A Multi-Dimensional Psychological Model of Driver Takeover Safety in Automated Vehicles: Insights from User Experience and Behavioral Moderators

Abstract

With the rapid adoption of automated driving systems, ensuring safe and efficient driver takeover has become a crucial challenge for road safety. This study introduces a novel psychological framework for understanding and predicting takeover behavior in conditionally automated vehicles, leveraging an extended Theory of Planned Behavior (TPB) model enriched by real-world driver experience. Drawing on survey data from 385 automated driving system users recruited in Shaoguan City, China, through face-to-face questionnaire administration covering various ADS types (ACC, lane-keeping, automatic parking), we demonstrate that driver attitudes, perceived behavioral control, and subjective norms are significant determinants of takeover intention, collectively explaining nearly half of its variance (R² = 48.7%). Importantly, our analysis uncovers that both intention and perceived behavioral control have robust, direct effects on actual takeover behavior. Crucially, this work is among the first to reveal that individual user characteristics—such as driving experience and ADS (automated driving system) usage frequency—substantially moderate these psychological pathways: experienced or frequent users rely more on perceived control and attitude, while less experienced drivers are more susceptible to social influences. By advancing a multi-dimensional psychological model that integrates personal, social, and experiential moderators, our findings deliver actionable insights for the design of adaptive human–machine interfaces, tailored driver training, and targeted safety interventions in the context of automated driving. Using structural equation modeling with maximum likelihood estimation (χ²/df = 2.25, CFI = 0.941, RMSEA = 0.057), this psychological approach complements traditional engineering models by revealing that takeover behavior variance is explained at 58.3%.

1. Introduction

Automated driving technology represents not only a significant technological advancement but also a transformative solution to critical transportation challenges [1]. This innovation shows tremendous potential: it can eliminate human driving errors—the most common cause of accidents—free drivers from the continual task of monitoring, greatly enhance the travel experience and efficiency, and achieve higher driving performance and extended range [2,3,4].

The Society of Automotive Engineers categorizes automated driving capabilities into six distinct levels: Level 0 (no automation), Level 1 (driver assistance), Level 2 (partial automation), Level 3 (conditional automation), Level 4 (high automation), and Level 5 (full automation) [5]. The increasing deployment of partially and conditionally automated vehicles has created new challenges for human–machine interaction in the driving domain. Unlike traditional manual driving, automated driving creates a dual-task paradigm where drivers may engage in non-driving-related tasks while monitoring system performance [6]. This paradigm shift fundamentally alters drivers’ roles from active operators to passive supervisors who must occasionally intervene—a challenge referred to as the “irony of automation” [7,8].

Level 2–3 autonomous vehicle adoption is accelerating globally with significant regional variations. Current penetration shows 63% of global vehicle sales forecast to feature Level 2+ autonomy in 2025, with North America achieving 36% Level 2 adoption and 70% overall ADAS penetration, while Level 3 sales are expected to exceed 25,000 units globally in 2024 [9]. Regionally, China leads growth with targets of 70% Level 2–3-equipped vehicles by 2025 and projections for Level 3 to reach 10% market share by 2028, while North America dominates current market share and Europe is projected to reach 80% advanced AV sales by 2040 [10]. The autonomous vehicle market is expanding rapidly from USD 68.09 billion in 2024 to USD 273.75 billion in 2025.

Nevertheless, like any emerging technology, autonomous vehicles present an innovative yet complex system characterized by uncertainty, dynamism, and imperfect reliability and safety [11]. In Level 3 automated driving systems, human drivers are still expected to retake control of the vehicle in critical scenarios—such as system malfunctions or when encountering situations beyond the system’s defined operational domain [12]. These systems differ from fully autonomous ones in that human intervention remains necessary under certain conditions. Studies have shown that the handover phase, during which control is transferred back to the driver, is particularly vulnerable to safety risks due to possible delays in driver reaction or inadequate takeover performance [13,14].

Emerging evidence has highlighted several pressing safety issues associated with the handover process in conditionally automated driving. Field operational trials have shown that drivers using Level 2 and Level 3 systems often engage in non-driving-related tasks, leading to diminished situational awareness and a notable increase in emergency reaction times—averaging 2 to 8 s longer than during conventional manual driving [15,16]. Similarly, findings from naturalistic driving data indicate that around 40% of users tend to overly rely on automation, frequently neglecting environmental monitoring and showing higher involvement in potentially distracting behaviors [17,18].

Safe takeover transitions are critical to the overall safety performance of automated driving systems, representing the most vulnerable point in human–automation interaction where driver behavior directly determines collision risk and occupant safety outcomes. Data derived from disengagement incident reports suggest that approximately 38% of disengagements happen in dense or unpredictable traffic environments requiring prompt driver action. Of these, about 22% escalated into near-collision events due to insufficient or delayed driver responses [19], highlighting the direct causal relationship between takeover performance and crash probability. Quantitative safety assessments indicate that properly executed takeovers can reduce collision risk by up to 74% compared to delayed or improperly performed transitions. Furthermore, simulator-based investigations have observed that as user trust in automated systems grows, individuals are more inclined to perform secondary tasks and simultaneously exhibit poorer takeover performance—characterized by excessive steering inputs, lane position variability, and critical safety-compromising maneuvers—an emerging phenomenon referred to as the “risk paradox” [20,21]. This creates a troubling safety conflict: as systems become more reliable, the critical safety buffer provided by human takeover capability is progressively compromised. Understanding the psychological factors that influence takeover safety is therefore essential for minimizing these risks and designing more effective human–machine interfaces that can maintain safety margins even as automation technology advances.

