Examining the impact of car-sharing on private vehicle ownership

Abstract

Car-sharing has experienced a significant boom in recent years, with estimates suggesting that car-sharing programs are now operating in over 30 countries worldwide, serving around five million users. The potential to reduce individual vehicle ownership rates is frequently cited as a motive for promoting car-sharing. While some previous studies have argued that customers are indeed willing to reduce the total number of vehicles owned after becoming car-sharing members, the reliability of these findings is tenuous given that many are based on self-selected samples of car-sharing users, resulting in biased estimates. In theory, the availability of car-sharing programs could have limited effect on the general public’s car ownership decisions–or at least have no effect on a large portion of travelers. Whether or not traveler decision processes are significantly influenced by specific attributes of different car-sharing options (e.g., access time, vehicle size, fuel type, etc.) remains an unanswered question, as there are limited quantitative studies on this issue. To contribute to filling this research gap, this paper discusses the findings of a study of 1,500 private households across major Australian cities. A nested logit model is used to investigate the impacts of car-sharing on respondents’ household vehicle ownership decisions. In contrast to the results of some previous studies, we find that the stated availability of car-sharing appears to have minimal impact on respondents’ decision to own a vehicle or not, leading to important policy implications. We agree with prior investigations that car-sharing could potentially reduce private car ownership. However, because this study finds limited impact of the availability of car-sharing on vehicle ownership, and because the majority of respondents did not self-identify as car-sharing users, education and awareness campaigns could be important factors in improving the general public’s preferences towards car-sharing and fully realizing car-sharing’s benefits to society.

Full text

Examining the impact of car-sharing on private vehicle ownership
Fan Zhoua,e, Zuduo Zhengb,∗
, Jake Whiteheadb,g, Robert K. Perronsc,f, Simon
Washingtonb, Lionel Paged
aZhijiang Institute for Statistics and Big Data, Zhejiang Gongshang University, China
bSchool of Civil Engineering, University of Queensland, Australia
cQueensland University of Technology, Brisbane, Australia
dEconomics Discipline Group, University of Technology Sydney, Australia
eInternational Business School, Zhejiang Gongshang University, China
fCentre for Strategy and Performance, University of Cambridge, United Kingdom
gThe UQ Dow Centre for Sustainable Engineering Innovation
Abstract
Car-sharing has experienced a significant boom in recent years, with estimates suggesting
that car-sharing programs are now operating in over 30 countries worldwide, serving around
five million users. The potential to reduce individual vehicle ownership rates is frequently
cited as a motive for promoting car-sharing. While some previous studies have argued that
customers are indeed willing to reduce the total number of vehicles owned after becoming
car-sharing members, the reliability of these findings is tenuous given that many are based
on self-selected samples of car-sharing users, resulting in biased estimates. In theory, the
availability of car-sharing programs could have limited effect on the general public’s car
ownership decisions—or at least have no effect on a large portion of travelers. Whether or
not traveler decision processes are significantly influenced by specific attributes of different
car-sharing options (e.g., access time, vehicle size, fuel type, etc.) remains an unanswered
question, as there are limited quantitative studies on this issue. To contribute to filling
this research gap, this paper discusses the findings of a study of 1,500 private households
across major Australian cities. A nested logit model is used to investigate the impacts of
car-sharing on respondents’ household vehicle ownership decisions. In contrast to the results
of some previous studies, we find that the stated availability of car-sharing appears to have
minimal impact on respondents’ decision to own a vehicle or not, leading to important policy
implications. We agree with prior investigations that car-sharing could potentially reduce
private car ownership. However, because this study finds limited impact of the availability
of car-sharing on vehicle ownership, and because the majority of respondents did not self-
identify as car-sharing users, education and awareness campaigns could be important factors
in improving the general public’s preferences towards car-sharing and fully realizing car-
sharing’s benefits to society.
Keywords: car-sharing, vehicle ownership, shared economy, shared autonomous vehicles
(SAVs)
Preprint submitted to Transportation Research Part A May 15, 2020

1. Introduction1
In recent years, shared mobility schemes, particularly car-sharing, have received signifi-2
cant interest from transport researchers, planners, and policymakers due to the challenges3
that communities face from the continued growth of vehicle ownership and usage, along with4
the associated consequences such as increased traffic congestion, parking congestion, and air5
pollution (greenhouse and particulate emissions). Car-sharing programs have the potential6
to reduce both personal vehicle usage and rates of ownership, as well as to encourage indi-7
viduals to use alternative modes of transport (e.g., public transport, cycling, walking, etc.)8
more frequently (Martin and Shaheen,2011).9
The ability to share the fixed costs of vehicle ownership represents the principal economic10
benefit of car-sharing (Duncan,2011). Through car-sharing programs, individuals gain the11
benefits equivalent (or close to equivalent) to owning a private vehicle, without the burden of12
complete ownership (e.g., insurance, maintenance, etc.). Particularly for those low-income13
households that cannot afford to own a private vehicle, they can benefit from increased14
mobility in highly automobile-dependent societies through the availability of car-sharing15
(Litman,2000). The group that could benefit the most from car-sharing can be easily16
quantified using census data. For example, according to the Australian Bureau of Statistics17
(ABS,2011), 8.4% of households in Australia did not own a private vehicle in 2011. In18
the U.S., 10% of households had no access to private vehicles as of 2000 (Duncan,2011).19
By implementing car-sharing services, non-car owners could benefit from increased options20
and/or flexibility. More specifically, these consumers will benefit from easy access to vehicles21
when they need them, without the financial burden of vehicle ownership, which is likely to22
increase their utility.23
In addition to carless households, the financial benefits of car-sharing also influence24
car owners who only drive occasionally. For instance, as shown by Litman (2000) and25
Prettenthaler and Steininger (1999), vehicle owners would be better off switching to car-26
sharing from private cars if they drive less than 10,000 km (15% of vehicles) and 15,000 km27
(69% of households) per year in the U.S. and Austria, respectively.28
One of the main streams in car-sharing research is to study its impacts on personal vehicle29
ownership. Although considerable effort has been made to measure the effects of car-sharing30
(Martin and Shaheen,2011;Martin et al.,2010;Huwer,2004;Zhou et al.,2017), there are31
several challenges. Firstly, prior studies have usually been drawn on data from existing32
car-sharing organizations or operators (i.e., revealed preference data). Thus all respondents33
were already car-sharing members, and most of them do not own a car (Martin et al.,34
2010). Self-selection bias could arise in this situation, more specifically, the early adopters35
of car-sharing who self-selected themselves into the group were also found to be more envi-36
ronmentally conscious and willing to commit to more sustainable behaviors (Costain et al.,37
2012). Therefore, the prior findings may be over-optimistic about the impacts of car-sharing38
on household vehicle ownership, and the availability of car-sharing programs could have only39
minimal or even no effect on vehicle ownership of general public.40
∗Corresponding author: Zuduo Zheng
Email address: zuduo.zheng@uq.edu.au (Zuduo Zheng)
2

Secondly, there is a lack of quantitative studies on the impacts of car-sharing on vehicle41
ownership. While prior vehicle ownership studies rarely incorporated car-sharing, the car-42
sharing focused research is insufficient due to the reason listed in the first point (i.e., sample43
selection biases). To expand the scope to be representative of the general public, stated44
preference analysis could be implemented, due to the low market penetration of car-sharing.45
Thirdly, there is a general lack of studies on the impacts of the shared autonomous46
vehicles (SAVs) on vehicle ownership from the perspective of consumers. It is argued that47
the advent of SAVs has the potential to reduce the level of car ownership, for example,48
(Schoettle and Sivak,2015) state that SAVs could reduce up to 43% of vehicle ownership.49
However, the full benefits could only be achieved if consumers are willing to give up their50
private vehicles for SAVs.51
The primary objective of this paper is to quantitatively investigate whether conventional52
car-sharing programs and SAVs have significant impacts on people’s vehicle ownership deci-53
sions, and whether previous studies have over-estimated the potential impact of car-sharing54
programs. Towards this end, a national online survey was developed and administrated cross55
major Australian cities between July and August of 2016. The respondents considered were56
a sample of the general public rather than only car-sharing members. A nested logit (NL)57
model is employed to analyze the data.58
The next section provides a review of the literature. Section 3presents the methodolog-59
ical design of this study including the survey plan, data collection, and the model construc-60
tion. Sections 4and 5describe the descriptive analysis and the choice modeling results,61
respectively. Finally, Section 6concludes by summarizing and discussing the main findings.62
2. Literature review63
Households’ decisions on vehicle ownership have been extensively studied in the litera-64
ture. Mainstreams of household vehicle ownership models include vehicle purchase (VP),65
vehicle holding (VH), and vehicle transaction (VT) models (de Franca Doria et al.,2009;66
Anowar et al.,2014). The VP models describe the likelihood that a household will decide to67
purchase a new vehicle with given attributes (e.g., price, make, performance, environmental68
characteristics, vehicle attributes, etc.) or predict the choice of the most recent purchased69
vehicles associated with the vehicle attributes. These models have successfully quantified70
consumers’ preferences toward private vehicles, which in turn produces market share esti-71
mates (Paleti et al.,2013b;Lave and Train,1979;Mannering and Mahmassani,1985;Choo72
and Mokhtarian,2004;Brownstone et al.,2000;Hensher and Greene,2001;Sierzchula et al.,73
2014;Rezvani et al.,2015;Gallagher and Muehlegger,2011). Moreover, the VH models74
explore the probability that a household with given characteristics will own a particular75
number of vehicles at some point of time. The households’ vehicle holding levels are found76
to be correlated with residential location, income, household size, and location (Bhat and Pu-77
lugurta,1998;Hanly and Dargay,2000;Baldwin Hess and Ong,2002;Whelan,2007;Paleti78
et al.,2013a). In addition, the VT models concern the change of the total number of vehi-79
cles owned by households, such as replacement and disposal decisions. For example, a nine80
year records of household vehicle holding data was analysed by Mohammadian and Miller81
3

