Joint factor and latent class ordered probit analysis of intent to use unstaffed automated buses

Abstract

By adopting a more environmentally friendly driving style and promoting sustainable transportation options, automated buses (ABs) could play a role in achieving long-term sustainability goals. This study explores the factors influencing the willingness to use ABs, based on survey data from 451 bus users in Scotland. Using latent class ordered probit analysis, distinct user groups were identified based on their intentions to use unstaffed ABs, along with the key factors influencing their class membership. Concerns about driving capabilities and the implementation of technology emerged as the strongest negative predictors for urban residents without health issues. For rural residents and those with health issues, use intentions were primarily influenced by expectations of the benefits associated with ABs. These insights are valuable for bus manufacturers, service operators, and transport authorities, informing the development of policies and strategies for the future deployment of SAE level 5 ABs.

Introduction

The past decade has seen extensive research into the potential benefits of fully automated private vehicles. Automated driving styles can be more precise, efficient, and consistent than human ones, thus improving safety, optimising fuel consumption and reducing emissions (both from combustion and tyre wear). Automated Vehicles (AVs) can enhance access to motorised mobility enabling a wide range of solutions and thus increasing the independence of people who do not drive¹. Clearly, most benefits will only be achieved when AVs become widely adopted, connected and affordable². The deployment of personal AVs may generate undesirable consequences such as increased number of vehicles on the road because of lower vehicle occupancy, additional trips made by people who are unable to drive due to factors such as age, disability or lack of driving skills, and additional, non-essential and/or longer trips associated with more productive in-vehicle time for people who would otherwise drive. Such consequences may subsequently result in more traffic congestion and emissions³. Furthermore, the increased utility of car trips for existing drivers may contribute to expansion of urban sprawl^4,5.

The implementation of automation in public transport can mitigate the non-sustainable consequences of the use of private AVs linked to number of vehicles and trip length, while still contributing to safety, energy consumption reduction and mobility supply enhancement. Incorporating Automated Buses (ABs) into a transportation system has the potential to create a more efficient, environmentally friendly, accessible public transport network and encourage sustainable mode choice, contributing to long-term sustainability goals. Public transport operating with no personnel on board, could offer further advantages including operating cost savings, a solution for the issue of staff shortages, and eliminating the constraints imposed by shift scheduling. Further advantages include the ability to extend routes and, for services not requiring guideways, increased flexibility, longer service hours and more frequent services (due to lower operating costs). Public transport operators can take advantage of automation to achieve their objectives of delivering a higher quality service while maintaining fare affordability, thereby increasing the appeal of bus services⁶.

Clearly the extent of the benefits generated by automation in public transport will be proportional to the use of the services. Implementing automated technology in public transport could make the abovementioned benefits happen only if it succeeds in promoting public transport use and attracting new users beyond the captive ones. There are numerous obstacles to the widespread adoption of AVs on public roadways, like reliable and safe operations in mixed traffic, regulatory approval, high upfront costs and system security. However, one of the primary challenges to achieving the benefits may be user acceptance of the technology⁷, especially during the initial stages of deployment. Considering this background, it is crucial to understand user attitudes towards Automated Public Transport Vehicles (APTV) to inform the roll out of future services, particularly in relation to the presence of staff on board.

Prior research focused on investigating perceptions and expectations of bus passengers suggests that services without any employees would be viewed less favourably by the public^8,9. As a result, this study examines the factors that influence the willingness to use ABs operating with no staff on board. The intention to use ABs is statistically analysed by means of latent class ordered probit models with allowances for class membership functions. The study results can contribute to the efforts towards enhancing the attractiveness of ABs by providing a more comprehensive understanding of the elements determining users’ decisions to ride in them. This understanding will assist bus manufacturers, service providers and transport authorities in developing policies, programs, and strategies to successfully move towards the introduction of SAE¹⁰ Level 5 ABs in the future.

Literature review

Theoretical framework

The research topic of AVs is presently undergoing substantial growth and has been the subject of extensive research in the previous decade with numerous studies examining the factors that impact the public’s willingness to accept AVs³. Theoretical models such as the Technology Adoption Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT) were designed to explore the acceptability of new technology in general. The UTAUT model is derived from a combination of eight acceptance models (including TAM). It proposes four constructs: performance expectancy, which relates to the perceived usefulness and the degree that using the system will enhance travel activities; effort expectancy, which measures the ease of use of the system; social influence, which reflects an individuals’ perceptions of the positive or negative attitudes of others towards the new technology, and facilitating conditions, which encompasses individuals’ beliefs regarding the availability of assistance and resources. The impact of these four constructs on intentions and use is moderated by individual differences of gender, age and experience. In UTAUT2¹¹ three additional constructs were introduced: hedonic motivation, which refers to the enjoyment derived from using the system; price value which is the balance between the system’s benefits and its monetary cost; and habit, which is the extent to which the individual is accustomed to using the system. In 2016, UTAUT3¹² was proposed, which utilises UTUAT2 as a baseline model and incorporates individual contextual factors (technology, task attributes, events) and higher level factors (environment, organisation and location).

Nordhoff et al.¹³ proposed the Multi-level model for Automated Vehicle Acceptance (MAVA) which integrates the factors of TAM and UTUAT as well as safety aspects, vehicle and service characteristics and perceptions of risks and benefits. The willingness to use AVs is determined by individual differences which moderate the meso-level factors. However, in contrast to UTUAT, MAVA suggests that individual differences can also have a direct impact on the willingness to use AVs.

Figure 1 illustrates the MAVA framework applied to the specific AB context. Despite MAVA comprising up to 70 elements³ which may render it too cumbersome for practical use by researchers, its robust and methodical basis in empirical research make it highly relevant for studies on AB adoption.

Fig. 1: AB acceptance framework.

The MAVA framework applied to an AB context, illustrating the factors considered in this research, including the meso-level AB system evaluation and intention to use ABs, as well as micro-level individual differences.

