How do app users select a free-floating bicycle? A stated preference survey investigating choice behaviour using generated screenshots
- Open Access
- 13.11.2025
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by (Link öffnet in neuem Fenster)
Abstract
Introduction
The first public free-floating bike-sharing (FFBS) system was introduced by Germany’s railway provider, Deutsche Bahn, in 2000 on a very small scale (Sun and Ertz 2021). Later in 2015, it was Chinese companies Mobike and Ofo that pioneered the expansion of large-scale, dockless and app-based bike-sharing systems in major Chinese cities (Zou et al. 2020). Technological advancements in GPS, mobile technology, and online payment systems have enabled the widespread adoption of FFBS services. They provide more flexibility than traditional, small-scale, station-based systems by simplifying the user experience. Users can easily unlock bikes by scanning a QR code with their smartphones, bypassing lengthy (physically present) registration processes (Ma et al. 2018). The integration of app-based platforms has significantly enhanced the flexibility of earlier bike-sharing generations, eliminating constraints related to station availability and docking space capacity (Chen et al. 2020). In China, this advancement has notably increased the adoption of shared bikes for ‘first and last mile’ trips in larger cities (Li et al. 2018, 2020; Xing et al. 2020).
The paradigm shifts in the way people travel and share transportation modes for urban trips have been enabled by mobile apps. They play a crucial role in the adoption of micromobility by providing real-time information on vehicle availability, allowing payment and therefore facilitating easy rentals (pick-up and return). Effective integration of these apps in travel plans with user-friendly interfaces and reliable data can enhance user satisfaction and increase usage rates (Forte and Darin 2017; Tao and Pender 2020). Discounted prices in the apps can also incentivize users to pick up bikes at fewer utilized areas within the network (Chen and Sakai 2022). At the same time, the reliance on exclusive app control may inadvertently prevent those who are less tech-savvy or do not possess smartphones from accessing FFBS systems (Chen et al. 2020). Providing alternative methods, such as IC cards, can help to integrate non-smartphone user groups (Li et al. 2018). However, this is only done in a few cases, since the main user groups of FFBS are mostly technophile young men or college students as studies from Switzerland, the US and Austria have revealed (Laa and Leth 2020; Reck et al. 2022; Shaheen and Cohen 2019). In the context of China, Li et al. 2018 present similar findings, highlighting that the probability of using an FFBS bike increases with higher income and education levels.
Anzeige
Our study was conducted during COVID-19, a prime time for FFBS usage. During that time, cycling gained global recognition as a key alternative to public transportation, which was perceived as crowded and posing a higher risk of infection. As a result, bike usage significantly increased (Paul et al. 2022; Teixeira et al. 2022; Xie et al. 2023). In China, cities like Nanjing saw a shift in user behavior with increased dependence on bike-sharing for daily commutes, especially among older adults and males, as part of a strategy to minimize virus transmission risks (Hua et al. 2020). Studies in Beijing and Wuhan further demonstrated that bike-sharing helped maintain urban mobility during lockdowns and adapted to changing demand patterns by shifting toward residential areas and longer trip distances (Chai et al. 2020; Li and Xu 2022). Shared bikes became a key substitute for public transport and a flexible response to quarantine-related disruptions (Shang et al. 2021; Xie et al. 2023). Behavioural responses included longer trips, more off-peak use, and greater concern for hygiene, comfort, and reservability. These shifts varied across China due to local lockdowns. We acknowledge this special context with local disparities as a limitation, yet our analysis focuses on in-app choices, even if some preferences partly reflect pandemic-driven changes in routines or modal choice.
Despite the socio-demographic knowledge about the users of FFBS and the importance of digital apps for booking, there is still a lack of empirical results about the momentum of preferred choice-making during app usage. When a user opens the app to plan a trip and potentially choose a bicycle, various factors are likely considered such as the distance of bikes from the user’s current location, the intended direction of travel, costs, etc. To address this, our study focuses on the trade-off between key factors that we suggest are likely to influence bicycle choice according to previous literature that is reviewed in the next section.
These insights are important to design and expand mobility services in line with user behaviour. Planning and management of FFBS must understand how users deal with the uncertainties associated with sharing schemes in general and free-floating schemes in particular. To improve the systems, it is important to further understand the choice influencing the traveller’s preference for bike pick-up. Therefore, a stated preference (SP) survey was conducted and analysed in this research. Following this introduction section, we first review related literature in the following section. We then explain the survey design with a focus on generating map-reading screenshots. Afterwards, we delve into the analysis of the survey results, employing multinomial logit (MNL), mixed multinomial logit (MMNL), and nested logit (NL) models to capture both systematic choice patterns and unobserved heterogeneity in user preferences and decision-making. Based on these results, we discuss the implications for FFBS design and outline directions for future research.
Literature review
This section reviews key factors that influence the booking decision during app usage. As research in app-based travel behavior is relatively rare, especially in the field of mode choice in FFBS, we concentrate on factors that are likely to influence app booking decisions such as distance, price etc. This literature review both aims to synthesize general patterns that are found to hold across app-based shared micromobility services in China and beyond. We structure this review by discussing the following six factors in subsections, that we hypothesize to be significant.
Anzeige
Distance to the destination
Micromobility services, such as shared bikes, have become a popular mode choice for first/last mile trips. In cities like Shanghai, a significant number of travelers use FFBS for short trips (Ma et al. 2018), especially when going to dine (Xing et al. 2020), reaching public transportation or commuting to work (Yu et al. 2021a, b). A study of 512 public transport users in Beijing found that walking for first/last mile trips decreased from 75.3% to 37.3%, while bike-sharing use increased to 45.9% after the introduction of an FFBS system (Fan et al. 2019). These and other research results (Reck et al. 2021; Tang et al. 2011) indicate that bike-sharing primarily replaces walking, public transport, or private bikes rather than private cars. A study by (Reck et al. 2022) also highlights that shorter distances tend to see a higher substitution of walking and public transport with micromobility options, whereas longer distances might lead to replacing car trips with personal e-bike rides.
Therefore, there appears to be an optimal trip distance when the distance to the destination increases, the likelihood of choosing a bike decreases. Instead, users might use more conventional modes of transportation such as public transit or personal vehicles, which are more efficient over longer distances (Limtanakool et al. 2006). However, there are also differences within the micromobility modes. For example e-bikes are used for longer and steeper trips than bikes without electric drive or e-scooters (Guidon et al. 2019; He et al. 2019). For longer distances, people tend to prefer personal e-bikes over shared ones, likely due to the added convenience and reliability of personal devices (Reck et al. 2022).
