total vehicle-mile travel in suburban areas.
Burkhardt and Millard-Ball (2006) advance that user attitudes towards sustainability and economic growth are the most
important factors influencing the success of carsharing programs. Celsor and Millard-Ball (2007) analyse GIS-based
carsharing user profiles and discover that household car ownership is highly correlated to carsharing frequency of usage.
Several contributions investigate the impact of carsharing on urban travel demand. The most interesting have been
proposed by Sacramento (Rodier and Shaheen, 2004), Seattle (Vance et al., 2005), Montreal (Morency et al., 2007) and the
University of Los Angeles (Zhou, 2012).
Unlike the previous descriptive approaches, different modelling approaches have been investigated to model membership
behaviours, frequency of usage and other choice dimensions. Most of them are ex-post analyses based on revealed
S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76 61
preferences or on real observed behaviours (service data-set), addressing medium-long term predictions and predicting
changes in individual car ownership, mode choice and carsharing usage.
Logistic regressions were explored (Shaheen, 1999), followed by binary Logit models (Cervero, 2003) to predict the use of
carsharing; the Multinomial Logit model was used to predict the likelihood of choosing carsharing as a travel model among
other travel modes and the Probit model was considered to examine factors influencing people’s acceptance of carsharing
(Zhou et al., 2011). Cervero et al. (2007) investigated the relative success of a carsharing program in the San Francisco
Bay area (RP and SP model). In particular, changes in car ownership, mode choice and daily vehicle miles travelled were
modelled through a Multinomial Logit Model. Morency et al. (2009) developed a dynamic econometric model to jointly
predict the probability of being an active member and the frequency of usage per month. Furthermore, they developed
an ordered probability model with Hidden Markov Chain in order to capture users’ behavioural dynamics. Habib et al.
(2012), jointly modelled activity persistence (monthly frequency of usage) and membership duration. Costain et al.
(2012) examined the administrative datasets of a carsharing service in Toronto proposing an econometric approach to model
membership duration, frequency of use, vehicle type and total vehicle kilometres travelled. Morency et al. (2012) developed
a probabilistic model of being active (using the system) and the monthly frequency of use (over the years); they used a
dynamic ordered Probit model. Stillwater et al. (2008) focussed on the relationship between built environment and
carsharing user activity using a GIS-based multivariate regression analysis. Ciari et al. (2013) introduced a new methodology
to estimate travel demand for carsharing based on activity-based micro-simulation. Shaefers (2013) explores carsharing
usage motives through a hierarchical means-end approach.
Contributions based on SP data mainly propose descriptive analyses and, in some cases, introduce a modelling approach.
Abraham (2000) discusses a SP survey on hypothetical carsharing contexts and estimates a Multinomial Logit model.
Shaheen and Wright (2001) presented the findings of a pilot study on a commuter-based carsharing program. They found
that carsharing can be a viable complementary mode to transit and feeder shuttles.
Huwer (2004) investigates mobility behaviour and the customer satisfaction of a combined service carsharing-transit
whereby users have access to the flexibility offered by the car, they do not need to buy a car and are more receptive to public
transport. Fukuda et al. (2005) investigates the potential of carsharing as an alternative mode based on a stated preference
survey conducted in Bangkok. Nobis (2006) explores different logistic regression models to investigate the awareness of car-
sharing, the acceptance of sharing the vehicle with others indicates that user attitudes and behavioural aspects are the most
important factors ensuring the success of any carsharing program. Zheng et al. (2009) analyse the potential market demand
for carsharing within the University of Wisconsin. They develop probabilistic models that take into account socio-economic
information, travel preferences, attitudes and knowledge about the concept of carsharing. Firnkorn and Müller (2011)
discuss the environmental effects of a free-floating carsharing system starting from an SP survey. The focus is on the total
number of car impacts but no modelling approach is proposed. Recently, Cascetta et al. (in press) specified and calibrated
a consumers’ choice model able to interpret and to model the potential demand for an urban car sharing service in which
conventional and EVs are supplied. A Binomial Logit model was specified and the ‘‘pure preference’’ in using electric vehicles
over traditional ones was quantified.
In conclusion, the following general considerations may be drawn:
(a) An effective carsharing program can increase transit modal share, reduce drive-alone modal share, reduce car
ownership, and reduce total vehicle-mile travel in suburban areas and traffic-related emissions in urban regions.
(b) Carsharing serves as a complementary mode to transit and feeder shuttles.
(c) The duration of membership as well as the presence of a carsharing service network in one’s neighbourhood of
residence has a profound impact on user activity persistency.
(d) Carsharing users are more concerned with personal utility than social or environmental benefit, and are motivated
more by convenience and less by affordability. In most studies the majority of carsharing users are transit users,
(e) Finally, users may be grouped in terms of the frequency of usage (frequent or occasional), favourite period of use
(weekday or weekend) and trip length. The most important influential factors are: levels of household ownership,
household income, education and professional condition (more likely to join carsharing programs), neighbourhood
(walkability) and transportation characteristics, familiarity of the project, attitudinal variables such as sensitivity to
congestion, willingness to experiment, concern for the environment, scheduling reliability, convenience and program
cost.
(f) Several studies have highlighted (Muheim, 1998) that carsharing users focus their mobility habits on public transport.
Thus, attractive public transport is very important and could be combined with carsharing services.
(g) Sometimes users are attracted to carsharing because of its good environmental image, and such image may be further
improved by adopting electric vehicles.
3. Methodological framework
Based on the previous considerations, the aim of the paper was to understand if an inter-urban carsharing program would
have been perceived and chosen, the secondary aim was to establish which modelling approach was the most effective and
finally, identify the main determinants of the phenomenon. In particular, carsharing was viewed as an alternative transport
62 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
mode that may be used to substitute the usually chosen transport mode. The proposed service was one-way, thus users
would pay what they will use, and incentives were contemplated (more details in Section 4).
The problem was investigated through the specification, calibration and validation of different modelling solutions
founded on the behavioural paradigm of the utility theory. In particular, the following modelling solutions were tested:
(a) Switching models conditional on the transport mode which is usually chosen (homoscedastic and heteroscedastic).
Conditional switching models are expected to be the most effective, but in order to be applied they require the
preliminary estimation of the consolidated transport modes’ market share.
(b) Unconditional switching models (homoscedastic and heteroscedastic).
Unconditional switching models require the preliminary estimation of the market share, but their calibration might
require a smaller number of observations. It was mainly investigated in order to verify if differences existed with
regard to the conditional models in terms of attributes, of relative magnitudes and of sensitivity.
(c) Holding models (homoscedastic, cross-correlated homoscedastic and heteroscedastic).
It is assumed that carsharing is a ‘‘transport mode’’ which is always available, and it can be assumed that users
systematically include carsharing in their own choice-set. Holding models overcome the limitations of switching models,
but might pay for the limitations in a choice context with consolidated transport modes and a completely new alternative.
Moreover, the holding approach makes it possible to explicitly simulate the competitiveness of the available transport
modes. Anyhow, holding models were calibrated in order to compare their goodness-of-fit and their generalisation capability
compared to switching solutions.
If the holding approach within the random utility paradigm is the most used in transportation mode choice issues
(Domencich and McFadden, 1975; Ben-Akiva and Lerman, 1985; Cascetta, 2009; Ortuzar and Willumsen, 2011), transporta-
tion behavioural modifications can count on a smaller number of contributions (Ben-Akiva and Morikawa, 1990; Cairns et al.,
2008; Fujii and Taniguchi, 2006; Garling and Fujii, 2009; Kearney and De Young, 1996) and have been mainly focussed on
transport mode choice, on route choice or on departure time choice.
