Drivers’ Safety Perception in Autonomous Vehicle Road Sharing: A Knowledge-Segmented TPB and Ordered Logit Analysis

Abstract

The large-scale deployment of autonomous vehicles (AVs) in mixed-traffic environments raises an important question: how do human drivers evaluate safety when interacting with AVs under real-world uncertainty? This study aims to examine how drivers’ objective knowledge of AVs shapes their perceived safety when sharing the road with AVs in mixed-traffic environments. Using survey data from 905 licensed drivers in Wuhan, China, this study treats perceived road-sharing safety as an interaction-level evaluative outcome rather than merely a precursor of adoption intention. Latent class analysis was first used to identify knowledge-based driver segments, structural equation modeling was then applied to estimate Theory of Planned Behavior (TPB)-related psychological constructs, and ordered logit regression was finally employed to examine the determinants of perceived safety across segments. The results indicate that behavioral intention consistently shows a positive association with perceived safety; however, attitude toward AVs exhibits a significant negative association among high-knowledge drivers. This attitudinal reversal challenges the implicit homogeneity assumption embedded in conventional TPB applications and suggests that cognitive familiarity may recalibrate, rather than amplify, technological optimism. Overall, the findings show that knowledge-based heterogeneity changes the psychological mechanisms underlying safety appraisal in mixed traffic. These insights carry important implications for differentiated communication strategies and trust calibration in transitional automated mobility systems.

1. Introduction

This section introduces the broader research background of autonomous vehicle road-sharing safety, outlines the key research gap concerning knowledge heterogeneity, and clarifies the theoretical motivation of the study. It also presents the study’s conceptual positioning, hypotheses, and empirical strategy, thereby providing a clear roadmap for the subsequent analysis.

1.1. Background: Mixed-Traffic Transition and Perceptual Uncertainty

The rapid advancement of intelligent transportation systems has positioned autonomous vehicles (AVs) as a transformative innovation in contemporary mobility systems. In China, this transition is no longer confined to small-scale demonstration projects. As of 2024, 20 cities (or city clusters) had been designated as pilot sites for intelligent connected vehicle “vehicle-road-cloud integration” applications, indicating that AV deployment has entered a broader policy-supported phase of real-world urban experimentation [1]. However, current deployment occurs within transitional mixed-traffic environments, where human-driven vehicles and AVs coexist. Recent review studies further suggest that such mixed-autonomy traffic is likely to remain a persistent transitional condition rather than a short-lived intermediate stage, and that interactions between human-driven vehicles and AVs have become a central issue in safety, trust, and traffic-operation research [2,3]. This transitional phase represents not merely a technological upgrade but a reconfiguration of human–machine interaction under uncertainty.

The practical relevance of this mixed-traffic context is particularly evident in Wuhan, the empirical setting of this study. Publicly reported figures indicate that by early 2025, Wuhan had opened more than 3379 km of AV test roads, accumulated over 2.5 million autonomous-driving service orders, and served more than 3.3 million passenger trips. AV–human vehicle coexistence in Wuhan is not hypothetical or marginal, but a repeatedly experienced component of the urban mobility environment. As AV deployment expands from pilot testing to routine urban service provision, the frequency of encounters between human-driven vehicles and AVs is likely to increase accordingly. Under such conditions, understanding how drivers evaluate road-sharing safety becomes an empirically urgent question rather than a purely anticipatory one, especially because recent evidence shows that perceived safety and acceptability in mixed-autonomy traffic remain highly sensitive to interaction context and driving behavior [4]. Recent experimental and simulator-based studies further show that human drivers do not respond to AVs as passive observers. Instead, they adapt their judgments and maneuvers according to the AV’s perceived driving style, yielding strategy, and behavioral predictability in specific interaction scenarios [5,6]. This suggests that safety perception in mixed traffic is not merely a generalized attitudinal stance toward automation, but a context-sensitive appraisal shaped by ongoing human–automation interaction.

Although AVs promise improvements in operational consistency and crash reduction, interaction asymmetries between algorithm-based driving logic and human adaptive behavior may generate perceptual uncertainty [7,8]. Drivers must interpret AV behavior without full transparency regarding system decision rules. Consequently, safety in mixed traffic is evaluated not only as an objective performance metric but as a subjective interpretive judgment formed under ambiguity.

Safety perception has repeatedly been identified as a central determinant of AV acceptance [2]. Drivers frequently report discomfort stemming from limited familiarity with AV operational logic and concerns regarding predictability [9]. Socio-demographic characteristics—including age, gender, and education—further moderate trust formation and safety appraisal [10]. In rapidly digitizing contexts such as China, where exposure to automation varies widely, understanding the cognitive mechanisms underlying safety perception becomes particularly important.

Thus, perceived safety in mixed traffic should be conceptualized as a situational evaluative process shaped by cognitive familiarity, psychological disposition, and structural heterogeneity.

1.2. Research Gap: Knowledge Heterogeneity and Segmented Psychological Pathways

Research on AV acceptance has evolved from descriptive perception studies toward theory-driven behavioral modeling. The Theory of Planned Behavior (TPB) has become a dominant explanatory framework, positing that attitudes, subjective norms, and perceived behavioral control shape behavioral intention [11].

However, two limitations remain. First, most TPB-based AV studies implicitly assume population homogeneity. Psychological pathways are estimated as if all drivers process technological uncertainty in similar ways. Yet technological familiarity is unevenly distributed. Drivers differ substantially in AV knowledge, system understanding, and exposure to pilot testing. Such cognitive heterogeneity may alter both baseline safety appraisal and the strength of TPB pathways. Second, safety perception is often embedded within adoption intention models. This conflates two analytically distinct constructs. Adoption intention reflects forward-looking commitment, whereas perceived safety represents an immediate evaluative judgment formed during interaction. From a cognitive appraisal perspective, safety perception may precede or even operate independently of intention formation.

Therefore, a key unanswered question is whether knowledge-based heterogeneity systematically restructures psychological mechanisms underlying perceived safety in mixed traffic.

1.3. Study Motivation and Theoretical Contributions

This study seeks to explain heterogeneity in drivers’ perceived safety when sharing roads with AVs, with a specific focus on knowledge-based segmentation.

We make three theoretical contributions.

First, we reposition perceived safety as an interaction-level evaluative outcome rather than a secondary attitudinal component within adoption intention models. This shifts the analytical emphasis from intention-based acceptance to situational safety appraisal. This theoretical repositioning directly engages with prior syntheses emphasizing perceived safety and risk as central determinants of AV adoption [2,8]. While existing research typically models safety as an explanatory variable of intention, we argue that in rapidly deployed Robotaxi environments—where exposure precedes formal adoption decisions—safety appraisal emerges as a primary evaluative process rather than a derivative construct. In such transitional contexts, drivers continuously interpret automated behavior during real-world encounters, making interaction-level safety cognition analytically distinct from long-term adoption commitment. This repositioning is also consistent with the trust-calibration literature, which emphasizes that human responses to automation depend not simply on positive or negative attitudes, but on whether trust is appropriately aligned with the system’s perceived competence, transparency, and controllability [2]. In mixed-traffic AV contexts, drivers continuously interpret the behavioral adequacy of automated systems in situ, which makes perceived safety closer to a dynamic interactional appraisal than to a stable intention construct.

Second, we introduce knowledge-based heterogeneity as a structural moderator of TPB pathways. Rather than assuming uniform psychological mechanisms, we examine whether attitudes, subjective norms, and perceived behavioral control operate differently across cognitively distinct driver segments. This extends TPB into a heterogeneity-sensitive framework applicable to emerging mobility technologies.

Third, the methodological integration serves a theoretical rather than merely technical function. Instead of combining TPB, Latent Class Analysis (LCA), and ordered logit for comprehensiveness, we embed latent cognitive segmentation within structural modeling to test whether psychological parameters remain invariant across informational strata. This moves beyond additive model stacking and enables direct examination of parameter instability under knowledge heterogeneity—a dimension largely overlooked in homogeneous structural equation modeling (SEM) applications within AV research.