Proximity to bike-sharing
Access distance to shared micromobility significantly impacts mode choice and short distances are shown to be a key factor for the success of sharing schemes (Heitz et al. 2018). Also car-sharing research has shown that access proximity is a key factor for choosing a (nearby) shared vehicle in the app (Niels and Bogenberger 2017).
When vehicles are conveniently accessible, users are more likely to choose them for short trips (Phithakkitnukoon et al. 2017). Early research results from Montreal already proved for station-based bike-sharing that the proximity of home to docking stations was the most significant factor influencing the likelihood of using a shared bicycle system (Bachand-Marleau et al. 2012). This suggests that users are more likely to choose shared bikes when docking stations are conveniently located near their residence or destination. More recent survey-based evidence further shows that walking access imposes a stronger disutility than in-vehicle micromobility time, underlining that reducing walking distances is more important for user uptake than reducing riding times (Zou and MacKenzie 2025). Reck et al. 2021 considered Swiss micromobility users’ willingness to walk to reach vehicles. They found that users of shared e-scooters will walk an average of 60 m, up to a maximum of 210 m, to access a vehicle. Users of shared e-bikes are willing to walk an average of 200 m, up to 490 m. Public transport users are found to be willing to walk even further, with an average of 400 m to reach their station. The type of vehicle also influences the willingness to walk. Reck et al. 2021 report that, as shared e-scooters are used for shorter trips, a 200 m walk is more burdensome relative to the overall distance.
Bicycle direction (detour to access the bicycle)
It is, however, not only the absolute distance that determines the willingness to walk to access a shared bicycle. The direction of travel often also dictates the distance, which in turn influences the mode selected (Hagenauer and Helbich 2017). Consider the case where a user needs to walk 100 m against the direction of his/her next destination to access a bicycle, and the case where a user passes a bicycle when walking to the destination. In the latter case the user is probably more likely to pick up the bicycle. Related to this, research comparing bike-sharing and taxi usage in Chicago found that travel distance and the number of parks and recreational facilities were critical spatial factors. The direction of travel was found to influence the mode choice. Users tended to prefer bikes in areas with more parks and stable bike speeds, whereas taxi speed varied by time and location (Zhou et al. 2019). A study by González et al. (2015) investigating station-based bike-sharing behaviour in a district in Santiago (Chile) found that the presence of bikeways and proximity to metro stops mostly influenced the destination and route choices for users. Most users took a shared bike for commuting to the Metro and from there to work. In studies with private-owned bike users (Akar and Clifton 2009; Stinson and Bhat 2003), the route choice was mainly influenced by route length, especially for work commuters (Broach et al. 2012), and the existence of bikeways.
Reservation availability and risk of non-availability in popular areas
We now turn to factors related to temporal availability. Allowing users to reserve bikes in advance ensures availability, especially during peak hours, and enhances the attractiveness of the scheme (Julio Castillo et al. 2022). A recent study with car sharing data found that users’ flexibility with reservation times is affected by socio-demographics and trip characteristics. Older adults, lower-income groups, employed persons, and those with a university degree tend to be less flexible (Suel et al. 2024). By integrating reservation systems, bike-sharing operators can balance demand across different times and locations. This also has a financial component as operational capacities for rebalancing are more predictable (Ghosh et al. 2017).
In popular areas, especially during peak hours, the demand for bikes often exceeds supply. This imbalance leads to a higher risk that bikes will not be available when needed. Popular areas typically include commercial districts, transportation hubs, and densely populated residential areas as well as green and park areas (Yu et al. 2021a, b). Research in analyzing spatio-temporal patterns in FFBS usage in Shanghai by Du et al. (2019) reported that the peak hour for FFBS rentals during weekdays is in the morning between 8:30–9:30 and late afternoon between 16:30–17:30. The rides during that time are mostly related to work/school/university commuting, whereas off-peak bike trips are associated with leisure activities. However, it cannot be completely asserted that micromobility is mainly associated with commuting, as research in other cities in the area of peak hours and locations does not come to the same results (Bai and Jiao 2020; Mathew et al. 2019). Research in other cities also indicates that morning peak traffic for shared bicycles typically flows from residential areas to business districts, while the pattern reverses in the afternoon (Bao et al. 2017; Brinkmann et al. 2016). In most public bike systems, usage is generally higher on weekdays than on weekends (Reck et al. 2021; Yu et al. 2021a, b).
Number of intersections
Furthermore, the built environment with its three central components (according to Handy et al. 2002) land use, transportation system, and urban design influence mode choice decisions (Rybarczyk and Wu 2014) and - so our hypothesis - also app-based mode decisions. The need to cross large streets and intersections can be a barrier and a trigger for choosing a bicycle that is further away. Several studies have examined the influence of street configurations on bike-sharing decisions. Research from Marshall and Garrick (2010) indicates that gridded street networks typically exhibit greater-than-average street connectivity and significantly higher street network density. These characteristics are linked to increased levels of walking and biking. A study conducted in Xiamen (China) found that higher intersection density is negatively correlated with bicycle-sharing trips (Lei et al. 2023). Cyclists perceive intersections as less convenient and associate them with increased risk of traffic accidents and riding insecurity (Scott-Deeter et al. 2023).
Cost
Finally, price plays a crucial role in users’ willingness to choose micromobility (Grigolon et al. 2025). It may also influence whether users are willing to walk a bit further in exchange for a lower fare, affecting their choice of which bicycle to use. There appear to be few studies directly on this trade-off,though. Studies from Belgrade (Serbia) on e-scooters have shown that people are willing to pay different amounts depending on the trip purpose and their monthly income (Glavić et al. 2023). Users who would use an e-scooter for daily commuting are more price-sensitive than those who would use it only occasionally for leisure or shopping. Gender differences have also been reported. Men are willing to spend more on commuting trips and other travel purposes than women. Promotions and special rates can significantly impact bike choice (Iranmanesh et al. 2017; Song et al. 2022) and bike trip behavior. Research by Jurdak (2013) on two U.S. cities (Boston and Washington D.C.) revealed cost-sensitive trends when bike-sharing users were charged extra fees after 30 min of usage, with some returning the bikes just minutes before the additional charges applied.
In summary, (1) access distance to the destination determines user convenience and willingness to choose a particular bike. (2) Bicycle direction and (3) proximity are crucial for user satisfaction, as bikes facing the desired direction are preferred. Close proximity reduces travel time and effort. (4) Reservation time and the risk of non-availability influence the likelihood of securing a bike. In particular, longer reservation times can mitigate uncertainty. We further explore whether (5) the number of intersections en route can affect perceived safety, travel efficiency and negative emotions. Fewer intersections are known to be preferred for smoother rides (Heitz et al. 2018) but their importance for access choice has, to the best of our knowledge, not been studied. Lastly, (6) cost can significantly influence the likelihood of choosing a particular bike (Song et al. 2022). As will be discussed, screenshots depicting potential bike pick-up locations were presented to respondents to evaluate these six factors in real-world scenarios.