Switching behaviour may happen in a static context or in a dynamic one. In the former, modelling switching behaviour
requires panel data (revealed or stated), in the latter cross-sectional revealed and/or stated intention. Switching behaviour
has been mainly investigated through the random utility, stationary or dynamic theories. In the latter, it is of interest in
conjunction with significant transportation system changes (e.g. a new transport mode), in this case, it is also of interest
if the dynamic evolution is of interest and/or if an information system exists. In this paper we considered a stationary choice
modelling framework.
Random utility theory is based on the hypothesis that every individual is a rational decision-maker, maximising utility
relative to his/her choices. The probability of selecting alternative jconditional on his/her choice set I, as the probability that
the perceived utility of alternative jis greater than that of all the other available alternatives: The perceived utility U
j
can be
expressed by the sum of the systematic utility V
j
and a random residual representing the (unknown) deviation of the utility
perceived by the user from the systematic utility. Systematic utility represents the mean or the expected value of the utilities
perceived by the decision-maker. It is supposed to be estimated by the analyst, and is usually expressed as a function of
attributes relative to the alternatives and the decision-maker. The function may be of any type, but for analytical and
statistical convenience, it is usually assumed that the systematic utility is a linear function in the parameters of the attributes
X
kj
or of their functional transformations: V
j
=
R
k
b
k
X
kj
+
R
q
b
q
f(X
qj
).
Various specifications of random utility models can be derived from the general hypotheses by assuming different joint
probability distribution functions for the random residual.
In this paper Multinomial Logit (MNL), Hierarchical Logit (HL), Cross-Nested Logit (CNL) and Mixed Multinomial Logit
(MMNL) models were investigated, but only MNL and MMNL resulted as statistically significant.
The Multinomial Logit (MNL) model is the simplest random utility model. It is based on the assumption that the random
residuals are independently and identically distributed (iid) according to a Gumbel random variable of zero mean and
parameter h. The independence of the random residuals implies that the covariance between any pair of residuals is null.
Under the assumptions made, the probability of choosing alternative jamong those available and belonging to choice set
Ican be expressed in closed form as:
p½j¼ expðV
j
=hÞ
P
m2I
expðV
m
=hÞ
The Mixed Multinomial Logit model (MMNL) is a highly flexible model that can approximate any random utility model
(McFadden and Train, 2000). The most straightforward formulation is based on random parameters, where the utility of each
decision-maker is specified as
U
j
¼X
k
b
k
X
kj
þX
h
c
h
X
hj
þ
e
j
8
j2I
where X
kj
are the explanatory attributes that relate to the alternative and decision-maker introduced before, b
k
is the generic
parameter of attribute k,
c
h
is the generic parameter of attribute hrepresenting the decision-maker’s taste that is supposed to
be distributed randomly with density f(
c
h
),
e
j
is the random term that is an iid Gumbel random variable of zero mean and
S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76 63
parameter h.Assuming that ~
c
and ~
bare the vectors of parameters, Mixed Logit probabilities are the integrals of standard
Logit probabilities over a density of parameters.
p½j¼Zexp½V
j
ð~
b=h;~
c
=hÞ
P
m2M
exp½V
m
ð~
b=h;~
c
=hÞ fð~
c
=hÞdð~
c
=hÞ
There are two sets of parameters in a Mixed Logit model: the parameters b
k
, and the parameters that describe densities of
parameters
c
h
. Mixed Logit does not exhibit independence of irrelevant alternatives, it allows for random taste variation,
unrestricted substitution patterns, and correlation in unobserved factors over time.
Systematic utility functions were linear in the attributes, but non-linear transformations were tested for continuous
attributes.
4. Case study, survey and descriptive results
As previously discussed, the investigated case study consisted in a carsharing program supplied among different
municipalities belonging to the metropolitan area of Salerno.
The proposed program was a one-way service, with dedicated parking slots in several attractive locations of each
municipality, and free parking and free access to the existent restricted traffic areas were guaranteed.
The geographical context taken into account was made up of one main municipality (Salerno – Campania Region,
Southern Italy) and of three smaller municipalities that significantly interact with Salerno and partially interact each other
(see Fig. 1 and Table 1). Salerno is the capital city of the Salerno Province, it is located 55 km from Naples, it has about
140,000 residents, and it is characterised by 10,000 daily commuters. The three considered municipalities belonging to
the metropolitan area of Salerno are: (1) Pontecagnano (25,000 inhabitants and 15 km from Salerno), (2) Baronissi
(20,000 inhabitants and 10 km from Salerno), (3) Cava dè Tirreni (53,000 inhabitants and 12 km from Salerno). All of them
are served by a transit system, are connected by highways, and commuting flows among each origin–destination pair mainly
travel by car, carpool or bus/train. Users travel on average 30 km a day, and the inter-urban travel demand between the
municipalities is not negligible and spreads over time periods longer than the peak hours and distributed over the whole
day. Furthermore, the car ownership rate is much higher than those municipalities in which the travel demand ends within
the municipality itself.
With regard to this type of context, the potentiality of a carsharing system was investigated through a stated preferences
survey.
The survey data were collected from a sample of 500 individuals aged 18 and over. Respondents were randomly selected
residents from the above mentioned municipalities. In particular, an intercept survey was conducted at the main sites of
each municipality (e.g. stations, squares and offices) and was carried out by students recruited (and trained) within a
research project financed by the University of Salerno. The survey was carried out in the spring of 2012, only residents
travelling for work between Salerno and the three municipalities (and vice versa) were considered. Respondents were
randomly selected to match census data (ISTAT 2011) proportions by gender (male, female), age (18–30; 31–60; >60 years
SA
12km
10km
15km
Salerno
Cava dè
Tirreni 3
Baronissi
2
Pontecagnano
1
Fig. 1. Case study. MapData, 2013ÓGoogle.
64 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
old) and type of occupation (employed, unemployed). Although no incentive was proposed, the response rate was greater
than 87%, thus, the non-response phenomenon was not a critical issue in the survey which was carried out.
Several precautionary strategies were taken into account during the interviewing process: respondents carried out the
exact same survey; they were briefed about the decision context and were also introduced to the options’ features and to
the possible benefits (with pictorial presentations). The whole interview lasted approximately 15 min.
Each respondent was presented with the same questionnaire, which consisted into two parts.
The first part aimed to gather information on users’ usual travel behaviour, on users’ geographical and socio-economic
characteristics; to investigate the general propensity to adhere to a carsharing program and to investigate the preferred
features of the service. Respondents were asked to describe their usual travel habits (transport, travel cost, travel time,
trip frequency, etc.); then he/she was introduced to the service and to the main qualitative characteristics (one way,
distance-based fees, dedicated parking slots). Users’ socio-economic characteristics were then collected, as well as users’ trip
characteristics (mode, activity duration, trip frequency, etc.), their interest in opting for the service (as it is, without knowing
the fees, the type of car or the parking location) and their main motivations.
The second part of the questionnaire aimed to investigate users’ switching behaviour through a specific stated prefer-
ences (SP) survey. Depending on the transport mode used to reach the usual destination, respondents were introduced to
the main characteristics of the carsharing program and to possible/realistic scenarios (3 or 4 scenario per user).
The proposed carsharing program was familiar to all the respondents, it was described in terms of the main features
(one-way service, dedicated parking slots, free parking, free access to restricted traffic areas). No details on the parking
locations and/or the type of fare (distance-based or time-based) were introduced to the respondents. These characteristics
were introduced in the SP scenarios as control variables. Indeed, the aim was to understand and to model users’ behaviour in
terms of the general level of service attributes, leaving the choice of parking location and fares to the decision maker who,
starting from observed/estimated behaviour, may design the most effective and efficient service.