(2003), Mohammadian and Rashidi (2007), and Rashidi and Mohammadian (2016) to study82
the transaction timing of consumers in Toronto area, the authors argue that the increased83
number of licensed drivers, member with higher education, employed workers is likely to trig-84
ger vehicle acquisition but a decrease in family members is likely to lead to vehicle disposal.85
These models are used to investigate the impacts of life-changes (e.g., increased/decreased86
adults or children) on vehicle transaction or to forecast the change of ownership and the87
duration of the transactions (Yamamoto,2008;Gilbert,1992;de Franca Doria et al.,2009).88
However, car-sharing has rarely been considered in existing vehicle ownership studies.89
Studies that focus on car-sharing’s impacts on vehicle ownership usually rely on real-90
world data from car-sharing operators. More specifically, based on investigations of car-91
sharing users only and by comparing the total number of households’ private vehicles before92
and after joining car-sharing programs. Prior studies argue that car-sharing effectively93
reduced respondents’ private vehicle ownership. For instance, Martin and Shaheen (2011)94
and Martin et al. (2010) surveyed car-sharing members in the U.S., asking about changes95
in private car ownership after joining car-sharing programs. The authors observe that the96
total vehicles owned by the 6,281 households surveyed decreased from 2,968 to 1,507. In97
the European context, it was also observed that car-sharing leads to a decrease in vehicle98
ownership, as summarized in Katzev (2003), a decline of 44%, 60%, and 25% are reported99
by car-sharing members in Netherlands, Switzerland, and Canada respectively. However,100
these findings only reflect the behavior of early adopters and are not representative of the101
general public, due to sample selection bias or self-selection bias (Heckman,1977;Rodier and102
Shaheen,2003;Willis and Rosen,1979). Car-sharing users that self-selected themselves into103
the car-sharing user group were also found to be more educated, wealthier, environmentally104
conscious, and innovative individuals who are willing to commit to social activities and try105
new products (Costain et al.,2012;Burkhardt and Millard-Ball,2006;Rodier and Shaheen,106
2003;Shaheen and Rodier,2005). Therefore, focusing on car-sharing users, prior studies107
expected too much of the impacts of car-sharing on vehicle ownership.108
As discussed previously, the majority of the studies focused on car-sharing are descriptive.109
Kim et al. (2017) is one of the very limited studies that quantitatively investigated car-110
sharing’s impacts on vehicle ownership. The authors considered vehicle purchasing decisions111
with car-sharing by adding ‘Buy 2nd car’ as an alternative to ‘Join car-sharing.’ However,112
the authors ignored the selling decision as an alternative, and it is therefore doubtful whether113
‘Buy 2nd car’ and ‘Join car-sharing’ are appropriate alternatives because these two options114
are not mutually exclusive. More specifically, car-sharing is complementary to our current115
transport systems, for instance, consumers can use car-sharing for shopping but drive their116
private car for other trips. Buying the second car does not mean that a consumer will not use117
car-sharing and similarly join car-sharing does not strictly prohibit a consumer from buying118
a second car. Therefore, our understanding of car-sharing’s impacts could be advanced119
by implementing a quantitative study that explains the effects of each car-sharing related120
attribute on respondents’ vehicle ownership decisions.121
Furthermore, regarding SAVs, one of the early assessments of SAVs is made by Fagnant122
and Kockelman (2018). The authors considered the potential impacts of SAVs, including re-123
duced crashes and accidents, thus fewer associated injuries, less congestion, and the reduced124
4

need for parking, etc. Thus, the benefits of each SAV per year is estimated to be about USD125
$2,960 (10% market share of SAV assumed) and up to $3,900 (90% market share of SAV126
assumed) in the American context. Schoettle and Sivak (2015) claim that multiple residents127
can be served by one SAV and they estimated that the SAVs could reduce vehicle ownership128
by up to 43% and increase the usage of each remaining vehicle by up to 75%. This, in turn,129
cuts the travel costs, reduces vehicle ownership cost and parking cost, and consequently130
leads to a reduction in parking spaces needed (Burns,2013;Fagnant and Kockelman,2015;131
Schonberger and Gutmann,2013). Several studies (i.e., Chen et al. 2016;Fagnant and Kock-132
elman 2014;Zhang et al. 2015) agree that SAVs lead to a decrease in vehicle ownership, in133
the meantime, maintaining an equal level of mobility supply. However, there is a lack of134
studies on the impacts of the SAVs on vehicle ownership from the consumers’ perspective.135
It is argued that the advent of SAVs has the potential to reduce the level of car ownership,136
however, the full benefits of SAVs can only be achieved if consumers are willing to give up137
their private vehicles and adopt these new services.138
3. Methodology139
3.1. Data collection140
This study is based on data collected in a large-scale survey conducted in late 2016.141
The online survey was implemented across major Australian cities–including Sydney, Mel-142
bourne, Brisbane, and Adelaide—which was completed by 1,433 respondents. Four responses143
(repeated SP experiment) were collected from each of the respondents, resulting in 5,732144
observed scenarios. It is known that asking too many questions could impose a significant145
cognitive burden to respondents, which in turn reduce both the response rate and reliability146
(Hensher et al.,2015). After two rounds of pilot test and discussion within the research147
team and external experts, we believe the workload of a survey contains four SP experi-148
ment, which requires 20 - 25 minutes on average to complete, is reasonable. Great efforts149
have been made to maximize the coverage across both the combinations of alternatives (i.e.,150
scenarios) and to maximize the variability of the alternative attributes, to force trade-offs.151
The objective was to collect 1500 responses, we should have approximately 11-12 completed152
responses per choice task if each respondent was presented with four unique choice tasks (i.e.,153
1500 ×4/512 = 11.72, 512 unique choice tasks were generated). Standard practice usually154
mandates around 20 completes per choice task, but given the complexity of the experiment155
design, we believe 11 completes per choice task should suffice for this study. The survey156
collected both revealed preferences (RP) and stated preferences (SP) on vehicle buy/sell de-157
cisions. RP information is useful in gaining an understanding of consumers’ current choices158
(or choices that have been made by the consumers in the past), and what factors are likely159
to influence their current choices; SP information, on the other hand, are used to obtain160
consumers’ reactions to choices presented to them about the future (e.g., self-driving vehi-161
cles). As suggested by Zheng et al. (2016), SP is popular due to the relatively low associated162
time and financial cost, and it is particularly useful when new attributes and alternatives163
are involved—like, for example, the self-driving capability and SAVs included in this study.164
5

Specifically, the survey consisted of four main parts. In the first part, household informa-165
tion, and each household member’s travel and activity patterns were collected. The second166
part obtains information regarding vehicles (e.g., car and two-wheelers) currently owned by167
respondents. The third part of the survey featured a stated choice experiment (described168
in detail in Section 3.3). The last part of the survey collected demographic and household169
background information of the respondents.170
A video clip was provided at the beginning of the SP experiment stage (the third part) to171
give essential and objective information on the new technologies (e.g., self-driving) and new172
mobility mode (e.g., car-sharing) to familiarize respondents with the terms before they were173
asked to make decisions on future vehicle choices. Respondents were required to watch the174
full video before proceeding to the SP experiment. The video was important in improving175
respondents’ understanding of the new technology and novel mobility options, including176
car-sharing, given it has been shown that the awareness of car-sharing is usually low. For177
instance, a car-sharing survey conducted in Germany in 2002 showed that only 15% of178
the respondents could describe what car-sharing was or name or refer to a car-sharing179
organization (Loose et al.,2006). Thus, providing this video introduction helped to improve180
the quality of the survey responses. Note that precaution was exercised to ensure only factual181
information was provided to respondents to avoid introducing any potential bias from the182
research team. Fig. 1a to Fig. 1c present a series of screenshots of the video1.183
Figure 1a: Video screenshot of booking a shared car
Prior to launching the main survey, two pilots and a focus group were conducted, and184
some questions in the final survey were modified according to comments and feedback re-185
ceived. The final survey was conducted between 12 July and 15 August 2016. The final186
survey was implemented across major Australian cities, which includes Melbourne, Sydney,187
Brisbane, etc. To ensure the representativeness of the dataset, gender and age cohorts were188
sampled to achieve consistency with the latest population census from the Australian Bureau189
of Statistics (ABS,2016).190
1The full video is available at https://youtu.be/8sD0ymSj4j0. A company called Explanimate!
(www.explanimate.com.au) was hired to convert the written script into an online animation.
6

Figure 1b: Video screenshot of booking a shared self-driving car
Figure 1c: Video screenshot of car-sharing service
In total, 2,015 respondents started the survey, and 1,433 of them completed it. The191
demographic details of the respondents are summarized in Table 1. As seen in this table,192
the gender-split is reasonably even between males and females. The majority of the decision-193
makers were between 25 and 54, which accounted for 68.8% of the respondents. Besides,194
most of them had a higher education degree (57.7%), which represented a relatively highly195
educated sample. The proportion of well-educated respondents was relatively high because196
only adults were included in this survey. Most of the respondents came from the capital cities197
in Australia including 33.6% from Sydney, 26.53% from Melbourne, 19.47% from Brisbane,198
11.67% from Perth, and the rest 8.78% of respondents came from other cities. The sample is199
reasonably representative of the national population in relation to gender, age, and education200
distributions.201
3.2. Data cleaning202
As mentioned previously, the total number of scenarios presented was 5,732. The choice203
scenarios in this study were personalized to increase the realism of the SP experiment (and204
in turn, the reliability of the SP responses). The maximum options of hypothetical vehicles205
that could be presented to respondents were two, thus the new vehicle alternative(s) could be206
zero, one, or two. Additionally, the number of current vehicles were subject to respondents’207
7

Table 1 Respondents profile
Gender Frequency Breakdown ABS (2016)
Male 745 51.99% 49.3%
Female 678 47.31% 50.7%
Other 4 0.28% -
Not answer 6 0.42% -
Age
Under 18 6 0.42% 24.8%
18-24 162 11.30% 6.7%
25-34 332 23.17% 14.4%
25-44 356 24.84% 13.5%
45-54 299 20.87% 13.3%
55-64 251 17.52% 11.8%
65-74 26 1.81% 8.9%
75 or older 1 0.07% 6.9%
Education
Postgraduate 254 17.73% -
Graduate 170 11.86% -
Bachelor 403 28.12% 22%(Bachelor and above)
Diploma 198 13.82% 8.9%
Certificate 185 12.91% 15.8%
None 223 15.56% -
City
Sydney 476 33.22% -
Melbourne 381 26.59% -
Brisbane 282 19.68% -
Perth 169 11.79% -
Other cities 125 8.72% -
Total 1433 100% -
actual vehicle holding reported, and the maximum number of vehicles recorded was five. As208
shown in Table 2, in only 1.9% (109) of the experiment, respondents stated they would like209
to sell more than two vehicles. In order to control the complexity of the model, these 109210
observations were removed, leaving an updated sample size of 5,623.211
Furthermore, as shown in Table 3, in only 2.93% (or 165) of the experiment, respondents212
stated they would like to buy two hypothetical vehicles. Similarly, to control the complexity213
of the model, these 165 observations were eliminated and resulted in a sample size of 5,458.214
Another 89 observations were removed because neither current vehicles nor hypothetical215
vehicles were available in these cases, that is, those respondents do not own private vehicles,216
and in the experiment, future vehicle option was not provided to them by design. Thus, the217
final sample size used in the analysis was 5,369.218
In the final sample, the maximum number of current vehicles and future vehicles are219
restricted at two and one, respectively. Therefore, respondents could decide to sell maximum220
8