Factors influencing willingness to use APTVs

In the context of APTV, both TAM and UTAUT have both been employed in exploring public perceptions and their influence on the adoption of this technology. The combination of TAM components, together with trust, social influence, emotional attitudes, and prior experience, have been recognised as indicators of willingness to use for both AV shuttles and ABs^14,15,16,17. Researchers have also integrated the UTUAT framework with variables including trust, perceived value, perceived risk, service quality and technology-task fit to examine acceptance of APTVs^6,18,19,20.

When considering sociodemographic characteristics, the impact of age on acceptance varies significantly across the research³. Five studies^{15,19,21,22,23} show no significant age effect on willingness to use APTVs. However, one study⁸ indicated that those aged less than 45 years of age are more likely to use ABs than those aged 45 and above. In contrast, two studies^20,24 suggest higher intentions to use APTVs among older individuals. Age has also been found to moderate perceived ease of use²⁵ and trust²⁴, with older participants valuing ease of use more, while trust has a greater impact on acceptance for younger age groups. Gender also plays a role in APTV acceptance, but the findings are similarly inconsistent. Two sources^8,26 show that men are more willing to use ABs, while three studies^19,23,26 found no significant gender differences in willingness. Another study²⁵ found that gender has a moderating effect on perceived usefulness.

Prior knowledge, awareness or experience of AVs has been associated with an increased willingness to use ABs⁸. Research involving an AB operating on a test route found that after the experience, users reported trust in the AV system and strong optimism about the future of ABs²⁷. Furthermore, research suggests that being exposed to ABs when a human safety operator is present encourages users to try unstaffed services²⁸.

A review conducted by Pigeon et al.³ emphasises several safety considerations related to the acceptance of APTV. These included concerns about the safety of the vehicle, apprehension about technology malfunction, fear associated with night-time operations of the services and concern regarding the decision-making algorithms. The absence of a staff on board raises concerns regarding resolving technical issues, provision of passenger information, assisting passengers with a disability, prevention of anti-social behaviour and the supervision of compliance with regulations^1,3,8,29.

In summary, the findings from studies that investigated the influence of sociodemographic characteristics and psychological factors on the intention to use APTV technology provide heterogeneous results^2,6. Furthermore, most of these studies primarily focus on AV shuttles, which are typically smaller vehicles (with a capacity of up to 12 passengers) that often operate on-demand on short routes, either in closed environments or as first/last mile solutions^15,19,22,30. Research on the acceptance of full-sized ABs operating scheduled services in mixed traffic conditions remains limited^2,9,31,32,33. Out of thirty-nine studies included in a literature review by Pigeon et al.³ on the acceptance of non-rail APTV, only four specifically focus on full-sized ABs. Similarly, this review found a limited number of AB studies, and Table 1 summarises the eight that were identified. One study found higher intentions to use AV Shuttles compared to full-sized ABs³⁴, which the authors attribute the public’s greater familiarity with AV shuttles. These findings highlight the need for further research into the public’s intention to use full-sized ABs.

Table 1 Overview of acceptance studies involving full-sized ABs

Study contributions

The long-term deployment of bus services without personnel on board necessitates an understanding of the initial perceptions towards them, particularly in the light of rapid advancements in the field. For instance, Level 4 ABs are currently operational on UK roads as part of Project CAVForth³⁵. Although the importance of driver control and supervision has been raised in some of the past studies, little is known about the intention to use full-sized ABs without any staff on board. Hence, the main objective of the present study is to understand differences in intentions to use unmanned ABs. For this purpose, MAVA¹³ is used as the theoretical framework. Moreover, most previous studies use data gathered from the general population to investigate people’s attitudes towards AV technology. Our study is distinct in that it relies on survey data obtained from existing bus users, who are more likely to be affected by the deployment of AB technology. This targeted approach allows us to gain a more extensive understanding of the specific concerns and preferences of individuals who regularly use buses. By focusing on this specific group, we can provide insights that may be overlooked in broader surveys of the general population.

From a methodological perspective, previous studies have employed several statistical approaches to analyse the relationship between dependent variables measuring intentions to use any public transit system and potential influential factors. In cases of ordered dependent variables (as, for example, with Likert-style questions), the ordered response modelling has been the most prevalent approach³⁶. The traditional ordered models impose a restrictive homogeneity assumption regarding the impact of exogenous variables on the ordered response levels³⁷. In addition, the effect of exogenous factors on the dependent variable is typically pre-assumed as constant across the sample; such a consideration may induce significant bias on statistical inference, as the—potentially significant—impact of unobserved heterogeneity on the statistical estimation is not acknowledged at all³⁸. To overcome such limitations, researchers have employed, among other approaches, latent class (or finite mixture) ordered models to accommodate heterogeneity stemming from unobserved groups of the population (classes). The approach has proven statistically and empirically robust and has received increasing attention in studies in the transportation field^36,39,40,41. Such models enable the estimation of class-specific parameters, which evidence the varying impact of the exogenous variables across the class. In this regard, this study aims to bridge this gap in AB literature by investigating the heterogeneity in the intention to use ABs among the potential users. To provide a more nuanced examination of the heterogeneity compared to previous evidence, two sources of unobserved variations were jointly considered: (i) the latent constructs of bus users’ perceptions and attitudes towards the AB systems; (ii) the latent classes of bus users across which the determinants of AB intentions may vary. To further understand the characteristics that define the latent classes of bus users, a flexible latent class formulation is employed, allowing for the identification of factors that determine the class membership, i.e. which sub-segments of the population are more likely to be represented by each latent class. The findings are expected to help transport operators and policymakers develop policies tailored to different user groups rather than generic policy measures, thus making efficient use of time, effort and money.

Thus, the following are the contributions of this study to the existing literature on AB acceptance: (1) the validation of the MAVA framework in the context of ABs, (2) the determination of the prevalence of intention to use ABs without staff on board, (3) an explanation of the heterogeneity of intentions in relation to attitudinal, behavioural and sociodemographic attributes of those surveyed and, in particular, (4) the identification of different user groups and their relative characteristics.