1.4. Literature Positioning: Socio-Demographics, Cognition, and AV Knowledge

Existing research identifies several sources of heterogeneity in AV perception.

Socio-demographic factors influence safety appraisal. Younger drivers often express greater technological optimism, potentially reflecting digital familiarity [12]. Gender differences are frequently mediated by trust and perceived system transparency rather than direct demographic causation. Education and income may proxy exposure to advanced vehicle technologies and informational environments.

Psychological determinants also play a central role. Recent AV studies increasingly suggest that trust should not be treated as a static psychological disposition, but as a dynamic and interaction-sensitive process that evolves with experience, system behavior, and contextual cues [2,13]. In this regard, perceived controllability is especially relevant: drivers’ evaluations of AV safety are shaped not only by whether they generally support automation, but also by whether they feel able to anticipate, interpret, and respond to AV maneuvers in real traffic situations [5,14]. Research on human–automation interaction further indicates that behavioral predictability, communication clarity, and context-appropriate driving styles are crucial for fostering both perceived safety and acceptance in mixed-autonomy traffic [6].

These insights suggest that AV safety perception is theoretically closer to trust calibration and situational controllability than to abstract acceptance alone. Accordingly, a literature review centered exclusively on TPB and adoption intention would be insufficient for positioning the present study, which explicitly conceptualizes safety as an interaction-level appraisal formed under uncertainty.

However, technological knowledge introduces an additional layer of cognitive differentiation. Drivers with limited familiarity may experience ambiguity sensitivity and heightened uncertainty, whereas more informed individuals may demonstrate stable but potentially more critical evaluations [2].

Latent Class Analysis (LCA) offers a probabilistic framework to identify unobserved knowledge-based segments [15,16]. Applications in transportation research show that preference heterogeneity substantially alters model estimates [17]. Yet few studies explicitly link knowledge segmentation to TPB pathway variation and ordinal safety outcomes.

Despite the growing body of AV acceptance research, few studies explicitly model how objective technological knowledge restructures psychological pathways rather than merely influencing mean-level attitudes. Existing latent segmentation approaches often rely on demographics or attitudinal clustering, leaving cognitive competence under-theorized. By operationalizing objective knowledge as the segmentation basis, this study isolates informational heterogeneity as a structural mechanism shaping risk cognition.

1.5. Theoretical Framework and Hypotheses: Knowledge-Segmented TPB Perspective

This study adopts the Theory of Planned Behavior (TPB) as a structured framework for organizing cognitive determinants of safety appraisal [11,18]. TPB has been extensively validated in transportation psychology and technology acceptance research, demonstrating strong explanatory capacity in modeling attitudes, subjective norms, perceived behavioral control, and behavioral intention [8]. It is important to clarify that the present study does not aim to test the canonical TPB intention–behavior sequence. Instead, TPB constructs are employed as a theoretically organized set of psychological predictors within a safety-appraisal framework. TPB is not adopted here as a self-contained intention model, but as a parsimonious organizational framework for structuring psychological predictors that are themselves informed by broader studies on trust calibration, perceived controllability, and human–automation interaction. Behavioral intention (BIU) is modeled as a concurrent evaluative orientation rather than as a mediating outcome in a hierarchical causal chain. Consequently, estimated coefficients for ATT, SN, and PBC in the ordered logit models represent partial associations conditional on BIU, and should not be interpreted as direct TPB pathway tests.

However, most TPB-based autonomous vehicle (AV) studies conceptualize behavioral intention as the ultimate dependent variable, implicitly assuming a hierarchical intention–behavior pathway. While analytically coherent, this structure may overlook interaction-level psychological processes occurring during real-world mixed-traffic coexistence. In early deployment stages, drivers are not merely forming adoption commitments; they are continuously evaluating the situational safety of interacting with AVs.

Accordingly, this study repositions perceived road-sharing safety as the primary dependent variable. Perceived safety represents an immediate affective–cognitive appraisal formed under uncertainty, shaped by trust calibration, perceived system competence, and social endorsement [19]. Behavioral intention toward AV use is retained as an evaluative orientation that may correlate with safety perception but is not treated as hierarchically superior. This repositioning aligns TPB constructs with risk-perception theory and addresses recent calls for expanding AV research beyond adoption intention [20].

Knowledge-Segmented Extension

Beyond theoretical repositioning, this study extends TPB by incorporating knowledge-based latent segmentation. Existing research indicates that technological familiarity and digital literacy significantly shape AV acceptance patterns [12]. Yet most TPB applications assume population homogeneity, potentially obscuring heterogeneous cognitive processing across drivers with varying AV knowledge levels.

By integrating segmentation into the TPB structure, we acknowledge that the magnitude and direction of psychological pathways may differ across knowledge-based subgroups [21]. This approach advances methodological integration and enhances theoretical precision by embedding cognitive heterogeneity within a structured behavioral framework.

Although the following directional hypotheses are formulated based on conventional TPB expectations, the knowledge-segmented design of this study allows empirical reassessment of their conditional validity. Given that psychological pathways may vary across informational strata, the proposed directions should be interpreted as baseline theoretical expectations rather than invariant structural assumptions.

Importantly, the contribution of the present study does not lie in proposing wholly novel directional hypotheses in isolation from prior TPB research. Rather, its theoretical contribution lies in re-situating these conventional expectations within a knowledge-segmented framework and testing whether their assumed directional stability holds under cognitive heterogeneity. In this sense, H1–H4 are not simple replications of earlier findings; they serve as theoretically grounded benchmark propositions against which segment-specific deviations, reversals, or instabilities can be identified.

Accordingly, the following hypotheses are formulated as theory-based baseline expectations, while the empirical design of this study allows us to evaluate whether these expected relationships remain valid, weaken, or reverse across knowledge-based latent segments.

• Hypotheses

Based on the knowledge-segmented TPB framework, we propose the following hypotheses:

H1: 

A more positive attitude toward autonomous vehicles is associated with higher perceived road-sharing safety [11].

H2: 

Stronger subjective norms supporting AV use are positively associated with perceived safety in mixed-traffic environments [22].

H3: 

Higher perceived behavioral control—reflecting confidence in one’s ability to interact safely with AVs—is positively associated with perceived safety [18].

H4: 

Higher behavioral intention toward AV use is positively associated with perceived road-sharing safety, reflecting alignment between evaluative orientation and situational appraisal [23].

Taken together, these hypotheses specify the expected direction of the main psychological associations under the conventional TPB logic, whereas the central analytical objective of the present study is to test whether these associations are invariant or knowledge-contingent. Furthermore, we expect that the strength and configuration of these relationships will vary across knowledge-based latent segments. Drivers with higher AV familiarity may exhibit stronger attitude–safety linkages, whereas drivers with lower familiarity may rely more heavily on subjective norms or perceived control cues.

By integrating TPB constructs with knowledge-based heterogeneity, this framework advances theoretical understanding of how cognitive familiarity conditions psychological determinants of safety perception in mixed-traffic AV environments.

1.6. Empirical Strategy and Study Context

To test these hypotheses, this study employs a multi-stage analytical framework. First, Latent Class Analysis segments drivers according to AV-related knowledge levels. Second, TPB-based Structural Equation Modeling estimates latent psychological constructs. Third, ordered logistic regression models graded safety perception outcomes across segments.

This integrated design captures structured heterogeneity, latent psychological pathways, and ordinal safety evaluation within a unified framework.

The empirical context is Wuhan, China, a major city with active AV pilot programs. While parameter estimates are context-specific, the underlying mechanism—cognitive heterogeneity reshaping psychological safety appraisal—holds broader analytical relevance for transitional mixed-traffic environments globally.

2. Methods

This section describes the methodological design of the study. It first presents the research design and data collection process, then introduces the measurement of key variables, followed by the validation of the measurement model, latent class segmentation, structural equation modeling, and the ordered logistic regression framework used to estimate drivers’ perceived safety in mixed-traffic AV environments.