In this study, a “takeover request” refers to a system prompt—either planned or emergency—requiring the human driver to resume manual control from the automated driving system [22,23]. “Takeover latency” is defined as the elapsed time between the takeover request and the driver’s first control input (e.g., steering and braking). “Takeover performance” refers to the overall quality and safety of driver actions during the transition, including metrics such as lane keeping, speed control, and collision avoidance [23].

Although research on driver takeover behavior has expanded considerably in recent years, key knowledge gaps persist. One major limitation is the dominance of engineering-oriented methodologies that emphasize objective indicators such as takeover latency, reaction time, and control input metrics [16,24]. While informative for assessing physical performance, these measures provide limited insight into the cognitive and psychological mechanisms that shape drivers’ decision-making during system disengagement events [25]. Another shortcoming lies in the disproportionate focus on external conditions—such as traffic complexity, time pressure, and secondary task engagement—when analyzing takeover quality [26,27]. This focus often neglects internal psychological determinants like drivers’ attitudes, belief systems, and self-efficacy, which are well-established predictors of behavior in broader driving research [28,29].

Furthermore, interactions between individual characteristics and psychological processes remain underexplored. While variables such as age, experience, and cognitive ability have been acknowledged [30], their moderating effects on the cognitive pathways involved in takeover remain insufficiently studied. Compounding this gap is the lack of robust theoretical models tailored specifically for automated driving handovers. As the authors of [31] observe, the field lacks integrative frameworks that synthesize psychological traits, personal factors, and contextual elements to explain variations in takeover performance and safety outcomes.

Drivers respond to takeover requests with differing levels of readiness and capability, often shaped by psychological influences such as acceptance of automation, perceived social expectations, and self-perceived control [32]. Simulator-based experiments have found that after every ten minutes of uninterrupted system operation, driver attentiveness declines by roughly 16%, while reaction time to intervention requests increases by about 0.8 s [20]. This inverse relationship between system reliability and driver preparedness creates significant risk, especially in unforeseen emergencies.

Addressing these issues calls for a more nuanced understanding of the psychological architecture underlying driver takeover behavior. This study aims to address the following research questions: (1) To what extent can the extended TPB model explain drivers’ intention and actual behavior in response to automated driving system disengagement? (2) How do individual differences such as driving experience and ADS use frequency moderate the psychological determinants of takeover performance?

The remainder of this paper is structured as follows: Section 2 reviews the relevant literature and presents the theoretical model and research hypotheses. Section 3 details the methodology, including participant recruitment, measures, and analytic strategy. Section 4 reports the empirical results, including model fit, hypothesis testing, and moderation analyses. Section 5 discusses the theoretical and practical implications of the findings, study limitations, and directions for future research. Section 6 concludes the paper. Therefore, this study aims to develop and validate a comprehensive psychological model of driver takeover behavior in conditionally automated vehicles using an extended Theory of Planned Behavior framework, while examining how individual characteristics moderate these psychological pathways.

The main contributions of this study are as follows: (1) Proposing a multi-dimensional psychological model for driver takeover in conditionally automated vehicles using an extended TPB framework; (2) empirically examining the moderating effects of driving experience and system use frequency; and (3) providing actionable insights for adaptive human–machine interface design and targeted driver training in the context of automated vehicles.

2. Literature Review

2.1. Driver Takeover Behavior After Automated Driving System Disengagement

The disengagement of automated driving systems refers to the transition in vehicle control from the system to the human driver, typically triggered by functional limitations, system malfunctions, or adverse environmental conditions [15]. These events are generally categorized into two types: planned disengagements, which occur when the system reaches its operational design limits and provides advance notice to the driver, and emergency disengagements, which arise unexpectedly due to technical faults or unforeseen circumstances demanding immediate driver intervention [33].

Driver takeover behavior involves the sequence of actions through which the human regains control of the vehicle following system disengagement. This process is inherently complex, encompassing several stages—including the detection of the takeover alert, the interpretation of the driving context, decision-making, and the execution of control maneuvers [22]. A growing body of research underscores the challenge of this transition, particularly as drivers are often required to switch from a passive supervisory role to active driving while potentially engaged in non-driving activities [26]. Scholarly attention has primarily concentrated on two measurable aspects of takeover behavior: takeover time and takeover quality. The former refers to the duration between the issuance of a takeover request and the driver’s successful re-engagement with vehicle control [23]. Empirical findings indicate considerable variability in takeover time, typically ranging from 2 to 7 s, influenced by contextual and environmental factors [33].

Takeover quality, on the other hand, evaluates how competently and safely drivers resume control. It is commonly assessed using performance indicators such as lane-keeping accuracy, control stability, and the appropriateness of evasive maneuvers [16]. This metric provides insight into drivers’ adaptive performance during the transition, which is known to be affected by variables including time pressure, traffic complexity, and the nature of the driver’s prior engagement [16,26]. While these studies have advanced the understanding of observable takeover performance, they offer limited insight into the internal cognitive and emotional mechanisms that drive such behavior. This is a notable shortcoming, as takeover is fundamentally a human–machine interaction involving interrelated cognitive, affective, and behavioral processes [34]. For instance, a driver’s response to a takeover request may depend not only on external conditions but also on psychological factors such as attitudes toward automation, normative beliefs, and perceived driving competence—areas that remain underrepresented in the literature.

While demographic and experiential factors such as age, prior driving experience, trust in automation, and cognitive functioning are known to influence takeover outcomes [24,35,36], prior studies have reported varying findings regarding the strength and direction of these effects. For example, trust and prior experience moderated takeover readiness, while research has shown that older age presents challenges for Level 3 automated vehicle control due to age-related physical, sensory, and cognitive decline [35], and higher trust in automation is associated with longer takeover response times [36]. Therefore, adopting psychological theories to investigate the determinants of takeover behavior holds considerable potential for constructing a more holistic and predictive framework for understanding this safety-critical phenomenon.