Methodology

Data collection

The factors that influence the intention to use automated buses with no human personnel on board were determined using data collected through a web-based survey undertaken during the course of the CAVForth project³⁵ which introduced a trial service of a fleet of full-size, single deck, automated (SAE¹⁰ level 4) buses travelling on a 20 km route on the Scottish public road network. The questionnaire was administered through the Qualtrics platform⁴² and disseminated to the Scottish passenger mailing list of the bus operator, Stagecoach. The survey was carried out during the last quarter of 2021, before the CAVForth trial began. The questionnaire, which was granted governance and ethical approval from Edinburgh Napier University, primarily sought to assess the attitudes and eagerness of passengers to use ABs. Of the 1,054 replies received, 603 were excluded from the analysis due to missing responses for at least one attitudinal (MAVA) item. As a result, 451 responses were used for the study. The sample size was further reduced during the modelling process when sociodemographic data was introduced, as additional ‘don’t know’ or missing responses were identified.

The questionnaire consisted of predominantly closed-ended questions that addressed the following subjects: sociodemographic characteristics (e.g. gender, age, employment status, educational qualifications, health issue affecting travel choices, financial concerns), and household characteristics (e.g. household income, number of cars available, number of adults and children). Additionally, the questionnaire sought information regarding travel behaviour (e.g. modes of travel, frequency of travel, satisfaction with bus services, car dependency, having a concessionary travel pass), general eagerness to adopt any new technology, familiarity with current Automated Driver Assistance Systems (ADAS), attitudes and feelings towards ABs and eagerness to use them. The items assessing perceptions of risks and benefits associated with ABs, as well as the level of comfort with ABs operating in different situations, were presented to respondents in a sequential order. Respondents were initially introduced to the concept of SAE Level 4 ABs through a brief overview of the CAVForth project, which highlighted that the buses have computer systems capable of performing all driving functions (such as steering, accelerating or braking) without direct driver input. However, all trial buses have trained human drivers behind the wheel at all times, whose role is to monitor the roadway and vehicle throughout the journey and to take control of the vehicle if necessary. The participants then answered a series of questions relating specifically to this type of service. Subsequently, participants were presented with a section on future services with fully (SAE level 5) automated buses where no driver would be present in the cab, highlighting that such services would not apply to the CAVForth trial. Respondents were then asked questions regarding their willingness to ride in an AB, with varying levels of staff involvement. The intention to use future AB services in the absence of a human driver to intervene and a steward to provide customer care was assessed using an ordered Likert-type scale that included five categories: ‘definitely yes’, ‘probably yes’, ‘might or might not’, ‘probably not’, and ‘definitely not’. The variable was used as the dependent variable in the analysis, as the intention to use a new technology is a critical predictor of its actual use.

The dependent variable (i.e. intention to use ABs without staff on board) was determined in light of the MAVA framework. MAVA is based on two theoretical frameworks: the Unified Theory of Acceptance and Use of Technology (UTAUT3)⁴³ and the Car Technology Acceptance Model (CTAM)⁴⁴. In studies on AVs, MAVA¹³ integrates acceptance factors that are allocated at two distinct levels: micro and meso. The exposure of individuals to AVs is captured at the meso-level, which includes instrumental domain-specific factors (performance and effort expectancy, facilitating conditions, safety and service, and vehicle characteristics), symbolic-affective factors (hedonic motivation and social influence), and moral-normative factors (perceived risks and perceived benefits)¹³. Individual differences (such as sociodemographics, personality, and travel behaviour) are micro-level factors that influence the meso-level factors. Although MAVA contains numerous factors, which could make its use by researchers occasionally cumbersome, the profound and systematic base of the framework in empirical research makes it very useful for AV acceptance studies.

Methodological approach

One of the primary methodological objectives of this study, as previously stated, is to simultaneously account for multiple strata of latent variations that may have pronounced impacts on the intention to use ABs. As such, two levels of statistical analyses were jointly carried out: (i) a series of factor analyses to identify the latent constructs underpinning the expectations and attitudes of bus users towards ABs in line with the MAVA framework; and (ii) latent class discrete choice models to comprehensively understand the determinants of the intention to use unstaffed ABs and their variations across unobserved groups of the bus users. In this study, a flexible formulation of the latent class model⁴⁵ is employed consisting of two model components: (i) a measurement model, which includes the determinants of the intention to use ABs; and (ii) a class membership model, which includes the variables determining which latent class each individual belongs to. To integrate these two sources of unobserved effects, the latent constructs are considered in the latent class models through various statistical formulations and inclusion in different model components.

The identification or confirmation of a limited number of factors or latent constructs from a large number of observed variables is achieved through the use of factor analysis⁴⁶. To reduce the dimensionality of the observed attitudinal variables (meso-level variables), an Exploratory Factor Analysis (EFA) was conducted as a preliminary step. The MAVA framework is employed as the underlying theory for the assignment of ‘meso’ items in the questionnaire under various constructs, as previously mentioned. A Confirmatory Factor Analysis (CFA) was implemented to evaluate the validity of the structure that was obtained following the EFA⁴⁷. The factor analysis allows not only to reduce the dimensionality of the data but also to validate the theoretical framework and the measurement instrument (the questionnaire). The latent meso-level constructs obtained after the factor analysis along with the micro-level variables are used as determinants in the latent class modelling.