We close this section by noting that other factors, such as weather and traffic conditions, as well as the quality of the bicycle infrastructure, clearly also impact the choice as to whether or not to take a shared bicycle (Corcoran et al. 2014; El-Assi et al. 2017; Kim 2018; Zheng et al. 2022; Zou and MacKenzie 2025). We do not review this in more detail as these factors do not appear to significantly impact the choice between free-floating bicycles significantly, which is the main focus of this study.
Methodology
The survey design overview
The first part of the survey seeks socio-demographic information as well as information about FFBS usage patterns. This is followed by the stated preference (SP) survey. We provide travellers with screenshots that resemble common apps of major FFBS providers. On such screenshots, a map of the street layout is provided, together with the traveller’s current location, the location of available bicycles, plus additional information, such as whether a bicycle is discounted due to its location. We emphasize that by controlling the location of bicycles and showing or hiding other information, such screenshots can help us also to exclude “survey bots” and respondents who do not pay attention to the questions. Following the literature review, we hypothesize that six attributes are essential when travellers are deciding whether and which bicycle to use.
1.
Distance to the destination: Longer distance may lead to higher tolerance in making a detour. Such distance will be given as a short scenario description at the beginning of each question.
2.
Proximity to the bicycle: Bicycles located closer to the traveller are assumed to be favourable. The distance will be calculated as the total sum of the blue arrows in Fig. 1. It is calculated based on the actual path instead of crow-fly distance.
3.
Direction: Bicycles located along the same direction toward the destination are assumed to be favourable. Travellers may prefer bicycles with less detour when heading towards destination. The ‘angle’ is calculated as the angle of the origin-bicycle vector minus origin-destination vector. Such angle ranges from \(\:180^\circ\:\) to \(\:-180^\circ\:\).
4.
Reservation time and risk of unavailability: Travellers may have higher preference for bicycle usage when longer reservation is allowed. In this research, travellers can make 0 min, 5 min, or 10 min reservation. Such information will be given in the short scenario description together with distance to the destination. The reservation time remains the same for all bicycle alternatives in this scenario.
Risk of unavailability is clearly related to reservation time, but we distinguish it in our analysis. Picking up bicycles without making a reservation can be risky and may lead to potential failure, especially when the desired bicycle is located in popular areas. Therefore, bicycles with lower risk are assumed to be favourable. In the presented screenshots, bicycles located in ‘popular spots’ are marked with a red flame in the legend. Scenario descriptions will also suggest that bicycles located in popular spots can be taken away in five minutes.
5.
Number of intersections: Bicycles are more favourable if no or fewer intersections are crossed on the way to pick them up.
6.
Cost: Service providers frequently use discounts as one way to intervene travellers’ choice behaviour. Bicycles with higher discount or less cost are assumed to be more favourable. In this study, the regular cost is two Chinese yuan (¥2), which is equivalent to about US$0.3 and a common price for FFBS in China. For some bicycles 50% discounts (− ¥1) and 100% discounts (− ¥2) are provided.
Fig. 1
Instructions for the map-reading figures given to respondents
Before presenting app-screenshot mimics and related choice questions, a short demonstration of the task is presented to the respondent, as shown in Fig. 1. Respondents are instructed to understand the map and legend, and the attributes they need to consider. All the ‘screenshots’ have the identical street map, the traveller’s origin and heading direction. Regarding “heading direction”, since the destination is presumed to be outside the local map, it is presented with an arrow (see middle top in Fig. 1 as well as Fig. 2). A legend is also provided for each figure. A short message describing the distance to the destination and the reservation time allowed in this scenario is presented together with each map figure. In the series of figures provided to the respondents, six bicycle locations are randomly generated with different locations and discounts attributes. Bicycles are only located along the street and in each scenario, the bicycles are generated in such a way that avoids cases where one bicycle is the dominant option, i.e. closer, cheaper and less popular. We define cases where a specific bicycle is the obvious choice as ‘anti-bot scenarios’, designed to filter out invalid responses and improve data accuracy. Figure 2 provides an example. Here, bicycle A is clearly the most attractive option. If the respondent repeatedly chooses a different bicycle, we judge that the respondent does not pay attention to the question or has not understood the choice task.
Fig. 2
Example of anti-bot questions and a demo example
After the instructions (Fig. 1), each respondent is presented with a total of 10 scenarios as in Fig. 2 accompanied by three questions. First, they decide whether they would like to use a bicycle or not. If they opt to use one and reservations are permitted in the scenario, they must decide which bicycle to choose and whether to reserve it. Therefore, a total of 13 alternatives are available in each scenario (walk + 6 bicycles reserved + 6 bicycles unreserved).
Each respondent is presented with two “anti-bot scenarios” and eight data-gathering scenarios. Respondents are randomly assigned two anti-bot questions from the anti-bot question pool and those unable to pick the dominant option are removed from the valid pool.
Sampling and sample profile
The online survey was conducted on the Chinese platform Tencent Questionnaire from January 27 to March 11, 2022, spanning 43 days. We chose this platform for several reasons: First, it targets Chinese respondents and is tailored to their preferences. China has the biggest bike-sharing market globally (Tang et al. 2017), and thus many respondents will be familiar with the SP scenarios. Second, the platform handles incentive distribution efficiently. By requiring social media login, incentives are sent directly to respondents’ accounts. This helps to reduce bot access. Third, besides the anti-bot questions, its analysis system detects responses from the same IP address and those completed too quickly. This further helps to disqualify fake respondents and remove them from the dataset. It should be noted that the platform limits manual disqualification to 40% to protect respondents’ rights. We further note that the platform shares common online survey issues, in that the sample profile bias in gender and age remains strong as we will present later in this section. The questionnaire takes 10 to 15 min to complete, justifying a 7.5 CNY incentive per qualified respondent, which we judged to be a sufficient motivation without encouraging repeated participation. Before respondents started the survey, we stated that only those with experience of FFBS eligible to answer.