Table 2 Number of vehicles respondents stated to sell
Car sold Frequency Percentage Cumulative
0 3529 61.57% 61.57%
1 1630 28.44% 90.00%
2 464 8.0% 98.10%
3 87 1.52% 99.62%
4 20 0.35% 99.97%
5 2 0.03% 100%
Total 5732 100% 100%
two current vehicles (i.e., zero, one or two) and/or buy at most one hypothetical vehicle.221
Table 3 Number of vehicle respondents stated to buy
Car bought Frequency Percentage Cumulative
0 4174 74.23% 74.23%
1 1284 22.83% 97.07%
2 165 2.93% 100%
Total 5623 100% 100%
3.3. Stated preference experiment222
The SP experiment aimed to investigate whether the availability of car-sharing affected223
households’ car ownership decisions. After watching the information video, respondents are224
required to answer a series of questions regarding their decisions on vehicle acquisition and/or225
disposal in four distinct scenarios. An example of the vehicle ownership choice SP experiment226
is demonstrated in Fig. 2. In this scenario, the respondents had one existing vehicle,227
which is a petrol-powered Green Corolla with an estimated current value of AU$5,500. The228
respondent is also presented with two future vehicles (i.e., vehicles hypothetically available229
on the market). Therefore, three alternatives in total are provided in this experiment. One230
of the future cars is a plug-in hybrid electric vehicle, and another is a conventional vehicle.231
Both the future cars are fully self-driving with the same operating cost (i.e., $6 per 10 km)232
and driving range (i.e., 500 km). However, the plug-in hybrid electric car is more expensive233
but also more environmentally friendly than the petrol-powered car.234
In addition to the private car alternatives, three other transport modes, which includes235
car-sharing, bike and walk, are also included in this the experiment along with their at-236
tributes; for example, the shared car in this example (see Fig. 2) is a very small conventional237
vehicle with no automatic features. Also, attributes such as access distance, waiting time,238
cost, etc. are included. To avoid the mutually non-exclusive issue, this kind of information239
(i.e., other modes) existed, by design, only to remind respondents of the availability and240
associated features of these modes, rather than treating them as alternatives. This is also241
9

an attempt to mimic the real world, where people are aware of alternative mobility modes242
and these alternatives are compatible with the travel modes they owned privately.243
Due to the complexity of the SP tasks, a tutorial-style instruction on how to respond to244
vehicle choice experiment is provided. As shown in Fig. 2, animated instructions with word245
descriptions were provided to guide respondents to complete the experiment step-by-step.246
Respondents are required to go through the example SP experiment before allowed to start247
doing the formal experiment.248
Figure 2: Sample screenshot of the stated preference experiment of car ownership choices
The level of the alternative-specific attributes is carefully designed to ensure each hypo-249
thetical scenario is both efficient and realistic. To improve the quality (or realism) of the SP250
experiment, whenever the real data are available (e.g., peak waiting time), the attributes’251
levels are pivoted from the revealed data by plus and minus a certain percentage (e.g., +/-252
50%). When the real data are unavailable (e.g., the purchase price of a hypothetical vehicle),253
10

levels are pivoted from comparable real data (e.g., the price of the current vehicle). The254
variations across different levels are sufficiently large in order to fully capture all potential255
responses from respondents.256
The selected attributes and associated levels are summarized in Table 4. As shown in this257
table, the price (or value) of current vehicles are defined by respondents, and the price for258
future vehicles ranged from AUD 20,000 to 60,000 in increments of AUD 10,000 (five levels).259
The average refuel distance also had five levels, ranging from 15 km to 100 km for both260
existing cars and hypothetical cars. Environmental impact is not applicable to respondents’261
existing cars, but there are four levels for the future vehicle. Level one denotes the least262
environmentally friendly cars, whereas level four indicates the most environmentally friendly263
options. Attributes such as self-driving capability are not applicable to current vehicles.264
The vehicle size attribute has four levels classified by the interior capacity of the vehicles,265
including micro vehicles (e.g., Mazda 2), compact vehicle (e.g., Ford Focus), Large vehicle266
(e.g., Toyota Camry), and SUV (e.g., Mitsubishi Outlander). The names will stick with the267
vehicle size, for example, respondents who encountered micro vehicle alternative will always268
see “micro e.g., Mazda 2”. We provided a real-world example for each vehicle size to avoid269
potential misunderstanding; otherwise, people might have a different interpretation of the270
vehicle size.271
Regarding car-sharing related attributes, “Car-sharing option” is an indicator variable272
that takes value one if car-sharing is presented to a respondent in the SP experiment as an273
option and zeroes otherwise (recall that the choice sets were personalized). Fig. 2 provides274
an example of the scenarios where car-sharing is presented to some respondents as an option.275
Also, a range of associated car-sharing attributes was provided to respondents if the car-276
sharing option was available. The “car-sharing driving range” has five levels ranged from277
100 km to 500 km. “car-sharing micro-car”, “car-sharing electric car” and “car-sharing278
self-driving capability” are all indicator variables that take value one, if the shared car is a279
micro-car (e.g., Mazda 2), electric car, or autonomous, otherwise zero. The average waiting280
time for car-sharing service is a continuous variable ranging from 3 min to 38 min.281
In addition, commute distance could be an important explanatory variable. On the282
other hand, however, it would be difficult for the respondents to estimate their commute283
distance if we require them to self-report it. Moreover, if we provide an attribute of commute284
distance with several levels in the experiments, some values could be unreal or irrelevant285
to some respondents. For example, some people may use vehicles only for other purposes286
like shopping, education, recreation, etc. Instead, some other variables such as ‘average287
operating cost’, ‘average distance from home’, and ‘average distance from work’ probably288
could be viewed as a proxy of the distance and time respondents travel, as in order to289
calculate their total cost, respondents might need to (explicitly or implicitly) recall the total290
distance they traveled.291
The last three attributes in Table 4are control variables. “Bought two-wheeler” and292
“Sold two-wheeler” indicate that a respondent also decided to change their two-wheeler293
ownership in an experiment. “Household size≥3” is a binary variable which takes the value294
of one if a household had three or more members, and zeroes otherwise.295
11

Table 4 Selected attributes and their levels
Attribute Current Vehicle (CV) Future Vehicle (FV)
Price Defined by respondents 20k; 30k; 40k; 50k; 60k
Average refuel distance 15km; 25km; 50km; 75km; 100km 15km; 25km; 50km; 75km; 100km
Environmental impact −∗1, 2, 3, 4#
Fuel type −Petrol; Petrol-like; Hybrid; Electric
Self-driving capability −Yes=1; No=0
Operating cost −$4, $5, $6, $7, $8 per 10km
Petrol powered vehicle −Yes=1; No=0
Policy: free registration fee −Yes=1; No=0
Policy: free EV public charging −Yes=1; No=0
Policy: $5000 tax rebate −Yes=1; No=0
Policy: 50% toll road discount −Yes=1; No=0
Policy: free use of transit + bus lanes −Yes=1; No=0
Vehicle size (interior capability) Micro 1, Compact 2, Large 3, SUV 4
Fuel type market share 2% - 80%
Car sharing option Yes=1; No=0
Car sharing access distance from home Yes=1; No=0
Car sharing access distance from work Yes=1; No=0
Car sharing average peak waiting 3min to 38min
Car sharing driving range 100km; 200km; 300km; 400km;500km
Car sharing policy: free use of transit + bus lanes Yes=1; No=0
Car sharing fuel type Petrol; Petrol-like; Hybrid; Electric
Car sharing self-driving capability Yes=1; No=0
Car sharing environmental impact 1, 2, 3, 4
Household size≥3 Yes=1; No=0
Bought two-wheeler Yes=1; No=0
Sold two-wheeler Yes=1; No=0
*: “−” means that this attribute is not applicable.
#: 1 denotes least environmental friendliness; 4 denotes most environmental friendliness.
Ngene2(ChoiceMetrics,2014) was used to design the SP scenarios, a balance fractional296
orthogonal design was used to construct 32 unique market structure scenarios. This is297
known as the availability design (as it determines what alternatives are available under a298
given scenario). Two properties were hold in this step: balance and orthogonality, meaning299
that each level corresponding to an alternative appears an equal number of times, and the300
availability of each alternative is independent of the availability of any other alternative.301
The attributes and their levels were kept balanced and independent based on the ideas302
of ‘utility balance’(Huber and Zwerina,1996). Utility balance refers to choice scenarios303
in which respondents have similar utilities for each of the available alternatives, and are304
therefore faced with a difficult choice. Difficult choices force respondents to make trade-offs305
that they would not have had to make in the case of easy choice tasks where one alternative306
is clearly preferred. Consequently, difficult choice tasks implicitly reveal more about the307
underlying structure to the respondent’s preferences.308
2“Ngene is software for generating experimental designs that are used in stated choice experiment for the
purpose of estimating choice models, particularly of the logit type”, Ngene 1.1.1 User Manual & Reference
Guide.
12

3.4. Car-sharing availability model and Car-sharing effect model309
Two models were estimated to investigate the impacts of car-sharing on households’ ve-310
hicle ownership: (1) Car-sharing availability model, and (2) Car-sharing effects model. The311
outcome variables (alternatives) are the same for the two models. The car-sharing avail-312
ability model (i.e., first model) is based on the full sample (i.e., N=5,369) to investigate313
the impacts of the car-sharing option on car ownership. In order to answer this question, a314
straightforward and effective way is to include a single variable related to car-sharing in the315
model: an indicator for car-sharing availability. Principal Component Analysis (PCA) was316
employed to construct the indicator, it also ensures car-sharing attributes are incorporated317
as much as possible. The PCA generated indicator could be used in the model to represents318
the car-sharing programs with different combinations of attributes levels. More specifically,319
range, environmental effect, refuel distance, cost, access distance from home, access distance320
from work, average waiting time, car-sharing market share are considered in PCA. As shown321
in Table 5, the results suggest that the first component with an eigenvalue of 5.777 explains322
72.2% of the total variance, which is very close to the 75% threshold (Kaiser,1960;Mor-323
rison,1990). To control the complexity of the model, we decided to only include the first324
component.325
Table 5 PCA of car-sharing attributes
Eigenvalue Cumulative
Component 1 5.777 .722
Component 2 .542 .79
Component 3 .424 .843
Component 4 .349 .886
Component 5 .323 .927
Component 6 .258 .959
Component 7 .219 .986
Component 8 .109 1
The car-sharing effect model (i.e., second model) focuses on the 2,707 scenarios with326
car-sharing option (recall that among the 5,369 SP experiment, the car-sharing option is327
presented to respondents in only 2,707 of the scenarios). This model tested whether the328
car-sharing related attributes (e.g., driving range, self-driving capability, wait time, etc.)329
affected respondents’ decisions on private vehicles.330
Instead of using two models, an alternative approach is to combine the two models by331
incorporating the car-sharing attributes into the first model mentioned above (car-sharing332
availability model), more specifically, a series of interactions between the car-sharing option333
dummy and the attributes could be included in this model. Compared to model 2 pro-334
posed above, this combined model makes use of the full data and the effect of car-sharing335
option is differentiated according to each attribute, however, this is incapable of answering336
the research question that whether the availability of car-sharing has impacts on private337
vehicle ownership directly. This question has seldom been quantitatively investigated in338
13