Since the dependent variable is ordered in nature, ordered probit models⁴⁸ are estimated. The latent class approach is used to address the occurrence of unobserved differences in the willingness to use unstaffed ABs. These differences may arise from the existence of unobserved groups within the target population (i.e. the bus users). Latent class models (LCM) assume that individual responses depend on observable attributes and on latent heterogeneity, which can be partially captured by hypothesising the existence of different, unobserved classes of respondents^45,49,50. However, these classes of respondents are internally homogeneous; in other words, the impact of the observable attributes is consistent across the members of each class, but varying across classes⁴⁵. For ordinal response, the probability of observing outcome k for individual i is a function of a latent variable y*. The latter is a linear combination of a vector of directly observed or derived variables ${{\rm{X}}}_{i}$, of a vector of estimable coefficients ${{\rm{\beta }}}_{m}$ specific to class $m$ and of a random error term ${\epsilon }_{i,m}$ representing unobserved characteristics for class m⁵⁰.

$${y}_{i,m}^{* }={{\rm{\beta }}}_{m}{{\rm{X}}}_{i}+{\epsilon }_{i,m}$$

(1)

Assuming a normal distribution for the error terms, the probability of observing outcome $k$ for individual i conditional on i belonging to class $m$ (measurement model) is:

$${P}_{i}\left({k|m}\right)=\varPhi \left({\mu }_{k,m}-{{\rm{\beta }}}_{m}{{\rm{X}}}_{i}\right)-\varPhi \left({\mu }_{k-1,m}-{{\rm{\beta }}}_{m}{{\rm{X}}}_{{\rm{i}}}\right)$$

(2)

where Φ represents the standard cumulative normal function and ${\mu }_{k,m}$ the estimated thresholds⁴⁹ of the ordered probability ranges. The probability that i belongs to m (class membership model) is defined according to a logit formulation as:

$${P}_{i}(m)=\frac{\exp \left({\gamma }_{m}{J}_{i}\right)}{{\sum }_{M}\exp \left({\gamma }_{m}{J}_{{\rm{i}}}\right)}$$

(3)

With ${J}_{i}$ being a vector of observed attributes determining the class membership, ${\gamma }_{m}$ is a vector of estimable parameters corresponding to $J$, and M is the hypothesised number of classes. Combining (2) and (3), parameters and thresholds can be estimated through the minimisation of the following log-likelihood function:

$$\mathop{\min }\limits_{{{\mu }_{k,m},\beta }_{m},{\gamma }_{m}}{\sum }_{I}{\mathrm{ln}}\left({P}_{i}\left(\hat{k}|m\right){P}_{i}(m)\right)$$

(4)

LCMs constitute a powerful modelling instrument because they probabilistically segment the potential users into groups of homogeneous behaviour, in a way that makes the interpretation of unobserved patterns easier, especially compared to other approaches that account for unobserved heterogeneity, such as mixed logit or latent variable models. Most typical LCMs are considered exploratory in nature, i.e. the number of classes is empirically explored and determined. The two-step modelling process employed by previous studies on LCMs³⁶ is also employed in this study. Therefore, an LCM with constant-only membership model (i.e. the probabilistic allocation of individuals to latent classes is fixed) is built first, and then, class membership models are constructed⁴⁵. Considering the class membership model, the individuals are probabilistically allocated to classes through a function of exogenous variables (class membership function).

To identify the nuances of latent heterogeneity to the best possible extent, we also tested different model specification approaches, which are defined by the type (micro-level vs. meso-level) of variables used in the measurement and class membership models. The investigation of different specification approaches can help determine whether the class membership depends on attitudes towards the AB system, or on individual sociodemographic and behavioural characteristics.

To ascertain the magnitude of the effects of the independent variables on the dependent variable, marginal effects were also estimated. Specifically, these effects illustrate how a one-unit change in the independent variable translates into a change in the probability associated with each outcome of the dependent variable. In line with previous research on latent class ordered models, the marginal effects are averaged across all the latent classes that turned out as significant in model estimation⁴⁵.

Figure 2 illustrates the methodology process flow from study design, questionnaire development and data collection to the various stages of analysis. The next section will present these approaches and the outcomes obtained.

Fig. 2: Flowchart of the study methodology and analytical approach.

Illustration of the research methodology process flow from study design, questionnaire development and data collection to the different stages of analysis.

Results

Demographic characteristics of the respondents

Sociodemographic information was analysed to assess whether the survey sample was representative of Scottish bus passengers as a whole. The Scottish Government publishes annual Scottish Transport Statistics⁵¹ (STS), which provide sociodemographic information on Scottish bus users, defined as those who reported using the bus on the day prior to being interviewed as part of the Scottish Household Survey. Data from 2019 was selected for the comparison as the 2020–21 dataset would be influenced by the significant disruptions to travel patterns caused by the COVID-19 pandemic⁵². Table 2 compares the STS data on Scottish bus users and the Stagecoach customer survey sample. It may be seen that the survey sample is broadly representative of STS bus users in terms of age, gender, and concessionary travel pass ownership. The distribution of employment status is also similar across both datasets, with equivalent proportions of full-time and part-time employees. However, the survey sample contains fewer retired respondents overall. A higher proportion of respondents from rural areas are present in the survey sample, which likely reflects the coverage of the Stagecoach network, where some urban areas are predominantly served by other operators.

Table 2 Comparison of Sociodemographic characteristics of Scottish bus users (STS) and Stagecoach survey sample

The summary descriptive statistics of the micro-level variables used at different stages of the modelling process are given in Table 3.

Table 3 Summary of micro-level variables used at different stages of modelling eagerness to use ABs

Intention to use ABs without staff On board

Participants were questioned regarding their intent to utilise future AB services without a human driver and without a steward to provide customer service. The distribution of intentions is given in Fig. 3. It may be seen that only a fifth (20.2%) of respondents would be likely to travel on a bus that does not have human staff on board.

Fig. 3: Distribution of the responses for the dependent variable.

Distribution of survey participants’ responses regarding their intent to utilise future AB services without a human driver and without a steward to provide customer service. The figure displays both the percentage of responses for each group and the cumulative percentage of responses.