As a matter of fact, the control variables were: the access time to parking sites, the travel time to the destination, the
service fare.
Once the control variables and the composition of the choice contexts to be proposed to the decision maker were
identified, the decision makers were presented with different choice contexts.
Each scenario was defined by a set of alternative options; each option was accompanied by some attributes defining its
characteristics. In the proposed choice contexts, the attributes vary between a prefixed numbers of values, or levels (see
Table 2). These levels were defined in absolute terms or proposed as percentage variations compared to the values of the
attributes for a real context previously experienced or known to the decision maker. The SP experiment was conducted
through a selection of all of the possible scenarios starting from the Full Factorial Design scenario and then a subset of
scenarios was generated introducing the partialisation techniques of the experiment known as Fractional Factorial Design
(Cascetta, 2009). This eliminates completely some scenarios while retaining orthogonal comparisons which allow for the
estimation of the main effects. If the resulting number of scenarios is still too high to be presented to a single decision-maker,
they can be further broken down into blocks by using the method described previously. Each of the 500 surveyed users
responded to 6 SP scenarios, thus 3000 observed behaviours were obtained.
Descriptive results indicated 73% of the intercepted users would be interested in the proposed carsharing program.
Interested users are mainly influenced by the inefficiencies of the public transport system (40%) and by the non-availability
of the car transport mode (25%). It is interesting to note the financial gain which is obtainable by adhering to the carsharing
program is not the main determinant in the decision process. Non-interested users are satisfied with the usual transport
Table 1
Case study: municipalities and demographic characteristics.
Municipality Inhabitants Extension
(kmq)
Density
(inhab./kmq)
Income
€/inhab.
#Car/
household
Systematic trips
toward Salerno
Systematic trips
from Salerno
Salerno 132,000 60 2237 12,700 1.50 – –
Cava dè Tirreni 54,000 36 1474 8300 1.92 1700 550
Pontecagnano 26,000 37 686 8000 2.23 1800 650
Baronissi 17,000 18 942 8400 2.13 1100 350
From census data (ISTAT 2011).
Table 2
Control attributes and values.
Attribute Values
Access time to carsharing parking 5 min 10 min 15 min 20 min
Travel time Equal to travel time by car
Travel cost wrt Bus (
D
)+0€+1 €+2 €+3 €+4 €
Travel cost wrt Car (
D
)30% 20% 10% = +10% +20% +30%
S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76 65
mode (53%). The remaining aliquot does not like to travel with other people, does not want to book the service in advance
and does not want to subscribe to the service in advance.
Among the interested users, 68% make use of public transport, 23% travel by car and only 9% take part in carpooling.
Non-interested users are mainly carpooling users (20%) and car users (44%). Gender and weekly trip frequency do not play
a significant role in the decision of being interested or not.
In analysing the preferred way to take part in the carsharing program, it is interesting to note that only 1% would prefer to
drive alone, 39% would be indifferent to people sharing the same car and 55% prefer to travel with known users. In particular,
female users prefer to travel with known people (66%), male users are more flexible. Among the 99% of users that would
prefer to travel with other people, more than 50% of the respondents stated that they were indifferent as regards being
drivers or simply passengers. Such a result is interesting, since carsharing plays a different role from that played in urban
contexts, it is not only a more flexible transport solution but, also, a realistic alternative that may compensate for public
transport inefficiencies. This percentage decreases for female respondents (48% indifferent) whereas it increases for male
respondents (55% indifferent). Moreover, non-indifferent users’ preferences change according to the gender: female users
prefer to be passengers (27%), male users prefer to be drivers (24%). The ratio between the number of cars and the number
of households is not a negligible determinant. In fact, respondents with a ratio less than 0.75 show a percentage of interest
greater than 70%; for ratios between 0.75 and 1.00 the percentage decreases to 58%; for ratios greater than 1, the percentage
is less than 50% (48%). In investigating the percentage of interest compared to the usually chosen transport mode, it is note-
worthy to indicate that public transport users are the easiest to influence with a percentage of interest equal to 86%; car and
carpool users follow with a percentage equal to 70%.
Finally, users were asked to respond on the preferred service features. With respect to the booking technology, 73%
responded internet, 13% by SMS and only the 14% declared to prefer a call center. As regards how long before book, 38%
accept to book one hour before, the 30% wish to book at least six hours before, whereas the 32% wish to book the day before.
Finally, about the 60% prefers automatic smartcard vehicle access (compared to human-based and the 74% prefers distance-
based fees, instead of time-based fees.
In conclusion, users seem to be interested in a carsharing program. They are not especially interested in a private share of
the car and, moreover, carsharing may attract users from all of the transport modes currently offered, but public transport
users seem to be more sensitive to the new transport alternative. Descriptive results confirm the potentiality of carsharing as
an alternative both to public transport and to the car transport mode.
5. Estimation results
In this section estimation results are showed for three different modelling approaches:
(i) Switching model conditional on the usually chosen transport mode to reach the final destination.
(ii) Unconditional switching model.
(iii) Holding model, assuming that the choice set includes carsharing.
The aim was threefold: (i) compare different approaches (switching vs. holding; conditional vs. unconditional);
(ii) investigate the main determinant of the choice process; (iii) investigate different modelling solutions within the random
utility framework.
All the tested attributes and those which are statistically significant are resumed in Table 3.
5.1. Conditional switching models
5.1.1. Car users
As introduced in Section 3, random utility switching models were specified and estimated (see Table 4). Both the
homoscedastic Logit binomial model and the heteroscedastic random parameter binomial model resulted in statistically
significant findings. The systematic utility functions (proposed below) consisted in five attributes: access Time to Carsharing
Parking (ATTcsp); Car Travel cost (CTC); Gender (Gen); the Car Frequency (FreqC).
Although travel time and several different socio-economic and activity based attributes were tested (age, income, trip-
chaining), only gender resulted in statistically significant findings. Travel time difference between the two transport modes
is too similar to be significant in users’ perception; socio-economic and activity based attributes have the same impact for
both transport modes, thus did not affect any change of behaviours.
In particular, the non-significance of users’ age counteracts the literature which has shown that specific ages segments
(usually referred to 25 year and 45 year old users) are more attracted by carsharing. Notwithstanding, it has to be
highlighted that the intercepted users are predominantly commuters which systematically use the car (or carpool). In this
case, the inertia towards the carsharing service, usually age dependent, plays a minor role. As a matter of fact, the systematic
user is more affected by the benefits than by the mistrust toward an innovative service and/or toward the need for driving a
different car.
Estimation results show the high significance of Access Time and Monetary Cost. In particular, the value of time – VOT –
(presumably walking time) is equal to 0.2 euros per minute (6 euros/h). The estimated magnitude, aside from being similar
66 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
to those estimated in different Italian case studies (Cantarella and de Luca, 2005), indicates the extreme importance of
parking location. Assuming that the average one-way travel monetary cost is equal to 3 €, 10 min walking time (about
700 m at 4 km/h) is more than half of the whole travel monetary cost.
Gender, equal to 1 if male, represents a disutility in the switching option and shows a certain attraction of male users
toward non-switching behaviour. This result has a socio-cultural interpretation in which male users usually own the car that
they drive, whereas females usually share the family car. Such a result may be extended to those geographical contexts
where the male is usually the owner and the prevalent user of the household’s car.
Table 3
Systematic utility attributes.