the literature (recall that previous studies usually compare the household car holdings be-339
fore and after joining car-sharing programs). Therefore, we believe that more information340
could be extracted from the data with two separate models. Nevertheless, the results of this341
alternative model are presented in Appendix A.342
A nested logit (NL) model is employed to model the data collected via the vehicle own-343
ership SP experiment. By allowing the random components to be correlated within each344
subset (or “nest”) of alternatives, the NL model (Ben-Akiva,1973) relaxes the restrictive345
independence of irrelevant alternatives (IIA) assumption of multinomial logit model (MNL).346
The NL model requires the alternatives to be constructed into a decision tree, in other words,347
the alternatives need to be split into levels (or groups). NL is convenient when the choice348
alternatives are multidimensional in nature, for instance, vehicle type and mode choices.349
3.5. Decision tree structure350
The maximum number of cars respondents can decide to sell or buy is two and one in the351
SP experiment, respectively. Therefore, there are in total six alternatives as shown in Table352
6: D2 - selling two current vehicles while buying no future vehicle, D1 - selling one current353
vehicle but buying no future vehicle, RPD1-selling two current vehicles while buying one354
future vehicle, RP - selling one current vehicle while buying one future vehicle, SQ (status355
quo)-doing nothing, and BUY - buying one future car without selling any current vehicles.356
Each alternative’s frequency in the survey and its consequence on the overall change of the357
vehicle ownership are also provided in Table 6.358
Table 6 Alternatives of the vehicle ownership model
Alternative Current car Future car Overall change Frequency Percentage
D2 Sell 2 No change Decrease 2 141 2.58%
D1 Sell 1 No change Decrease 1 896 16.42%
RPD1 Sell 2 Buy 1 Decrease 1 157 2.88%
RP Sell 1 Buy 1 No change 731 13.39%
SQ No change No change No change 3137 57.48%
BUY No change Buy 1 Increase 1 396 7.26%
Different decision trees for the NL model can be constructed depending on how the359
unobserved attributes are shared between the alternatives in the error terms. Four possible360
two-level NL models are considered in this paper, as presented in Fig. 3to Fig. 6. In the361
NL model A (see Fig. 33; Model A hereon), the alternatives are classified according to their362
buy/sell decisions, which gives rise to three branches at the higher level: Sell, Replace, and363
Buy. The first branch includes the decisions that do not involve buying behavior, including364
D2, D1 and SQ; the second branch includes all replacement behaviors: RPD1 and RP; the365
third branch is a degenerated branch, which is a branch or nest contains only one alternative.366
3D2 stands for decrease two cars, D1 stands for decrease one car, SQ stands for status quo, RPD1 stands
for sell two cars and buy one car, RP stands for replacement, and BUY stands for buy one car.
14

This model tests whether selling behaviors (e.g., D2 and D1) are correlated and whether367
replacement behaviors (RPD1 and RP) are also correlated.368
As shown in Fig. 4, the NL model B (Model B hereon) is modified based on model A369
by separating the SQ alternative from the Sell branch. Thus, the first branch of model B370
consists with pure selling behaviors, D2 and D1; the second branch contains all replacement371
decisions, RPD1 and RP; and the rest two of the branches are degenerative branches. This372
model also tests whether selling behaviors share common components in the error terms and373
whether RPD1 and RP are correlated with each other.374
While Model A and Model B are constructed according to respondents’ behavior, Model375
C and Model D are based on the change of the total number of vehicles owned in a household.376
Model C (see Fig. 5) has four branches. The first branch consists with all decisions that377
lead to a decreasing number of total vehicles; the rest three branches are all degenerated378
branches that representing replacement while sustains the number of vehicles, doing nothing,379
and buying a new vehicle.380
Models D (see Fig. 6) is similar to model C as both models test whether the behaviors381
that lead to a decreased number of vehicles are correlated. Besides, model D also grouped382
the two alternatives that resulted in no changes to vehicle holding level: SQ and RP. The383
buying decision itself is a degenerated branch.384
Figure 3: Nest Structure A
Figure 4: Nest Structure B
4. Analysis and results385
4.1. Model description and selection386
As mentioned previously, a series of NL models were estimated using the stated choice387
data. In this section, we present the process of selecting the models. Four nested structures388
15

Figure 5: Nest Structure C
Figure 6: Nest Structure D
(see Fig. 3– Fig. 6) for the car-sharing availability model and four for car-sharing effects389
model, thus eight models in total, are estimated using Nlogit (Greene,2016), a statistical390
package specialized in discrete choice modeling. The summary statistics for the different NL391
models, which represent the model performance, are given in Table 7and Table 8. Walds392
tests were used to test the logsum values, more specifically, the null hypothesizes of Wald393
test one and two are the logsum values are statistically equal to zero and one, respectively.394
As shown in Table 7, one or both of the logsum values in models B, C, and D is sta-395
tistically greater than one, suggesting the inconsistency with the theoretical derivation, and396
these models are rejected. In contrast, the logsum values of the two nests in Model A are397
between zero and one, which implies the correlation among the alternatives within the nests.398
In addition, amongst all the four models, Model A has the largest R square, and the NL399
model base on structure A statistically outperformed MNL model according to the likelihood400
ratio test. The likelihood ratio test uses the Chi-squared test statistic that can be obtained401
through −2×[LLM N L −LLN L]. The MNL model can be rejected if the value is greater than402
the critical value of with degrees of freedom equal to the number of logsum parameters for403
the Chi-square distribution. The Chi-squared test statistics for this model is 27.4954, which404
is greater than the critical value of 5.99 (95%), thus rejecting the MNL model. Therefore,405
Model A was selected as the best model for modeling car-sharing availability’s impact on406
vehicle ownership. The modeling results are presented later.407
Regarding the car-sharing effects models, the summary statistics are given in Table 8.408
Both decision tree A and B are valid because the logsum values of the branches are all409
4The log-likelihood value of the base MNL model is −5164.8475 and the log-likelihood value for the NL
model A is −5151.1. Chi-squared test statistics = −2×[(−5164.8475) −(−5151.1)]
16

Table 7 Summary statistics for different car-sharing availability models
Model R square Nests Logsum Wald test 1 Wald test 2
MNL 0.11 - - - -
NL model A 0.532 Nest 1 0.527 3.3 -2.96
Nest 2 0.688 12 -5.47
NL model B 0.413 Nest 1 1.087 15.42 1.24
Nest 2 0.836 14.67 -2.88
NL model C 0.411 Nest 1 1.181 20.92 3.21
NL model D 0.475 Nest 1 1.244 17.15 3.34
Nest 2 1.168 10.48 1.51
statistically between zero and one, however, specification A is selected due to the highest410
R-squared value. Model C is rejected because the null hypothesis that the logsum is not411
different from one could not be rejected, thus collapses to MNL model Model D is also412
rejected due to the same reason (first logsum value equals one). Similarly, the performance413
of NL model A is statistically better than the MNL model because the likelihood ratio test414
statistic of this model is 11.195, which exceeds the critical value of 5.99 (95%). The modeling415
results will be presented below.416
Table 8 Summary statistics for different car-sharing effects models
Model R square Nests Logsum Wald test 1 Wald test 2
MNL 0.109 - - - -
NL model A 0.527 Nest 1 0.377 1.61 -2.66
Nest 2 0.701 7.23 -3.08
NL model B 0.407 Nest 1 1.138 9.03 1.1
Nest 2 0.943 8.73 -0.53
NL model C 0.408 Nest 1 1.142 13.6 1.69
NL model D 0.471 Nest 1 1.028 11.82 0.32
Nest 2 0.628 4.03 -2.38
4.2. Model performance417
As shown in Table 7and Table 8structure A (see Fig. 3) is the best performing spec-418
ification, with the logsum parameters significantly between zero and one, for the two NL419
models. The overall performances of of both the models are statistically significant at a 99%420
confidence level, with the reasonably good McFadden Pseudo R-squared values of 0.5318421
(see Table 11) and 0.527 (see Table 12), repetitively, as it is suggested that a Pseudo R2422
value greater than 0.3 is an indication of a decent model fit for a discrete choice model and423
5The log likelihood value of the base MNL model is −2627.201 and the log likelihood value for the NL
model A is −2621.605. Chi-squred test statistics = −2×[(−2627.201) −(−2621.605)]
17

a Pseudo R2 between 0.4 and 0.6 reflect an excellent model fit (Dissanayake and Morikawa,424
2002;Jovicic and Hansen,2003;Ortuzar and Willumsen,2002). For both models, the logsum425
values of both Sell and Replace branches are statistically between zero and one, suggesting426
the non-zero correlation among the alternatives within the same nest, which thus validates427
the NL structure (Koppelman and Bhat,2006).428
In addition to the R-squared, another useful tool to determine model performance is the429
contingency table, which effectively compares the choice outcomes at an aggregated level430
(Hensher et al.,2015). Table 9and 10 present the contingency tables for the two models in431
this study. The columns represent the frequency of choices that are predicted to be made432
for each alternative, while the rows represent the number of choices actually made by the433
respondents for each alternative. The elements on the diagonals of the two tables counted434
the total number of times that the choices have been correctly predicted by the models for435
each alternative, whereas the off-diagonals numbers represent the total number of incorrectly436
predicted choices for each alternative. We can simply sum up the diagonal elements and437
divide them by the total number of choices made to generate the rates of correct predictions.438
The frequencies that the two models correctly predicted the choices made by respondents439
are 60.83% and 60.55%, respectively. Given the novelty of some alternatives in this study440
(e.g., AVs and SAVs), it is should not be surprising to see that the percentage of correct441
predictions are not high. Besides, as one of the early studies in this direction, the main442
motive is to provide some insights and advance our understanding on consumers’ preference443
towards the new mobility options.444
Table 9 Contingency table for car-sharing availability model
D2 D1 SQ RPD1 RP BUY Total
D2 619 92 1 20 3 141
D1 1 270 527 4 84 10 896
SQ 1 80 2817 0 107 43 3048
RPD1 0 7 109 237 2 157
RP 0 71 499 0 143 18 731
BUY 1 9 300 3 56 27 396
Total 9 456 4344 10 447 103 5369
4.3. Impact of car-sharing availability (car-sharing availability model)445
Table 11 summarizes the results of the model based on the decision tree structure A.446
Doing nothing (SQ) is set as the base alternative, the coefficients of bought two-wheeler,447
sold two-wheeler and age are estimated as generic. As shown in this table, the households’448
decisions on the ownership of two-wheelers (both buy and sell) are positively correlated with449
the alternatives, suggesting that households that change their two-wheeler ownership are also450
more likely to change their households’ vehicle ownership. As expected, age is negatively451
associated with all behaviors, the elder respondents are less likely to change the ownership452
status, suggesting that the probability is low for elder respondents to neither purchase nor453
sell their households’ vehicles.454
18