2.2. Theory of Planned Behavior and Its Application in Driving Behavior Research

The TPB, developed by [37], extends the earlier Theory of Reasoned Action by incorporating the concept of perceived behavioral control to better account for behaviors that may not be entirely voluntary. Within the TPB framework, behavior is directly shaped by behavioral intention—defined as the individual’s motivational readiness to perform a specific action. This intention is influenced by three principal components: attitude (the individual’s evaluation of performing the behavior), subjective norm (the perceived social pressures regarding the behavior), and perceived behavioral control (the perceived ease or difficulty of enacting the behavior). Notably, when perceived behavioral control accurately mirrors actual control, it may also exert a direct influence on behavior, particularly in situations requiring specific competencies or resources. TPB has consistently demonstrated high predictive validity across a range of behavioral domains and has been extensively applied within the realm of traffic safety. In driving research, TPB has been used to predict a variety of risky and safety-related behaviors, including speeding [38], impaired driving [39], aggressive driving [29], mobile phone use while driving [40], and seat belt compliance [41]. These investigations have consistently shown that drivers’ attitudes, perceived norms, and control beliefs significantly predict their intentions and actual behavior on the road.

For instance, ref. [38] reported that TPB variables accounted for 54% of the variance in drivers’ intentions to adhere to speed limits and 67% of the variance in self-reported speeding. Likewise, ref. [29] found that TPB constructs explained 47.2% of the variance in intentions to engage in driving violations. These findings underscore the theoretical robustness of TPB in explaining a broad spectrum of driver behaviors. With the growing integration of advanced technologies into transportation, researchers have increasingly turned to TPB to explore user acceptance of driving-related innovations. Using an extended TPB model, ref. [42] examined intentions to use GPS navigation, identifying perceived usefulness, ease of use, and subjective norm as key predictors of technology adoption. Ref. [43] compared several theoretical models and concluded that TPB outperformed technology-centered models in explaining drivers’ acceptance of advanced driver assistance systems.

Despite its demonstrated utility, TPB has seen limited application in the context of driver takeover behavior during automated driving, leaving a notable gap in the literature. A summary of critical research gaps is presented in

Table 1. Summary of critical research gaps.

Gap Category	Specific Limitation	Impact on Field
Methodological	Dominance of engineering approaches [16,24]	Limited psychological understanding
Theoretical	No integrative psychological frameworks [31]	Fragmented knowledge base
Individual Differences	Underexplored moderating effects [30]	Cannot personalize interventions
TPB Application	Minimal use in automated driving contexts	Missing proven theoretical foundation
Internal vs. External	Focus on conditions vs. psychological determinants [26,27] vs. [28,29]	Incomplete behavioral prediction

. This is a missed opportunity, given TPB’s strong alignment with the characteristics of takeover behavior. First, the act of taking over is a deliberative, volitional process well-suited to TPB’s scope. Second, it involves attitudinal judgments, perceived social expectations, and assessments of one’s ability to manage control—factors precisely represented in TPB’s core structure. Third, TPB allows for the inclusion of context-specific variables, such as trust in automation or previous takeover experience, enhancing its adaptability to the automated driving domain.

In addition, TPB offers a theoretical lens through which individual differences—such as age, trust levels, or driving experience—can be understood in terms of their influence on the underlying psychological determinants of behavior. Applying TPB to the study of takeover decisions in automated driving thus holds promise for advancing our understanding of the cognitive and motivational mechanisms underpinning this crucial safety-related action. It may also support the design of tailored interventions aimed at improving driver readiness and response quality during system handovers. Compared with previous TPB-based studies, our study uniquely integrates safety-oriented measurement and a multi-group SEM framework to examine experience-based moderation in a Chinese context.

Two individual difference variables were selected as moderators in the present study: driving experience and frequency of automated driving system use. Driving experience has been widely recognized as a key determinant of driving performance, risk perception, and decision-making in both manual and automated vehicle contexts [24,26]. More experienced drivers tend to exhibit greater situational awareness and are better able to anticipate and manage dynamic traffic scenarios during takeover events. The frequency of ADS use serves as an indicator of users’ familiarity and behavioral adaptation to automation. Prior studies [36] have suggested that higher usage frequency may enhance perceived control but can also foster over-reliance or complacency, potentially altering the psychological mechanisms underlying takeover behavior. Despite the acknowledged importance of these factors, previous research has seldom examined how driving experience and ADS use frequency may moderate the psychological pathways specified by the Theory of Planned Behavior in the context of automated driving takeovers. The present study addresses this gap by empirically testing their moderating effects within an extended TPB framework. Trust in automation, while important, was not included in the present model in order to maintain model parsimony and focus on the psychological pathways specified by the Theory of Planned Behavior.

The integration of TPB with automated driving takeover behavior research represents a natural theoretical convergence. The volitional nature of takeover decisions, combined with the attitude–behavior relationships established in driving research, makes TPB particularly suitable for understanding the psychological mechanisms underlying takeover performance. This theoretical foundation provides the basis for our extended model that incorporates individual characteristics as key moderators of these psychological pathways.

2.3. Theoretical Framework

This study adopts the TPB as the foundational theoretical model to explore the psychological processes influencing driver behavior following the disengagement of automated driving systems. In this context, TPB offers a structured and integrative lens to examine how drivers’ attitudes toward taking over control, perceived normative expectations regarding timely takeover, and their self-assessed capability to perform takeover maneuvers collectively shape both their intention and actual behavior during such transitions. Figure 1 presents our theoretical model developed for this research, which is an adaptation of the classical TPB framework tailored specifically to the domain of driver takeover behavior in conditionally automated driving environments.