Attribute Acronym Unit Min Max Mean
Alternative specific constant ASC – 1 1 1
Travel time TT Minutes 13 43 28.5
Travel cost Car 4 10 5.9
Carsharing TC Euro 4 13 5.8
Bus 38 2
Carsharing travel time CSTT Minutes 13 43 28.5
Access time ATT Minutes 5 20 9.5
Access time to carsharing Parking ATTcsp Minutes 5 20 9.9
Access time to bus stop ATTbs Minutes 5 15 8.7
Age
x–y
Age
x–y
Equal to 1 for age within interval [x,y]01–
Gender Gen Equal to 1 for male users 0 1 0.52
Car frequency FreqC Number of weekly trips made by car 0 5 2.0
Frequency Freq Number of weekly trips 2 5 4.3
Go straight back home GSBack Equal to 1 for home-based trips 0 1 –
City area
x
City area
1, 2 or 3
Equal to 1 for users belonging to trip origin type x0 1 0.4
Car availability CarAV No. of household vehicles/no. of household members 0 1.25 0.55
Table 4
Estimation results for conditional switching models.
⁄
In parenthesis the t-student test value.
S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76 67
Car trip weekly frequency shows a positive sign in the ‘‘NoSwitching’’ systematic utility, meaning that systematic car
users have systematic car availability, thus they do not perceive any benefit in using a carsharing service. Finally, it should
be noted that the alternative specific constant plays a significant role in switching propensity. The interpretation is twofold:
(i) its value measures the immeasurable users’ choice determinants; (ii) it represents the propensity to change the usual
choice when a new alternative is proposed in a stated preference context.
Finally, CarAv shows a positive sign in the ‘‘NoSwitching’’ alternative. This attribute may be interpreted as a proxy of
income and/or as a measure of the user’s car availability. In both interpretations, it can be easily concluded that the
probability of switching decreases as the easiness of using the household’s car and/or as the household’s income increase.
However, it should be noted that CarAv plays a marginal role. Indeed, its monetary equivalent value is slightly greater than
0.5 €with a CarAV value equal to one.
Heteroscedasticity was investigated through both the random parameter and error components formulations. The former
resulted in statistically significant findings and led to slightly better goodness-of-fit. The only distributed parameter was the
monetary travel cost, with a not negligible standard deviation. This result makes it possible to conclude that heterogeneity
among users mainly depends on the perception of travel costs, whereas access time seems to be more clearly perceived
and/or not significantly distributed among users.
5.1.2. Carpool users
Systematic utility functions are reported below (see Table 4). Attributes which were statistically significant were the
same as car users, except for car trip frequency. The results are coherent with the expectations and give robustness to both
models. The simplicity of systematic utility functions is comprehensible since carpooling users already adhere to a sort of
self-organised carsharing, thus their switching propensity mainly depends on the level of service attributes (access time
and travel cost).
First of all, it is interesting to note that the ratios between access time, alternative specific constant and travel cost
parameters are the same as those estimated for car users. The result confirms, as expected, that car and carpool users have
similar behaviour.
As for car users, CarAv shows a positive sign in the ‘‘NoSwitching’’ alternative. The monetary equivalent value continues
to be smaller than 1 €with CarAV value equal to one. In this case, if carpool users had a car available, they would not to
switch to carsharing, but to the car or would continue carpooling.
The only remarkable difference is for gender parameter which increases its role. In fact female carpool users seem more
attracted by a potential carsharing service. As introduced for car-users, female travellers usually show a smaller car
ownership rate (or access to the household’s car), thus they usually are passenger in a carpool. In this case, carsharing allows
the access to a sort of private use of the car.
Unlike the car users’ model, car trip frequency is not significant. This result is coherent with the interpretation that
carpool users normally have a trip frequency greater than other users and similar among them.
5.1.3. Bus users
Unlike previous models, the bus users switching model shows different systematic utility specifications (see Table 4). In
particular, the number of socio-economic and activity-related attributes increases (age, gender, going straight back home) as
well as the travel related level of service attributes (access time to parking, access time to bus stop, on-board travel time,
travel monetary cost).
In terms of socio-economic attributes, female and older users show a higher switching propensity. Home-based trips
increase the systematic utility of non-switching behaviours.
Female are usually attracted by transit alternatives for socio-cultural reasons (low car ownership rates, low availability of
the family’s car), but in the presence of an available car-based system they would change. However, the estimated parameter
has an absolute value 10 times smaller than 1 equivalent euro.
Differently from car and carpool users, users’ age plays a role not negligible. Utility of switching increases with age, and it
increases up to 1.5 equivalent euros for an age interval greater than 40 years old. If on the one hand this result may appear in
contrast with existing evidences, on the other hand it should be reminded that model refers to bus users only, to systematic
users and to an inter-urban context. In this case, it is reasonable that older commuting users, due to their greater experience
and the fact that they are more disillusioned with the transit system, would be more inclined to switch.
Users travelling home after work are less interested in switching to a more flexible transport mode. The attribute’s role is
meaningful, its parameter’s value is equal to 0.5 equivalent euros, confirming that carsharing may effectively support after
work activities. In fact, users taking part in a trip chain after work prefer to adhere to the carsharing service since they may
count on the same flexibility granted by a car, and they can do different or more complex trip chains.
With regards to the level of service attributes, it is noteworthy how access time to parking and access time to the bus stop
are perceived and weighted differently. The corresponding value of time are sensibly different, 0.23 euros/minute (2.3 per
10 min) for the access time to parking and 0.07 €/minute (0.7 €per 10 min) for the access time to bus stop. Such a result
indicates that the two attributes should be considered separately and that parking location is a crucial design issue.
Surprisingly, travel time has a small value and its corresponding value of time (0.85 €/h) is quite small compared to values
existing in mode choice case studies. Overall, the result highlights the need for a service different from that supplied by the
transit system. The interpretation may be twofold: (i) travellers that have to choose between bus and carsharing may be not
68 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
much interested in a gain in terms of a travel time saving, but in having an alternative to the bus. In other words users desire
reliability and efficiency, not solely shorter travel times. (ii) A limited willingness to pay due to the socio-economic back-
ground of the investigated users (see also the aggregate economic indicators in Table 1).
As for the previous models, CarAv continues to show a positive sign in the ‘‘NoSwitching’’ alternative. The monetary
equivalent value compared to car and carpool users models increases up 1.2 €. It continues to play a minor role, but it is
interesting to conclude that if bus users had a car available, they would not to switch to carsharing, but they would switch
to the car.
As for the previous models, heteroscedasticity was modelled by a mixed Binomial Logit model and only the random
parameter turned out to be statistically significant. Among the level of service attributes, only travel monetary cost resulted
as normally distributed, though with a small standard deviation. The result is coherent and gives further robustness to the
behavioural interpretation drawn for car users’ model.
5.2. Unconditional switching models
In this section estimation results for the unconditional model are proposed (see Table 5).
As shown in Table 5, systematic utility functions are much more complex in the attributes since they have to model/inter-
pret behaviour of travellers using different transport modes. In terms of the unconditional models, socio-economic, travel-
related and level of service attributes were used. Homoscedastic and heteroscedastic modelling solutions were tested and
non-linear transformations were investigated for the level of service attributes.
Overall, the specification of a unique switching model for all the users required a segmentation of travellers with respect
to the type of trip origin (city area) and to the number of car per household.
Estimation results confirm most of the comments proposed in the previous sections: being female increases the proba-
bility of switching (+0.32 €eq), home-based trips increase the probability of not-switching (0.43 €eq). Access time to parking
and to bus stops should be estimated separately; in fact access time to bus stops shows a VOT equal to 0.1 €/min (5.4 €/h),
whereas the VOT of access time to carsharing parking is slightly less than 0.3 €/min (17 €/h). Travel time, though statistically
significant, shows a VOT of 2.9 (€/h) coherently with conditional models estimation results and coherently with estimation
carried out on previous RP surveys.
It is worthwhile noting the role of the alternative specific constant with a positive value in switching behaviour utility.
This highlights how travellers are attracted by new alternatives.