Table 10 Contingency table for car-sharing effect model
D2 D1 SQ RPD1 RP BUY Total
D2 29 48 2 6 4 71
D1 3 125 282 3 45 6 464
SQ 1 51 1425 1 33 15 1526
RPD1 0 3 54 812 7 84
RP 0 17 252 7 68 10 354
BUY 0 10 150 5 32 11 208
Total 6 215 2211 26 196 53 2707
The possibility of selling the existing vehicles increases as the total vehicle interior capa-455
bility increase, this is understandable, and for example, if two vehicles owned by a household456
are large/family vehicles, they are likely to reduce their vehicle holding levels. Besides, the457
likelihood of reducing one vehicle is higher for more educated respondents, and households458
with more private vehicles garaged are more willing to reduce only one but not two of their459
current vehicles. The average refuel distance is negatively associated with D1, suggesting460
that the probability of selling one existing vehicle decreases if respondents are live in a dis-461
advantaged area where refuel infrastructures are far from their home and not convenient to462
use. Also, larger families with members greater or equal to three are unlikely to dispose463
their vehicle.464
In terms of the replacement behaviors including RPD1 (sell two and buy one) and RP465
(sell one and buy one). As seen in this table, if the estimated resale price of the current466
vehicle(s) exceeds the purchase price of a new vehicle, respondents are more likely to replace467
one or two of their current vehicle(s) with a new car. When considering to streamline or468
upgrade the fleet of family cars (i.e., RPD1 or RP) respondents are less likely to choose to469
replace the current vehicle(s) with a petrol-powered vehicle. The potential policies such as470
free registration fee, free EV public charging, and free use of transit and bus lanes could471
effectively increase the probability of vehicle transaction behaviors. Also, the education level472
is positively correlated with the likelihood of vehicle replacement, and the respondents tend473
to replace a current vehicle with a relatively smaller future vehicle. Moreover, in a more474
developed market with higher car-sharing market share, consumers are willing to stick with475
their current fleet of vehicles, probably because their occasional use of other types of vehicles476
could be fulfilled by car-sharing.477
As expected, households with more family members (e.g., three or more members) are478
more likely to acquire a new vehicle, whereas, families with a larger number of vehicle479
holdings are less likely to add an additional car to their fleets. Moreover, the likelihood480
of purchasing a new car is negatively correlated with both vehicle purchasing price and481
operating cost. Interestingly, Australian respondents are environmentally friendly as they482
are more willing to purchase a greener car. Besides, it could be expected that the government483
incentives, which includes tax rebate, toll road discount and free use of transit and bus lanes484
could increase the new car sales significantly.485
19

More importantly, the car-sharing option has no impact on respondents’ private vehicle486
decision. It can be seen in this table, whether the car-sharing information is presented to487
respondents or not does not have a statistically significant influence on respondents’ vehicle488
ownership decisions including acquiring, disposing, or replacing vehicles.489
Table 11 The nested logit model with full sample (N=5369)
Variable Coefficient z-value P-value
D2 -3.671*** -8.53 .0000
Bought two-wheeler 1.58*** 14.33 .0000
Sold two-wheeler 1.108*** 7.76 .0000
Age -.195*** -7.67 .0000
Interior capability .557*** 13.83 .0000
Total private car -.388** -2.47 .0136
Car sharing option .026 .69 .4913
D1 -1.66*** -8.5 .0000
Bought two-wheeler 1.58*** 14.33 .0000
Sold two-wheeler 1.108*** 7.76 .0000
Age -.195*** -7.67 .0000
Interior capability .557*** 13.83 .0000
Education .11*** 4.17 .0000
Household size ≥3 -.235*** -2.67 .0077
CV average refuel distance -.55*** -12.58 .0000
Total private car .23*** 4.28 .0000
Car sharing option .011 0.64 .5222
RPD1 -1.197*** -5.64 .0000
Bought two-wheeler 1.58*** 14.33 .0000
Sold two-wheeler 1.108*** 7.76 .0000
Age -.195*** -7.67 .0000
Difference in price: CV price - FV price .004*** 2.87 .0041
FV petrol powered car -.552*** -2.24 .0251
FV free registration fee .685*** 2.72 .0064
FV free EV public charging .788** 2.16 .0304
Car sharing option .034 .81 .4168
RP -.709** -2.6 .0092
Bought two-wheeler 1.58*** 14.33 .0000
Sold two-wheeler 1.108*** 7.76 .0000
Age -.195*** -7.67 .0000
Education .177*** 4.77 .0000
FV petrol powered car -.287*** -1.97 .0488
Difference in price: CV price - FV price .007*** 3.22 .0013
Interior capability: FV - CV -.372*** -8.7 .0000
FV free registration fee .378*** 2.21 .0269
20

FV free EV public charging .48* 1.9 .0571
FV free use of transit + bus lanes 0.384** 2.37 .018
Car-sharing market share -.088* -1.71 .0867
Car sharing option -.026 -.94 .3475
BUY -.303 -.84 .4029
Bought two-wheeler 1.58*** 14.33 .0000
Sold two-wheeler 1.108*** 7.76 .0000
Age -.195*** -7.67 .0000
Household size ≥3 .443*** 3.78 .0002
FV price -.022*** -4.18 .0000
FV environmental friendliness .109*** 1.72 .0848
FV average operating cost -.096** -2.05 .0404
Total private car -.294*** -3.86 .0001
Tax rebate .357** 2.31 .0207
Toll road discount .323** 2.1 .0356
Free use of transit + bus Lanes 0.35** 2.17 .0299
Car sharing option .014 0.56 .5744
Sell .527*** 3.34 .0008
Replace .688*** 12.02 .0000
BUY 1
Full information maximum likelihood (FIML) Nested Multinomial Logit model.
Log-likelihood= -5151.1; Restricted log-likelihood = -11001.7953.
McFadden Pseudo R-square = 0.5318; AIC = 10384.2, AIC/N=1.934.
Normalization type = RU1. Number of parameters estimated = 41
4.4. Impact of car-sharing attributes(car-sharing effect model)490
The car-sharing availability model, based on the full sample (N=5,369) above, suggests491
that the car-sharing option appears to have no impact on respondents’ vehicle ownership492
decision-making process. However, the influence of each car-sharing attribute remains un-493
known. Specifically, when making ownership decisions, would respondents completely ignore494
the car-sharing factors or would they consider some specific factors? In this section, the car-495
sharing effects model, focusing on the scenarios in which the car-sharing related information496
is always provided (N=2707), is presented to explicitly examine whether respondents are497
affected by some of the car-sharing attributes.498
The results of the car-sharing effects model based on specification A (see Fig. 3) are499
summarized in Table 12. Results of this model are consistent with the car-sharing availabil-500
ity model (see Table 11) based on the full sample. Respondents’ vehicle ownership decisions501
including purchasing, reselling, replacing or doing nothing are largely dependent on the ve-502
hicle specific attributes and households’ socio-demographic factors, however, the car-sharing503
specific variables appear to have moderate impacts on respondents’ decision. Similarly, re-504
spondents who change ownership status of two-wheelers are also more likely to change their505
21

household vehicle ownership, and the elder respondents are more likely to stick with their506
current vehicles.507
As seen in this table, respondents are more likely to sell two vehicles if the total vehicle508
interior capability is higher. More educated respondents or families with a larger number509
of vehicles are more willing to dispose one of their households’ vehicle(s). Households with510
more family members are less likely to reduce their vehicle holdings by two, as more vehicles511
are needed to fulfill their mobility demand, and the probability of reducing a vehicle is low for512
respondents in disadvantaged areas. In addition, car-sharing with long access distance from513
home significantly reduce customers’ willingness to us, which in turn reduce the possibility of514
disposing current vehicles. However, respondents in Australia appear to favor large or petrol-515
powered shared vehicles, as these attributes are negatively associated with their probability516
to reduce their current cars.517
Regarding replacement behavior, when the resale price of the current vehicles exceeds518
the purchase price of a considered vehicle or the future vehicle comes with a registration fee519
exemption, respondents are willing to replace two current vehicles with a new car; however,520
they are unlikely to replace their current vehicle with a traditional petrol-powered vehicle.521
Similarly, respondents tend to replace one of their current cars with a smaller one and the522
purchase price of future cars has negative impacts on their likelihood of vehicle replacement.523
Concerning the effects of car-sharing, if the shared car is allowed to use transit and bus524
lane for free, respondents are unlikely to reduce their vehicle ownership, and the access525
distance from work appears to have negative impacts on the respondents’ replace with526
disposal decision. Moreover, respondents are more likely to replace (and decrease) their527
vehicles if SUV or micro vehicles are available from the car-sharing operators.528
In addition, respondents’ intention of buying an additional vehicle is negatively associ-529
ated with the purchase price and operating cost, however, is positively associated with the530
environmental friendliness. As expected, households with three or more family members531
are more likely to buy an FV, whereas families with more current vehicles are less likely532
to acquire an additional one. Moreover, car-sharing operating cost and driving range are533
negatively associated with new vehicle purchase behavior, and respondents are more likely534
to buy their new car if the shared car available to them is a micro-size vehicle.535
In terms of the household-specific variables, more educated respondents are more likely536
to resell or replace one of their current vehicles. Respondents who had longer average refuel537
distance are likely to keep their current vehicles and less likely to purchase new vehicles than538
those who lived in the advantaged area. As expected, households with three or more family539
members are more likely to buy an FV and less likely to reduce one of their current vehicles.540
In summary, as shown in Table 12, respondents’ decisions are likely to be influenced541
by some of the car-sharing specific factors including operating cost, market share, driving542
range, fuel type, vehicle type, and access cost. Interestingly, car-sharing cost and driving543
range are found to have statistically significant impacts on respondents’ vehicle purchasing544
decision, however, had no impacts on disposal decisions. While the availability of car-sharing545
had no impacts on respondents’ behavior, the car-sharing specific attributes presented had546
a moderate influence on the choice probabilities, suggesting that the prior studies may have547
overestimated the impact of car-sharing availability.548
22

