Carsharing: the impact of system characteristics on its potential to replace private car trips and reduce car ownership

Since we aim to investigate the impact of carsharing on car ownership, it makes sense to narrow the research subjects down to potential consumers of cars. Therefore, our target population is people who have a driver’s license and either own a car or intend to buy a car within the following 3 years. In addition, we only include respondents whose intended purchase is a new car for private use. People who plan to acquire second-hand cars or company cars are excluded because these decisions may involve different considerations (e.g. company car may not be financed by the user).

We used an existing Dutch national panel (Panelclix) to recruit respondents. These panel members fill out questionnaires on a regular basis for a small reward. The members who are invited to participate in our survey are selected at random from the Panelclix list. Those who choose to participate, first answered a series of filter questions and only people who fit our above requirements were asked to finish the entire survey. The data was collected in June 2016 and the final sample consists of 1003 respondents.

Sample characteristics are listed in Table 1. Comparing our sample to the Dutch car owner data,³ we can see that our sample is fairly representative regarding employment status and age, while being slightly over-represented by females (due to survey distribution quota aiming to reach gender balance among respondents), and people with relatively low income, which shall be taken into account when interpreting the results.

Corresponding to the two questions in each choice task, we have two dependent variables. The first is an ordinal one measuring the extent to which the respondent is willing to replace private car trips by carsharing. Although in each choice task we collect responses for up to⁴ four common trips of different purpose (commuting, grocery shopping, other shopping and leisure), we assume that all influential factors have the same effect on these four responses and use a single model to describe these effects. The second variable is dichotomous and denotes the choice whether to proceed with a planned car purchase (or between keeping or shedding the current car). These two dependent variables are indicators for the latent utility of each level of replacement intensity or each choice.

Regarding the trip replacement intensity, we explore how its utility is determined by the attributes of both the carsharing system and own car. We have already elaborated upon the carsharing system attributes in “Questionnaire design” section. The own car attributes we concern are fuel costs and walking distance to the parking location. Utility is also expected to be dependent on the trip characteristics as carsharing may be more feasible and suitable for some trips than others. The characteristics we investigated include trip frequency, duration, staying time at location and the purpose of trip.

As for the choice of car ownership, the utility of choosing carsharing is also assumed to be dependent on the attributes of carsharing system and own car. Although more own car attributes are expected to be influential in this decision: apart from fuel cost and distance to parking location, we also explore the effect of car price, monthly maintenance cost and availability of own car.

The effects of these attributes and factors on utility are expected to be heterogeneous among people. Therefore, we assume that the entire population consists of several classes: these effects are homogeneous within each class and vary between different classes.

Finally, we are also interested in the role individual variables play in determining class membership. In addition to the common socio-economic and socio-demographic variables, we also investigated the influence of frequency of using bikes and public transport and attitudes towards carsharing. Figure 2 is an illustration of the conceptual models for both trip replacement and car ownership.

We applied latent class models to implement the above conceptualization. To be more specific, we estimated a latent class ordinal regression model for modelling trip replacement and a latent class choice model for the car ownership model.

Let \( y_{it} \) denote the response of respondent i in choice task t and m represent a specific category of all possible responses. In the case of trip replacement intensity, m can range from never (1) to all (5). In the case of car ownership choice, the respondent can choose either car purchase or carsharing, therefore m can take two values. The final stated responses \( y_{it} \) are indicators of \( \eta_{{m|z_{it} }} \) which indicates the latent systematic utility of each category (trip replacement intensity) or alternative (car choice) of the response variable for subject i in choice task t.

In car ownership model, the value of this latent utility has the following form:in which \( \beta_{m} \) and \( \beta_{mk} \) denote the alternative-specific constant and attribute effects respectively. \( z_{itmk}^{att} \) represents the value of attribute k of alternative m in choice task t for subject i. In latent class models, the entire sample population is assumed to belong to K different latent classes which differ in their taste parameters. Therefore, the utility function of members from class x iswhich implies that a different set of \( \beta_{m} \) and \( \beta_{mk} \) will be estimated for each class x. The conditional probability for the response follows the multinomial logit function:

In the trip replacement model, the dependent variable is of ordinal level and the response probability function is exactly the same as (3) while the “utility” function becomeswhich applies the function of an adjacent-categories ordinal logit model (Agresti 2002). \( z_{itq}^{pred} \) denotes the explanatory exogenous factor q which differs between choice tasks. These factors are usually termed as “predictors” in latent class regression models. \( \beta_{xm0} \) and \( \beta_{xq} \) are class-specific intercepts of level m and effects of predictor q on utility which need to be estimated.