Furthermore, the ratio between the number of vehicles per household increases the utility of not switching. The same
comments made for the conditional models hold.
Finally, it is interesting to note that the geographical attribute city area increases the switching probability, meaning that
users living in the outskirts are more motivated to take part in carsharing. As a matter of fact, they usually have to deal with
poor (not frequent or not existing) transit services; therefore, on the one hand they will be surely more car dependent, but on
the other hand they be more inclined to switch to a carsharing service. In this case, carsharing will allow reducing the car-
ownership rate.
As shown in Table 5, Box–Cox transformation of travel cost resulted in statistically significant findings and allowed for a
slight increase of the model’s goodness-of-fit. The Box–Cox parameter is smaller than one, meaning that marginal travel cost
disutility decreases as travel cost increases. Unlike conditional models, Heteroscedastic formulation was successfully cali-
brated. Such a result is comprehensible since a same model was calibrated independently from the transport mode. In par-
ticular, the Mixed-Logit random parameter model was estimated assuming travel time and travel cost was distributed
normally. The model’s goodness-of-fit did not noticeably increase, but standard deviation values were not negligible, other
than allowing a further interpretative hint. The result regarding the travel cost confirm what it has already been observed for
the conditional models; as concerns the travel time, the result is reasonable since the unconditional choice model was cal-
ibrated on users travelling by different transport modes and, presumably, characterised by much different travel time
perceptions.
5.3. Holding models
The holding model explicitly simulates the choice among the possible transport modes. If on the one hand the conditional
switching approach requires the preliminary estimation of the current transport market shares, on the other hand, the hold-
ing approach overcomes such a limit by directly estimating the market shares for all the transport modes. Moreover, the
holding approach may allow to explicitly simulate the competitiveness between the available transport modes, thus it will
allow the simulation of planning scenarios in which the carsharing alternative is a stable alternative, and/or in which the
level of service of the other transport modes might change.
To this aim, the choice model was calibrated using the combined results of Stated Preferences (SP) and Revealed Prefer-
ences (RP) surveys. In fact, SP surveys should be considered as complementary to traditional RP surveys and the combined
use of the two can balance reciprocal merits and shortcomings. Experimental evidence indicates that the combined use of RP
and SP data for estimating the parameters usually results in an improvement in statistical precision and in more reasonable
parameter values.
S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76 69
From an operational point of view, specific models were individually calibrated for actual mode choice behaviour (choice
set I
RP
: car, bus, carpool) and stated choice behaviour (choice set I
SP
: car, bus, carpool, carsharing), then combined by cali-
brating the scale parameters (Ben-Akiva and Morikawa, 1990; Cascetta, 2009; Ortuzar and Willumsen, 2011). Along with
the attributes used in the switching models (see Table 3), the inertia attribute was introduced to represent the conditioning
of the generic SP decision-maker with respect to the alternative actually chosen (RP). Inertia was modelled as a dummy var-
iable equal to one if the generic user chose an alternative present in the RP context, zero otherwise.
Assuming the random residuals for RP and SP models (
e
RP
and
e
SP
) as i.i.d. Gumbel variables of parameters h
SP
and h
RP
respectively, the probability of choosing the generic alternative assumes the form of a Multinomial Logit model for both
the RP and the SP models. Furthermore, as usual, to take into account the possible difference of the variances of the residuals
e
RP
and
e
SP
, a scale factor
l
,equal to the ratio between the parameters, h
SP
and h
RP
, of the two random vectors, was introduced
and calibrated.
In particular, the systematic utility functions are shown in Table 6.
Estimation results (see Table 7) indicated that the typical level of service attributes were statistically significant. In par-
ticular the corresponding VOTs were 10 €/h for travel time and 0.30 €/min (18 €/h) for the access time to bus stop or to park-
ing locations. In this case was not possible to distinguish the access time to bus stop or to parking locations, however results
show, once more, how access time is a crucial design parameter both for carsharing and transit systems.
Weekly travel frequency (Freq) took a positive value in the systematic utilities of current transport modes. In particular,
Freq’s equivalent monetary value is equal to about 2 €equivalent, meaning that as weekly travel frequency increases, the
disutility of choosing carsharing may increase up to 8–10 €equivalent. This result allows several interpretations: (i)
although the trip purpose is home-to-work, the trip frequency varies among users; (ii) the attribute is a sort of measure
of the inertia (the more weekly trips, the more inertia to change the consolidated transport modes); (iii) carsharing may
be a potentially solution for systematic users, but users with smaller trip frequency are more inclined to change the cur-
rent transport mode.
To confirm the previous interpretation, the inertia attribute turned out highly correlated with the Freq attribute,
and was statistically significant at a level smaller than 75%. If on the one hand, such a result indicates that the inertia
Table 5
Estimation results for unconditional switching models.
⁄
In parenthesis the t-student test value.
70 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
phenomenon exists and cannot be neglected, on the other hand the inertia may be better interpreted in the light of the
trip frequency.
In terms of socio-economic attributes, the only statistically significant was age. In fact, carsharing systematic utility
increases, on average, as age increases. Different age segments were tested and those proposed in Table 7 are the most sig-
nificant ones. As also noted for the switching models, users with more travel experience are more inclined and/or motivated
to choose carsharing.
Finally, the role of the alternative specific constants should be noted. They are statistically significant for carsharing and
carpool alternatives. Both take values not negligible: positive and equal to 22 equivalent €for carsharing; negative and equal
to 25 equivalent €for the carpool alternative. These results leads to two main conclusions: first, the holding approach,
compared to the switching approach, needs to be supported by alternative specific constants, indeed the holding approach
is not able to reproduce users’ choice through observable and/or measurable attributes; secondly, the choice probabilities
will be more rigid with regard to the level of service attributes. In conclusion, the holding approach should be better adopted
in ex-post analyses, where users are aware of the real choice-set, and the competitiveness among transport modes can be
more easily observed and modelled in terms of the level of service features of each mode.
6. Cross comparison between modelling solutions
In this section the different approaches and the different modelling solutions within the single approach are compared. In
particular, the models’ goodness-of-fit were validated through consolidated statistic tests and through specific indicators.
Along with the Percent-Right indicators, a models comparison was carried out through the validation protocol proposed
by de Luca and Cantarella (2009). The protocol, developed to compare discrete choice models based on different theoretical
paradigms, introduces several indicators able to highlight the models’ effectiveness and goodness-of-fit:
Table 6
Systematic utility functions for RP and SP models.
RP model SP model
V
car
=b
1
TT + b
2
TC + b
3
Freq b
1
TT + b
2
TC + b
3
Freq + b
3
Inertia
V
carpool
=b
1
TT + b
2
TC + b
3
Freq + b
9
ASC
RP
b
1
TT + b
2
TC + b
3
Freq + b
9
ASC
SP
+b
3
Inertia
V
bus
=b
1
TT + b
2
TC + b
3
Freq + b
4
ATT b
1
TT + b
2
TC + b
3
Freq + b
4
ATT + b
3
Inertia
V
carsharing
=– b
1
TT + b
2
TC + b
5
ATT + b
6
Age
18–25
+b
7
Age
26–40
+b
8
Age
>40
Table 7
Estimation results for the holding model.