Latent meso-level construct

Exploratory Factor Analysis (EFA) was conducted on the meso-level variables using Principal Component Analysis (PCA) with Varimax rotation. The following criteria were used to extract factors/constructs from observed attitudinal variables using PCA: (i) the Eigenvalue must be greater than 1; and (ii) the factorial loads must have exceed 0.5⁵³. Several variables were systematically eliminated from the analysis because of their low communality, to enhance the alpha values or due to large cross-loadings. This approach aligns with the methodology employed in prior research^2,19. Bartlett’s Test of Sphericity was significant (df = 406, Chi-Square = 11603.347, p < 0.001) and the Kaiser–Meyer–Olkin measure of sampling adequacy was 0.966, which is above the recommended threshold of 0.80⁵⁴, support the suitability of the data for factor analysis. The EFA carried out using SPSS⁵⁵, identified five factors, and they explained 72.4% of the variation in the data. Cronbach alpha values for all constructs were within the threshold of 0.6⁵⁶. The average variance extracted (AVE) for all constructs exceeded the 0.50 threshold⁵⁷. The factor scores of all construct items are presented in Table 4 along with the mean and standard deviations for each item.

Table 4 The five-factor model and loadings obtained after EFA

SPSS AMOS⁵⁵ was utilised to conduct a Confirmatory Factor Analysis (CFA). The factor loadings of each item were evaluated as part of the CFA, and all factors demonstrated satisfactory loadings that exceeded the 0.5 threshold. Where feasible, the error terms that were high were covaried to minimise the discrepancy. An analysis of Modification Indices (MI) was conducted to improve the fit of the model. The threshold for the MI was set at a value of 10. The MI pertaining to the ‘driving capability on public roads with other vehicles’ exhibited a high value, leading to its exclusion from the domain-specific safety construct.

The models’ overall goodness-of-fit was evaluated using well-established goodness-of-fit metrics shown in Table 5 and all values were found to be within the commonly accepted levels. The data analysis revealed that the five-factor model, consisting of Safety, Experience, Affective Motivation, Perceived Benefit, and Perceived Risk, demonstrated a strong fit. Specifically, the following model fit metrics were estimated as: - CMIN/df = 1.859, CFI = 0.974, TLI = 0.969, SMRS = 0.034, RMSEA = 0.044.

Table 5 CFA Model fit statistics

Cronbach’s alpha and composite reliability were used to evaluate construct reliability. The Cronbach alpha values for each construct in the study exceeded the minimum threshold of 0.70⁵⁶. The composite reliabilities, which ranged from 0.724 to 0.969, all exceeded the established criterion of 0.70⁵³. The average variance extracted (AVE) was employed to estimate the convergent validity of the scale items. The AVE exceeded the 0.50 threshold⁵⁷. Therefore, the instruments utilised in the present study exhibit the necessary convergent validity. The Heterotrait-Monotrait (HTMT) Ratio was employed to assess discriminant validity. HTMT compares the correlations of indicators within a construct to the correlations of indicators across constructs. The specified limit of 0.85 was not exceeded by any of the ratios that were observed⁵⁸. The results of discriminant validity are presented in Table 6.

Table 6 The five-factor model and discriminant validity

The factors obtained correspond to the theoretical categories of factors proposed in the MAVA framework. The five-factor model is given in Table 7. The normalised scores on each construct are used when the specific constructs are included as potential independent variables in the LCM. The construct ‘Concerns with Driving Capabilities’ encompasses attitudes towards various driving and safety aspects of ABs defined by different operational, road and external conditions, such as weather, lighting, intersection presence, other road users and so on. The construct ‘Expectations about Benefits’ focuses on operational, travel and environmental benefits that ABs are anticipated to bring forth. Another construct arises from respondents’ concerns about implementation of technology, such as risk of hacking, low reliability of technology or lack of situational awareness of the unstaffed AB systems. Affective Motivation emerged as a distinct construct reflecting both users’ enjoyment of ABs and their sense of feeling safe. The role of affective aspects in determining willingness to use ABs was also evidenced in previous studies². Additionally, research on AV acceptance⁵⁹ suggests that feeling safe is closely related to reduced anxiety, which, in turn, enhances user’s enjoyment of the technology.

Table 7 The five-factor model and loadings obtained after CFA

The last construct derived from the CFA is the ‘experience with AV technology’, which captures the level of experience and exposure to features of advanced automated systems. Such a construct serves as a proxy of respondents’ tech-savviness, and its influence on attitudes towards ABs has been documented in previous studies⁹. Although latent variables typically benefit from having at least three indicators for reliable measurement, research⁴⁷ suggests that it is possible to retain a factor with only two items under certain circumstances. In the case of the Experience construct, the two items (‘Experience with ADAS’ and ‘Experience with warning driver assistance systems’) composite reliability of 0.728 was above the 0.7⁵³ threshold and the HTMT Ratio confirmed good discriminant validity well below the of 0.85⁵⁸ limit, showing that Experience is distinct from other constructs. There is also strong theoretical support for including Experience as a construct, as the MAVA framework¹³ specifically highlights experience as a crucial factor in influencing AV acceptance.

Comparisons across models

A key consideration of the LCM specification stems from the type of factors that will be included in the measurement and class membership components. This is an important point, as it will eventually determine which are the direct contributors to the intention to use unstaffed ABs, as well as the characteristics that define distinct, yet latent groups of respondents with homogeneous attitudes towards ABs.

To ascertain whether the latent classes of respondents should be defined in terms of micro- or meso-level variables, the best possible models were estimated considering the following five approaches:

• 1. Traditional Ordered Probit model without latent classes (OP) – this model served as a reference basis.

• 2. LCM with latent meso-level constructs in the measurement model and micro-level variables defining the class membership model (LCM1).

• 3. LCM with micro-level variables in the measurement model and latent meso-level constructs defining the class membership model (LCM2).

• 4. LCM with fixed class probabilities using a combination of latent meso-level constructs and micro-level variables (LCM3).

• 5. LCM featuring a combination of latent meso-level factors and micro-level variables in both measurement and class probability models (LCM4).

