Identifying heterogeneity in university students’ transport mode choice

College students are argued to have some interest in adopting alternative mobility solutions. Even though previous works have examined the transport mode choice of college students, the inclusion of shared mobility services is scarcely found. Thus, this study aims to examine the transport mode choice of university students by revealing the heterogeneity in transport mode usage frequency incorporating socio-demographic and travel characteristics as covariates in the latent class cluster analysis (LCCA) framework. The analysis results in four clusters. Clusters of public transport users and walkers are mostly characterized by lower-income students who can reach the university in less than 11 minutes. Private car users mostly fall into a cluster where students generally come from higher-income families. Policies supporting sustainable transportation among college students are discussed, including the mobility points around the university and public transport stops.


Introduction
Travel behavior of younger generations, including university students, is different from the preceding generations, where young individuals are most likely to be more flexible in transport mode choices [1,2].They are proven to be the early adopters of alternative mobility solutions like shared mobility services [3,4].This condition might be caused by the fast adoption of recent technologies among young individuals and consequently affect their travel behavior [1].
Previous studies have investigated the transport mode choice among university students considering several factors (e.g., socio-demographic, travel characteristics, attitudes, and built environment) using logit models [3,[5][6][7].These studies examine conventional transport modes, such as private cars, bikes, walking, and public transportation (PT), but little attention has been paid to new mobility services.Information technology is continuously developing, and it brings effects to transportation sectors, where travel options are more varied [8].Shared transport modes are more available in cities worldwide [9].Furthermore, by examining the heterogeneity in transport mode choice, we could differentiate the students' travel behaviour.Once the travel behaviour is clustered, we could draw specific policies based on the identified clusters to raise the use of environmentally friendly modes, instead of providing general policies.
Considering the preceding background, this study aims to cover the gaps by examining the heterogeneity of transport mode choice among college students using the latent class cluster analysis (LCCA).The extra layer of heterogeneity is presented by including a wide range of covariates in the modeling framework.

Literature review
Several studies have examined the essential factors affecting college students' mode choices, including socio-demographic characteristics.Male students are argued to be motorcycle users, while their counterparts are more likely to travel to school by walking [6].Different findings state that male students are more likely to be active travelers [7] and use PT than shared taxis and private cars [5,10].Regarding income, students from high-income families are more inclined to use private cars [5] and are not interested in traveling by active modes and PT [6].In the case of travel characteristics, increasing travel time leads to a higher likelihood of using bicycles and cars [3].In this study, socio-demographic and travel characteristics are included in the model as covariates.
Meanwhile, previous works have examined the heterogeneity in transport mode choice preferences.A study argues that educated young people with high technology literacy are more likely to be in a group of mobility-on-demand early adopters [11].Another study states that the multimodal travel group is related to highly educated young individuals and small households [12].Individuals who live in urban areas are more likely to be active travelers and PT users [1].Mentioned studies prove that several factors significantly influence travel mode choice, but mostly include conventional transport modes (e.g., PT, car, active modes).
Even though previous works have investigated transport mode choices among college students, most studies neglect shared transport modes.Thus, this study includes shared mobility services as indicators in the model.Furthermore, socio-demographic and travel characteristics are incorporated as covariates, giving an extra layer of heterogeneity.

Method
This study was realized in Budapest, Hungary's capital and most populated city.The metropolitan area of Budapest covers 525.5 km 2 , with around 1.7 million inhabitants.The PT services in Budapest integrate metro, tram, bus, and trolley buses under one system and can be accessed through a single travel pass.Emerging mobility services are present in Budapest such as car-sharing, bike-sharing, and e-scooter-sharing.
An online survey was conducted among the Budapest University of Technology and Economics (BME) students from April 2022 to May 2022.The survey consists of three parts.The first part aims to capture the socio-demographics of the participants.The second part is about travel behavior, including their frequency of using the transport modes, ranging from (1) "never" to (8) "several times per day".The number of valid responses is 687.It is worth noting that BME can be easily accessed by all means of urban transport modes, with a metro station, tram stops, and bus stops around the university.Collection points for bike-sharing and e-scooter-sharing are present in the surrounding area of the university.LCCA is performed to identify the heterogeneity of travel patterns among students.LCCA is a statistical method following a probabilistic approach to reveal different unobserved subgroups of populations that share specific observed characteristics [13].LCCA is chosen since the method has advantages, such as the opportunity to include mixed types of variables and statistical criteria to identify the optimal number of clusters [14].The probability of observing response patterns is formulated as follows [15]: Figure 1 represents the modeling framework of this study, inspired by [12], with different indicators and covariates.Let   denote the response of participant  on indicator .The latent class  has  categories and intervenes between indicators  and covariates    .Two kinds of probabilities are present in LCCA: the probability of the membership model (|  ) and the probability of the measurement model (  |,    ).The optimal number of clusters is examined using Bayesian Information Criterion (BIC), where the low values are associated with better model fit [16].