2.1. Study Design and Data Collection

This study adopts a cross-sectional survey design to examine drivers’ perceived safety when sharing the road with autonomous vehicles (AVs) in mixed-traffic environments. Data were collected in Wuhan, China, between 17 June and 31 July 2024. Wuhan constitutes an appropriate empirical context due to its government-authorized AV pilot programs operating under real-world urban traffic conditions.

Eligibility criteria required participants to (i) hold a valid driving license issued in China, (ii) have actively driven within the previous 12 months, and (iii) report direct real-world exposure to autonomous vehicle operations while driving. Exposure was operationalized as at least one of the following: (a) encountering a Robotaxi (Society of Automotive Engineers (SAE) Level 4) during on-road interaction, (b) riding as a passenger in an autonomous taxi or shuttle service. Respondents who reported only indirect exposure (e.g., media consumption without traffic interaction) were excluded.

Data collection followed a mixed-mode approach combining online distribution (via driver associations and transportation-related digital platforms) with offline recruitment at vehicle registration offices, parking facilities, and AV pilot operation zones. Participation was voluntary and anonymous. Of 962 collected questionnaires, 905 met eligibility and completeness criteria and were retained for analysis. The final sample comprised urban Wuhan residents with prior AV exposure. Ethical approval was secured from the Institutional Review Board of Wuhan Polytechnic University (BME-2024-1-28). All analyses were conducted using StataMP 17. The complete questionnaire is provided in Appendix A, including eligibility screening, exposure verification, TPB measurement items, objective AV knowledge items, response scales, and socio-demographic coding categories. To improve transparency and reproducibility, Appendix A also serves as the direct reference for the operationalization of the variables used in Equations (1)–(5). The questionnaire explicitly operationalizes “Survey Instrument for Measuring Drivers’ Road-Sharing Safety Perception and Knowledge of Autonomous Vehicles” as interaction during real-world driving rather than hypothetical product adoption.

2.2. Measures

This subsection explains how the core variables used in the empirical analysis were operationalized. It includes the measurement of the dependent variable, the TPB-based psychological constructs, and the objective AV knowledge indicators used for latent segmentation. Together, these measures provide the foundation for the subsequent validation and modeling procedures.

2.2.1. Dependent Variable: Perceived Road-Sharing Safety

Perceived safety when sharing the road with AVs was measured using a five-point ordered scale (1 = very unsafe; 5 = very safe). Given its ordinal structure, cumulative logit specifications were employed in subsequent modeling.

2.2.2. Psychological Constructs

Psychological determinants were operationalized using four latent constructs derived from the Theory of Planned Behavior (TPB): Attitude (ATT), Subjective Norm (SN), Perceived Behavioral Control (PBC), and Behavioral Intention to Use (BIU) [11,18].

Although TPB traditionally models behavioral intention as the primary outcome, this study repositions perceived safety as the dependent variable and treats TPB constructs as parallel structured predictors within a safety-appraisal framework. This repositioning is consistent with research emphasizing that cognitive orientations shape risk evaluation processes in emerging technology contexts.

Each construct was measured using three reflective indicators rated on five-point Likert scales (1 = strongly disagree; 5 = strongly agree). Items were adapted from validated TPB instruments and contextualized to AV mixed-traffic coexistence scenarios. A pilot test (n = 30) ensured linguistic clarity and construct coherence.

Table 1. Measurement Items, Literature Source.

Latent Construct	Code	Measurement Item	Literature Source
Attitude (ATT)	ATT1	I have a favorable overall impression of driving in traffic environments where autonomous vehicles operate.	[24]
	ATT2	Driving alongside autonomous vehicles in regular traffic conditions is beneficial from my perspective.	[11]
	ATT3	Sharing road space with autonomous vehicles represents a desirable traffic environment.	[11]
Subjective Norm (SN)	SN1	People whose opinions matter to me think that interacting with autonomous vehicles in traffic is acceptable.	[11]
	SN2	My close social circle generally supports my willingness to drive in areas where autonomous vehicles operate.	[25]
	SN3	Observing others comfortably driving near autonomous vehicles influences my own readiness to do so.	[25]
Perceived Behavioral Control (PBC)	PBC1	I feel confident in my ability to manage driving situations involving autonomous vehicles.	[24]
	PBC2	Handling traffic scenarios that include autonomous vehicles would be manageable for me.	[26]
	PBC3	I believe I can effectively adapt my driving behavior when encountering autonomous vehicles.	[26]
Behavioral Intention (BIU)	BIU1	I am willing to continue driving in traffic environments where autonomous vehicles are present.	[20,25]
	BIU2	If given the choice, I would not avoid routes where autonomous vehicles operate.	[27]
	BIU3	I would encourage others to feel comfortable sharing the road with autonomous vehicles.	[25,27]

reports the original wording (English translation), literature source, and theoretical alignment of each TPB indicator to enhance transparency and replicability. All items were reformulated to explicitly refer to mixed-traffic interaction contexts rather than general technology adoption, thereby ensuring conceptual alignment between TPB predictors and the dependent variable of road-sharing safety perception.

2.2.3. AV Knowledge Indicators

To capture cognitive heterogeneity in AV-related understanding, five knowledge indicators were constructed covering technical, operational, and contextual domains: core AV components, automated navigation principles, global pilot deployment awareness, emergency response mechanisms, and urban application scenarios. For analytical purposes, the five knowledge items (see Appendix A, Questions 4.1–4.5) were coded as que1–que5 in the latent class modeling procedure.

The inclusion of objective knowledge measures is theoretically grounded in the literature demonstrating that domain-specific familiarity and cognitive sophistication systematically shape risk perception, trust calibration, and technology acceptance [12]. From a broader cognitive-processing perspective, knowledge functions as a structuring mechanism influencing how individuals interpret uncertainty [28]. Variations in technological literacy have also been shown to moderate attitudinal and intentional pathways in innovation adoption research [29].

Each indicator was coded dichotomously (0 = incorrect/no knowledge; 1 = correct knowledge). Binary operationalization was adopted for two reasons. First, it reduces ambiguity associated with self-assessed knowledge scales. Second, categorical indicators are well-suited for latent class modeling, which estimates subgroup membership based on discrete response patterns [30]. Model selection relied primarily on the Bayesian Information Criterion (BIC), consistent with simulation evidence supporting its performance in class enumeration [16].

By incorporating objective AV knowledge measures, the study advances beyond demographic segmentation and directly models cognitive heterogeneity as a structural moderator of psychological mechanisms. To establish the psychometric adequacy of the objective knowledge indicators, preliminary diagnostics were conducted prior to latent class modeling. Item-level discrimination was assessed using point-biserial correlations with the total knowledge score, all exceeding 0.30, indicating acceptable discriminatory power. Entropy statistics from the selected two-class solution exceeded 0.80, suggesting high classification precision and limited ambiguity in class assignment. These diagnostics strengthen the construct validity and external interpretability of knowledge-based segmentation beyond demographic stratification.

2.3. Measurement Model and Validity Assessment

Construct validity was evaluated using confirmatory factor analysis (CFA):

$$\mathbf{x}=\mathsf{\Lambda }\mathit{\eta }+\mathit{ϵ}$$

(1)

where $\mathbf{x}$ represents observed indicators, $\mathrm{\Lambda }$ factor loadings, $\mathit{\eta }$ latent constructs, and $\mathit{ϵ}$ measurement error. In substantive terms, the CFA model specifies how each observed questionnaire item reflects its underlying TPB construct.

Internal consistency was assessed using Cronbach’s α and Composite Reliability (CR). Convergent validity was evaluated via Average Variance Extracted (AVE ≥ 0.50). Discriminant validity followed the Fornell–Larcker criterion [31]. Model fit was assessed using Comparative Fit Index (CFI), Tucker–Lewis Index (TLI), Root Mean Square Error of Approximation (RMSEA), and Standardized Root Mean Square Residual (SRMR), following established guidelines [32].

2.4. Measurement Invariance Across Knowledge Segments

Because the study compares structural relationships across knowledge-based latent classes, the equivalence of the measurement model must be established prior to cross-group structural estimation. Without demonstrating measurement invariance, observed differences in structural parameters may reflect discrepancies in construct operationalization rather than substantive heterogeneity [33,34]. Accordingly, measurement invariance was examined using multi-group confirmatory factor analysis (CFA).