Table 12 The nested logit model for the sample with car-sharing option (N=2707)
Variable Coefficient z-value P-value
D2 -2.756*** -8.15 .0000
Bought two-wheeler 1.558*** 10.82 .0000
Sold two-wheeler .952*** 5.21 .0000
Age -.203*** -5.95 .0000
Interior capability .188*** 4.62 .0000
Household size ≥3 -.273** -2.31 .0209
CS access distance from home -.044*** -3.23 .0013
CS large vehicle -.817** -2.37 .0178
D1 -1.169*** -4.43 .0000
Bought two-wheeler 1.558*** 10.82 .0000
Sold two-wheeler .952*** 5.21 .0000
Age -.203*** -5.95 .0000
Interior capability .188*** 4.62 .0000
Total private car .393*** 5.37 .0000
Education .116*** 3.17 .0015
CV average refuel distance -.535*** -9.00 .0000
CS petrol powered vehicle .571*** 4.17 .0000
RPD1 -1.267*** -2.97 .0030
Bought two-wheeler 1.558*** 10.82 .0000
Sold two-wheeler .952*** 5.21 .0000
Age -.203*** -5.95 .0000
Difference in price: CV price - FV price .012*** 2.60 .0092
FV petrol powered car -1.274*** -3.23 .0012
FV free registration fee .568* 1.68 .0936
CS free use of transit + bus lanes -.523** -1.98 .0482
CS access distance from work -.033* -1.95 .0509
CS SUV 1.155*** 2.91 .0036
CS micro vehicle .834*** 2.62 .0089
RP -1.173*** -2.91 .0036
Bought two-wheeler 1.558*** 10.82 .0000
Sold two-wheeler .952*** 5.21 .0000
Age -.203*** -5.95 .0000
Education .164*** 2.95 .0032
FV petrol powered car -.49** -2.28 .0228
Interior capability: FV - CV -.265*** -4.42 .0000
FV price .013*** 3.01 .0026
CS micro vehicle .465** 2.44 .0146
CS environmental friendliness .175* 1.85 .0637
BUY 1.051* 1.68 .0930
Bought two-wheeler 1.558*** 10.82 .0000
23

Sold two-wheeler .952*** 5.21 .0000
Age -.203*** -5.95 .0000
Total private car -.223** -2.13 .0335
Household size ≥3 .519*** 3.15 .0016
FV price -.032*** -4.43 .0000
FV environmental friendliness .211** 2.36 .0184
FV operating cost -.14** -2.12 .0341
CS operating cost .111* 1.69 .0919
CS driving range -.12* -1.92 .0554
CS micro vehicle .308* 1.82 .0689
Sell .377 1.61 .1083
Replace .701*** 7.26 .0000
BUY 1
Full information maximum likelihood (FIML) Nested Multinomial Logit model.
Log-likelihood= -2621.605; Restricted log-likelihood = -5541.782.
McFadden Pseudo R-square = .527; AIC = 5321.2, AIC/N=1.966.
Normalization type = RU1; Number of parameters estimated = 39
5. Policy implications549
The findings presented here usefully build upon the existing literature by underscoring550
the crucial role of policy. Previous studies that focus on car-sharing users only usually551
find a positive influence of car-sharing on vehicle ownership (i.e., consumers are likely to552
reduce household vehicles after joining car-sharing). However, it is intrinsically difficult to553
reduce car use and car ownership (Graham-Rowe et al.,2011), and comparing the results554
of the two models in this study allows us to identify the possibly over-optimistic results in555
existing studies. As evidenced by the present study, the general consumers are unlikely to556
change their car ownership status simply due to the availability of car-sharing (i.e., the first557
model), whereas the second model shows that if car-sharing attributes are put in front of558
the respondents closely, the consumers’ behavior might be nudged or implicitly influenced559
by car-sharing related factors. Based on this, we believe it is reasonable to expect that560
education and awareness campaigns might facilitate car-sharing adoption, and thus increase561
car-sharing’s impact on car ownership. This is not surprising, as in the real-world, even if562
the car-sharing option is available, most people tend to not consider it while making car563
ownership decisions. Therefore, beyond merely making car-sharing programs available, our564
evidence suggests that there is an urgent need for policies that actively make car-sharing565
schemes attractive to consumers. Some specific policy implications can be considered from566
this study.567
Firstly, prior studies have shown that respondents with car-sharing experience are likely568
to significantly reduce their car ownership (Martin and Shaheen,2011;Martin et al.,2010;569
Katzev,2003). On the one hand, policy makers should recognize that the results of these570
prior studies may tend to over-estimate the impact of car-sharing availability on vehicle571
24

ownership given potential self-selection bias, for example, up to 60% decrease in private572
vehicles was reported in Katzev (2003).573
On the other hand, this present study highlights the importance of the car-sharing ex-574
perience, given car-sharing’s minimal impact on private car ownership may simply be the575
result of respondents’ unfamiliarity with car-sharing. Based on the results of the two models576
in this study, it could be argued that education and awareness campaigns might facilitate577
car-sharing adoption. More specifically, while our car-sharing availability model shows no578
significant impact of car-sharing availability on general public’s car ownership decision, our579
car-sharing effect model reveals that when we focus on the SP scenarios in which car-sharing580
is available, several car-sharing attributes appear to have significant impact on their car own-581
ership decisions, which implies that if we can help the general public to know more about582
car-sharing (e.g., how it works, what are its benefits to individuals and society, etc.) through583
some well-designed education and awareness campaigns, car-sharing could exert more influ-584
ence on people’s car ownership decision than what is currently revealed in our car-sharing585
availability model. Also, as shown in Jensen et al. (2013), a significant change of consumers’586
preferences for the electric vehicle could be observed after they had experienced driving587
an electric vehicle. The same effect could be expected in car-sharing and SAVs. Besides588
providing car-sharing experience to potential consumers, education campaigns designed to589
promote the general public’s awareness might be another effective way to increase the gen-590
eral public’s preference towards car-sharing and to fully realize car-sharing’s benefits to the591
society. More specifically, through education, consumers might better understand the un-592
derlying costs of travel and become more appreciative of the low car-sharing costs. For593
instance, in Nordlund and Garvill (2003)’s field study, compared to a control group, a policy594
intervention that increased the awareness of alternative travel modes and the awareness of595
the context of the pre-planned trips was applied to the experimental group. The results596
indicated that respondents in the experimental group effectively reduced their travel times.597
Similar effects were also found in Alcott and DeCindis (1991), whereby psychological in-598
tervention (i.e., education and awareness rising) facilitated the adoption of car-pool. Also,599
Anable et al. (2004) argued that education and awareness raising lead to reduced automobile600
use. Hence, appropriate policies and education programs to provide consumers with virtual601
experience and increased awareness can potentially facilitate the adoption of car-sharing and602
SAV, leading to reductions in both car use and car ownership.603
In addition to education and awareness campaigns, pricing is also important. Car-sharing604
programs are available in many cities around the world with various degrees of success. A big605
challenge that policy makers face is to encourage a mode shift from private vehicles to car-606
sharing. A possible way to increase car-sharing’s attractiveness and eventually induce mode607
shift is to adopt MaaS (The Economist,2016) model (Mobility as a Service) which combines608
different travel modes and seamlessly integrates them into a subscription program. The609
MaaS program can significantly further improve car-sharing’s flexibility and affordability.610
For example, with a similar cost of vehicle ownership (e.g., $400/month), one can get access611
to a wide range of mobility options including public transport, taxi, car-sharing, ferry, etc.612
This kind of innovative solution can be appealing to consumers, and possibly lead to reduced613
private car ownership and increased use of car-sharing.614
25

The elasticity theories back this up, the intuition behind these subscription programs is615
to decrease the cost of a bundle of alternative travel modes. As shown by Hensher and King616
(1998), an increase of 10% in public transit fare (train and bus) causes 1.96% increase in car617
sales. In the U.S. context, Mcmullen and Eckstein (2011) argue that the long-run elasticities618
between public transit and private vehicle usage are 0.0228 (higher transit usage, fewer car619
travels). Based on the results of the stated choice experiments, Haboucha et al. (2017)620
concluded that increasing customers’ awareness of SAVs and the associated benefits, along621
with raising the costs of regular cars, could facilitate SAV adoption. Other factors including622
fuel price (-0.2 to 0.0), income (0.75 to 1.25), taxation (-0.08 to -0.04) population density623
(-0.7 to -0.2), etc. were also appear to influence consumer car stock choices, calculated624
elasticities in the parentheses (Litman,2019). The results of this study are consistent with625
the literature, for example, based on the estimation of elasticities, 1% increase in new vehicle626
purchase price, future vehicle operating cost, or car-sharing driving range is likely to reduce627
0.79%, 0.56% 0.21% reduction in the probability of buying a new car (other factors remain628
constant), respectively; and an increase of 1% car-sharing access distance could possibly629
reduce the possibility of disposing two private cars by 0.44%. Therefore, changing the fare630
(e.g., operating and parking cost), quality (e.g., premium car or access distance), ease of631
use (e.g., park in non-designed lots) of car-sharing, or implement policies that increase the632
cost of vehicle ownership (e.g., congestion fee) could facilitate car-sharing adoption and car633
stock reduction.634
6. Conclusion635
Pronounced growth in rates of vehicle ownership around the world continues to perpet-636
uate a series of economic and environmental issues. Car-sharing has received substantial637
interest in the last few decades as a potential solution to these issues. Despite car-sharing’s638
promise on these fronts, there have been limited quantitative studies investigating the im-639
pacts of car-sharing and SAVs on vehicle ownership. Most prior vehicle ownership studies640
have largely failed to consider car-sharing and SAVs, and those few studies that have been641
specifically focused on car-sharing (Martin et al.,2010;Fagnant and Kockelman,2014;Chen642
et al.,2016) have tended to focus on RP data, or been based on the simulated cost-saving643
analysis. For example, by comparing the size of households’ fleet of private vehicles before644
and after joined car-sharing, some studies have concluded that a significant reduction in645
private vehicles can be achieved thanks to car-sharing (Firnkorn and M¨uller,2012;Martin646
and Shaheen,2011). However, studies that rely on RP data solely, for example, data from647
the car-sharing operators, could suffer from sample selection bias or self-selection bias, thus648
leading to less robust results. In addition, studies drawing solely on RP information are not649
capable of investigating the impacts of emerging technologies like autonomous vehicles in650
car-sharing651
This paper contributes to this knowledge gap by providing population-based estimates of652
the impact of car-sharing program availability on personal vehicle ownership. Towards this653
end, car-sharing’s impacts on personal vehicle ownership was analyzed from two aspects:654
firstly, the impact of car-sharing program availability on personal vehicle ownership; and655
26