2.4. Research Hypotheses

Based on these research questions, we propose the following hypotheses:

H1: 

Drivers’ attitudes toward takeover behavior have a significant positive effect on takeover intention.

Attitude refers to an individual’s overall evaluation of engaging in a behavior, encompassing both instrumental (e.g., beneficial vs. detrimental) and affective (e.g., pleasant vs. unpleasant) components. Within the TPB framework, more favorable attitudes toward a given behavior are expected to enhance the formation of behavioral intentions [37]. In the context of automated driving, if drivers perceive that prompt takeover enhances safety or reduces crash risks, they are more likely to form strong intentions to assume control when needed. This relationship is consistent with prior findings across driving-related behaviors [29,38] and technology adoption [42].

H2: 

Subjective norms have a significant positive effect on drivers’ takeover intention.

Subjective norms represent perceived social expectations—both injunctive (what others believe one should do) and descriptive (what others actually do). TPB posits that individuals who sense stronger normative pressure are more inclined to intend to perform the expected behavior [37]. Applied to takeover scenarios, drivers who believe that significant others (e.g., family and peers) or society at large expect them to respond promptly to system disengagements are more likely to form corresponding intentions. Empirical support comes from studies examining normative effects on rule compliance [29] and adoption of new driving technologies [43].

H3: 

Perceived behavioral control has a significant positive effect on drivers’ takeover intention.

This construct captures a person’s belief in their ability to perform a behavior, integrating aspects of self-efficacy (confidence in capability) and controllability (perceived influence over external conditions). According to TPB, greater perceived control enhances behavioral intention [37]. In takeover contexts, drivers who believe they possess sufficient driving skills and can manage the demands of regaining control are more likely to form a strong intention to do so. This hypothesis aligns with findings on speeding behavior [38] and sustainable driving practices [44].

H4: 

Perceived behavioral control has a significant positive effect on drivers’ actual takeover behavior.

Unlike attitude and subjective norm, perceived behavioral control is theorized to influence behavior both indirectly (via intention) and directly—particularly in scenarios where behavior involves limited volitional control and control perceptions align with actual ability [37]. Since takeover performance often requires immediate action and specific competencies, perceived control may directly predict successful takeovers regardless of intention strength. Prior studies have confirmed such direct effects in various driving behaviors [38,39].

H5: 

Takeover intention has a significant positive effect on drivers’ actual takeover behavior.

TPB identifies intention as the most proximal determinant of behavior. Stronger intentions typically correspond with a greater likelihood of behavior enactment [37]. In the automated driving context, drivers with clear intentions to take over promptly are expected to demonstrate higher takeover performance. This intention–behavior link has been repeatedly validated in driving safety research [29,38].

H6: 

Individual driver characteristics (driving experience and automated driving system use experience) have significant moderating effects on TPB constructs.

Personal traits may influence how TPB variables interact by altering the salience of attitudinal, normative, or control factors. For instance, more experienced drivers might rely more on perceived behavioral control and less on normative cues compared to novice drivers. Similarly, familiarity with automated systems may enhance confidence, thereby strengthening the control–intention link. This moderating role of experience is supported by prior TPB-based meta-analyses [45] and studies on driver performance during takeovers [24].

Together, these hypotheses form a comprehensive analytical framework for investigating the psychological underpinnings of driver takeover behavior. They also provide the basis for empirical testing and the development of targeted interventions to improve takeover readiness and safety outcomes in automated driving contexts.

3. Research Methods

3.1. Participants

Participants for this study were recruited via in-person questionnaire administration conducted in the central urban area of Shaoguan City, Guangdong Province. The target population comprised drivers with prior experience using automated driving systems. To enhance the representativeness and relevance of the sample, inclusion criteria were defined as follows: (1) possession of a valid driver’s license for a minimum of one year; (2) engagement with at least one automated driving function—such as adaptive cruise control or lane-keeping assistance—within the preceding six months; and (3) prior experience in manually resuming control following automated system disengagement.

3.2. Measurement Instruments

The questionnaire consisted of three main sections: demographic information, automated driving system usage, and TPB construct measurement scales. The development process followed rigorous methodological procedures. First, an initial item pool was generated based on a comprehensive literature review. This draft questionnaire was then reviewed and revised by three experts in human factors engineering and traffic psychology to ensure content validity. Following expert review, the questionnaire was pretested with 30 participants who met the inclusion criteria. Based on their feedback regarding question clarity, relevance, and format, the questionnaire was further optimized to improve face validity and user experience.

The questionnaire measured five constructs based on the TPB, adapted to the automated driving takeover context with a specific focus on safety dimensions. The scales for Attitude toward Takeover Behavior (5 items) were designed to capture drivers’ evaluations of takeover behavior’s safety implications, with items specifically addressing safety perceptions (e.g., “Taking over promptly when the automated driving system disengages is safe”). The Subjective Norm Scale (4 items) assessed perceived social expectations regarding safe takeover practices, including safety recommendations from authorities. Perceived Behavioral Control Items (5 items) measured drivers’ confidence in their ability to execute safe takeovers, particularly focusing on control capabilities that directly impact takeover safety outcomes.

The Takeover Intention Scale (3 items) assessed drivers’ readiness to respond promptly to system disengagements—a critical safety precursor. Finally, the Takeover Behavior Scale (5 items) incorporated self-reported measures of safety-critical performance dimensions during takeovers, including vehicle stability control, appropriate response to traffic conditions, and overall takeover quality. These safety-related behavioral indicators were informed by previous research on takeover performance metrics associated with collision avoidance and safe maneuvering [16,24].