Following the hierarchical framework widely adopted in structural equation modeling research, increasingly restrictive invariance constraints were imposed sequentially. First, configural invariance was assessed by estimating the baseline model simultaneously across groups without equality constraints. This step evaluates whether the same factorial structure holds across segments, thereby establishing conceptual equivalence of latent constructs.

Second, metric invariance (weak invariance) was tested by constraining factor loadings to equality across groups. Metric invariance ensures that latent constructs are scaled equivalently, permitting meaningful comparison of structural path coefficients across segments [33,34]. Absent metric invariance, differences in regression paths may arise from differential item functioning rather than true psychological variation.

Third, scalar invariance (strong invariance) was examined by additionally constraining item intercepts. Scalar invariance is required for valid comparison of latent means, as it establishes equivalence in both scale units and origin points of measurement [32,35].

Given the sensitivity of chi-square difference tests to sample size, model comparisons were based on changes in approximate fit indices. Consistent with simulation-based recommendations, invariance was considered supported when ΔCFI ≤ 0.01 [36], supplemented by ΔRMSEA ≤ 0.015 as a robustness criterion [37]. These criteria provide a more stable basis for evaluating cross-group equivalence in moderate-to-large samples.

Establishing at least metric invariance allows structural coefficients to be compared across knowledge segments with interpretive validity. Where full scalar invariance is not achieved, partial invariance procedures may be applied to preserve theoretical comparability while maintaining statistical adequacy [35].

By explicitly testing measurement invariance prior to multi-group SEM estimation, the present study ensures that any observed differences in structural pathways across AV knowledge segments reflect substantive psychological heterogeneity rather than measurement non-equivalence. This step strengthens the internal validity of cross-segment inference and enhances the methodological rigor of the heterogeneity analysis.

2.5. Latent Class Analysis

This step was designed to identify unobserved heterogeneity in AV-related knowledge before estimating group-specific structural and ordinal-response models.

Latent Class Analysis (LCA) was used to identify unobserved knowledge-based subgroups. The likelihood for an individual $i$ is:

$$P(y_i)=\sum _{c=1}^C{\pi }_c\prod _{j=1}^{J}P(y_{ij}\mid C_i=c)$$

(2)

where $y_i$ denotes the full vector of observed knowledge responses for respondent $i$; $C$ is the total number of latent classes; ${\pi }_c$ is the prior probability that respondent $i$ belongs to latent class $c$; $J$ is the number of observed binary knowledge indicators; and $P(y_{ij}\mid c)$ is the conditional probability of respondent $i$’s response to item $j$, given membership in class $c$. In the present study, the five observed indicators correspond to the five objective AV knowledge items (que1–que5), each coded as 0 = incorrect/no knowledge and 1 = correct knowledge. Thus, Equation (2) models the probability that an individual response pattern is generated by a specific latent knowledge class, rather than assigning respondents deterministically to observed categories.

Model selection relied on the Bayesian Information Criterion (BIC):

$$BIC=-2\ln \left(L\right)+p\ln \left(n\right)$$

(3)

where $L$ is the maximized likelihood of the estimated latent class model, $p$ is the number of free parameters to be estimated, and $N$ is the sample size. The BIC penalizes model complexity while rewarding goodness of fit; therefore, a lower BIC value indicates a more parsimonious and empirically preferable class solution. In this study, Equation (3) was used to compare models with one to five latent classes, and the two-class solution was retained because it yielded the minimum BIC value.

Following recommended practice, a three-step approach was adopted to reduce bias from classification uncertainty when incorporating class membership into subsequent models [30]. Class membership was incorporated into subsequent models using Vermunt’s improved three-step procedure, which corrects for classification uncertainty by adjusting for misclassification probabilities [38]. This approach avoids bias associated with naive modal assignment while preserving interpretability of class-specific parameter estimates.

2.6. Structural Equation Modeling

After identifying latent knowledge segments, SEM was used to examine whether the internal TPB structure remained stable across cognitively distinct driver groups.

Structural relationships among TPB constructs were estimated using SEM:

$$BI{U}_i=\mathsf{\alpha }+{\beta }_{1}AT{T}_i+{\beta }_{2}S{N}_i+{\beta }_{3}PB{C}_i+{\zeta }_i$$

(4)

where $BI{U}_i$ denotes the latent behavioral intention score for respondent $i$; $AT{T}_i$, $S{N}_i$, and $PB{C}_i$ denote the corresponding latent scores for attitude, subjective norm, and perceived behavioral control, respectively; $\alpha$ is the intercept term; ${\beta }_1$, ${\beta }_2$, and ${\beta }_3$ are structural coefficients representing the partial associations of attitude, subjective norm, and perceived behavioral control with behavioral intention; and ${\zeta }_i$ is the residual disturbance term capturing unexplained variance in behavioral intention. In substantive terms, Equation (4) tests whether the three TPB antecedents remain statistically associated with behavioral intention after accounting for their simultaneous effects. Because the SEM is estimated separately by knowledge segment, these coefficients are allowed to vary across the low- and high-knowledge groups, thereby enabling a direct assessment of structural heterogeneity.

Multi-group SEM was conducted to test whether structural coefficients differed across knowledge segments. Nested model comparisons were performed by constraining path coefficients equal across groups and evaluating model fit deterioration. It should be noted that, given the cross-sectional design, structural paths are interpreted as conditional associations rather than causal effects. The theoretical directionality follows TPB logic and risk-appraisal theory; however, reverse causation cannot be entirely ruled out. To mitigate endogeneity concerns, psychological constructs were estimated prior to inclusion in ordered logit models, and multicollinearity diagnostics (VIF < 3) indicated no critical inflation of standard errors.

2.7. Ordered Logistic Regression Model

Finally, ordered logistic regression was used to link psychological constructs and socio-demographic characteristics to the ordinal outcome of perceived road-sharing safety.

Perceived road-sharing safety was modeled using a cumulative logit (proportional odds) specification, a widely established approach for ordinal outcome variables with ordered response categories [39,40].

Let $Y_i$ denote the perceived safety level reported by driver $i$, measured on a five-point scale (1 = very unsafe; 5 = very safe). The ordered logit model estimates the cumulative probability that perceived safety falls at or below category $j$:

$$\log \left({\displaystyle \frac{P\left(Y_i\le j\right)}{P\left(Y_i>j\right)}}\right)={\tau }_j-X_i^{\prime }\beta$$

(5)

where $P(Y_i\le j)$ is the cumulative probability that respondent $i$ reports a safety level at or below category $j$, and $P(Y_i>j)$ is the complementary probability of reporting a category above $j$. The term ${\tau }_j$ denotes the estimated threshold (cut-point) separating adjacent ordered safety categories, with one threshold for each cumulative split of the five-point outcome. $X_i$ is the vector of explanatory variables for respondent $i$, including psychological constructs and socio-demographic characteristics, and $\beta$ is the corresponding vector of regression coefficients. A positive coefficient indicates that higher values of the predictor are associated with greater odds of reporting a higher safety category, whereas a negative coefficient indicates lower perceived safety.

In the present study, $X_i$ includes the TPB-based latent constructs (ATT, SN, PBC, and BIU) together with demographic and driving-experience covariates. The cumulative logit specification is appropriate because the dependent variable is ordinal rather than continuous, and the proportional-odds assumption implies that the effect of each predictor is constant across all cumulative splits of the outcome categories. Separate models were estimated for the full sample and for each latent knowledge segment in order to examine whether the estimated coefficient vector $\beta$ varies systematically across cognitive strata.

All explanatory variables were treated as fixed effects. Accident history was included as a categorical predictor to account for experiential heterogeneity in safety appraisal. Because each respondent contributed a single safety evaluation and no hierarchical nesting structure (e.g., repeated measures or clustered sampling) was present, a multilevel specification was not required.

The proportional odds assumption—implying that predictor effects are constant across cumulative splits of the outcome—was evaluated using the Brant test. Test results did not indicate systematic violation of the proportional odds assumption. As a robustness check, alternative model specifications including dummy-variable coding of accident history and reduced predictor sets were estimated, yielding substantively consistent results.