For each subject i, the probability of belonging to a class x is predicted by its individual characteristics \( \varvec{z}_{\varvec{i}}^{{\varvec{cov}}} \varvec{ } \) which are termed “covariates”. This probability function also takes the form of a multinomial logit model:

Hence, for each class an intercept (\( \gamma_{x0} \)) and a set of regression coefficients (\( \gamma_{xr} \)) are estimated. However, some individual-specific variables are dependent on other common covariates (such as socio-economic characteristics) and thus cannot be considered as “truly independent”; in contrast to the active covariates, these variables can be included as “inactive” covariates. The name implies that these covariates do not affect the probability of class membership and are not included in the model estimation. Instead, we calculate the distribution of inactive covariates for each class, which provides a richer profile of different classes. In this study, urban density is included as an inactive covariate.

Finally, the probability of observing a certain sequence of responses can be written as

The latent class regression and the latent class choice model were each estimated separately for one-way and roundtrip carsharing, hence, four models were estimated in total. For the trip replacement model, we pooled and stacked the responses for each of the four trips with different purpose in one dataset in order to estimate a single model for all trips as we mentioned in conceptualization.

We used LatentGold (Vermunt and Magidson 2013) to estimate all four models. Effects coding was used for all parameters of categorical variables. We used several criteria in order to determine the optimal number of classes: first, statistical measures including ρ2 values and the Bayesian Information Criterion (BIC), which take both model quality and parsimony into account; second, the interpretability of the estimated model, such as the sign and size of coefficients; third, avoid solutions with classes which are not essentially different from other classes; According to all the above criteria, for the latent class ordinal regression model of trip replacement, we arrived at a 3-class structure; and we chose the 2-class solution for the latent class discrete choice model of car ownership.

This section looks at people’s choice regarding whether they will use carsharing to replace their expected car purchase or current car. Table 6 presents the estimated choice model and Table 7 presents the distributions of covariates within each class. We found that a two-class model structure best describes the behavior. We first estimate a full model, and in the final model we constrain those parameters which are not significantly different across classes to be equal. The final model fit is high and the improvement from basic multinomial logit model is also significant. However, since most attributes are non-significant, the model fit is mostly contributed by the constants. This is mainly caused by non-trading behavior which will be discussed later in detail.

For both one-way and roundtrip carsharing, Class 1 and 2 are labeled as “Ownership-Oriented” and “CS-Oriented” according to their choice patterns. The choice responses are rather extreme for both classes: when answering whether to obtain or give up ownership of a current (or intended) car if carsharing becomes available, Class 1 choose to keep the car or go through the planned car purchase in over 97% of responses, while Class 2 opt for carsharing and forego the planned car purchase or replace one of their current cars in the vast majority (over 70% for one-way carsharing and 85% for roundtrip) of choice tasks. This implies that non-trading behavior is prevalent in the sample. Some research suggests that these observations shall be discarded (Hess et al. 2010) which can improve model fit (Wardman and Ibáñez 2012); however, it can be an expression of genuine preferences (Börjesson et al. 2012): given the attribute range in choice experiment design, when none of the other alternatives are more attractive than the alternative which the respondent sticks to, non-trading behavior is observed.

Comparing the two models, Class 1 takes a dominant share in both models (78.3% and 82.5%); while Class 2 of one-way carsharing (21.7%) is slightly larger in size than that of roundtrip carsharing (17.5%). The model results therefore suggest that the potential of both types of carsharing in reducing car ownership is on par with each other.