All scales were developed following [46] measurement guidelines with adaptations from previous driving behavior research [29,38], specifically targeting safety-relevant aspects of takeover behavior. Items for the automated driving context drew from work on technology acceptance and research on advanced driver assistance systems [42,43], while the takeover behavior measures were informed by takeover performance metrics that have been linked to safety outcomes in simulator studies. Previous research has established that these metrics—including response time, control stability, and appropriate maneuver selection—are significantly associated with collision risk during takeover transitions.

All TPB construct items used a 5-point Likert scale (1 = strongly disagree; 5 = strongly agree). In this study, “takeover request” refers to both planned and emergency disengagements. The questionnaire items were intentionally designed to capture driver intentions and behaviors in response to either type of system disengagement, regardless of whether advance notice was provided. The safety-oriented framing of items across all constructs enables a comprehensive assessment of the psychological factors that influence safe takeover behavior in automated driving contexts.

3.3. Data Collection Procedure

Data collection was conducted through a face-to-face questionnaire survey from November to December 2024. Several strategies were implemented to enhance the participation rate and data quality: All participants were screened for eligibility by requiring them to report a valid driver’s license and prior use of automated driving system functions. Responses from individuals who did not meet these criteria were excluded. Multiple attention check questions were embedded throughout the questionnaire to identify careless responding and ensure data quality. Responses failing these checks were excluded from the final analysis. Participants who completed valid questionnaires received modest compensation appropriate to the survey length, enhancing motivation without introducing potential response bias. Before completing the questionnaire, participants read and agreed to an informed consent form detailing the study purpose, confidentiality measures, voluntary participation, and right to withdraw at any point.

The survey was designed to take approximately 15 min to complete, reducing the likelihood of respondent fatigue. Data collection occurred over a six-week period to ensure adequate sample size and diversity. Participants were recruited at vehicle management offices and car dealerships in the urban area of Shaoguan City, Guangdong Province, China. The initial sample consisted of 407 respondents. After excluding 22 cases due to incomplete responses or failure on attention check items (e.g., instructed response or inconsistent answers), the final valid sample comprised 385 participants. The refusal rate was approximately 5.4%. Demographic data (age, gender, driving experience) were collected to assess sample representativeness; the distribution is broadly consistent with the local licensed-driver population, although generalizability to other regions may be limited. All participants provided informed consent prior to completing the questionnaire. The final sample of 385 participants exceeded the minimum sample size requirement for structural equation modeling [47], providing sufficient statistical power for the analyses, ensuring robust statistical power for model estimation.

3.4. Data Analysis

All absolute skewness and kurtosis values were within acceptable ranges (|skewness| < 2, |kurtosis| < 7), indicating approximate univariate normality. No missing values or extreme outliers were detected, and thus no data transformation was necessary. Since no severe non-normality or outliers were detected, original (untransformed) scores were used in the SEM. The model was estimated using the maximum likelihood (ML) method.

Data analysis was conducted using SPSS 26.0 for descriptive statistics and reliability analysis, and AMOS 24.0 for confirmatory factor analysis and structural equation modeling. Figure 2 illustrates the data analysis process flow. The analysis began with descriptive statistical analysis to provide an overview of the data distribution and initial relationships between constructs. Mean, standard deviation, and correlation coefficients were calculated for all variables. This was followed by reliability and validity analysis to assess the measurement model. The internal consistency of each scale was assessed using Cronbach’s α coefficient, with values above 0.70 considered acceptable [48]. Construct validity was assessed using confirmatory factor analysis (CFA), which encompassed evaluations of both convergent validity—via factor loadings, average variance extracted (AVE), and composite reliability (CR)—and discriminant validity, by comparing the square root of each construct’s AVE with the corresponding inter-construct correlations.

Upon establishing the adequacy of the measurement model, SEM was employed to examine hypotheses H1 through H5, focusing on the directional relationships among the TPB constructs. The analysis was conducted using the maximum likelihood estimation method. Model fit was evaluated using a set of established indices to ensure robustness against measurement error. The following fit criteria were used to assess model adequacy: χ²/df < 3, GFI > 0.90, CFI > 0.90, TLI > 0.90, RMSEA < 0.08, SRMR < 0.08 [48]. This combination of fit indices provides a comprehensive assessment of model fit, considering both absolute fit and comparative fit measures.

To test hypothesis H6, which posits moderating effects of individual characteristics, multi-group SEM analysis was performed. Before conducting multi-group analyses, measurement invariance was assessed across groups (e.g., by driving experience or ADS use frequency). Three levels of invariance—configural, metric, and scalar—were tested sequentially following established guidelines. The criteria for invariance were based on changes in model fit indices (ΔCFI < 0.01; ΔRMSEA < 0.015).

Participants were categorized into subgroups based on driving experience (≤6 years vs. >6 years) and frequency of automated system use (several times per month or less vs. several times per week or more). Chi-square difference testing was applied to compare nested models (constrained vs. unconstrained) and assess whether structural path coefficients varied significantly across groups. This approach allowed for an examination of how individual differences might influence the strength of relationships between TPB constructs in the context of automated driving takeover behavior.

4. Research Results

4.1. Descriptive Statistics and Correlation Analysis

A total of 385 valid questionnaires were collected after excluding incomplete responses and those failing attention checks. The demographic characteristics of the sample are shown in

Table 2. Sample characteristics (N = 385).

Characteristic	Category	Frequency	Percentage (%)
Gender	Male	226	58.7
	Female	159	41.3
Age	18–25 years	87	22.6
	26–35 years	156	40.5
	36–45 years	98	25.5
	46 years and above	44	11.4
Driving Experience	1–3 years	112	29.1
	4–6 years	143	37.1
	7–10 years	76	19.8
	More than 10 years	54	14.0
Frequency of Automated Driving System Use	Almost daily	98	25.4
	Several times a week	145	37.7
	Several times a month	108	28.1
	Rarely use	34	8.8
Type of Automated Driving System Used *	ACC (adaptive cruise control)	342	88.8
	LKA (lane-keeping assist)	289	75.1
	APA (automatic parking assist)	196	50.9
	HDA (highway driving assist)	165	42.9
	Others	43	11.2

* Note: multiple responses were allowed; percentages will not sum to 100%. Source: own elaboration based on survey data.