Separate ordered logit models were estimated for the full sample and for knowledge-based latent segments identified through LCA, allowing assessment of parameter heterogeneity across cognitive strata.

3. Analysis

This section presents the empirical analysis in a stepwise sequence designed to align with the study’s theoretical and methodological framework. We first report descriptive statistics to summarize the socio-demographic composition of the sample and the baseline distribution of key variables. We then apply latent class analysis (LCA) to identify unobserved heterogeneity in drivers’ objective AV knowledge and to derive knowledge-based segments for subsequent subgroup analysis. Next, structural equation modeling (SEM) is used to examine whether the internal TPB structure differs across these latent knowledge groups. Finally, ordered logit models are estimated for the full sample and for each latent segment to assess how psychological constructs and socio-demographic characteristics jointly shape perceived road-sharing safety. This workflow enables a systematic progression from sample profiling to cognitive segmentation, to structural validation, and ultimately to segment-specific safety-perception modeling.

3.1. Descriptive Statistics

As the first step of the analytical workflow, this subsection describes the socio-demographic characteristics of the sample and the baseline distribution of key study variables prior to latent segmentation and structural modeling.

The final sample consisted of 905 licensed drivers residing in Wuhan, China, all of whom reported real-world exposure to autonomous vehicles (AVs) while driving. As shown in

Table 2. Socio-economic Characteristics of Survey Respondents.

Variable	Category	Frequency	Percentage
Gender	Male	425	47.0%
	Female	480	53.0%
Age	≤29	314	34.7%
	30–39	290	32.0%
	40–49	174	19.2%
	50–59	71	7.8%
	≥60	56	6.2%
Education	Middle school and below	153	16.9%
	College (Associate degree)	454	50.2%
	Bachelor’s degree	219	24.2%
	Master’s degree or above	79	8.7%
Household Income (CNY/month)	≤3000	125	13.8%
	3001–5000	409	45.2%
	5001–10,000	224	24.8%
	10,001–20,000	89	9.8%
	≥20,001	58	6.4%
Employment Status	Full-time	780	86.2%
	Part-time	28	3.1%
	Student	30	3.3%
	Retired or unemployed	67	7.4%
Household Size	1	110	12.2%
	2	316	34.9%
	3	290	32.0%
	4	123	13.6%
	5 or more	66	7.3%
Car Ownership	Yes	800	88.4%
	No	105	11.6%
IC Card Ownership	Yes	784	86.6%
	No	121	13.4%
Primary Mode of Transport	Walk	139	15.4%
	Bicycle (electric bike and motorbike)	240	26.5%
	Car (taxi)	148	16.4%
	Public transit (bus, train)	378	41.8%
Accident History (Car)	0 accidents	671	74.1%
	1–2 accidents	127	14.0%
	3–4 accidents	69	7.6%
	5 or more accidents	38	4.2%

, the sample exhibits a relatively young age structure, with approximately 66.7% of respondents under 40 years of age. Female drivers constitute a slight majority (53.0%). Educational attainment is comparatively high: over 83% possess at least a college degree, reflecting the urban and relatively educated composition of the sample.

These socio-demographic characteristics are analytically relevant because they help contextualize potential variation in AV safety perception. A younger sample profile may imply greater familiarity with digital technologies and lower baseline resistance toward automation, whereas older drivers may rely more heavily on accumulated driving experience and risk-avoidance heuristics when evaluating AV interactions. Similarly, higher educational attainment may be associated with stronger information-processing capacity and greater exposure to technological discourse, both of which can influence how AV behavior is interpreted under uncertainty. Thus, the demographic composition of the sample is not merely descriptive; it provides an important context for understanding why perceived safety may vary systematically across drivers.

Car ownership is prevalent (88.4%), although public transit remains the most frequently reported primary travel mode (41.8%), suggesting multimodal mobility patterns. Regarding traffic safety experience, 74.1% reported no accident involvement within the past year, while approximately 25.9% experienced at least one accident, forming a meaningful subgroup with direct crash exposure.

This mobility profile is also substantively important. Vehicle ownership and primary travel mode may shape both the frequency and context of AV encounters, as well as the degree to which respondents evaluate AVs from the perspective of active vehicle control versus broader urban mobility participation. Likewise, prior accident experience is theoretically meaningful because it may heighten risk salience, alter trust calibration, and shape comparative evaluations of human-driven and automated driving behavior. In this sense, accident history is not merely a background control, but an experiential factor that may condition perceived road-sharing safety.

Taken together, these characteristics indicate that the sample represents an urban, AV-exposed driving population with moderate demographic diversity and substantial mobility resources. More importantly, the observed variation in age, education, mobility patterns, and accident experience provides a theoretically meaningful basis for subsequent heterogeneity analysis. These factors may influence drivers’ technological familiarity, perceived control, trust formation, and risk sensitivity, and therefore help explain why safety perception toward AV road-sharing is unlikely to be uniform across the sample.

This descriptive profile also supports the need for latent segmentation, as observed socio-demographic diversity may coexist with deeper cognitive heterogeneity in AV-related knowledge and safety appraisal.

3.2. Latent Class Model Analysis

As the second step of the analysis, Latent Class Analysis (LCA) was employed to uncover unobserved heterogeneity in drivers’ objective knowledge of autonomous vehicle (AV) technology and to establish cognitively distinct subgroups for subsequent modeling. LCA is a person-centered statistical method widely used to categorize individuals into mutually exclusive subgroups based on observed response patterns. It is especially effective when no prior assumptions exist about the number of subpopulations, and it enables probabilistic class assignments with minimal measurement error [41].

In the present study, LCA is theoretically appropriate because AV-related knowledge is unlikely to be a simple linear trait that can be fully represented by a summed score. Instead, knowledge about AVs is multidimensional, encompassing technical principles, operational logic, emergency response, and application awareness. Drivers may therefore exhibit qualitatively different response profiles even when their total knowledge scores appear similar. A person-centered latent class approach is better suited to capturing such patterned cognitive heterogeneity than a variable-centered additive index. In this sense, LCA allows the analysis to identify substantively meaningful knowledge configurations rather than imposing a priori continuity on heterogeneous knowledge structures.

Models with one to five latent classes were estimated in STATA 17.0. Model selection was guided by the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC). As shown in

Table 3. Model Fit Indices for Latent Class Solutions.

Number of Classes	N	Log-Likelihood	df	AIC	BIC
1	905	−3028.33	5	6066.650	6090.690
2	905	−2036.05	11	4094.108	4146.995
3	905	−2033.88	16	4099.757	4176.684
4	905	−2029.59	20	4099.174	4195.333
5	905	−2025.11	27	4104.218	4234.033

Note: Lower AIC and BIC values indicate better model fit relative to complexity. The two-class solution was retained because it achieved the minimum BIC and yielded substantively interpretable class profiles.

, the two-class model yielded the lowest AIC (4094.108) and BIC (4146.995), suggesting the best balance between fit and parsimony [16].

Maximum likelihood estimation identified two distinct knowledge profiles among the 905 respondents: Class 1 (n = 440, 48.62%) and Class 2 (n = 465, 51.38%). Conditional response probabilities for each class are reported in

Table 4. Conditional Response Probabilities and Class Prevalence.

Indicator	Response	Class 1	Class 2
Basic Knowledge (que1)	No	95.45%	0.86%
	Yes	4.55%	99.14%
Operational Principle (que2)	No	69.77%	76.99%
	Yes	30.23%	23.01%
Global AV Testing (que3)	No	99.32%	2.37%
	Yes	0.68%	97.63%
Emergency Response (que4)	No	49.09%	58.06%
	Yes	50.91%	41.94%
Application Scenarios (que5)	No	99.32%	3.87%
	Yes	0.68%	96.13%

Note: Class labels are assigned based on the overall conditional-response profile across the five AV knowledge indicators rather than on any single item.

.