Now we briefly discuss the taste parameter for service attributes. The fuel type of the shared car does not have any significant impact on the final choice for both classes under both one-way and roundtrip carsharing. Except for the registration fee of one-way carsharing, none of the taste parameters for carsharing system attributes are significant for Class 2, while access time is significant for Class 1 both for roundtrip and one-way carsharing. In addition, monthly membership cost and car availability also have significant impact under roundtrip carsharing. Similar to the trip replacement model, these non-significant parameters may be a true reflection of people’s preferences: when considering whether to use carsharing and forego a planned car purchase (or shed an owned car), carsharing system attributes genuinely do not play an important role as long as they are not extremely high or low. It may also be explained by two other reasons: first, Class 1 hardly trade-off between attributes across choice tasks; second, the size of Class 2 is limited.

All variables with respect to the current (or expected) car (car price, fuel cost, maintenance cost and access time to one’s own car) are significant for at least one class in the model for one-way or roundtrip carsharing. This implies that these factors influence the decision regarding whether to use carsharing and reduce car ownership. For Class 2 in both models, people who (are expected to) have a more expensive car are less likely to forego their ownership: this suggests that an expensive car may be more than a tool for transport and bears a symbolic value, which was revealed by previous studies (Steg 2005).

Several socio-economic variables account for preference heterogeneity (see Table 7). Both under roundtrip and one-way carsharing, employed people are more likely to be a member of Class 2. Under one-way carsharing, Class 2 is also younger. Gender, education, household composition, income are non-significant predictors for class membership in both models. There is no significant difference between the two classes regarding urban density distribution either. Although higher education and higher income can lead to a higher possibility to join and use carsharing, they are also found to have positive impact on the probability to maintain car ownership (Le Vine and Polak 2017), which may explain why their effects in our model become non-significant. In contrast, (Le Vine and Polak 2017) also found that people with children are more likely to join carsharing and reduce car ownership: although our Class 2 under one-way carsharing still has a much larger share of people with children, we do not find this strengthened impact. Under both one-way and roundtrip carsharing, people who are expected to buy a new car have a higher probability of belonging to Class 2 than those who do not. This was expected since the potential buyers are asked whether they will forego the planned purchase while the others answer whether they will shed a current car. It is certainly easier to give up a purchase which has not been materialized than giving up a car one already owns.

People’s travel patterns also have substantial influence on class membership. Class 2 has more frequent public transport users under one-way carsharing and contains more intensive bike users under roundtrip carsharing. This suggests that a multi-modal person is more likely to forego car ownership when carsharing becomes available.

Of the three attitudinal items, only the hedonic attitude has a significant effect on class membership, which shows that the decision of giving up car ownership is influenced by the individual’s attachment to car ownership, while one’s perception regarding the environmental friendliness and symbolic image of carsharing do not have much impact.

A main motivation for car ownership is to conduct trips by car. If carsharing becomes available and can also fulfill this functional use of car ownership, it is likely to reduce the need for car ownership. Therefore, a plausible conjecture is that the intention of reducing car ownership is related to people’s willingness to replace their private car trips by carsharing. In electric vehicle adoption research, several studies assume that the acceptance of EV is strongly related to the inconvenience caused by EV which stems from the mismatch of the limited driving range of EV and the travel pattern (such as long trips) of individuals (Tamor et al. 2015; Tamor et al. 2013). Similarly, as we mentioned in the introduction, there have been previous research effort to measure the potential of carsharing by calculating how many people’s current travel patterns are economically compatible with carsharing (Duncan 2011; Schuster et al. 2005): the inherent assumption is that the choice of giving up car ownership depends on the extent to which carsharing can cover (replace) one’s current trips. In this section, we aim to investigate whether people who intend to use carsharing to replace more of their private car trips are more willing to give away (one of) their car.

We assign each respondent to a class in both classifications (trip replacement and car ownership) according to the posterior probabilities of belong to a particular class based on their responses and individual characteristics, and explore whether there is any relation between these two class memberships. Tables 8 and 9 display the cross tables of people’s membership under the two classifications for one-way and roundtrip carsharing. We can derive some interesting insights from the two tables: It is expected that one-way free-floating carsharing is the most popular scheme due to its flexibility comparing to one-way station-based and roundtrip carsharing. Our results show that indeed a larger group of car drivers is interested in using one-way carsharing for some of their car trips compared to roundtrip; however, one-way carsharing does not show significantly higher potential in reducing car ownership. The relation between usage intensity and car ownership reduction is also different for the two types of carsharing systems. Furthermore, the perks of free-floating carsharing—free parking at all public parking spots—do not show any effect on the probability of joining or using carsharing. We expect carsharing (especially one-way) is most likely to be operated within big cities with a certain level of urban density to be profitable. People living in highly urbanized areas may also have a higher interest in carsharing due to the limited parking and expensive parking fee in these areas. However, we discerned no significant difference in the overall preferences for carsharing between people living in areas with different levels of urban density (the urban density level variable is non-significant in the membership function).