. Male participants represented 58.7% (n = 226) of the sample, while females accounted for 41.3% (n = 159). The largest age group was 26–35 years (40.5%, n = 156), followed by 36–45 years (25.5%, n = 98), 18–25 years (22.6%, n = 87), and 46 years and above (11.4%, n = 44). Regarding driving experience, 37.1% (n = 143) had 4–6 years of experience, 29.1% (n = 112) had 1–3 years, 19.7% (n = 76) had 7–10 years, and 14.0% (n = 54) had more than 10 years. In terms of automated driving system use frequency, 37.7% (n = 145) reported using such systems several times a week, 28.1% (n = 108) used them several times a month, 25.5% (n = 98) used them almost daily, and 8.8% (n = 34) rarely used them.

Table 3. Descriptive statistics and correlation analysis of main variables (N = 385).

Variable	Mean	SD	AT	SN	PBC	TI	TB
Attitude (AT)	4.11	0.63	1
Subjective Norm (SN)	3.77	0.74	0.359 **	1
Perceived Behavioral Control (PBC)	3.87	0.71	0.402 **	0.380 **	1
Takeover Intention (TI)	4.03	0.71	0.465 **	0.419 **	0.501 **	1
Takeover Behavior (TB)	3.85	0.70	0.375 **	0.343 **	0.494 **	0.515 **	1

Note: ** p < 0.01; source: own elaboration based on survey data.

presents the descriptive statistics and correlation analysis results of the main variables. The mean scores for the TPB constructs ranged from 3.77 (subjective norm) to 4.11 (attitude), indicating generally positive evaluations of takeover behavior among the participants. The standard deviations (ranging from 0.63 to 0.74) suggest moderate variability in responses across the sample. The correlation analysis revealed significant positive correlations between all TPB constructs (p < 0.01), which is consistent with theoretical expectations. The correlation coefficient between takeover intention and takeover behavior was the highest (r = 0.515), followed by the correlation between perceived behavioral control and takeover intention (r = 0.501). The correlation coefficient between perceived behavioral control and takeover behavior was also relatively high (r = 0.494), supporting the TPB hypothesis that perceived behavioral control directly influences behavior in addition to its indirect effect through intention.

4.2. Reliability and Validity Analysis

Model fit indices for the final structural equation model are presented in

Table 4. Model fit indices.

Fit Index	Value	90% CI
χ2/df	2.18	–
CFI	0.945	–
TLI	0.937	–
RMSEA	0.055	0.046–0.064
SRMR	0.048	–
R2 (TI)	0.487	–
R2 (TB)	0.583	–

. All fitness indices met the requirements. The reliability and validity analysis results for each scale are presented in

Table 5. Reliability and validity analysis results.

Variable	Item	Factor Loading	Cronbach’s α	CR	AVE
AT	AT1	0.807	0.829	0.880	0.594
	AT2	0.762
	AT3	0.783
	AT4	0.767
	AT5	0.733
SN	SN1	0.829	0.816	0.879	0.645
	SN2	0.796
	SN3	0.792
	SN4	0.794
PBC	PBC1	0.749	0.838	0.886	0.608
	PBC2	0.814
	PBC3	0.805
	PBC4	0.773
	PBC5	0.755
TI	TI1	0.855	0.794	0.879	0.709
	TI2	0.847
	TI3	0.823
TB	TB1	0.749	0.823	0.876	0.587
	TB2	0.793
	TB3	0.817
	TB4	0.722
	TB5	0.745

Note: CR = composite reliability; AVE = average variance extracted; source: own elaboration based on survey data.

. All constructs demonstrated good internal consistency, with Cronbach’s α coefficients ranging from 0.794 (takeover intention) to 0.838 (perceived behavioral control). Confirmatory factor analysis indicated that the five-factor model fit the data well (χ²/df = 2.18, GFI = 0.921, CFI = 0.945, TLI = 0.937, RMSEA = 0.055, SRMR = 0.048). All items exhibited factor loadings greater than 0.70 (ranging from 0.722 to 0.855), AVE values greater than 0.50 (ranging from 0.587 to 0.709), and CR values greater than 0.80 (ranging from 0.876 to 0.886). These results provide evidence of good convergent validity for the scales. Discriminant validity was also confirmed, as the square root of AVE for each construct exceeded the correlation coefficients between that construct and other constructs in the model.

All TPB constructs and takeover behavior items were measured using self-report scales framed in a safety-oriented context. To assess potential common method variance (CMV), Harman’s single-factor test was performed on all self-report items using principal component analysis. The first unrotated factor accounted for 33.8% of the total variance, which is below the recommended threshold of 50%, indicating that common method bias is unlikely to be a significant concern in this study.

4.3. Hypothesis Testing

Based on the TPB theoretical framework, we constructed a structural equation model of driver takeover behavior after automated driving system disengagement. The model fit indicators demonstrated good fit with the data (χ²/df = 2.25, GFI = 0.918, CFI = 0.941, TLI = 0.934, RMSEA = 0.057, SRMR = 0.051). The results revealed that attitude (β = 0.324, p < 0.001), subjective norm (β = 0.215, p < 0.001), and perceived behavioral control (β = 0.302, p < 0.001) all had significant positive effects on takeover intention, supporting hypotheses H1, H2, and H3. Furthermore, takeover intention (β = 0.534, p < 0.001) and perceived behavioral control (β = 0.296, p < 0.001) both had significant positive effects on takeover behavior, supporting hypotheses H4 and H5. The model explained 48.7% of the variance in takeover intention and 58.3% of the variance in takeover behavior.