Table 1
The sample profile for this survey
Attribute | Attribute level | Percentage | Attribute | Attribute level | Percentage |
|---|---|---|---|---|---|
Age | 12 ~ 19 | 7.7% | Occupation | Student | 20.5% |
20 ~ 34 | 72.3% | White-collar Employee | 25.2% | ||
35 ~ 49 | 18.7% | ||||
50 ~ 64 | 1.2% | Blue-collar worker | 11.1% | ||
65 ~ 74 | 0.0% | Freelancer | 10.1% | ||
> 75 ~ | 0.0% | Professional | 7.3% | ||
Gender | Male | 64.9% | Public servant | 3.7% | |
Female | 35.1% | Company manager | 4.8% | ||
Monthly income (CNY) | 0 ~ 500 | 3.2% | Service worker | 6.3% | |
500 ~ 1500 | 10.7% | Contractor | 6.5% | ||
1501 ~ 3000 | 15.6% | No occupation | 4.5% | ||
3001 ~ 5000 | 19.1% | Education level | Middle school | 2.7% | |
5001 ~ 8000 | 28.2% | High school | 13.3% | ||
8001 ~ 10,000 | 12.8% | College | 24.9% | ||
1001 ~ 20,000 | 7.9% | Undergraduate | 53.2% | ||
> 20,000 | 2.6% | Graduate and above | 5.9% |
The online questionnaire was accessed 3857 times and 2034 (52.7%) of these were completed with an average completion time of 542 s. However, only 1189 were considered as valid (58.5% of 2034 completed questionnaires). Among all 2034 respondents, 553 failed to pass the anti-bot test (27.3%). Another 292 (14.4%) respondents wereare excluded because of different reasons such as completing the questionnaire too quickly, making identical choices no matter which question they were assigned to, inconsistencies in age, education level, and occupation.
The sample profiles of the remaining 1189 questionnaires are provided in Table 1. A large number of responses were obtained from a population group that is also known in the literature to be particularly engaged with free-floating bike-sharing: Young, male adults, predominantly aged 20 to 34, who are well-educated, with most holding at least an undergraduate degree. Most participants fall within the middle-income bracket (earning 5001 to 8000 CNY). They align with the typical user group of bike-sharing services (young, mostly male, economically active urbanites), which offers a cost-effective and flexible mobility solution for them. Not displayed is the geographic distribution of the respondents. We did not add any requirement for a specific city. The top three locations which respondents stated as their home were: 12% (159) from Guangdong Province, 11% (146) from Shandong Province, and 8% (99) from Jiangsu Province. All provinces in China are covered exclude Tibet. This diversity and lack of knowledge on the detailed places where respondents have experienced FFBS made it difficult to analyse specific city impacts in more detail.
We further survey attitudes towards FFBS, physical activity, as well as app usage. The motivation for these questions has been to understand if these variables correlate with the choice of the bicycles. As we do not find much significant correlation we only discuss some key observations and report the detailed response distributions in Tables A1 to A4 in Appendix A: Respondents show generally a positive attitude towards FFBS usage. Median in most of the questions ranks 5 or even higher in a six-level scale question. Respondents primarily use FFBS for commuting (66.4%), followed by shopping (17.9%) and entertainment (15.6%). Overall, they express a positive perception of FFBS’s advantages across various usage purposes.
Behavioural responses to situations when there is no bicycle displayed to be available are reported in Table A4. We find that 69.5% of respondents will keep checking for newly available bicycles while walking towards their destination. We further gathered their frequency of checking. 22.7% of respondents report that they will only check their smartphone when necessary or when waiting at traffic lights, while 24.4% of respondents report they will stay on the same page of the APP and check anytime. The other respondents (37.3%) will check at regular intervals. Though we did not find the frequency of app usage to be significant in the subsequently reported choice models, we suggest these statistics support the assumption that reserve-ability is an important criterion for the attractiveness of FFBS.
We close this section by noting that the influence of the COVID-19 epidemic was also asked about. More respondents report to have increased shared cycle usage frequency (29.4%) than decreased the usage frequency (23.4%), while the usage frequency most of the respondents remain unchanged (47.2%). This result appears to confirm some of the findings from other literature (mentioned before) that report increased usage of active modes during the COVID-19 pandemic due to fear of using public transport. We now return to the main purpose of this paper, the explanation of when a bicycle is attractive and why a certain bicycle is preferred.
Choice model results
MNL and MMNL models with “walking” as reference category
The panel data sample composed of 9512 observations (1189 \(\:\times\:\) 8) is used for the construction of conditional logit models with panel characteristics.
We firstly construct a model to find factors which encourage the usage of any of the bicycles compared to not taking one. We denote with \(\:J=\left\{C+w\right\}\) the set of alternatives which includes the set of 12 cycling options C, and the non-cycling option which we denote by w for “walking”, even though we do not specify which option the respondents would choose if preferring not to take a FFB. Let \(\:i\) denote the alternative, \(\:n\) the individual and \(\:t\) the particular choice occasion in the set \(\:{T}_{n}\). The indirect utility of alternative \(\:i\), \(\:{U}_{int}\), consists of a random component \(\:{\epsilon\:}_{int}\) and a deterministic component \(\:{V}_{int}\) which is estimated with a set of attributes \(\:{\varvec{X}}_{int}\) representing the attributes for the bicycle options \(\:i\). In line with the above discussion these are walking distance to the bicycle, cycling distance to the destination from the location of the bicycle, number of intersections to cross when walking to the bicycle, reservation time, price, risk of a bicycle not being available upon arrival and finally “angle”. Reservation time we define as zero for non-reservable bicycles in our initial models. “Risk” we define as a binary variable, with value one given to those bicycles that are indicated to be at a popular spot (see Fig. 1 or 2). “Angle” denotes explicitly whether the bicycle is in the direction of the destination. A value of zero indicates that the bicycle can be found on a straight line to the top left corner and the larger the absolute value of the angle the more the location of the bicycle deviates from the straight line. For example, in Fig. 2, bicycle B has a low angle whereas bicycles C and E have a large angle.