secondly, the influence car-sharing availability and attributes on personal vehicle ownership.656
Data collected from a national survey of respondents from major Australian cities was used,657
and to account for correlations amongst different transport modes, a series of Nested Logit658
(NL) modles were estimated and compared. Unlike many prior studies, this survey is based659
on a sample of the general public rather than car-sharing user groups. To avoid the problem660
caused by offering as alternatives car-sharing and buying a 2nd car as incompatible (Kim661
et al.,2017), the present study establishes car-sharing as background information to remind662
respondents as to the availability and characteristics of the car-sharing or SAVs. This design,663
which more closely mimics reality, is expected to be more realistic, as respondents’ car-664
sharing behavior should be indirectly related to their vehicle ownership choices. In addition,665
the impact of shared autonomous vehicles on vehicle ownership has also been investigated666
from the consumers’ perspective.667
Based on the modelling results of car-sharing availability and car-sharing effects, we668
found that the availability of car-sharing appear to have no influence on respondents’ vehicle669
ownership decisions, while its associated attributes had a moderate impacts. Overall, our670
findings differ with those reported in prior studies, however, this is not surprising given that671
many previous studies have analyzed data collected from early adopters of car-sharing (i.e.,672
either car-sharing users or operators). This group of respondents is generally found to be673
more environmentally conscious, less automobile-dependent, and more willing to commit to674
sustainable activities than the average population (Costain et al.,2012;Martin et al.,2010).675
In contrast, this study examined a sample drawn from the general public, and consequently,676
did not detect a significant impact of car-sharing availability on private car ownership.677
The results of this study suggest that policy-makers or related stakeholders should use678
the existing forecasts with caution. Education programs could be provided to consumers679
to increase their awareness of car-sharing and to experience the service, based on advice680
from other studies. Also, to facilitate car-sharing adoption, more innovative and creative681
ways (e.g., MaaS) are expected to attract consumers. Combing necessary education, trial682
programs and more efficient and optimized schemes, the adoption of car-sharing might be683
effectively increased.684
Despite the fact that this paper sought to address methodological shortcomings in prior685
investigations, we acknowledge that it has weaknesses of its own. Firstly, it should be rec-686
ognized that the results of this current study are based on stated preference data; which687
have been criticised for their lack of reliability as people’s actual choices might be inconsis-688
tent with stated choices. Additionally, limitations of the nested logit model, with “nests”689
determined by researchers, may result in clusters of alternatives associated in unobserved690
ways. Specifically, a five alternative choice set can have a maximum of 50 potential two-level691
nesting structures, whereby the analyst’s judgment affects the model structure. In addition,692
as mentioned above the respondents were asked to do the experiments repeatedly, however,693
the panel impacts stay unanswered because the NL model cannot represent heterogeneity in694
respondents’ preferences, thus, fails to capture customers’ potential taste variations. More-695
over, as mentioned above, transaction timing is considered one of the important factors696
influencing consumers’ vehicle ownership decisions, however, our data does not contain such697
information, which makes it impossible to incorporate it in our models.698
27

Furthermore, we also tried the multivariate formulation, which is an efficient way to deal699
with correlations among alternatives, at the early stage including simultaneous probit model700
(Mallar,1977) and the multivariate probit model (Chib and Greenberg,1998), these models701
address the correlated alternatives by allowing the error terms to be freely correlated, thus702
can be used to estimate a group of correlated binary outcomes (buy/sell) simultaneously.703
These models have been frequently and successfully used in the transport research literature704
(Choo and Mokhtarian,2008;Viswanathan et al.,2000;Golob and Regan,2002). However,705
these models could not present the change in vehicle ownership (e.g., increase or decrease).706
In addition, to incorporate panel impacts and to explore the potential taste variation, we707
also tried mixed logit model; however, due to the large sample size and complexity of the708
experiments, it is time consuming and computationally intensive to estimate the model.709
Therefore, we adopted the NL model at the end since it better describes the correlated710
buy/sell behaviors and the change of vehicle holdings with a nest structure.711
While our study suggests that the availability of car-sharing and SAVs do not have712
significant impacts on respondents’ vehicle ownership choices, there is significant evidence713
to support the substantial potential impact these new forms of mobility could have on714
households’ mode choices (Zhou et al.,2020). Further work is needed to understand how715
consumers can be better engaged and informed in order to trial car-sharing programs, and716
how government and private sector transport operators can work collectively to develop717
attractive pricing models that, combined with awareness campaigns, help cities capitalize718
on the significant potential benefits of these new forms of mobility.719
Acknowledgements720
Financial support was received from the Australian Federal Government and Malaysia721
Automotive Institute through the Cooperative Research Centre for Advanced Automotive722
Technology (AutoCRC): Project Market Intelligence and Technology Assessment - EEV To-723
wards a Green Mobility Solution for ASEAN and Australia (Washington et al.,2016a,b).724
The authors would like to thank Professor Joffre Swait, Professor John Rose, and Dr. Ak-725
shay Vij for their help in stated preference experiment design and choice modelling. The726
authors would also like to thank two anonymous reviewers from this journal for comments727
made on earlier versions of this paper.728
References729
ABS (2011). Census of population and housing.730
ABS (2016). 9309.0: Motor Vehicle Census, Australia.731
Alcott, R. and DeCindis, M. M. (1991). Clean air force campaign 1989-1990: programs, attitudes, and732
commute behavior changes. Transportation Research Record, 1321:34–44.733
Anable, J., Kirkbride, A., Sloman, L., Newson, C., Cairns, S., and Goodwin, P. (2004). Smarter choices?734
changing the way we travel. case study reports. UCL (University College London), Department for735
Transport.736
Anowar, S., Eluru, N., and Miranda-Moreno, L. F. (2014). Alternative modeling approaches used for737
examining automobile ownership: a comprehensive review. Transport Reviews, 34(4):441–473.738
28

Baldwin Hess, D. and Ong, P. (2002). Traditional neighborhoods and automobile ownership. Transportation739
Research Record: Journal of the Transportation Research Board, pages 35–44.740
Ben-Akiva, M. E. (1973). Structure of passenger travel demand models. PhD thesis, Massachusetts Institute741
of Technology.742
Bhat, C. R. and Pulugurta, V. (1998). A comparison of two alternative behavioral choice mechanisms for743
household auto ownership decisions. Transportation Research Part B: Methodological, 32(1):61–75.744
Brownstone, D., Bunch, D. S., and Train, K. (2000). Joint mixed logit models of stated and revealed745
preferences for alternative-fuel vehicles. Transportation Research Part B: Methodological, 34(5):315–338.746
Burkhardt, J. and Millard-Ball, A. (2006). Who is attracted to carsharing? Transportation Research Record:747
Journal of the Transportation Research Board, pages 98–105.748
Burns, L. D. (2013). Sustainable mobility: a vision of our transport future. Nature, 497(7448):181–182.749
Chen, T. D., Kockelman, K. M., and Hanna, J. P. (2016). Operations of a shared, autonomous, electric750
vehicle fleet: Implications of vehicle & charging infrastructure decisions. Transportation Research Part751
A: Policy and Practice, 94:243–254.752
Chib, S. and Greenberg, E. (1998). Analysis of multivariate probit models. Biometrika, 85(2):347–361.753
ChoiceMetrics (2014). Ngene 1.1.2 User manual & reference guide. Australia.754
Choo, S. and Mokhtarian, P. L. (2004). What type of vehicle do people drive? the role of attitude and lifestyle755
in influencing vehicle type choice. Transportation Research Part A: Policy and Practice, 38(3):201–222.756
Choo, S. and Mokhtarian, P. L. (2008). How do people respond to congestion mitigation policies? a multivari-757
ate probit model of the individual consideration of three travel-related strategy bundles. Transportation,758
35(2):145–163.759
Costain, C., Ardron, C., and Habib, K. N. (2012). Synopsis of users’ behaviour of a carsharing program: A760
case study in toronto. Transportation Research Part A: Policy and Practice, 46(3):421–434.761
de Franca Doria, M., Boyd, E., Tompkins, E. L., and Adger, W. N. (2009). Using expert elicitation to define762
successful adaptation to climate change. Environmental Science & Policy, 12(7):810–819.763
Dissanayake, D. and Morikawa, T. (2002). Household travel behavior in developing countries: Nested logit764
model of vehicle ownership, mode choice, and trip chaining. Transportation Research Record: Journal of765
the Transportation Research Board, pages 45–52.766
Duncan, M. (2011). The cost saving potential of carsharing in a us context. Transportation, 38(2):363–382.767
Fagnant, D. J. and Kockelman, K. (2015). Preparing a nation for autonomous vehicles: opportunities,768
barriers and policy recommendations. Transportation Research Part A: Policy and Practice, 77:167–181.769
Fagnant, D. J. and Kockelman, K. M. (2014). The travel and environmental implications of shared au-770
tonomous vehicles, using agent-based model scenarios. Transportation Research Part C: Emerging Tech-771
nologies, 40(Supplement C):1 – 13.772
Fagnant, D. J. and Kockelman, K. M. (2018). Dynamic ride-sharing and fleet sizing for a system of shared773
autonomous vehicles in austin, texas. Transportation, 45(1):143–158.774
Firnkorn, J. and M¨uller, M. (2012). Selling mobility instead of cars: new business strategies of automakers775
and the impact on private vehicle holding. Business Strategy and the environment, 21(4):264–280.776
Gallagher, K. S. and Muehlegger, E. (2011). Giving green to get green? incentives and consumer adoption777
of hybrid vehicle technology. Journal of Environmental Economics and management, 61(1):1–15.778
Gilbert, C. C. (1992). A duration model of automobile ownership. Transportation Research Part B: Method-779
ological, 26(2):97–114.780
Golob, T. and Regan, A. (2002). Trucking industry adoption of information technology: A multivariate781
discrete choice model. Transportation Research Part C: Emerging Technologies, 10:205–228.782
Graham-Rowe, E., Skippon, S., Gardner, B., and Abraham, C. (2011). Can we reduce car use and, if so,783
how? a review of available evidence. Transportation Research Part A: Policy and Practice, 45(5):401–418.784
Greene, W. H. (2016). NLOGIT: Version 6.0: Reference Guide. Econometric software.785
Haboucha, C. J., Ishaq, R., and Shiftan, Y. (2017). User preferences regarding autonomous vehicles. Trans-786
portation Research Part C: Emerging Technologies, 78:37–49.787
Hanly, M. and Dargay, J. (2000). Car ownership in great britain: Panel data analysis. Transportation788
Research Record: Journal of the Transportation Research Board, pages 83–89.789
29