To verify hypothesis H6, we conducted a multi-group analysis to examine the moderating effects of driving experience and automated driving system use experience. Moderating Effect of Driving Experience: The sample was divided into a low driving experience group (≤6 years, n = 255) and a high driving experience group (>6 years, n = 130). A chi-square difference test showed a significant difference between the constrained model and the unconstrained model (Δχ² = 18.27, Δdf = 5, p < 0.01), indicating that driving experience has a moderating effect on TPB construct relationships.

In the high driving experience group, the effects of perceived behavioral control on takeover intention (βhigh = 0.358 vs. βlow = 0.249, p < 0.05) and takeover behavior (βhigh = 0.343 vs. βlow = 0.253, p < 0.05) were stronger. Conversely, in the low driving experience group, the effect of subjective norm on takeover intention was stronger (βlow = 0.265 vs. βhigh = 0.172, p < 0.05) (see

Table 6. Multi-group analysis results of the moderating effect.

Moderating Effect of Driving Experience
Path	Low Driving Experience Moderation (≤6 years, n = 255)	High Driving Experience Moderation (>6 years, n = 130)	Δχ2	p-Value
SN → TI	0.265	0.172	5.34	<0.05
PBC → TI	0.249	0.358	6.21	<0.05
PBC → TB	0.253	0.343	5.87	<0.05
TI → TB	0.518	0.546	2.11	>0.05
AT → TI	0.317	0.336	1.92	>0.05
Moderating Effect of ADS Use Frequency
Path	Low ADS Use Frequency Moderation (≤several times a month, n = 142)	High ADS Use Frequency Moderation (≥several times a week, n = 243)	Δχ2	p-Value
AT → TI	0.279	0.375	5.76	<0.05
SN → TI	0.287	0.156	7.32	<0.01
PBC → TI	0.258	0.347	6.13	<0.05
PBC → TB	0.275	0.311	2.43	>0.05
TI → TB	0.521	0.542	1.89	>0.05

Note: Δχ² indicates the chi-square difference test value between constrained and unconstrained models for each path. Source: own elaboration based on survey data.

). These results suggest that as driving experience increases, drivers rely more on their perceived ability to execute takeover maneuvers and less on social norms when forming takeover intentions.

Moderating Effect of Automated Driving System Use Experience: The sample was divided into a low-frequency group (several times a month or less, n = 142) and a high-frequency group (several times a week or more, n = 243). The chi-square difference test showed a significant difference between the constrained model and the unconstrained model (Δχ² = 16.54, Δdf = 5, p < 0.01).

In the high-frequency group, the effects of attitude (βhigh = 0.375 vs. βlow = 0.279, p < 0.05) and perceived behavioral control (βhigh = 0.347 vs. βlow = 0.258, p < 0.05) on takeover intention were stronger. In contrast, in the low-frequency group, the effect of subjective norm on takeover intention was stronger (βlow = 0.287 vs. βhigh = 0.156, p < 0.01) (see Table 6). The results indicated that configural, metric, and scalar invariance were established across the relevant groups. Thus, subsequent group comparisons were justified. These findings indicate that as drivers gain more experience with automated driving systems, their intentions to take over control become more influenced by their attitudes and perceived behavioral control, and less by subjective norms.

These results collectively support hypothesis H6, demonstrating that individual driver characteristics have significant moderating effects on the relationships between TPB constructs in the context of automated driving takeover behavior. Figure 3 illustrates the moderating effects of driver characteristics on the relationships between TPB constructs. As shown in the above panels, high driving experience strengthens the influence of PBC on both takeover intention and behavior, while diminishing the impact of subjective norms. Similarly, the below panels demonstrate that frequent users of automated driving systems exhibit stronger effects of attitude and perceived behavioral control on takeover intention, with reduced reliance on subjective norms compared to infrequent users.

Figure 4 shows the structural equation model analysis results, including standardized path coefficients and explained variance (R²).

Table 7. Standardized path analysis results of the structural equation model.

Hypothesis	Path	Standardized Path Coefficient (β)	p-Value	95% CI	Result
H1	AT → TI	0.324	<0.001	[0.245, 0.403]	Supported
H2	SN → TI	0.215	<0.001	[0.137, 0.293]	Supported
H3	PBC → TI	0.302	<0.001	[0.224, 0.380]	Supported
H4	PBC → TB	0.296	<0.001	[0.218, 0.374]	Supported
H5	TI → TB	0.534	<0.001	[0.456, 0.612]	Supported

Note: source: own elaboration based on survey data.

and

Table 8. The results of hypothesis 6.

Hypothesis	Path	Standardized Path Coefficient (β)	p-Value	Result
H6	SN → TI	Driving Experience Moderation: −0.093	<0.05	Supported
		ADS Use Frequency Moderation: −0.131	<0.01
	PBC → TI	Driving Experience Moderation: 0.109	<0.05
		ADS Use Frequency Moderation: 0.089	<0.05
	PBC → TB	Driving Experience Moderation: 0.09	<0.05
	AT → TI	ADS Use Frequency Moderation: 0.096	<0.05

Note: source: own elaboration based on survey data.

clearly show the test results of the five main hypotheses (H1–H6), and all paths significantly support the theoretical hypotheses of this study.