Carsharing Car Carpool Bus
ASC
RP
1.87 –––
(+5.12)
ASC
SP
––1.72 –
(+11.72)
ATT 0.0251 – – 0.0251
(4.32) (4.32)
TT 0.0139 0.0139 0.0139 0.0139
(2.76) (2.76) (2.76) (2.76)
TC 0.0833 0.0833 0.0833 0.0833
(2.32)(2.32)(2.32)(2.32)
Age
18–25
0.223 – ––
(+1.98)
Age
26–40
0.367 –––
(+2.32)
Age
>40
0.777 –––
(+2.10)
Freq – 0.432 0.432 0.432
(+17.87)(+17.87)(+17.87)
Inertia
⁄
– 0.0866 0.0866 0.0866
(0.63)(0.63)(0.63)
Scale parameter
l
0.689
(2.47)
Init log-likelihood 3001
Final log-likelihood 1632
Rho-square 0.456
Adjusted rho-square 0.454
⁄
In parenthesis the t-student test value.
S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76 71
– MSE ¼P
i
P
k
ðp
sim
k;i
p
obs
k;i
Þ
2
=N
users
P0 mean square error between the user observed choice fractions and the simulated
ones, over the number of users in the sample, N
users
(SD is the corresponding standard deviation, which represents
how the predictions are dispersed if compared with the choices observed.); If different models have similar MSE errors,
the one with the smallest SD is preferable.
– MAE ¼P
i
P
k
jp
sim
k;i
p
obs
k;i
j=N
users
P0 mean absolute error.
–FF¼P
i
p
sim
i
=N
users
2½0;1;with FF = 1, Fitting Factor (FF). This is the ratio between the sum over the users in the sample of
the simulated choice probability for the mode actually chosen, p
sim
user
2[0,1], and the number of users in the sample, N
users
.
FF = 1 means that the model perfectly simulates the choice actually made by each user (p
sim
user
= 1).
– %Right: It is common practice to compare different models through the %right indicator, that is the percentage of users in
the calibration sample whose observed choices are given the maximum probability (whatever the value) by the model.
– %Clearlyright(t) percentage of users in the sample whose observed choices are given a probability greater than threshold t
by the model.
– %Clearlywrong(t) percentage of users in the sample for whom the model gives a probability greater than threshold tto a
choice alternative different to the observed one.
All the indicators were computed on the calibration data set and, in order to compare the conditional switching proba-
bilities with the unconditional ones, resulting switching probabilities were estimated from the conditional switching prob-
abilities estimated for each transport mode.
Moreover, direct elasticities were computed by introducing the variation of the attributes equal to 20% of the initial value.
Comparison results for the switching models are proposed in Table 8.
Regarding conditional switching homoscedastic models, the bus users switching model shows goodness-of-fit which is
better than car and carpool users; simulating heteroscedasticity slightly improves goodness-of-fit of all models. Differences
among the models decrease. The unconditional model, compared with the resulting switching probabilities, shows all of the
indicators values dominated. In particular, the Fitting Factor, the MSE and its standard deviation are significantly different
and %clearlyright and %clearlywrong state that the conditional modelling approach is advisable.
Comparing elasticity (Tables 9 and 10), it can be noted that elasticity values (for shared attributes) are quite different for
the three conditional switching models. In particular,
– The main effects of the attributes variation can be seen in car users switching models with regard to the cost attributes
and to the carsharing parking access time (see Table 9).
– In the case of carpool switching models, all attributes variations have a significant effect on the choice probability (see
Table 9).
– In the case of Bus switching models the variations of carsharing parking access time, the carsharing travel costs and the
bus travel costs have a significant effect on alternative choices (see Table 9).
Table 8
Validation protocol for conditional and unconditional switching models.
Models FF% MSE SD MAE %Right %Clearly
Right 0.9 Wrong 0.9
Conditional S
witching
Homo Car users 64 0.350 0.157 0.711 76 12 2
Carpool users 74 0.262 0.161 0.520 77 35 2
Bus users 77 0.144 0.021 0.453 76 23 3
Hetero Car users 66 0.346 0.190 0.676 76 17 3
Carpool users 78 0.136 0.021 0.432 76 26 3
Bus users 78 0.397 0.584 0.439 79 73 19
Resulting S
witching
Homo 74 0.199 0.061 0.516 75 21 2
Hetero 76 0.201 0.096 0.489 76 39 6
Unconditional S
witching
Homo linear 61 0.494 0.347 0.790 65 21 9
Homo non-linear 54 0.663 0.481 0.925 55 21 16
Hetero 60 0.540 0.436 0.797 64 27 13
Table 9
Direct elasticities: conditional switching models.
Car users Carpool users Bus users
Attribute ATTcsp TCcarsharing TCcar ATTcsp TCcarsharing TCcarpool ATTcsp ATTbs TCcarsharing TCbus TTbus
Homo S
witching
0.210 0.565 – 0.368 0.833 – 0.563 – 0.587 – –
NS
witching
–– 0.452 – – 0.792 – 0.069 – 0.292 0.063
Hetero S
witching
0.238 0.678 – 0.948 0.999 – 0.589 – 0.600 – –
NS
witching
–– 0.548 – – 0.998 – 0.073 – 0.298 0.065
72 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
Bus users’ switching models, as expected, are much more affected by access travel time to parking. Elasticity with regard
to carsharing travel costs (TCcarsharing) is similar for car and bus users’ switching models, and sensitivity increases for
carpool users. Moreover, elasticities with regard to the travel costs of used transport modes (car, carpool and bus travel costs)
are different and confirm the need for conditional models. Heteroscedastic models lead to slightly greater elasticity values.
With regard to unconditional switching models, the variation of travel costs and travel time has a significant effect on the
probability of the switching choice (Table 9). However, elasticity values, except for travel time, are about ten times smaller
than the conditional switching models. This result, together with the validation indicators, makes it possible to conclude that
the unconditional switching model cannot be an effective modelling solution.
With respect to the holding model, it is meaningless to compare the indicators from the validation protocol (their values
are shown in Table 11), whereas it is interesting to compare the direct elasticities.
As for the switching models, direct elasticities were computed by introducing the variation of the attributes equal to 20%
of the initial value (see Table 11). The more significant variations of attributes are in bold. Except for a few attributes, elas-
ticity values show values much smaller than those estimated for switching models. It should be said that the holding model
might show smaller elasticity due to the jointly calibration of RP and SP data, anyhow the elasticity values are significantly
different and significantly smaller the switching models.
As introduced in the previous section, the holding approach leads to rather rigid choice probabilities, thus it leads to low
generalisation capabilities. As for the validation indicators, the switching approach can be confirmed as being the most effec-
tive and advisable.
7. Conclusions
Although carsharing has become a consolidated transport alternative in many urban contexts, carsharing behaviours have
been mainly analysed through ex-post analysis and in terms of vehicle usage and/or ownership rate. In this paper carsharing
behaviour was investigated with regard to an inter-urban context and through an ex-ante approach based on a stated pref-
erences survey, and in terms of mode choice phenomenon within the random utility paradigm.
The aim was fourfold: (i) to analyse the feasibility of a inter-urban carsharing program; (ii) to investigate the main deter-
minants of the choice behaviour; (iii) to compare different approaches (switching vs. holding; conditional vs. unconditional);
(iv) to investigate different modelling solutions within the random utility framework (homoscedastic, heteroscedastic and
cross-correlated closed-form solutions).
Overall, results highlighted that the inter-urban carsharing service may be a substitute of the car transport mode, but also
a complementary alternative to the transit system.
The proposed models and the achieved insights indicate potential market segments and a way to model potential user’s
behaviour. The obtained results should be interpreted as a step towards understanding the ‘‘potential use of the service’’,
could be transferred to other similar case studies and may support viability/feasibility analyses (technical, economic and
Table 11
Validation protocol and direct elasticities for the holding model.