This study aims to investigate car drivers’ intention of replacing private car trips by carsharing and reducing car ownership when a carsharing scheme is available, which shed light on the potential of carsharing among all car drivers. Latent class models are estimated to identify groups with different preference profiles. We found that for both intentions of trip replacement and car ownership reduction, people vary significantly with respect to their preferences for carsharing in general, its system attributes and how the characteristics of their own car affect their preferences. In total around 40% of the entire sample (CS-leaning and CS-enthusiasts) indicate that they may be willing to use carsharing to replace at least some of their private car trips. About 20% (CS-oriented) are likely to give up a planned car purchase or shed a current car when a suitable carsharing system becomes available. These numbers can be regarded as an upper limit of the potential for carsharing in replacing private car trips and reducing car ownership since models calibrated by stated preference data tend to overrate the preference for new products (see “Limitations and recommendations for research” section). This variance of preference may be attributed to the difference in socio-economic condition, travel pattern and carsharing-related attitudes, confirming similar findings in previous studies.

We also examined the impact of carsharing system attributes on these two intentions. As for the fuel type of the shared cars, it does not make any difference in the decision of giving up car ownership. Regarding the trip replacement decision, EVs are even preferred to gasoline vehicles by some classes; however, EVs with limited range is less preferred when carsharing is mostly used for long trips, but a driving range of 200 km is enough to compensate in this case. Based on these findings we may conclude that consumers in general do not show resistance and even demonstrate preference for electric vehicles. Regarding other system attributes such as costs and availability, if the current performance level is already acceptable (within the range of our experiment), a further improvement of the system performance in these aspects do not seem to have a significant impact on facilitating carsharing to replace car ownership or private car trips.

As for the relation between the decision of trip replacement and forgoing car ownership, people who use carsharing to replace more trips are not necessarily more willing to give up their car, which indicates that these two effects of carsharing must be studied separately. By looking at the two decisions together, we arrive at a more detailed classification of the population and a richer picture of people’s preference profiles regarding carsharing. We reveal groups such as heavy carsharing users who still want the guarantee of their own car and people who are willing to give up their car even when carsharing cannot replace most of their car trips (which they may conduct by other travel modes instead).

We can derive several policy implications from the findings. First, our study reveals that the potential for car sharing is quite large: as explained above 40% of the entire sample indicate that they may be willing to use carsharing to replace at least some of their private car trips, and 20% are likely to give up a planned car purchase or shed a current car when a suitable carsharing system becomes available. This implies that policies stimulating car sharing can have substantial societal relevant advantages, related to owning and using cars, as explained in the introduction. Policies to stimulate car sharing can, for example, be the provision of designated parking facilities (pull) but also the introduction of more paid parking in residential areas (push). Secondly, deploying electric vehicles has no negative or even slightly positive impact on increasing carsharing use, which confirms the potential of carsharing in reducing car trip emissions. This not only is relevant because shared vehicles can be EVs reducing the environmental pressure of car use, but it is also relevant because an increase in EV sales in the fleets of shared vehicles can stimulate EV sales in the entire car fleet, because due to scale effects (more sales) prices of EVs will go down, and the relatively high purchase costs are a barrier for many people to buy an EV (Liao et al. 2017). Thirdly, the potential of CS to reduce car ownership reduces the environmental impact of car ownership—note that producing cars also results in environmental pressure. And parking pressure can to some extent be reduced via increased levels of sharing. So, policies that stimulate CS might have environmental benefits via reduced ownership levels. Fourth, for one-way carsharing systems, in contrast to station-based setting, free-floating setting usually require much more government cooperation since it demands access to public parking spots. However, our results suggest that consumers do not really appreciate the extra flexibility brought by free-floating. This suggests that a station-based one-way carsharing system (such as Autolib in Paris) is a better option which is easier to implement and does not reduce utility for its users. Fifth, reducing user costs or increasing the availability of shared cars seem to have little or no impact on mid-term decisions such as the extent of replacing private car trips and reducing car ownership. Therefore, these strategies are probably not useful if the goal is to facilitate more trip replacement and car ownership reduction. Sixth, our results reveal that potential consumers’ preferences regarding carsharing are highly heterogeneous. Certain groups have more favorable attitudes and preferences towards carsharing and may be more susceptible to carsharing promotion policies/strategies, thus it is recommended that they are given higher priority in such promotion. Furthermore, since the groups which intend to use carsharing to replace more private car trips do not necessarily overlap with the group which is more willing to reduce car ownership, campaigns and advertisements promoting carsharing should choose target groups depending on their specific goal. Seventh, because our study shows that young people are more than average inclined to become users of CS systems, such systems may lead to postponed car ownership, or even to an overall reduction of the desirability of owning a car, as debated in the literature on ‘peak car’ (Goodwin and van Dender 2013).