Note that there cannot be an alternative-specific constant in the model for the bicycle options i as these do not have a meaning per se and hence an “inherent preference” would not be explainable. In Eq. (1) we treat walking as an unspecified alternative and estimate all parameters relative to it.
$$\:{U}_{int}=\left\{\begin{array}{cc}0&\:\text{i}\text{f}\:i=w\\\:\:\varvec{\beta\:}{\varvec{X}}_{int}+{\epsilon\:}_{int}&\:\text{i}\text{f}\:i\in\:C\end{array}\right.$$
(1)
Instead, in Eq. (2) we consider the alternative explicitly as the walking option and estimate a dummy variable for this option and consider the given distance to the destination, denoted by D, as walking impedance. Parameter \(\:{\beta\:}_{w}\) is presumed to be the same as the coefficient for access walking to the bicycles.
$$\:{U}_{int}=\left\{\begin{array}{cc}{\delta\:}_{w}+{\beta\:}_{w}D+{\epsilon\:}_{int}&\:\text{i}\text{f}\:i=w\\\:\:\varvec{\beta\:}{\varvec{X}}_{int}+{\epsilon\:}_{int}&\:\text{i}\text{f}\:i\in\:C\end{array}\right.$$
(2)
Assuming that the error term \(\:{\epsilon\:}_{int}\) is independently and identically distributed according to a type I extreme-value distribution, leads to the well-known MNL formulation. To account for correlated errors resulting from multiple observations for the same individual, it is necessary to allow some parameters to vary randomly across choices. This idea leads to a multinomial logit model with random effects where choice probabilities for repeated observations of the same individual share the same unobserved random effects (Gönül and Srinivasan 1993; Train 2009). The logit probabilities are then obtained with Eq. (3) where \(\:{y}_{int}=1\) if option i is chosen by individual n in choice t. This leads to the sequence probability for person n as in (4) that is required for the log likelihood estimation.
$$\:{P}_{nit}(i\mid\:\varvec{\beta\:})=\frac{exp\left(\varvec{\beta\:}{\varvec{X}}_{int}\right)}{{\sum\:}_{j\in\:C}exp\left(\varvec{\beta\:}{\varvec{X}}_{jnt}\right)}$$
(3)
$$ P_{n} \left( {\user2{y}_{n} {\mid }\user2{\beta }} \right) = \prod\nolimits_{{t = 1}}^{{T_{n} }} {\mathop \prod \limits_{{i \in C}} \left[ {P_{{nit}} \left( {i\user2{\beta }} \right)} \right]^{{y_{{int}} }} } $$
(4)
In addition to the two “base MNL models” we also estimate mixed multinomial logit choice models. We allow all parameters to vary across the individuals with a normal distribution for which we estimate the distribution parameters \(\:\varvec{\sigma\:}\). Considering this “mixing of each \(\:\beta\:\)” leads to the multidimensional integral Eq. (5) that is approximated with simulation draws. We further omit the details of this well known model for brevity and refer instead to e.g. Train (2009).
$$\:{P}_{n}({y}_{n}\mid\:\varvec{\sigma\:})=\int\:{P}_{n}({y}_{n}\mid\:\varvec{\beta\:})f(\varvec{\beta\:}\mid\:\varvec{\sigma\:})d\varvec{\beta\:}$$
(5)
Apollo choice modelling (version 0.2.7) is adopted for the estimation of our model (Hess and Palma 2019) and the results are reported in Table 2.
Table 2
Estimated coefficient results for conditional logit models
Independent Variables [Unit] | Model 1: Using a FFBS choice relative to Not Using FFBS | Model 2: Not using FFBS as “Walking Choice” | |||||||
|---|---|---|---|---|---|---|---|---|---|
MNL | MMNL | MNL model | MMNL model | ||||||
Coef. (Std. Coef.) [\(\:{10}^{-3}\)] | t-value | Coef. (Std. Coef.) [\(\:{10}^{-3}\)] | t-value | Coef. (Std. Coef.) [\(\:{10}^{-3}\)] | t-value | Coef. (Std. Coef.) [\(\:{10}^{-3}\)] | t-value | ||
Non FFBS choice dummy [Binary] | – | − 847.7** (79.9) | − 10.6 | − 506.5**(47.7) | − 4.86 | ||||
Access distance [m] | \(\:\beta\:\) | − 1.8** (0.056) | − 28.3 | − 3.98**(− 3.56) | − 17.5 | − 1.38** (0.043) | − 32.2 | − 3.32**(− 2.97) | − 30.4 |
\(\:\sigma\:\) | – | 4.84**(2.87) | − 27.8 | – | 3.0**(1.78) | 29.9 | |||
Cycle distance [m] | \(\:\beta\:\) | 0.65**(− 0.038) | 30.93 | − 0.293**(− 0.093) | -6.58 | − 0.55** (0.032) | -17.4 | − 0.213**(− 0.068) | − 4.79 |
\(\:\sigma\:\) | – | 0.370*(0.752) | 2.87 | – | 0.78**(1.59) | − 15.5 | |||
Number of intersections | \(\:\beta\:\) | − 102.0 *(7.14) | − 2.19 | − 95.8(− 38.0) | − 1.63 | − 562.9** (39.62) | − 14.2 | − 79.8(− 31.66) | − 1.46 |
\(\:\sigma\:\) | – | 484.8**(82.5) | 4.53 | – | 860.3(146.4) | 19.8 | |||
Reservation time [min] | \(\:\beta\:\) | 41.0** (3.53) | 9.98 | 100.4**(86.0) | 5.20 | 47.84** (4.13) | 11.6 | 78.7**(67.4) | 5.19 |
\(\:\sigma\:\) | – | 488.4**(450.0) | 18.9 | – | 453.1**(417.6) | − 21.8 | |||
Cost [CNY] | \(\:\beta\:\) | − 15.69** (1.00) | − 4.39 | − 1853.0**(− 2202) | − 12.8 | − 248.9** (16.01) | − 15.5 | − 199.2**(− 236.7) | − 14.1 |
\(\:\sigma\:\) | – | 990.9**(1208.7) | 7.78 | – | 1131**(1379.6) | − 11.2 | |||
Risk [Binary] | \(\:\beta\:\) | − 199 ** (17.09) | − 7.07 | − 559.6**(− 493.2) | − 12.7 | − 323.7** (27.97) | − 11.6 | − 511.2**(− 450.5) | -13.2 |
\(\:\sigma\:\) | – | 672.2**(665.2) | 10.8 | – | 659.2**(652.3) | − 11.4 | |||
Angle \(\:\theta\:\) [Degree] | \(\:\beta\:\) | − 2.1** (− 1.49) | − 4.18 | − 6.08**(− 7.30) | − 6.58 | 0.62 (0.44) | 1.41 | − 6.63**(− 7.96) | − 12.3 |
\(\:\sigma\:\) | – | 7.49**(3.54) | − 7.75 | – | 3.85**(1.83) | − 4.44 | |||
Sample size (individuals) | 1189 | ||||||||
Observations | 9512 | ||||||||
Adjusted \(\:{R}^{2}\) | 0.0565 | 0.1951 | 0.0766 | 0.1826 | |||||
The results show all parameters are significant. The only exceptions are “number of intersections” in the MMNL version of Model 1 and “Angle” in the MNL version of Model 2. The heterogeneity captured in the mixed logit version adds significantly to the model fit and demonstrate the inter-person variability observed in our sample. We tested a range of alternative model formulations to capture these. Additional explanatory variables that we could extract from the survey, in particular socio-demographics, however, would only lead to marginal improvements in model fit, and on further examination, not very stable results. We acknowledge therefore that some of the discussion on detailed parameter values would need to be verified with additional surveys if used for practical policies. Nevertheless, all signs are general according to our expectations and we suggest that the relative impact of our variables is fairly consistent. Even considering the wide range of parameters obtained among the models this leads to useful insights and demonstrates the usefulness of our survey approach.