These five models were compared considering their statistical fit as a key criterion. Since the statistically significant variables in the best model of each approach were different, the number of observations (N) and the degrees of freedom (${\rm K}$) were also different in each case. The variability of N is attributed to differences in the extent of missing information across the independent variables. Hence, the adjusted McFadden’s pseudo ${R}_{{adj}}^{2}$ was used for the evaluation of the statistical fit and the comparison of the different models. This ${R}_{{adj}}^{2}$ metric is defined as follows⁵⁰:

$${R}_{{adj}}^{2}=1-\frac{{LL}\left({\boldsymbol{\beta }}\right)-{\rm K}}{{LL}(c)}$$

(5)

Where ${LL}\left({\boldsymbol{\beta }}\right)$ is the log-likelihood at convergence, ${LL}\left({\boldsymbol{\beta }}\right)$ is the log-likelihood value at constant, and K is the number of estimable parameters in the model. This metric accounts for the number of observations (indirectly through the restricted log-likelihood at constant of the model) and the number of parameters used to estimate the models. The model yielding the highest R² value (it can range between 0 and 1) is considered as the approach with the best statistical fit.

The LCM approach with micro-level variables in the class membership function and meso-level variables in the measurement model (LCM1) was found to outperform the others (Fig. 4). The plot shows that LCM1 provides the highest adjusted pseudo R² value across all compared modelling approaches. The Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) were also computed to evaluate the statistical performance of the competing models. Among them, the LCM1 model was again found to yield the lowest values for both AIC and BIC, further supporting its superiority in terms of statistical fit and explanatory power. Thus, the introduction of classes consisting of micro-level variables was found to significantly improve the model fit. The next section presents the estimation results of the LCM1 model.

Fig. 4: Comparison of five modelling approaches.

Comparison of five modelling approaches based on adjusted pseudo R² values. The plot illustrates that LCM1, which incorporates micro-level variables in the class membership function and meso-level variables in the measurement model, provides the highest adjusted pseudo R² value.

The results of the Latent Class Ordered Probit model, which was estimated using the econometric software Nlogit6, are presented in Table 8. A specification with two latent classes was identified as the most robust one, as the inclusion of additional classes did not statistically improve the model fit. The class membership model is a function of two micro-level variables, namely the ‘area of residence’ (coded as 1: Rural, 0: Urban) and ‘limiting health issues’ (coded as 1: Yes, 0: No). A Chi-Square Test of Independence indicated that there was no significant relationship between area type and health status (χ² = 0.040, df = 1, p = 0.841).

Table 8 Estimation results of the best latent class model (LCM1)

The average marginal effects of the explanatory variables included in the measurement model of the Latent Class Ordered Probit model are provided in Table 9.

Table 9 Marginal effects of the variables included in the best latent class model (LCM1)

The model suggests a negative relationship between people residing in rural areas and likelihood of membership in Class 1. Similarly, the negative coefficient of ‘limiting health issue’ reveals that people having physical or mental health issues/disabilities that affect their travel choices have a lower likelihood to belong in Class 1. In other words, Class 1 is more likely to include urban dwellers and individuals with no health issues, whereas the opposite is observed for Class 2. The positive sign of the constant term indicates the larger propensity of individuals to belong to Class 1 compared to Class 2.

All the attitudinal constructs are statistically significant (at a greater than 95% level of confidence) in the measurement model of Class 1. The strongest predictor is ‘concerns with driving capabilities’; as shown in Table 8, this factor exerts the strongest effects across all outcomes of the dependent variables and is overall associated with a lack of willingness to ride an unstaffed AB. In contrast, the factor ‘expectations about benefits’ is found to significantly increase the willingness of respondents to ride an unstaffed AB. Specifically, the likelihoods of ‘3-Might or Might Not’, ‘4-Probably Yes’, and ‘5-Definitely Yes’ rise by 0.0872, 0.0741 and 0.017, respectively, as indicated by the corresponding marginal effects. Experience with automation was statistically significant only for Class 1. The relevant coefficient is positive, which means that people without any experience in advanced systems in Class 1 are less likely to use ABs without staff on board. Only two factors were statistically significant in Class 2 (consisting of rural dwellers and those with health issues), with the strongest being the ‘expectation about benefits’. ‘Concerns with driving capabilities’ was another factor affecting intentions of Class 2 members, but to a lesser degree when compared to those in Class 1.

Discussion

The study found an observable minority (20.2%) of participants would be willing to travel on a bus with no employees on board. This suggests that the general public is not convinced of the capability of automated technology, and they may also value the non-driving tasks provided by bus drivers, such as access to information, ensuring passenger safety, handling technical problems, preventing of anti-social behaviour and assistance to people with disabilities^1,3,8. Attitudinal constructs were identified by means of EFA and CFA. The components of the constructs are aligned with those proposed by the MAVA framework. This can be considered a confirmation of the effectiveness of both the framework for studying the acceptance of ABs and the questionnaire used as to measure intentions to use them. Different approaches were used to estimate the latent class models by considering different variable types (micro or meso, or both) in both the measurement model and class membership function. While reviewing the trends and usage of latent class modelling in transportation, researchers³⁶ observed that numerous travel behaviour studies have demonstrated that attitudes are key determinants of latent population segments. However, 72% of the 142 reviewed papers using membership variables had considered sociodemographic variables as the segmentation basis³⁶. The latent class model with micro-level variables (measuring sociodemographic and travel behaviour) defining class membership was found to have the highest statistical and explanatory power in this study. This nuanced approach clearly delivers target groups that can inform tailored policy recommendations.