Class 1 was labeled the Low AV Knowledge Group because respondents in this class showed consistently low probabilities of answering correctly on several core items, especially basic AV knowledge (que1), global AV testing awareness (que3), and application scenarios (que5). Although this group was not uniformly uninformed, its members still showed some correct responses on selected items, such as emergency response, and its overall response pattern indicates fragmented and limited understanding rather than coherent technological knowledge.

Class 2 was labeled the High AV Knowledge Group because respondents in this class displayed very high probabilities of correct responses on the foundational and contextual items that are most indicative of broad AV familiarity, including basic knowledge, global testing awareness, and urban application scenarios. While not all items reached uniformly high correct-response probabilities, the overall pattern reflects a substantially more informed and internally consistent knowledge structure. The distinction between the two classes is therefore not based on a single item, but on the full conditional-response profile across the five knowledge indicators.

In practical terms, the two latent classes represent cognitively distinct driver profiles rather than merely statistical partitions. The Low AV Knowledge Group can be interpreted as drivers who have encountered AVs in real traffic but possess only a partial or fragmented understanding of how these systems operate. Such drivers may recognize AVs as a visible mobility phenomenon without having a stable cognitive model of their technical logic or operating constraints. By contrast, the High AV Knowledge Group represents drivers with a broader and more internally coherent understanding of AV technology, including its functional principles, deployment context, and application scope. In real-world mixed traffic, these two groups are likely to differ not only in what they know but also in how they interpret uncertainty, evaluate system predictability, and form safety judgments during AV interaction.

3.3. Structural Equation Modeling Analysis

Following knowledge-based segmentation, structural equation modeling (SEM) was conducted as the third analytical step to examine whether the internal TPB structure differed across the identified latent classes. SEMs were estimated separately for three groups: the full sample (Model 1), Low AV Knowledge Group (Model 2), and High AV Knowledge Group (Model 3). This stratified approach enables the detection of psychological differences across subpopulations.

According to TPB, behavioral intention is shaped by three latent constructs: attitude (ATT), subjective norm (SN), and perceived behavioral control (PBC). These constructs have been validated in the AV adoption literature [16]. Convergent validity was assessed using standardized factor loadings, Cronbach’s alpha, Composite Reliability (CR), and Average Variance Extracted (AVE). Acceptable thresholds include factor loadings ≥ 0.70, Cronbach’s alpha and CR ≥ 0.70, and AVE ≥ 0.50 [42].

Convergent validity results are presented in

Table 5. Convergent Validity of the Latent Constructs.

Latent Variables	Code	Means	SD	Standardized Factor Loading	Cronbach’s	CR	AVE
All sample
ATT	ATT1	3.811	1.078	0.714	0.791	0.755	0.507
	ATT2	3.482	1.100	0.688
	ATT3	3.648	1.082	0.733
SN	SN1	3.991	1.035	0.730	0.789	0.753	0.504
	SN2	3.669	1.101	0.693
	SN3	3.862	1.054	0.707
PBC	PBC1	4.146	1.019	0.761	0.839	0.811	0.588
	PBC2	3.848	1.103	0.754
	PBC3	3.983	1.089	0.786
BIU	BIU1	3.761	1.134	0.782	0.843	0.816	0.596
	BIU2	3.510	1.180	0.743
	BIU3	3.636	1.152	0.791
The low AV knowledge group
ATT	ATT1	3.841	1.075	0.720	0.633	0.754	0.506
	ATT2	3.482	1.076	0.679
	ATT3	3.686	1.049	0.734
SN	SN1	4.005	1.017	0.734	0.602	0.752	0.503
	SN2	3.632	1.076	0.709
	SN3	3.884	1.034	0.683
PBC	PBC1	4.159	0.991	0.753	0.704	0.800	0.572
	PBC2	3.784	1.099	0.744
	PBC3	3.957	1.109	0.772
BIU	BIU1	3.780	1.100	0.808	0.868	0.831	0.621
	BIU2	3.505	1.151	0.756
	BIU3	3.627	1.178	0.800
The high AV knowledge group
ATT	ATT1	3.783	1.082	0.709	0.684	0.756	0.508
	ATT2	3.482	1.124	0.697
	ATT3	3.611	1.113	0.731
SN	SN1	3.978	1.052	0.727	0.655	0.755	0.507
	SN2	3.703	1.125	0.681
	SN3	3.841	1.073	0.728
PBC	PBC1	4.133	1.046	0.771	0.750	0.822	0.607
	PBC2	3.908	1.104	0.765
	PBC3	4.009	1.071	0.800
BIU	BIU1	3.744	1.166	0.760	0.845	0.802	0.575
	BIU2	3.516	1.207	0.731
	BIU3	3.645	1.128	0.783

Note: Standardized factor loadings above 0.70, CR above 0.70, and AVE above 0.50 indicate acceptable convergent validity.

. In this study, most factor loadings exceeded 0.70, indicating acceptable item-level. Cronbach’s alpha values ranged from 0.602 (SN, low-knowledge group) to 0.868 (BIU), suggesting good internal consistency overall. Similarly, CR values ranged from 0.752 to 0.831 across constructs, while AVE values were consistently above 0.50, confirming satisfactory convergent validity [31]. It should be noted that Cronbach’s α for the Subjective Norm (SN) construct in the low-knowledge segment was 0.602, slightly below the conventional 0.70 threshold. However, Cronbach’s α is known to underestimate reliability when the number of indicators is small and when factor loadings are heterogeneous [16]. Composite Reliability (CR = 0.752) exceeded the recommended 0.70 criterion, indicating acceptable internal consistency at the construct level.

Further inspection of the item–total correlations suggested that the comparatively lower alpha may reflect contextual semantic variation rather than construct instability. Given that AV exposure and social discourse differ across knowledge strata, subjective norm items may be interpreted more variably among less informed drivers. No item deletion materially improved reliability, and therefore, the construct was retained to preserve theoretical completeness.

Discriminant validity was confirmed using the Fornell–Larcker criterion. The square root of each construct’s AVE exceeded its correlations with other constructs across all models (see

Table 6. Discriminant Validity (Fornell-Larcker Criterion).

Construct	ATT	SN	PBC	BIU
All sample
ATT	0.712
SN	0.398	0.710
PBC	0.430	0.456	0.767
BIU	0.332	0.410	0.392	0.772
The low AV knowledge group
ATT	0.711
SN	0.466	0.709
PBC	0.409	0.451	0.756
BIU	0.362	0.466	0.415	0.788
The high AV knowledge group
ATT	0.713
SN	0.343	0.712
PBC	0.450	0.463	0.779
BIU	0.304	0.353	0.371	0.758

), indicating good discriminant separation [32].

Discriminant validity was assessed using the Fornell-Larcker criterion, wherein the square root of the AVE for each latent construct is compared to its inter-correlations with other constructs (Table 6). Discriminant validity is deemed adequate when the square root of AVE for each construct surpasses its correlations with all other latent constructs. In this study, diagonal values, representing the square root of AVE, were consistently higher than their corresponding off-diagonal correlation coefficients. For example, in the full sample group, the square root of AVE for ATT was 0.712, exceeding its correlations with SN (0.398), PBC (0.430), and BIU (0.332). Similar patterns were consistently observed across both low and high AV knowledge groups, thus providing clear evidence of robust discriminant validity among the constructs [32].

The SEMs were estimated using maximum likelihood estimation, and model fit was assessed using four widely accepted goodness-of-fit indices: the Comparative Fit Index (CFI), Tucker–Lewis Index (TLI), Root Mean Square Error of Approximation (RMSEA), and Standardized Root Mean Square Residual (SRMR). The recommended thresholds for these indices are: CFI and TLI values ≥ 0.90, RMSEA ≤ 0.08, and SRMR < 0.08, as established in SEM methodology literature [43].

As shown in

Table 7. Summary of Goodness-of-Fit Indices for Theory of Planned Behavior (TPB)-Based Structural Equation Modeling (SEM) Models.