We suggest Model 2 is easier to interpret as it assigns utilities to all options but suggest Model 1 leads to additional insights. In the MNL version of Model 1, we find a reversed sign for “cycling distance” compared to Model 2 which is explainable. Compared to non-cycling options, FFBS becomes attractive the longer the distance to the destination. We remind that our distances to destination range from 1 km to 5 km, hence all distances are cyclable. In the MMNL version of Model 1 the parameter is though negative with a large sigma indicating high interpersonal variability. Also in Model 2 the negative parameter indicates that the farther a particular bicycle is away compared to the other options, the less likely this option will be chosen. Further “angle” is significant in all three models except for the MNL version of Model 2. Compared to not taking a bicycle, cycling becomes more attractive, the less detour on the way to one’s destination has to be made to find a bicycle. For a particular bicycle option the angle might not be significant as the access distance and the cycling distance already jointly explain the detour factor. We tested a range of different specifications for Model 2 such as considering specific ranges of angles but did not find a model specification where Angle becomes consistently significant also in Model 2. We therefore omit angle in the Nested Model discussion in the following section as it builds on Model 2.
We continue by deepening our discussion for Model 2. For the access and cycle distance parameters we note that the estimation of walking access distance has a consistently larger absolute value than cycling distance, which suggests that the walking distance is more influential than the remaining cycling distance to the destination once a person reaches the bicycle. In the MNL model the walking factor is 2.5 times larger than the cycling distance factor, which is reasonable as this is in line with the speed difference between these two forms of transport. Also considering the standardized factors yields the same conclusion. In the MMNL model the difference between the two variables is even larger though the absolute values are still not as large as one might expect. For example, given two bicycles with identical attributes, a 100 m change in walking access distance is associated with an 8.3% chance reduction in choosing the option.
When travellers are deciding which bicycle to pick up, they further prefer to choose bicycles with fewer intersections to cross, lower cost and lower risk. The intersection estimate is surprisingly highly significant. The MMNL result suggests that, for a fixed \(\:\beta\:\), the odds to choose a particular bicycle reduce by 14% if one more junction has to be crossed. The possibility to reserve a bicycle for longer time will lead to a higher preference for using these bicycles, with this variable also being highly significant. Regarding the price, the overall cost of FFBS is low also in a Chinese context, nevertheless we find that it is significant. In the MMNL model a 1 CNY increase is equivalent to nearly 3 min additional reservation time or 60 m of additional walking time. These estimates might help operators in understanding the values of specific service attributes.
Exploring the order of choice with nested logit models
We further extend Model 2 into three-level nested logit (NL) models. Our rationale is partly based on findings from the MNL model that reservation time and risk are important factors; we therefore test whether the fact that a bicycle is reservable or not is a criterion which travellers might use to preselect a bicycle. Furthermore, as discussed, the SP survey gives users the choice between taking a bicycle option with or without reservation. As this means that users have the choice between sets of two similar options, this also suggests that a nested structure could reflect the decision-making process better. We test different model structures to study whether users will choose between reserving a bicycle first or location first. We further aim to confirm whether app users will consider all available bicycles as one choice compared to the walking choice. All of the hereafter proposed models have a three-level structure that can be formulated as follows and is illustrated in Fig. 3.
Our nesting structure has three levels. Alternative \(\:i\) falls into the nest \(\:{O}_{m}\) on the lowest level of nesting, which itself is a member of nest \(\:m\) on upper level of nesting, with \(\:m\) being in the root nest. The notations \(\:{\lambda\:}_{m}\) and \(\:{\lambda\:}_{{O}_{m}}\) are adopted as the nesting parameter for the first two levels. Therefore, the probability of person \(\:n\) choosing alternative in choice situation \(\:t\) is then given by:
$$ \begin{gathered} P_{{i,n,t}} = P_{{m,n,t}} \times P_{{\left( {O_{m} {\text{|}}m} \right),n,t}} \times P_{{\left( {i{\text{|}}O_{m} } \right),n,t}} \hfill \\ = \frac{{{\text{exp}}\left( {\lambda _{m} I_{{m,n,t}} } \right)}}{{\mathop \sum \nolimits_{{l = 1}}^{M} {\text{exp}}\left( {\lambda _{l} I_{{l,n,t}} } \right)}} \times \frac{{{\text{exp}}\left( {\frac{{\lambda _{{O_{m} }} }}{{\lambda _{m} }}I_{{O_{m} ,n,t}} } \right)}}{{\mathop \sum \nolimits_{{l_{m} = 1}}^{{M_{m} }} {\text{exp}}\left( {\frac{{\lambda _{{l_{m} }} }}{{\lambda _{m} }}I_{{l_{m} ,n,t}} } \right)}} \hfill \\ \times \frac{{{\text{exp}}\left( {\frac{{V_{{i,n,t}} }}{{\lambda _{{O_{m} }} }}} \right)}}{{\mathop \sum \nolimits_{{j \in O_{m} }}^{{}} {\text{exp}}\left( {\frac{{V_{{j,n,t}} }}{{\lambda _{{O_{m} }} }}} \right)}} \hfill \\ \end{gathered} $$
(6)
With the logsums of the two nests being expressed by
$$\:{I}_{m,n,t}=\text{l}\text{o}\text{g}\sum\:_{{l}_{m}=1}^{{M}_{m}}\text{e}\text{x}\text{p}\left(\frac{{\lambda\:}_{{l}_{m}}}{{\lambda\:}_{m}}{I}_{{l}_{m},n,t}\right)$$
(7)
$$\:{I}_{{O}_{m},n,t}=\text{l}\text{o}\text{g}{\sum\:}_{j\in\:{O}_{m}}^{}\text{e}\text{x}\text{p}\left(\frac{{V}_{j,n,t}}{{\lambda\:}_{{O}_{m}}}\right)$$
(8)
For normalisation, we set the nesting parameter of root equals to one. We would then expect that 0 < \(\:{\lambda\:}_{{O}_{m}}\) ≤ \(\:{\lambda\:}_{m}\) ≤ 1.