Area of residence and limiting health issues were the two micro-level variables that affected the allocation of potential users into classes. Disability status was also found to significantly impact public acceptance of APTV in other research⁶⁰. The result suggested that campaigns targeted at improving the public acceptance of ABs should be configured differently in rural and urban areas and for people with disabilities. The constant of the measurement model of Class 1 is significantly positive, whereas that of Class 2 is not significantly different from zero. This suggests a weaker propensity of members of Class 2 to use services with no staff on board, strengthening the importance of paying careful consideration to accessibility/usability of ABs, especially for people with physical and/or mental health issues. The measurement model suggested that the latent meso-level constructs significantly influence the intention to use ABs. These findings are in accordance with previous studies, which considered psychological and behavioural characteristics as determinants of user intentions^{8,15,18,22,61}. This reiterates the finding that attitudinal factors need to be consistently included in studies on public acceptance of ABs².

People without limiting health issues and urban dwellers were more likely to belong to the most prevalent class. All five latent meso-level constructs influenced the intention to use ABs in this class. Among all the significant factors, concerns with AB driving capability and implementation of technology were found to be the strongest predictors negatively affecting the use intentions. Thus, bus operators and stakeholders can positively influence the intentions of urban dwellers by addressing these concerns. Furthermore, at this stage in the technology’s development, there is a need for wider successful deployment of AB services to provide the public with evidence of ABs driving capability. Similar concerns regarding the self-driving capabilities and the potential for hacking were expressed by respondents in a recent study in Jakarta, Indonesia⁶⁰. The findings were in line with those from another study in Virginia, USA, in which participants voiced their concerns about poor interaction between self-driving shuttles and other road users, especially the vulnerable ones (e.g. other vehicles, pedestrians, and motorcyclists) as well as the system’s ability to respond in unfavourable circumstances, such as poor weather conditions or traffic collisions⁶².

Thus, in urban areas, it is important to inform and reassure people as to driving capabilities of AB under various situations, such as driving at night, in bad weather conditions, while boarding and alighting at bus stops, and on busy roadways with pedestrians and other vehicles. Relevant authorities should establish frameworks for security standards and regulations on the operation of ABs, and these measures should be widely communicated to people as a means of mitigating against concerns about system failure and cyberattacks. Certain studies have also proposed countermeasures, such as a stop button or a communication device for emergency situations, vehicle malfunction, or passenger antisocial behaviour³. Manufacturers and stakeholders ought to allocate greater resources towards the advancement of technology and innovation for bus automation to mitigate the collision risk of ABs and improve reliability and predictability of AB systems¹⁸. The enhancement of user confidence and acceptance is subject to adequate awareness of AB passengers about the processes underpinning the vehicles' functioning and intended manoeuvres. Bus stewards could proactively engage passengers during AB trials, providing clear and concise explanations regarding the operation of sensors, control, and navigation systems². Designing AB systems that are more suited to the requirements of passengers may assist in strengthening their confidence.

The significance of Affective Motivation suggests that AB service providers should find ways to make the on-board experience of passengers pleasurable, a theme common to the use of all forms of public transport. The results corroborate the findings of previous studies indicating that positive emotions can influence users’ behavioural intention. Users who derive enjoyment and pleasure from riding in APTVs are likely to perceive more value in using them due to its emotional utility⁶. This can be assured by incorporating, for example, Wi-Fi connectivity, spacious seats, refreshments and entertainment screens on the back of seats in the vehicle design. Public transport operators should make sure that AB trials provide a high-quality, enjoyable travel experience to passengers to enhance favourable attitudes towards AB.

People residing in rural areas and those having physical or mental health issues/disabilities are more likely to belong to the less prevalent Class 2. Even though operation of AVs and their implementation are challenging in rural environments, it is interesting to note that the willingness-to-use ABs of the individuals belonging in this class was mainly affected by their expectations about benefits. The attitude towards perceived benefits of APTVs, such as improved safety and efficiency, was more likely to influence the acceptance of APTVs than the attitude towards the implementation concerns in a previous study⁶⁰. This is an important hint for transport operators and policy makers, since it should inspire them to design bus services that prioritise the benefits that can be generated for users, society, and environment rather than focusing on the commercial or financial interests for the operators. The mobility challenges faced by rural communities are unique and include restricted accessibility, reliance on cars, social isolation, and concerns regarding road safety⁹. The benefits of ABs, such as lower costs of operation, expanding service hours, the flexibility to modify routes and timetables and affordable fares, are expected to solve at least some of these issues, therefore, these benefits are of particular interest to rural dwellers. Potential travellers should be informed that the introduction of ABs would provide better service quality in their areas, hence fostering favourable intentions towards their use. In rural areas, Government should introduce public campaigns, especially through social media, TVs, and radio to introduce ABs and to make the clear and distinct benefits of ABs to potential passengers in rural areas. They should also emphasise the ‘economic (e.g. saving money, potentially reduced fares), operational (e.g. improved safety, convenience, comfort, speed and reliability), social and environmental (e.g. decreased traffic congestion, air and noise pollution), and emotional utilities (e.g. exciting, fun and enjoyable) of APTV⁶. Additionally, particular attention should be directed towards those with health conditions or disabilities. Advocates could be appointed to reassure these people about the accessibility, convenience, and comfort of ABs in order to encourage their use among them. These segments of the population potentially face barriers to their accessibility, and the capability of ABs to improve bus services, especially in deprived or remote areas, can assist in addressing these barriers. As such, the acceptance of these groups of users (who are often captive to public transport⁶³) is a prerequisite for ABs to realistically contribute towards mitigating mobility inequalities concerning disadvantaged users. However, issues such as provision of support for wheelchair users remain unresolved in relation to how unstaffed AB systems are designed and operated and should be further examined through future design of AB vehicles and technology. Table 10 provides a summary of key findings from the study, along with specific recommendations for stakeholders.

Table 10 Summary of key findings and recommendations for different stakeholders

Conclusions

This research adds to the existing literature on attitudes toward ABs by using data obtained through an online survey from Stagecoach bus passengers in Scotland. ABs have already been introduced in the area through large-scale pilot implementations; as such, a thorough understanding of user intentions and the mechanisms shaping such intentions is essential in the initial stages of deployment to frame effective policies aimed at reaping maximum benefits from the service. The investigation of attitudes towards ABs was informed by the MAVA framework. The results of the exploratory and confirmatory factor analysis of the survey questions provide tangible evidence supporting the suitability of the framework for ABs. The latter demonstrates that the views of respondents in relation to domain-specific, symbolic-affective, and moral-normative constructs shape their intention to use ABs.