Model	CFI	TLI	RMSEA	SRMR
Threshold	≥0.90	≥0.90	≤0.08	<0.08
Model 1: All samples	0.996	0.994	0.021	0.021
Model 2: Low AV Knowledge Group	0.993	0.990	0.027	0.028
Model 3: High AV Knowledge Group	0.995	0.993	0.023	0.028

, all three models demonstrate excellent fit to the data. Model 1 (full sample) achieved a CFI of 0.996, TLI of 0.994, RMSEA of 0.021, and SRMR of 0.021. Model 2 (low knowledge group) and Model 3 (high knowledge group) similarly exceeded the recommended thresholds, indicating strong internal validity and model adequacy across subsamples.

These results validate the structural integrity of the TPB-based framework across heterogeneous driver segments and justify the use of these latent constructs in subsequent ordinal logistic regression models to assess safety perception when sharing roads with AVs. Moreover, the subgroup-specific models help uncover potential psychological asymmetries between digitally literate and digitally disadvantaged drivers—differences that could be critical for designing targeted educational and communication strategies [44].

4. Results

Building on the descriptive, segmentation, and structural analyses reported in Section 3, this section presents the ordered logit results for the full sample and the two knowledge-based driver segments.

Table 8. Results of Ordered Logit Models Across Driver Segments.

Variables	All Sample		Low AV Knowledge Group		High AV Knowledge Group
	Coefficient	z-Value	Coefficient	z-Value	Coefficient	z-Value
ATT	−0.228	−2.28	−0.124	−0.83	−0.304	−2.16
SN	0.118	1.1	0.055	0.35	0.170	1.17
PBC	−0.136	−1.29	−0.152	−0.96	−0.144	−0.99
BIU	0.171	2.00	0.218	1.63	0.134	1.17
gender	−0.016	−0.26	−0.066	−0.72	0.037	0.43
29 years and below	−0.118	−0.63	0.090	0.23	−0.186	−0.78
30–39 years	0.066	0.36	0.338	0.89	−0.146	−0.61
40–49 years	−0.021	−0.11	0.169	0.43	−0.113	−0.44
50–59 years	0.486	2.45	0.510	1.31	0.493	1.95
3000 CNY and below	−0.158	−0.76	−0.050	−0.15	−0.223	−0.8
3001–5000 CNY	−0.651	−4	−0.359	−1.39	−0.903	−4.13
5001–10,000 CNY	−0.237	−1.58	−0.171	−0.78	−0.286	−1.37
10,001–20,000 CNY	0.599	2.84	1.072	3.33	0.393	1.37
High School or lower	0.338	1.86	0.604	2.02	0.198	0.84
Associate degree	0.168	1.12	−0.249	−1.11	0.520	2.45
Bachelor’s degree	−0.100	−0.6	−0.562	−2.32	0.242	1.05
Full-time	−1.031	−4.89	−1.133	−2.26	−0.906	−3.55
Part-time	0.071	0.21	−0.291	−0.35	0.152	0.38
Student	0.729	1.97	0.965	1.48	0.632	1.27
Car ownership	−0.075	−0.69	0.073	0.44	−0.196	−1.33
Car accident	0.360	3.88	0.603	4.14	0.249	2.03
N	905		440		465
Pseudo R2	0.0561		0.0709		0.0615

presents the results of the ordered logit models for the full sample, Low AV Knowledge Group, and High AV Knowledge Group. Each column displays estimated coefficients and corresponding z-values for psychological and demographic predictors across segments.

Table 8 presents the results of the ordered logit models for the full sample, Low AV Knowledge Group, and High AV Knowledge Group.

Among psychological variables, behavioral intention (BIU) consistently exhibited a positive effect on perceived safety across all groups, though only statistically significant in the full sample (β = 0.171, z = 2.00). In contrast, attitude (ATT) showed a statistically significant negative association with safety perception in both the full sample (β = −0.228, z = −2.28) and the High Knowledge Group (β = −0.304, z = −2.16), suggesting possible dissonance between general AV attitudes and perceived safety during real-world interactions. Subjective norm (SN) and perceived behavioral control (PBC) did not reach conventional significance thresholds (p < 0.05) in any group. This may imply that these constructs play a lesser role in shaping safety perceptions compared to intention and attitude in the AV context.

Regarding socio-demographic factors, age, income, and education revealed heterogeneous effects across groups. For instance, respondents aged 50–59 showed significantly higher perceived safety (β = 0.486, z = 2.45 for the full sample), whereas middle-income drivers (3001–5000 CNY) in the High Knowledge Group reported lower safety perceptions (β = −0.903, z = −4.13). Education effects were also group-specific: in the High Knowledge Group, holding an associate degree significantly increased perceived safety (β = 0.520, z = 2.45), while a bachelor’s degree showed no significant effect.

Employment status emerged as a key differentiator. Full-time employment was a strong negative predictor across all groups, with the largest effect observed in the Low Knowledge Group (β = −1.133, z = −2.26). Conversely, student status was positively associated with perceived safety, particularly in the full sample (β = 0.729, z = 1.97).

Finally, traffic accident experience consistently predicted higher safety perceptions, with the strongest effect observed in the Low Knowledge Group (β = 0.603, z = 4.14). This finding supports the hypothesis that prior crash experience may elevate safety awareness and increase trust in AV technologies.

The proportional odds assumption underlying the cumulative logit specification was formally evaluated using the Brant test [45]. The global test statistics for the full-sample and subgroup models were not statistically significant (p > 0.05), indicating no systematic violation of the parallel-lines assumption. These results support the appropriateness of the proportional odds framework for modeling ordered safety perception in this context.

For substantive interpretation, estimated coefficients were exponentiated and reported as odds ratios (ORs), facilitating direct interpretation in terms of cumulative odds. In the full-sample model, a one-unit increase in Behavioral Intention (BIU) was associated with an 18.6% increase in the cumulative odds of reporting a higher safety category (OR = 1.186), holding all other variables constant.

To further enhance interpretability beyond log-odds metrics, average marginal effects (AMEs) were computed. Among high-knowledge drivers, the negative coefficient of Attitude (ATT) corresponded to an approximate 6.2 percentage-point decrease in the predicted probability of reporting the highest safety category (“very safe”), evaluated at sample means of covariates. This effect illustrates that the attitudinal reversal observed in the regression coefficients translates into substantively meaningful differences in predicted safety perception.

Robustness checks using alternative codings of accident history and reduced predictor specifications yielded consistent marginal patterns, confirming the stability of the reported associations.

5. Discussion

This study set out to examine whether drivers’ objective knowledge of autonomous vehicles reshapes the psychological mechanisms underlying perceived road-sharing safety in mixed-traffic environments. The results provide a clear affirmative answer. First, knowledge-based latent segmentation identified two cognitively distinct driver groups, indicating that AV-related understanding is not uniformly distributed even among drivers with real-world AV exposure. Second, the ordered logit results show that the psychological determinants of perceived safety are not structurally invariant across these groups. Most notably, attitude toward AVs exhibited a significant negative association with perceived safety among high-knowledge drivers, whereas behavioral intention retained a positive role across models. Taken together, these findings suggest that knowledge heterogeneity does not merely shift average safety perceptions; it reconfigures the evaluative architecture through which AV interaction is interpreted.

Importantly, one of the central contributions of this study lies in demonstrating that objective AV knowledge is not simply a background covariate, but a meaningful basis for segmenting drivers into cognitively differentiated groups with distinct patterns of safety appraisal.

5.1. Interpreting the Attitude–Safety Divergence Among High-Knowledge Drivers

One of the most theoretically intriguing findings of this study is the negative association between attitude and perceived safety among high-knowledge drivers. This pattern departs from conventional TPB expectations [11] and from much of the AV acceptance literature, where positive attitudes are typically associated with favorable adoption outcomes [25]. This finding is particularly important because the positive direction of H1 might appear theoretically self-evident under a conventional TPB interpretation. However, the empirical results show that once objective knowledge heterogeneity is incorporated, attitude does not uniformly translate into greater perceived safety. Instead, the sign reversal observed among high-knowledge drivers demonstrates that even seemingly intuitive TPB-based expectations require empirical verification under conditions of uneven cognitive familiarity.