Indicators FF% MSE SD MAE %Right %Clearly
Right 0.9 Wrong 0.9
Holding 57 0.731 0.137 1.021 62 4 9
Attributes TCbus TCcarpool TCcar TCcarsharing ATTbs ATTcsp TTbus TTcarpool TTcar TTcarsharing
Alternatives Carsharing – – – 0.015 – 0.071 – – – 0.208
Car – – 0.008 – – – – – 0.020 –
Carpool – 0.011 – – – – 0.021 – –
Bus 0.768 ––– 0.197 –0.032 – –
Table 10
Direct elasticities: unconditional switching models.
Attribute ATTcsp ATTbs TCcarsharing TCothermodes TT
Homo Linear S
witching
0.0185 – 0.0420 – –
NS
witching
–0.0785 – 0.4937 0.3193
Non-linear S
witching
0.0250 – 0.0562 – –
NS
witching
–0.0773 – 0.4897 0.3158
Hetero S
witching
0.0097 – 0.0254 – –
NS
witching
–0.0935 – 0.5726 0.4112
S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76 73
environmental). However, they do not ensure the viability/feasibility of the investigated carsharing service, especially with
regard to the operator point of view, but it was out of the scope of this research.
Concerning the modelling approaches, all led to statistically significant results.
In the conditional switching approach, different systematic utility functions resulted in statistically significant findings.
Car and carpool users were characterised by systematic utility functions which were simpler than those of bus users, these
included different types of attributes: age, trip type (round trip vs. trip chain) and access time to buses. This result indicates
that if the usually chosen transport mode is similar to the carsharing alternative, the level of service attributes are the main
determinants of the choice process. On the other hand, bus users’ switching behaviour is more affected by the user’s specific
characteristics.
In the unconditional switching approach, geographical (location of trip origin) and economic (number of cars per house-
hold) segmentations have to be introduced to gain a statistically significant model. This result is coherent with expectations,
since a single choice model had to interpret different users’ choice processes. Nevertheless, though the goodness-of-fit was
comparable to the conditional models, the unconditional model showed worse generalisation capabilities. Indeed, the sen-
sitivity to the level of service attributes was much less than the conditional switching models. Therefore, it is possible to
conclude that the conditional switching approach is the approach to be pursued.
In the holding approach, the mix of RP and SP lead to the best validation indicators. Compared to the switching models,
age and trip frequency can be confirmed as being statistically significant, but the role of dummy or constant attributes
increased (alternative specific constants and inertia attribute) and, above all, validation indicators showed worse good-
ness-of-fit and smaller sensitivity to the level of service attributes. Thus, it is possible to conclude that the holding approach,
though statistically significant, is not the most effective solution for modelling carsharing choice behaviour in ex-ante sce-
narios. It should be better pursued in ex-post scenarios, where carsharing is already a perceived and known transport
alternative.
Among the investigated modelling solutions within the random utility theory and within each of the proposed
approaches, random parameter Mixed Multinomial Logit formulation was statistically significant. Depending on the specific
approach, only a two attributes, such as travel cost or access time to parking, turned out to be randomly distributed. How-
ever, explicit simulation of taste variation among users, where significant, did not lead to significant gain in the models’
goodness-of-fit, except for the holding models.
With regard to the attributes’ relevance, although in altering the modelling solution, the systematic utility functions
changed, estimation results highlighted the great importance of travel monetary cost and access time to carsharing parking
slots. Gender, age, trip frequency, car availability and the type of trip (home-based) influenced the probability of choosing
carsharing or not. Elasticity results for the conditional switching models showed that access time to the parking slots
predominantly influences choice probability for bus and carpool users; change in carsharing travel costs mainly affects
carpool users; change in travel costs of the usually chosen transport mode mainly affects car and carpool users. In
conclusions, (i) access time to parking slots is the most important design attribute; (ii) the carsharing service’s characteristics
should be specifically designed in order to deal with specific target groups (e.g. car users vs. bus users); (iii) socio-economic
and activity based attributes play a significant role compared to level of service attributes.
Finally, although the case study refers to a specific geographical context, the obtained results allow drawing some
operational conclusions.
First of all, it exists a market segment made by inter-urban commuters, usually not considered in carsharing market
analyses, that perceives and would choose carsharing as an alternative to the existing transport modes.
A carsharing service may give the proper flexibility to those demand flows not served by the transit service (not frequent
and continuous), thus it could be a complementary alternative to the transit system in those time periods in which the
service is not guaranteed or efficient.
The proposed methodological insights may be transferred to those contexts in which a carsharing service should be
beforehand designed.
Furthermore, the determinants of the choice process, if on the one hand are case-specific in the estimated values, on the
other hand do emerge a set of attributes which is slightly different from those acknowledged in urban contexts and which
may be generalised to similar inter-urban contexts.
Some research perspectives seem worthy of interest: investigating choice determinants in the case of non-systematic use
of carsharing, investigating the effect of the reliability of the service, investigating the effect of the non-availability of the
service due to the overbooking phenomena and the possible effects of carsharing on mobility behaviours (trip frequency
and trip-chaining). Finally, the implementation of different theoretical paradigms for modelling switching behaviours could
be of interest.
Acknowledgement
The author greatly appreciated the comments and suggestions received from the anonymous referees.
74 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
References
Abraham, J.E., 2000. Carsharing: a survey of preferences in carsharing. J. World Transport Policy Practice.
Barth, M., Shaheen, S.A., Fukuda, T., Fukuda, A., 2006. Carsharing and station cars in Asia. J. Transport. Res. Records 1986, 106–115.
Ben-Akiva, M., Lerman, S., 1985. Discrete Choice Analysis: Theory and Application to Travel Demand. The MIT Press, Cambridge, Massachusetts.
Ben-Akiva, M., Morikawa, T., 1990. Estimation of switching models from revealed preferences and stated intentions. Transport. Res. Part A 24 (6), 485–495.
Burkhardt, J.E., Millard-Ball, A., 2006. Who’s attracted to carsharing. Transport. Res. Record 1986, 98–105.
Cairns, S., Sloman, L., Newson, C., 2008. Smarter choices: assessing the potential to achieve traffic reduction using ‘soft measures’. Transport Rev. 28 (5),
593–618.
Cantarella, G.E., de Luca, S., 2005. Multilayer feedforward networks for transportation mode choice analysis: an analysis and a comparison with random
utility models. Transport. Res. Part C 13 (2), 121–155.
Cascetta, E., 2009. Transportation Systems Analysis: Models and Applications. Springer Verlag, US.
Cascetta, E., Cartenì, A., de Luca, S., 2014. A random utility model for park & carsharing services and the pure preference for electric vehicles. Res. Transport.
Econom. (in press).
Celsor, C., Millard-Ball, A., 2007. Where does carsharing work? using GIS to assess market potential. Transport. Res. Record 1992, 61–69.
Cervero, R., 2003. City CarShare: near-term travel-behavior impacts. J. Transport. Res. Records 1839, 159–166.
Cervero, R., Golob, A., Nee, B., 2007. City CarShare: longer term travel demand and car ownership impact. Transport. Res. Record 1992, 70–80.
Ciari, F., Schuessler, N., Axhausen, K.W., 2013. Estimation of carsharing demand using an activity-based microsimulation approach: model discussion and
some results. Int. J. Sustain. Transport. 7 (1), 70–84.
Costain, C., Ardron, C., Habib, N.K., 2012. Synopsis of users’ behavior of a carsharing program: a case study in Toronto. Transport. Res. Part A 46 (3), 421–434.
de Luca, S., Cantarella, G.E., 2009. Validation and comparison of choice models. In: Saleh, W., Sammer, G. (Eds.), Success and Failure of Travel Demand
Management Measures. Ashgate, UK, pp. 37–58.
Doherty, M.J., Sparrow, F.T., Sinha, K.C., 1987. Public use of autos: mobility enterprise project. J. Transport. Eng. 113 (1), 84–94.