Heckman, J. J. (1977). Sample selection bias as a specification error (with an application to the estimation790
of labor supply functions).791
Hensher, D. and Greene, W. (2001). Choosing between conventional, electric and lpg/cng vehicles in single-792
vehicle households in hensher, da (ed) the leading edge of travel behaviour research.793
Hensher, D. A. and King, J. (1998). Establishing fare elasticity regimes for urban passenger transport: Time-794
based fares for concession and non-concession market segmented by trip length. Journal of Transportation795
Statistics, 1.796
Hensher, D. A., Rose, J. M., and Greene, W. H. (2015). In the beginning, page 3–15. Cambridge University797
Press, 2 edition.798
Huber, J. and Zwerina, K. (1996). The importance of utility balance in efficient choice designs. Journal of799
Marketing research, 33(3):307–317.800
Huwer, U. (2004). Public transport and csar-sharing—benefits and effects of combined services. Transport801
Policy, 11(1):77–87.802
Jensen, A. F., Cherchi, E., and Mabit, S. L. (2013). On the stability of preferences and attitudes before803
and after experiencing an electric vehicle. Transportation Research Part D: Transport and Environment,804
25:24–32.805
Jovicic, G. and Hansen, C. O. (2003). A passenger travel demand model for copenhagen. Transportation806
Research Part A: Policy and Practice, 37(4):333–349.807
Kaiser, H. F. (1960). The application of electronic computers to factor analysis. Educational and psychological808
measurement.809
Katzev, R. (2003). Car sharing: A new approach to urban transportation problems. Analyses of Social810
Issues and Public Policy, 3(1):65–86.811
Kim, J., Rasouli, S., and Timmermans, H. (2017). Satisfaction and uncertainty in car-sharing decisions: An812
integration of hybrid choice and random regret-based models. Transportation Research Part A: Policy813
and Practice, 95:13–33.814
Koppelman, F. S. and Bhat, C. (2006). A self instructing course in mode choice modeling: multinomial and815
nested logit models. Prepared For US Department of Transportation Federal Transit Administration.816
Lave, C. A. and Train, K. (1979). A disaggregate model of auto-type choice. Transportation research part817
A: Policy and Practice, 13(1):1–9.818
Litman, T. (2000). Evaluating carsharing benefits. Transportation Research Record: Journal of the Trans-819
portation Research Board, 1702:31–35.820
Litman, T. (2019). Understanding transport demands and elasticities. Victoria Transport Policy Institute.821
Loose, W., Mohr, M., and Nobis, C. (2006). Assessment of the future development of car sharing in germany822
and related opportunities. Transport Reviews, 26(3):365–382.823
Mallar, C. D. (1977). The estimation of simultaneous probability models. Econometrica: Journal of the824
Econometric Society, pages 1717–1722.825
Mannering, F. and Mahmassani, H. (1985). Consumer valuation of foreign and domestic vehicle attributes:826
econometric analysis and implications for auto demand. Transportation Research Part A: Policy and827
Practice, 19(3):243–251.828
Martin, E. and Shaheen, S. (2011). The impact of carsharing on public transit and non-motorized travel:829
an exploration of north american carsharing survey data. Energies, 4(11):2094–2114.830
Martin, E., Shaheen, S., and Lidicker, J. (2010). Impact of carsharing on household vehicle holdings:831
Results from north american shared-use vehicle survey. Transportation Research Record: Journal of the832
Transportation Research Board, 2143:150–158.833
Mcmullen, S. and Eckstein, N. (2011). Relationship between vehicle miles traveled and economic activity.834
Transportation Research Record: Journal of the Transportation Research Board, 2297(1).835
Mohammadian, A. and Miller, E. (2003). Dynamic modeling of household automobile transactions. Trans-836
portation Research Record: Journal of the Transportation Research Board, pages 98–105.837
Mohammadian, A. and Rashidi, T. (2007). Modeling household vehicle transaction behavior, competing838
risk duration approach. Transportation Research Record: Journal of the Transportation Research Board,839
pages 9–16.840
30

Morrison, D. F. (1990). Multivariate statistical methods., volume 3. McGraw-Hill.841
Nordlund, A. M. and Garvill, J. (2003). Effects of values, problem awareness, and personal norm on842
willingness to reduce personal car use. Journal of environmental psychology, 23(4):339–347.843
Ortuzar, J. d. D. and Willumsen, L. G. (2002). Modelling transport, volume 3. Wiley.844
Paleti, R., Bhat, C., and Pendyala, R. (2013a). Integrated model of residential location, work location,845
vehicle ownership, and commute tour characteristics. Transportation Research Record: Journal of the846
Transportation Research Board, pages 162–172.847
Paleti, R., Bhat, C., Pendyala, R., and Goulias, K. (2013b). Modeling of household vehicle type choice ac-848
commodating spatial dependence effects. Transportation Research Record: Journal of the Transportation849
Research Board, pages 86–94.850
Prettenthaler, F. E. and Steininger, K. W. (1999). From ownership to service use lifestyle: the potential of851
car sharing. Ecological Economics, 28(3):443–453.852
Rashidi, T. H. and Mohammadian, K. (2016). Application of a nested trivariate copula structure in a853
competing duration hazard-based vehicle transaction decision model. Transportmetrica A: Transport854
Science, 12(6):550–567.855
Rezvani, Z., Jansson, J., and Bodin, J. (2015). Advances in consumer electric vehicle adoption research: A856
review and research agenda. Transportation research part D: transport and environment, 34:122–136.857
Rodier, C. and Shaheen, S. A. (2003). Carsharing and carfree housing: predicted travel, emission, and858
economic benefits. In Proceedings of the 83 th Annual Meeting of the Transportation Research Board.859
Schoettle, B. and Sivak, M. (2015). Potential impact of self-driving vehicles on household vehicle demand860
and usage.861
Schonberger, B. and Gutmann, S. (2013). A self-driving future: At the intersection of driverless cars and862
car sharing. Signline Institute.863
Shaheen, S. and Rodier, C. (2005). Travel effects of a suburban commuter carsharing service: Carlink case864
study. Transportation research record: Journal of the transportation research board, pages 182–188.865
Sierzchula, W., Bakker, S., Maat, K., and van Wee, B. (2014). The influence of financial incentives and866
other socio-economic factors on electric vehicle adoption. Energy Policy, 68:183–194.867
The Economist (2016). Transport as a service: It starts with a single app. Economist, [electronic edn].868
Viswanathan, K., Goulias, K., and Jovanis, P. (2000). Use of traveler information in the puget sound region:869
Preliminary multivariate analysis. Transportation Research Record, 1719:94–102.870
Washington, S., Page, L., Perrons, R., Zheng, Z., Whitehead, J., Hew, A., Zhou, F., McGrath, J., and871
Aminmansour, S. (2016a). Emerging trends in the vehicle and mobility markets of the asia-pacific region:872
What the experts think. Technical report, the Cooperative Research Centre for Advanced Automotive873
Technology (AutoCRC).874
Washington, S., Page, L., Zheng, Z., Perrons, R., Whitehead, J., Li, Z., Aminmansour, S., Zhou, F., and875
Hew, A. (2016b). Current and emerging trends in the mobility markets of the asia-pacific. Technical876
report, the Cooperative Research Centre for Advanced Automotive Technology (AutoCRC).877
Whelan, G. (2007). Modelling car ownership in great britain. Transportation Research Part A: Policy and878
Practice, 41(3):205–219.879
Willis, R. J. and Rosen, S. (1979). Education and self-selection. Journal of political Economy, 87(5, Part880
2):S7–S36.881
Yamamoto, T. (2008). The impact of life-course events on vehicle ownership dynamics: The cases of france882
and japan. IATSS research, 32(2):34–43.883
Zhang, W., Guhathakurta, S., Fang, J., and Zhang, G. (2015). Exploring the impact of shared autonomous884
vehicles on urban parking demand: An agent-based simulation approach. Sustainable Cities and Society,885
19:34–45.886
Zheng, Z., Washington, S., Hyland, P., Sloan, K., and Liu, Y. (2016). Preference heterogeneity in mode887
choice based on a nationwide survey with a focus on urban rail. Transportation Research Part A: Policy888
and Practice, 91:178–194.889
Zhou, F., Zheng, Z., Whitehead, J., Perrons, R., Page, L., and Washington, S. (2017). Projected prevalence890
of car-sharing in four asian-pacific countries in 2030: What the experts think. Transportation Research891
31

Part C: Emerging Technologies, 84:158–177.892
Zhou, F., Zheng, Z., Whitehead, J., Washington, S., Perrons, R. K., and Page, L. (2020). Preference893
heterogeneity in mode choice for car-sharing and shared automated vehicles. Transportation Research894
Part A: Policy and Practice, 132:633 – 650.895
Appendix A.896
Table A.1 shows the model outputs of the model based on only observations that car-897
sharing presented (column 2) and the model with full sample (column 3), both models aim898
to answer research question two (RQ 2). As seen from this table, two models are highly899
consistent, in terms of both the magnitude and significance of the coefficients. However,900
this combined model dose not allow us to directly answer RQ 1: whether the appearance901
of CS has impacts on private vehicle ownership or not? Therefore, we believe that more902
information could be extracted from the data with two separate models.903
Table A.1 The nested logit model regarding RQ 2 (investigating the impact of CS attributes)
Variable Model 2 above Model with
(Table 11) interactions
N=2707 N=5369
D2 -2.756*** -4.51***
Bought two-wheeler 1.558*** 1.545***
Sold two-wheeler .952*** 1.068***
Age -.203*** -.199***
Interior capability .188*** .561***
Household size ≥3 -.273** -.215**
CS access distance from home -.044*** -.032***
CS large vehicle -.817** -.949***
D1 -1.169*** -1.711***
Bought two-wheeler 1.558*** 1.545***
Sold two-wheeler .952*** 1.068***
Age -.203*** -.199***
Interior capability .188*** .561***
Total private car .393*** .258***
Education .116*** .112***
CV average refuel distance -.535*** -.584***
CS petrol powered vehicle .571*** .565***
RPD1 -1.267*** -1.279***
Bought two-wheeler 1.558*** 1.545***
Sold two-wheeler .952*** 1.068***
Age -.203*** -.199***
Interior capability .188*** -
Difference in price: CV price - FV price .012*** .00443***
FV petrol powered car -1.274*** -.734***
32

FV free registration fee .568* .469**
CS free use of transit + bus lanes -.523** -.519**
CS access distance from work -.033* -.021
CS SUV 1.155*** .898**
CS micro vehicle .834*** .661**
RP -1.173*** -.87***
Bought two-wheeler 1.558*** 1.545***
Sold two-wheeler .952*** 1.068***
Age -.203*** -.199***
Education .164*** .177***
FV petrol powered car -.49** -.358**
Interior capability: FV - CV -.265*** -.37***
FV price .013*** .008***
CS micro vehicle .465** .346*
CS environmental friendliness .175* -.002
BUY 1.051* -.07
Bought two-wheeler 1.558*** 1.545***
Sold two-wheeler .952*** 1.068***
Age -.203*** -.199***
Total private car -.223** -.301***
Household size ≥3 .519*** .43***
FV price -.032*** -.021***
FV environmental friendliness .211** .099
FV operating cost -.14** -.115**
CS operating cost .111* -.033
CS driving range -.12* -.093
CS micro vehicle .308* .262
Sell .377 .436***
Replace .701*** .657***
BUY 1
*, **, and *** for 90%, 95%, and 99% confidence levels respectively.
33