Importantly, the negative marginal effect of attitude among high-knowledge drivers should not be interpreted as a theoretical inconsistency but as evidence of cognitive recalibration. In several cross-national AV studies, greater prior knowledge or exposure has not uniformly increased adoption willingness; in some cases, it has produced more cautious or even skeptical evaluations [12]. This pattern aligns with expectation–disconfirmation theory, whereby higher baseline expectations amplify sensitivity to perceived performance gaps.

From a risk-calibration perspective, knowledge does not necessarily reduce perceived uncertainty; rather, it may refine individuals’ understanding of operational design domain constraints, edge-case failures, and system limitations. As knowledge increases, evaluative criteria become more stringent, potentially weakening the positive association between general technological attitudes and situational safety appraisal.

Thus, the observed negative attitude–safety linkage reflects a knowledge-conditioned evaluative shift rather than attitudinal contradiction. It suggests that cognitive sophistication transforms the meaning of “positive attitude” from diffuse technological optimism to critical performance-based assessment.

Rather than interpreting this divergence as a statistical anomaly, we argue that it reflects a structural shift in evaluative processing under conditions of cognitive familiarity.

Drivers with greater AV knowledge are more likely to understand technological limitations, boundary conditions of operational design domains, and unresolved ethical or algorithmic dilemmas [8]. Increased knowledge may therefore recalibrate risk thresholds rather than amplify optimism. In such cases, positive attitudes toward technological innovation coexist with heightened awareness of system fragility.

This interpretation aligns with risk-perception theory, which distinguishes between abstract endorsement and embodied safety appraisal [45]. Knowledge does not uniformly reduce perceived risk; in certain domains, it increases sensitivity to low-probability but high-consequence events.

A second mechanism may involve expectation asymmetry. High-knowledge drivers may hold more demanding performance standards for AV systems. When real-world pilot deployment falls short of anticipated technological maturity, safety appraisal may decline despite generally favorable attitudes toward innovation.

Taken together, the negative partial effect of attitude suggests that cognitive familiarity transforms the evaluative frame through which AV interaction is judged. Safety perception thus becomes more stringent—not less—among informed drivers. This finding contributes to a more nuanced understanding of technological acceptance under uncertainty.

5.2. Beyond Methodological Innovation: Knowledge as a Structural Axis of Inequality

While the methodological integration of LCA, SEM, and ordered logistic regression enhances analytical rigor, the substantive innovation lies in theorizing knowledge as a structural axis shaping AV safety perception.

The contribution of the knowledge-based segmentation should therefore not be understood as a purely methodological refinement. Its significance is substantive and theoretical. By distinguishing between low- and high-knowledge drivers, the study shows that heterogeneous AV familiarity is associated with different interpretive logics of safety evaluation. In other words, segmentation reveals that drivers do not simply differ in how much they know, but in how that knowledge structures trust calibration, uncertainty interpretation, and evaluative judgment in mixed traffic.

Most prior AV studies treat knowledge either as a control variable or as part of demographic segmentation [12]. However, knowledge of digitally mediated mobility systems is not evenly distributed. It intersects with educational access, digital literacy, occupational exposure, and socioeconomic resources—elements often described within the broader framework of the digital divide.

By demonstrating that TPB pathways differ across knowledge segments, this study shows that psychological determinants of AV safety perception are not socially neutral. Instead, they are conditioned by access to technological literacy.

This reframes AV acceptance from an individual attitudinal phenomenon to a stratified cognitive process embedded within broader social structures. The findings therefore extend TPB beyond individual-level psychology and situate it within a context of informational inequality.

Accordingly, we advance a refined theoretical proposition: In transitional automation environments, cognitive familiarity operates not merely as an individual attribute but as a structurally distributed resource that conditions the evaluative architecture of safety perception.

This perspective bridges behavioral theory and digital inequality scholarship, expanding the originality of the study beyond methodological triangulation.

5.3. Toward a Conditional Theory of Automation Acceptance

The above findings are theoretically consequential precisely because they emerge from the knowledge-based segmentation of drivers. Without distinguishing between latent knowledge groups, the reversal and instability of TPB-related associations would remain obscured in aggregate models. The cumulative findings support a conditional extension of TPB.

First, perceived safety should be conceptualized as an interaction-level appraisal rather than a downstream consequence of intention.

Second, parameter homogeneity cannot be assumed under uneven knowledge diffusion.

Third, cognitive segmentation reveals that acceptance processes are stratified, with different evaluative logics operating across informational tiers.

This conditional theory of automation acceptance contributes to a more precise understanding of how individuals process technological uncertainty and how cognitive resources mediate that process.

6. Conclusions and Implications

This section summarizes the core findings of the study and discusses their broader implications. It first highlights the main contribution of the research, then elaborates on the theoretical and practical implications, and finally outlines the study’s limitations and promising directions for future research.

6.1. Main Contribution

This study contributes to high-level AV behavioral modeling by advancing a conditional framework of safety appraisal under informational heterogeneity. Rather than assuming psychological parameter invariance, we demonstrate that objective knowledge can alter both the magnitude and direction of attitudinal effects. This finding challenges homogeneous TPB applications and underscores the necessity of incorporating cognitive stratification into behavioral modeling of emerging technologies. The study therefore contributes not by asserting entirely new directional hypotheses but by demonstrating that conventionally expected TPB relationships cannot be assumed to remain stable once cognitive heterogeneity is explicitly modeled. In particular, the reversal of H1 in the high-knowledge group shows that knowledge segmentation changes not only the magnitude but also the direction of psychological associations.

The central contribution is theoretical: knowledge functions as a structural moderator of psychological pathways, transforming the direction and strength of attitudinal effects. This insight challenges the assumption of parameter invariance commonly embedded in behavioral adoption models.

By situating cognitive familiarity within broader socio-structural contexts, the study expands AV acceptance theory toward a stratified and inequality-sensitive framework.

6.2. Theoretical Implications

The findings imply that behavioral theories applied to emerging technologies must incorporate informational asymmetry as a core structural dimension.

Rather than treating knowledge as a background variable, future research should examine it as a conditioning mechanism that reshapes evaluative hierarchies.

Moreover, the divergence between generalized attitudes and situational safety perception suggests that acceptance models must distinguish between symbolic endorsement and embodied interactional judgment.

These theoretical refinements extend beyond AV research and are relevant for broader AI-enabled systems undergoing public integration.

6.3. Practical Implications: From Insight to Action

The findings generate several actionable implications.

For AV developers, communication strategies should differentiate between informational segments. High-knowledge users require transparency regarding system limitations, failure modes, and operational design boundaries. Providing detailed technical briefings and scenario-based disclosures may mitigate expectation-driven skepticism.

For policymakers, AV deployment should be accompanied by structured public education initiatives. For low-knowledge drivers, simplified explanatory materials—such as interactive simulations, community demonstration programs, and visual scenario-based education modules—may reduce interpretive uncertainty and build baseline familiarity.

For educators and public institutions, integrating AV literacy into driver training curricula could address structural informational inequality before widespread deployment.

Crucially, public trust-building should not rely solely on aggregate safety statistics. It must address the cognitive frames through which drivers interpret automation risk.

6.4. Limitations and Future Directions

The study is limited by its cross-sectional design and single-city context. Longitudinal studies could examine how knowledge acquisition dynamically alters evaluative thresholds.

Future research may incorporate direct measures of digital literacy, media exposure, and risk propensity to disentangle cognitive familiarity from dispositional traits.

Comparative cross-national studies would clarify whether knowledge-conditioned heterogeneity persists under different institutional trust environments.

A further limitation concerns the range of driving-related operational variables included in the present study. Although accident history was incorporated as a broad indicator of prior traffic-safety experience, the survey did not collect more fine-grained operational measures such as driving frequency, annual mileage, congestion exposure, proportion of night driving, or recent near-miss experience. As a result, the current models cannot fully distinguish how routine driving exposure and interaction intensity in complex traffic environments may shape perceived AV road-sharing safety. Future research should incorporate these operational variables in order to provide a more behaviorally grounded account of how real-world driving experience conditions safety appraisal in mixed-traffic AV environments.