Fig. 3
Illustration of the three nested model structures
Nested model 1: “first choosing bicycle, then whether to reserve it”
This nesting structure is shown on top of Fig. 3. The presumption is that there is significant correlation between all bicycle options which is expressed with the upper nest. With the lower nest it is presumed that the correlation between the two options relating to the same bicycle, whether reserved or non-reserved, needs to be considered. One might hence say, that the location choice takes precedence over the reservation choice. We remark that, though we ask respondents to consider all attributes together including whether a bicycle is reservable or not, this is also the order of choice in the survey. To avoid a choice between a long list of 13 options, respondents are first asked to choose among the bicycles and the walking option, before then indicating whether they would reserve the bicycle. Further note, that in this NL model, considering the alternatives in the middle level are the index of identical bicycles which should not bear any preference, we further assume \(\:{\lambda\:}_{{O}_{m}}\) to be identical for all alternatives \(\:{O}_{m}.\)
Nested model 2: “first choosing reservation, bicycle choice as lower nest”
In this alternative structure the upper nest is identical, but with the lower nesting we express the hypothesis that there is significant correlation between all reservable options. Despite the survey design we suggest this model structure indicated in Fig. 3 is also reasonable. Since previous literature and the MNL results showed that reservation is important and since many respondents indicated that they keep checking the app for available options, the security not to “walk in vein” might be an attribute of high importance to those intending to use a FFB. It suggests that the reservable attribute shared among bicycles is more important than other bicycle specific attributes.
Nested model 3: first choosing reservation, bicycle choice as lower nest
We further test a model with the reserved options not being structured. This model can be seen as an intermediate model compared to the MNL model and the 2nd nested structure model. It implies that there is significant correlation in the perception of the non-reserved options but not among the reserved options.
Table 3
Estimation results for three-level NL model (\(\:\varvec{N}=9512\))
Independent variables | NL Model 1 | NL Model 2 | NL Model 3 | |||
|---|---|---|---|---|---|---|
Coefficient | t value | Coefficient | t value | Coefficient | t value | |
Access distance [100 m] | − 0.05 (− 0.1524) | − 9.786 | − 0.05 (− 0.1627) | − 10.1 | − 0.05 (− 0.1651) | − 10.259 |
Cycle distance [100 m] | − 0.007 (− 0.0977) | − 5.814* | − 0.008 (− 0.1026) | − 5.73 | − 0.0081 (− 0.1045) | − 5.806 |
Number of intersections to cross | − 0.117 (− 0.0613) | − 8.386* | − 0.123 (− 0.0648) | − 8.45 | − 0.122 (− 0.0661) | − 8.645 |
Reservation time [Minutes] | 0.0203 (0.0839) | 6.795* | − 0.003 (− 0.0133) | − 1.356 | − 0.003 (− 0.0145) | − 1.202 |
Cost [CNY] | − 0.0696 (− 0.0273) | − 8.226* | − 0.059 (− 0.0230) | − 7.792 | − 0.058 (− 0.0238) | − 8.091 |
Risk [Binary] | − 0.0870 (− 0.0207) | − 7.237* | − 0.071 (− 0.0169) | − 7.052 | − 0.07 (− 0.0173) | − 7.158 |
\(\:{\delta\:}_{Not\:use}\) | − 1.792 | − 18.864* | − 1.845 | − 20.1 | − 1.837 | − 20.246 |
\(\:{\lambda\:}_{\text{t}\text{o}\text{p}}\): Use FFB yes/no, | 0.25 | 9.457* | 0.299 | 4.488 | 0.326 | 9.604 |
\(\:{\lambda\:}_{mid}:\) Bicycles | 0.462 | 6.998* | – | |||
\(\:{\lambda\:}_{mid}:\) All non-reserved options | – | 0.236 | 9.408 | 0.233 | 9.747 | |
\(\:{\lambda\:}_{mid}:\) All reserved options | – | 0.327 | 9.569 | – | ||
Sample size | 1189 | |||||
Observations | 9512 | |||||
LL(0) | − 22353.11 | |||||
LL (final) | − 20485.15 | − 20449.17 | − 20449.27 | |||
Adjusted \(\:{R}^{2}\): | 0.0766 | 0.0847 | 0.0848 | |||
The estimation results of the three NL models are presented in Table 3. The coefficients for NL Model 1, are all negative except for the reservation time. Comparing to the MNL model in Table 4, most of the estimates of the NL model are lower, which can be expected given the capturing of correlations. The ratio between walk distance and cycle distance increases from 2.5 times in the MNL to 6.0 times in the NL model, which suggests the influence of physical fatigue is emphasized if the correlation between the reserved and unreserved bicycle are considered. The ratio between cost and reservation time, on the contrary, has decreased from 5.2 times to 3.4 times in the NL model. This can also be interpreted as an increase of ‘value of reservation time’ from 0.96 to 1.46 CNY per five minutes.
The value of the two nesting parameters are estimated as \(\:{\lambda\:}_{\text{t}\text{o}\text{p}}=0.245\) and \(\:{\lambda\:}_{\text{m}\text{i}\text{d}.}=0.462\) and are both found to be highly significant. The ratio of \(\:{\lambda\:}_{\text{m}\text{i}\text{d}.}/{\lambda\:}_{\text{t}\text{o}\text{p}}=1.886\) does, however, violate common expectation as values smaller than 1 are expected. Exceptions are, however, known and discussed to be reasonable in among others, (Lee 1999; Train et al. 1987). (Börsch-Supan 1990) demonstrated local sufficiency conditions that permit values of larger one. The Daly-Zachary-McFadden condition of the validity of stochastic utility maximization in nested logit models is shown to be in some cases to be an unnecessarily strong condition. Also, Kling and Herriges (1995) and Herriges and Kling (1996) provide tests of consistency of nested logit with utility maximization when the ratio is larger than one.
A ratio of smaller than one indicates that travellers would prefer to substitute within the bottom level, rather than making changes in the middle choice level. Hence, a value larger \(\:1\) suggests that there is a greater substitution across the middle level (in this case, alternatives A ~ F) than between the two options in the lowest level (Use / Not use reservation). In other words, \(\:\lambda\:>1\) suggests that if reservation is not allowed for a specific chosen bicycle, travellers are more willing to change their choice made on the alternatives ‘A ~ F’ rather than changing on the decision of ‘Use / No use reservation’. The result confirms the importance of the reservation feature.