This study has several limitations. First, since carsharing is still a niche market, despite the fact that we collected a sample of average size, the number of respondents who are potentially interested in carsharing is rather limited; we also observed the prevalence of non-trading behavior among the general population. This may lead to statistical insignificance of some attributes, predictors and covariates. If we wish to have better estimates of the preference coefficients of the potentially interested group in order to fine tune the carsharing scheme services, we need a sample which is more targeted towards the potentially interested customers. However, this was not the main aim of this study, which was to examine the potential of carsharing among the general population of car drivers. Second, stated preference method is known to result in inflated willingness-to-pay for some socially desirable behaviors (Axsen et al. 2015), and the online survey we used for data collection is known to result in even more positive responses than other types of surveys such as face-to-face interviews (Efthymiou and Antoniou 2016). Therefore, our results may be over-optimistic in evaluating the potential of carsharing. Thus, while we find that the carsharing potential is rather limited in the general car driver population, it may even be more limited than we find here. Third, in this explorative study we simplified some aspects of the choice problems: for example, we did not consider the uncertainty of remaining range when someone takes an electric shared car with limited range. Neither did we consider more flexible pricing structure (such as different price for driving and parking). Finally, a large part of our respondents resides in rural areas. Although they seem to have no significant difference in terms of the intention of trip replacement and car ownership reduction and a fair share of them seem to be quite positive towards carsharing, it shall be kept in mind that some of the service attribute levels we used in the experiment are not economically feasible to be realized in those areas.

More future research is needed in order to better investigate people’s preferences and the possible benefits of all types of carsharing. Comparing the usage pattern of roundtrip and one-way carsharing is an interesting direction which is of high practical relevance. For example, if one-way carsharing scheme is especially often used for shopping trips, then the carsharing scheme can set up more stations (parking docks) around shopping centers. The potential of peer-to-peer (P2P) carsharing in rural area is also worth investigating: our results show that people living in rural area seem to be as interested in carsharing as people from urban areas; however, as we mentioned in the limitations, the carsharing systems in our experiment may not be feasible or profitable in rural area and P2P carsharing may be the only option. Therefore, it is important to examine people’s preference for P2P carsharing. Furthermore, if we wish to arrive at a more realistic forecast of the potential of carsharing, we may combine revealed preference data with stated choice data in the model estimation. Finally, the introduction of shared autonomous vehicles will also further complicate or even completely change the entire picture. Many researcher, planners and policy makers now envision a prospect in which car ownership is vastly reduced because people on a large scale will make use of shared autonomous cars. However, our results pose doubt on this prospect: most people prefer to remain owning a car and only intend to make limited use of carsharing to replace their trips, and this preference is not very sensitive to improvements of carsharing systems. It is more likely that as long as cars are affordable and parking regulations with respect to car parking do not dramatically change, people will continue to own and use private cars even when shared autonomous cars become available on a large scale. Therefore, more behavioral research is needed to investigate the feasibility and possibility of the rosy future scenario promised by the introduction of shared autonomous vehicles.