Results
The descriptive statistics of the participants are given in Table 1.Most participants are male (68%), undergraduate students (65%), and aged under or equal to 23 (73%).The sample is close to being representative since around 69% and 61% of BME students are male and undergraduate students [17].The majority of respondents come from higher-income households (69%) and have at least one car in their households.Regarding transport mode usage, Figure 2 shows that more than half of the participants use PT several times per day.Private cars are used occasionally, as 25% use them several times per week.The figure indicates that shared transport modes are not popular among respondents.We tested ten models to examine the most suitable number of clusters.The BIC values continue to decline until the 4-cluster model.Thus, the 4-cluster model is selected for further analysis.Table 3 presents the parameters of the final model.The class-specific parameters indicating the (un)attractiveness of the indicators are given in the measurement model.The most significant positive sign for cluster-1 is walking.For cluster-2, PT is mostly preferred by the respondents.Meanwhile, participants in cluster-3 tend to have higher shared mobility usage frequency.Finally, college students in cluster-4 are private car users.Moving on to the structural model, female, younger, and lower-income students are more likely related to cluster-1 and cluster-3, while male, older, and higher income students are positively associated with cluster-3 and cluster-4.More cars lead to higher positive parameters in cluster-4.Similarly, students having a driving license are generally related to cluster-4.Additionally, college students who spend less than 11 minutes going to the university are primarily associated with cluster-1.In comparison, cluster-4 is more related to college students who spend more than 30 minutes to reach the university.The socio-demographic and travel characteristics of clusters are presented in Figure 4.The WPC cluster and PW cluster are represented by female students (31% and 42%) who are aged under 23 (86% and 68%) and come from lower-income households (36% and 35%).Meanwhile, the BS cluster and C cluster are characterized by male students (82% and 71%) who are aged more than 23 (45% and 47%) and come from higher-income households (79% and 84%).In the case of travel characteristics, students in the WPC cluster have two cars in their households (45%) and spend less than 11 minutes going to the university (26%).The most prominent characteristics of students falling into the PW cluster are those who do not have cars (47%) and a driving license (32%).College students in the BS cluster have one car in their household (39%), and take around 11-30 minutes to reach the university (63%).Finally, the C cluster is represented by students who have more than two cars in their households (32%), and spend more than 30 minutes going to the university (64%).

Discussion
This study investigates the heterogeneity in transport mode choice among university students, where the LCCA is performed to reveal cluster membership.Four clusters are identified, including WPC (group of students who use PT, walking, and car), PW (group of students who use PT and walking), BS (groups of students who use the bike and shared mobility services), and C (group of students who use private cars).The WPC and PW clusters mostly consist of female students, which aligns with studies stating that female students are more likely to walk [6] and dominate the PT users class [18].The majority of lower-income students belong to the WPC and PW clusters, which aligns with the study implying lowincome millennials are related to the most frequent PT users [1].The BS cluster is dominated by male students, which resonates with a study arguing that male students are likelier to bike to the university [7].A significant percentage of college students from higher-income households is present in the BS and C clusters.This is aligned with a study in Australia revealing that individuals with relatively high incomes increase the probability of bike-sharing memberships [19].Similar findings are reported by a Dutch study confirming that higher-income people are more inclined to use car-sharing [20].
In terms of travel characteristics, college students in the WPC cluster can reach the university in less than 11 minutes.Most probably, they live near the university.College students in this cluster are occasional users of private cars.The PW cluster consists of college students who have no cars in their households.Conversely, the C cluster is characterized by students who have more than two cars in their households.It resonates with a study arguing that students with cars available to their families are more likely to commute to the university by car [5].Meanwhile, university students who spend around 11-30 minutes to reach the university are more likely to fall into the BS cluster.This is understandable since most probably they do not live near the university, while their families only have one car.Thus, besides PT, using bikes or shared mobility services is the most reasonable choice for them.
As for policies to the shift to more environmentally friendly transport modes, we focus more on the C cluster.While it is clear that college students using private cars are mostly present in this cluster, they also show some interest in car-sharing and e-scooter-sharing services.Thus, placing more shared mobility points around the university and PT stops could be a good way to decrease private car use.The multimodal classes (WPC, PW, BS) already show sustainable travel behavior.Thus, a breakthrough mobility solution, such as Mobility as a Service (MaaS) could target college students as its potential users since the service promises more accessible ways to use a wider range of transport modes through one single system [21], and college students show a high interest in adopting the service [22].Since college students in WPC and PW classes come from lower-income households, the mobility companies could support them in favor of enhancing the use of shared sustainable transport modes by giving discounts.
The first limitation of this study is related to the use of self-reported travel pattern data.While the self-reported travel pattern data is generally faster and more affordable, the reliability is questionable [23].Future studies could use travel data from the government or mobile applications with GPS trackers.Since this study focuses on college students, the results are not necessarily generalizable to other groups of individuals.

Conclusion
This study investigates the heterogeneity in transport modes of university students, including shared mobility services.The LCCA method is used, where the level of transport mode usage serves as indicators, and socio-demographic and travel characteristics are included as covariates.Four clusters are identified: WPC, PW, BS, and C. The WPC cluster captures students who mostly walk to the university and use PT and a car occasionally.The PW cluster consists of lower-income college students who use PT several times per day.The BS and C clusters consist primarily of higher-income students.College students who use bikes and shared mobility services are in the BS cluster, while those who use private cars mostly fall into the C cluster.Responsible institutions could utilize the current study to set the policies encouraging the shift to more environmentally friendly transport modes.

Figure 2 .
Figure 2. Distribution of participants' usage level of transport modes.

Figure 3 .
Figure 3. Distribution of the usage level of transport mode within clusters.

Table 1 .
Descriptive statistics of respondents.

Table 2 .
Model fit of the LCCA models.

Table 3 .
Parameters of the LCCA model with 4 clusters.

Table 4 .
The within-cluster distribution of indicators.