To better understand the rationale for this argument we provide NL Model 2, where the choice nesting between bicycles and reservation is reversed. Three lambdas are estimated in this model. Same as in Model 1, \(\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{F}\text{F}\text{B}}\) is adopted to denote the top-level lambda, while the two \(\:{\lambda\:}_{\text{N}\text{o}\text{R}\text{e}\text{s}}\) and \(\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{R}\text{e}\text{s}}\) are introduced to capture correlation between the non-reserved and the reserved choice options. We find that the same problem remains for an unconstrained \(\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{R}\text{e}\text{s}}\), as \(\:{\lambda\:}_{\text{N}\text{o}\text{R}\text{e}\text{s}}<{\lambda\:}_{\text{U}\text{s}\text{e}\text{B}\text{i}\text{c}}\) while \(\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{R}\text{e}\text{s}}>{\lambda\:}_{\text{U}\text{s}\text{e}\text{B}\text{i}\text{c}}\) remains. We further find that the difference between \(\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{R}\text{e}\text{s}}\) and \(\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{B}\text{i}\text{c}}\) is insignificant with a t-ratio of 0.41.
This leads us to our preferred model structure NL Model 3. Therefore, we set \(\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{R}\text{e}\text{s}}\equiv\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{B}\text{i}\text{c}}\) during the regression. This structure is shown in Fig. 3 (bottom). The log-likelihood of this structure (-20449.27) is slightly improved compared to that of the survey-logic (−20485.15), and the lambdas are estimated as \(\:{\lambda\:}_{\text{U}\text{s}\text{e}\text{B}\text{i}\text{c}}={\lambda\:}_{\text{U}\text{s}\text{e}\text{R}\text{e}\text{s}}=0.326\) and \(\:{\lambda\:}_{\text{N}\text{o}\text{R}\text{e}\text{s}}=0.233\). The estimation results can be interpreted as follows:
Assume one bicycle in the nest of ‘No Res.’ becomes unavailable, then the probability of another unreserved bicycle being chosen will increase at a higher rate than a reserved bicycle being chosen. However, if a preferred reserved bicycle is assumed to become unavailable, the probability of other bicycles in the same nest will increase equivalently to the unreserved ones. In other words, we suggest a higher correlation is observed among the non-reserved alternatives: When the desired bicycle becomes unavailable, travellers will behave differently according to their current reservation state. For those travellers who already choose bicycles without reservation, they are more likely to choose that with the same reservation state. However, for those who choose bicycles with reservation, they do not hold a strong preference as they will evenly consider bicycles with or without reservation. We suggest hence that some travellers would like to stick to non-reserved bicycles, possibly because of the discount provided and the non-commitment.
Conclusions
For the design of micromobility sharing schemes and specifically free-floating schemes, the question of how the distribution of the vehicles influences the scheme attractiveness is important. Our study provides valuable insights into the factors influencing user preferences in FFBS systems considering the location of bicycles compared to a person’s current location. Compared to existing literature, our study investigates deeper how travellers choose between various free-floating bicycles in their vicinity during app usage. We quantify the importance of close access and the value of bicycles being reservable. As long reservation times also create costs for the system, as shown in previous research (Yao and Schmöcker 2023), this can support an analysis as of optimal reservation times. We suggest our study is further one of the first showing that person’s trade off walking time and cycling time and consider the directness of the total path including walking to the bicycle and cycling to the destination.
Our nested model analysis led to additional insights, specifically with respect to the importance of bicycles being reservable. The model’s structure suggests that users first decide whether to use a bicycle and then consider the reservation option before deciding on a particular bicycle. The results indicate a higher correlation in non-reservation alternatives compared to those with reservations, emphasizing the importance of the reservation feature in user decision-making. Cost, although low in typical FFBS schemes, still exerts a measurable influence on bicycle choice, indicating that even small price discounts can induce travelers to walk farther or select different vehicles. Furthermore, the number of intersections between the user and the bicycle also plays a significant role. More intersections raise perceived inconvenience and safety concerns, reducing the likelihood of a bike’s selection. Although bicycle “direction”—measured as the deviation from a direct route—can matter at the broader decision stage (i.e., “to bike or not to bike”), its importance becomes less pronounced when comparing multiple bikes located in different directions.
We suggest our results highlight the importance of detailed locational decisions where shared bicycles are placed. Especially in urban contexts with many competing mobility options, the profit obtained by a shared bicycle operator might depend on detailed factors such as which side on the street the bicycles are placed. For selecting shared bicycle parking locations our results suggest that not only the access to the parking spot from the last activity should be considered but that an operator should also consider in which direction the travellers likely wants to move. A further implication of our study is that guaranteed availability, for example by allowing to reserve, is an important feature. This favours operational strategies where bicycles are distributed over a smaller area. The detailed trade-off between distance coverage and nearby availability is, however, beyond the scope of this study.
To ensure the validity of the choices made, considerable effort has been invested in developing a survey design that maintains maximum explanatory power. Besides the specific findings, we suggest this survey design, based on mimicked app screenshots, introduces new features in the design of the choice experiment. The screenshots allowed us to display a range of attributes for the bicycles and the street environment which provided the foundation to study the impact of the decision process. Our survey design also included efforts to present multiple alternatives as attractive, preventing any from appearing as the clearly superior option. At the same time, including some options with clearly dominant options helped us to filter out respondents that did not pay attention to the survey, which is particularly a problem for online surveys.
Finally, we remind that the survey was conducted during the COVID-19 pandemic which brought increased recognition of cycling as a low-risk travel option. As stated before we do not have a specific hypothesis as to how this would impact the ways users interact with FFBS apps. However, as some users appear to have increased their preference for cycling, one COVID-19 impact might be that the ASC for walking and other modes in the choice model might have been overestimated and that people might be even more sensitive to the attributes discussed in this study after the pandemic, when public transport is again a more attractive alternative. For this reason and as we obtain very high inter-person variability, a future research direction is clearly to validate the results obtained here with revealed preference data. For this, information on the app usage and information displayed to users alongside with rental records need to be obtained. We also suggest that the role of traffic characteristics, visibility of the bicycles, the quality of the walking environment should be further examined. If this research direction is followed, it would clearly also be an advantage to obtain a larger sample from a single or a few cities to be able to understand the impact of these factors on app interaction. As part of this, the role of socio-demographic characteristics on choices should also be further studied. In initial analysis, we did find some interaction terms of attributes with gender or age to be significant. Especially for age, we concluded, however, that our results are not reliable as our sample of senior persons is too limited. A possible reason for this might again be our too diverse sample masking some of the socio-demographic impacts.
Declarations
Conflict of interest
The authors declare no competing interests.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Dr. Dong Zhang
Dr. Dong Zhang is currently Associate Professor at School of Maritime Economics and Management, Dalian Maritime University, China. He previously held research and teaching positions at the Dalian University of Technology. His research interests focus on the behavioral aspects in urban public transportation systems, as well as maritime decarbonization.