Intentions to use were statistically analysed through a latent class model using as independent variables socioeconomic and demographic characteristics and the attitudinal constructs of MAVA. To identify the latent class formulation with the highest explanatory power, different combinations were investigated, considering interchangeably micro-level and meso-level factors in the measurement and class probability model components. The results revealed the existence of different classes of respondents, distinguished by the micro-level factors of the MAVA framework, and specifically, area of residence and health status. Interestingly, the relative impact of the attitudinal constructs on intentions to use varies across the two latent groups of users: concerns with the driving capability of ABs and implementation of technology are found to be the most influential predictors of intentions to use ABs among people in urban areas and without health issues. Expectations about individual and societal benefits of ABs emerged as the most pronounced determinant for rural dwellers and those with long-term mental or physical health problems. These findings highlight the need for tailored policy actions that effectively disseminate the benefits and operational capabilities of ABs across both latent groups of users to booster the user confidence towards the AV technology, and subsequently, foster the adoption of ABs, encouraging the adoption of sustainable transport modes and contributing to long-term sustainability goals.

Our work has some limitations and calls for further research. A review⁶⁴ examining AV acceptance studies found that they often suffer from biases, including systematic errors and sentiment bias. Firstly, data were gathered through a web-based survey of existing bus passengers, which introduces potential self-selection bias. Individuals with a high interest in self-driving technology may be more likely to engage with the survey after receiving the email invitation about AB research. Also, participants were required to be Stagecoach customers who had previously interacted with the company online and were registered for the Stagecoach mailing list. This sampling strategy may have led to an overrepresentation of individuals who are more comfortable with technology and online platforms, excluding groups such as older adults or individuals with lower technological proficiency who may have limited access to or familiarity with online surveys. Occasional bus users who purchase tickets on board rather than online may have also been excluded. However, the use of email invites to customers allowed for greater control over the sample compared to a broader open-access web-based survey. A more inclusive sampling method involving on-board surveys, or a combination of online and offline data collection, could have captured occasional bus users and those who have not signed up to the Stagecoach mailing list. Initially, the methodology design involved collecting data via face-to-face through on-board surveys. However, due to the ongoing COVID-19 restrictions in 2021, face-to-face data collection was not feasible.

In terms of the representativeness of the survey sample in relation to the broader Scottish bus using population, Table 2 suggest that while the survey sample is representative of bus users in several key areas (e.g. age, gender, employment and concessionary pass ownership), certain demographic groups, such as retired individuals and urban residents, are underrepresented. These differences should be taken into account when interpreting the findings of the Stagecoach customer survey. It is also important to note that the sample is specific to Scottish bus users and may not reflect the characteristics of bus passengers in other countries, where sociodemographic profiles and the availability and quality of bus services could differ significantly. The fact that the participants are bus users may mean, on the one hand, that they have different views from those of the general population. On the other hand, being acquainted with bus services, their opinion regarding the importance of the presence of staff on board may be less arbitrary and informed by previous experience. In any case, a future survey of the general public would provide more comprehensive evidence as to the validation of our results.

Secondly, the questionnaire design presented items within each construct in a fixed, rather than randomised, order, which raises the potential for order effects such as fatigue or primacy bias. However, no such effects were observed in the measurement model: neither early nor late-position items were systematically removed during the EFA, and the internal consistency of the constructs remained unchanged when later items were excluded. Nonetheless, it is recommended that future research randomise item order within constructs.

Thirdly, the survey question was a specific context involving willingness to use ABs without a steward to provide customer care and without a driver. However, the survey did not ask respondents about their willingness to use different types of unstaffed AB services, such as those with remote assistance available from a control centre or with a robot assistant available on board. Future research should include different types of unstaffed AB service provision.

Fourthly, the development of the latent class models is constrained by the number of available observations. A larger sample size would help segment the population at a more disaggregate level. Within-class variations may not be fully captured by the employed latent class approach; future work could explore the estimation of within-class random parameters, which is, however, expected to increase significantly the complexity of the estimation process⁴⁵. Fifthly, the separate estimation of the latent factors and the latent class ordered probit model may lead to potential parameter bias due to measurement error in the latent factors. Although the flexibility of the latent class models with class membership functions can partially mitigate this bias, future research could explore full joint estimation of the latent constructs and the latent class discrete choice model to allow for a more rigorous treatment of measurement error.

Sixthly, the ‘Experience’ construct included two items related to respondent experience with automated vehicle features. Future research should consider incorporating additional experience-related variables, such as participation in AV trials to strengthen the construct further. Seventhly, the survey captured the opinions and expectations of bus users without any exposure to the technology or any services featuring it. Understanding the perceived intentions to use is still important to inform and ensure a successful launch of AB services. However, the actual experience with the AB service is expected to transform attitudes and behaviour in different directions, and hence, studying the public acceptance of ABs following their actual use constitutes a key research direction for the future.

The relevance of the views regarding the effect of the technology on environmental and road safety performance calls for these aspects to guide the efforts of technology developers and regulators. These impacts should be attentively monitored and communicated in future research and industry endeavours. Our research corroborates the importance of considering the needs of people with chronic health problems or disabilities affecting their capacity to travel, highlighting the need for further research to enhance the broader acceptance of ABs. It should be noted that this study exclusively focuses on the acceptance of unstaffed ABs. However, varying staffing conditions (e.g. with safety driver only, with bus stewards or both personnel present) may exert disparate effects on intentions to use the service⁹. In this context, a better understanding of how these staffing setups influence attitudes and preferences towards ABs can inform the design of more inclusive services tailored to the needs and special characteristics of different population segments.