Domencich, T.A., McFadden, D., 1975. Urban Travel Demand: A Behavioral Analysis. American Elsevier, New York.
Firnkorn, J., Müller, M., 2011. What will be the environmental effects of new free-floating car-sharing systems? the case of car2go in Ulm. Ecol. Econom. 70,
1519–1528.
Fujii, S., Taniguchi, A., 2006. Determinants of the effectiveness of travel feedback programs – a review of communicative mobility management measures for
changing travel behavior in Japan. Transport Policy 13, 339–348.
Fukuda, T., Fukuda, A., Todd, M., 2005. Identifying potential market of carsharing users through modal choice behaviour and socio-economic perspective: a
case study on Bangkok Metropolis. In: Proceedings of the 84th Annual Meeting of Transportation Research Board, Washington DC.
Garling, T., Fujii, S., 2009. Travel behavior modification: theories, methods, and programs. In: Kitamura, R., Yoshii, T., Yamamoto, T. (Eds.), The Expanding
Sphere of Travel Behaviour Research. Emerald, UK, pp. 97–128.
Habib, N.K., Morency, C., Tazul, I.M., Grasset, V., 2012. Modelling users’ behaviour of a carsharing program: application of a joint hazard and zero inflated
dynamic ordered probability model. Transport. Res. Part A 46 (2), 241–254.
Haefeli, U., Matti, D., Schreyer, C., Maibach, M., 2006. Evaluation Car-Sharing. Federal Department of the Environment, Transport, Energy and
Communications, Bern.
Huwer, U., 2004. Public transport and car-sharing-benefits and effects of combined services. Transport Policy 11, 77–87.
Katzev, R., 1999. Carsharing Portland: Review and Analysis of its First Year. Department of Environmental Quality, Portland, OR.
Katzev, R., 2003. Car sharing: a new approach to urban transportation problems. Anal. Social Issues Public Policy 3 (1), 65–86.
Kearney, A., De Young, R., 1996. Changing commuter travel behavior: employer-initiated strategies. J. Environ. Syst. 24, 373–393.
Lane, C., 2005. PhillyCarShare: first-year social and mobility impacts of carsharing in Philadelphia, Pennsylvania. J. Transport. Res. Board 1927, 158–166.
Litman, T., 2000. Evaluating carsharing benefits. J. Transport. Res. Board 1702, 31–35.
McFadden, D., Train, K., 2000. Mixed MNL models for discrete response. J. Appl. Econom. 15, 447–470.
Meijkamp, R., 1998. Changing consumer behaviour through eco-efficient services: an empirical study of carsharing in the Netherlands. Business Strategy
Environ. 7 (4), 234–244.
Morency, C., Trepanier, M., Agard, B., Martin, B., Quashie, J., 2007. Car sharing system: which transaction datasets reveal on users’ behaviors. In: Proceedings
of the 2007 IEEE Intelligent Transportation Systems Conference. pp. 284–289.
Morency, C., Habib, N.K., Grasset, V., Tazul, I.M., 2009. Application of a dynamic ordered probit model for predicting the activity persistency of carsharing
member. In: Proceedings of the 88th Annual Meeting of Transportation Research Board, Washington, DC.
Morency, C., Habib, N.K., Grasset, V., Tazul, I.M., 2012. Understanding members’ carsharing (activity) persistency by using econometric model. J. Adv.
Transport. 46 (1), 26–38.
Muheim Peter, 1998. CarSharing-der Schlüssel zur kombinierten Mobilität, Bundesamt für Energie, Energie 2000, Bern.
Nobis, C., 2006. Car sharing as a key contribution to multimodal and sustainable mobility behaviour – the situation of carsharing in Germany. Transport.
Res. Record 1986, 89–97.
Ortuzar, J.D.D., Willumsen, L.G., 2011. Modelling Transport, fourth ed. Wiley and Sons Ltd., UK.
Rodier, C.R., Shaheen, S., 2004. Carsharing and carfree housing: predicted travel, emission and economic benefits. In: Proceedings of the 83th Annual
Meeting of Transportation Research Board, Washington, DC.
Shaefers, T., 2013. Exploring carsharing usage motives: a hierarchical means-end chain analysis. Transport. Res. Part A 47, 69–77.
Shaheen, S., 1999. Dynamics in Behavioral Adaptation to a Transportation Innovation: A Case Study of CarLink-A Smart Carsharing System. Institute of
Transportation Studies UCD-ITS-RR-99-16.
Shaheen, S., Cohen, A., 2007. Growth in worldwide carsharing: an international comparison. J. Transport. Res. Board 1992, 81–89.
Shaheen, S., Cohen, A., 2013a. Carsharing and personal vehicle services: worldwide market developments and emerging trends. Int. J. Sustain. Transport. 7
(1), 5–34.
Shaheen, S., Cohen, A., 2013b. Innovative Mobility Carsharing Outlook: Carsharing Market Overview, Analysis, and Trends. Transportation Sustainability
Research Center at the University of California, <http://tsrc.berkeley.edu/node/629>.
Shaheen, S., Rodier, C.J., 2005. Travel effects of a suburban commuter-carsharing: a carlink case study. Transport. Res. Record 1927, 182–188.
Shaheen, S., Wright, J., 2001. The carlink II pilot program: testing a commuter-based carsharing program. In: Proceedings of IEEE Intelligent Transportation
System Conference, Oakland, USA. pp. 25–29.
Shaheen, S., Wright, J., Dick, D., Novick, L., 1999. Carlink-A smart carsharing system field test report. California Partner for Advanced Transit and Highways
(PATH) Research Report UCB-ITS-PRR-2000-10.
Shaheen, S., Cohen, A., Roberts, J.D., 2006. Carsharing in North America: market growth, current developments, and future potential. J. Transport. Res. Board
1986, 116–124.
Shaheen, S., Cohen, A., Chung, M., 2009. North American carsharing: 10-year retrospective. J. Transport. Res. Board 2110, 35–44.
Steininger, K., Vogl, C., Zettl, R., 1996. Car-sharing organization: the size of the market segmentation and revealed change in mobility behaviour. Transport
Policy 3 (4), 177–185.
Stillwater, T., Mokhtarian, P.L., Shaheen, S., 2008. Carsharing and built environment: a GIS-based study of one US operator. In: Proceedings of the 87th
Annual Meeting of Transportation Research Board, Washington, DC.
Vance, R., Rutherford, G., Anderson, C., 2005. Flexcar Seattle: Evaluation of Carsharing Program. In: Proceedings of the 84th Annual Meeting of
Transportation Research Board, Washington, DC.
S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76 75
Wagner, C., Shaheen, S., 1998. Carsharing and mobility management: facing new challenges with technology and innovative business planning. World
Transport Policy Practice 4 (2), 39–43.
Walb, C., Loudon, W., 1986. Evaluation of the short-term auto rental service in San Francisco, California. Prepared for the Urban Mass Transportation
Administration, Research and Special Programs Administration. Cambridge Systematics, Cambridge, MA.
Zheng, J., Scott, M., Rodriguez, M., Sierzchula, W., Platz, D., Guo, J., Adams, T., 2009. Carsharing in a university community: assessing potential demand and
distinct market characteristics. Transport. Res. Record: J. Transport. Res. Board 2110, 18–26.
Zhou, J., 2012. An analysis of university employee car-sharers in Los Angeles. Transport. Res. D: Tr. E. 17 (8), 588–591, ISSN 1361-9209.
Zhou, B., Kockelman, K.M., Gao, R., 2011. Opportunities for and impacts of carsharing: a survey of the Austin, Texas Market. Int. J. Sustain. Transport. 5 (3),
135–152.
76 S. de Luca, R. Di Pace / Transportation Research Part A 71 (2015) 59–76