5. Discussion

5.1. Discussion of Research Results

This study constructed a prediction model for driver takeover behavior after automated driving system disengagement based on TPB, providing a novel perspective for understanding the psychological decision-making processes underlying takeover behavior. The findings confirm the predictive utility of the TPB constructs in explaining driver takeover behavior. Specifically, attitude, subjective norm, and perceived behavioral control collectively accounted for 48.7% of the variance in takeover intention, while takeover intention in conjunction with perceived behavioral control explained 58.3% of the variance in actual takeover behavior. These levels of explanatory power are consistent with prior research applying TPB to driver behavior and technology acceptance [38,42], underscoring the framework’s suitability for modeling behavioral responses in automated driving scenarios.

Among the TPB constructs, attitude exerted the most substantial influence on takeover intention (β = 0.324, p < 0.001), followed by perceived behavioral control (β = 0.302, p < 0.001) and subjective norm (β = 0.215, p < 0.001). These results suggest that drivers’ favorable evaluations of takeover behavior are the primary driver of their intention to assume control following system disengagement. This pattern mirrors findings from [42], who emphasized the central role of attitude in the acceptance of advanced driver assistance technologies. The significant impact of perceived behavioral control further highlights the importance of drivers’ self-assessment regarding their ability to execute a safe and timely takeover, corroborating evidence from [16], which links perceived ability with actual takeover performance. Although the effect size for subjective norm was comparatively smaller, it remained statistically significant, indicating that perceived social expectations and normative pressures also contribute to the formation of takeover intentions.

This study confirmed the direct effects of takeover intention (β = 0.534, p < 0.001) and perceived behavioral control (β = 0.296, p < 0.001) on takeover behavior, with takeover intention having a stronger effect. This result supports the core hypothesis of TPB that behavioral intention is the most direct antecedent of behavior and also indicates that perceived behavioral control can directly predict behavior, especially when behavior execution is limited by ability and resources [37]. This is consistent with previous international research demonstrating that drivers who report higher levels of self-efficacy or perceived control tend to exhibit quicker and more effective takeover responses [24,26]. In the context of automated driving system disengagement, takeover behavior often needs to be completed within a short time frame, with high requirements for drivers’ abilities; thus, the direct effect of perceived behavioral control on takeover behavior becomes more prominent.

Furthermore, individual driver characteristics demonstrated significant moderating effects on TPB construct relationships. Specifically, for drivers with high driving experience (>6 years), perceived behavioral control had stronger effects on takeover intention (βhigh = 0.358 vs. βlow = 0.249, p < 0.05) and takeover behavior (βhigh = 0.343 vs. βlow = 0.253, p < 0.05). This might be because rich driving experience enhances drivers’ confidence and sense of control over their own takeover ability [22]. In contrast, drivers with low driving experience (≤6 years) were more influenced by subjective norms (βlow = 0.265 vs. βhigh = 0.172, p < 0.05), indicating that drivers with less driving experience rely more on social expectations and normative guidance in takeover decisions.

Similarly, for frequent users of automated driving systems (several times a week or more), the effects of attitude (βhigh = 0.375 vs. βlow = 0.279, p < 0.05) and perceived behavioral control (βhigh = 0.347 vs. βlow = 0.258, p < 0.05) on takeover intention were stronger. Infrequent users (several times a month or less) were more influenced by subjective norms (βlow = 0.287 vs. βhigh = 0.156, p < 0.01). This finding aligns with [24] research on the effect of experience on drivers’ takeover performance, indicating that use experience can influence takeover behavior by changing takeover attitudes and perception of control.

Notably, some moderation effects showed borderline significance, particularly for the attitude–intention relationship across experience groups (p = 0.06). This suggests that experience-based differences in attitudinal influences may be more nuanced than our categorical analysis revealed. Future research using continuous moderation techniques might uncover threshold effects or non-linear relationships.

The present study’s results align with a growing body of international evidence on the psychological determinants of driver behavior during automation disengagement. Similarly to studies conducted in Europe and North America [24,26], we found that both intention and perceived behavioral control are key drivers of safe takeover performance, regardless of cultural or regulatory context. This suggests that the TPB framework can be robustly applied across different geographic settings and supports the generalizability of our findings beyond the Chinese sample.

5.2. Theoretical Implications

The application of the TPB to driver takeover behavior following automated system disengagement broadens its theoretical relevance within the domains of traffic safety and human–machine interaction. The present findings substantiate the explanatory and predictive capacity of TPB in this emerging context, thereby extending the theoretical framework into the evolving landscape of automated mobility. While TPB has traditionally been employed to model conventional driving behaviors such as speeding and impaired driving, its deployment in automated driving scenarios remains relatively underexplored. This study highlights TPB’s adaptability and utility in addressing behavioral responses to advanced vehicle technologies, contributing to the diversification of its applications across novel technological environments.

Building on the traditional TPB framework, this study explored the moderating effects of individual driver characteristics, enhancing the explanatory power of the theory. The significant moderating effects of driving experience and use frequency on the relationships between TPB constructs indicate that the relative influence of attitude, subjective norm, and perceived behavioral control varies with individual characteristics. This provides empirical support for refining TPB theory in the context of human–automation interaction and suggests that TPB models should incorporate individual difference factors when applied to complex technological behaviors.

This study explored the influencing factors of driver takeover behavior from a psychological decision-making perspective, complementing existing research that primarily focuses on the external manifestations of takeover behavior while neglecting internal psychological processes. By quantifying the relationships between psychological constructs and takeover behavior, the research offers a more nuanced understanding of how internal cognitive and affective processes shape drivers’ responses to automated system disengagement. This approach provides a new perspective for understanding the psychological mechanisms of driver–automation system interaction, advancing theoretical development in human factors engineering and traffic psychology.

The observed correlation structure between TPB constructs (with correlations ranging from 0.343 to 0.515) further supports the theoretical coherence of the TPB model in this context. The moderate correlation between perceived behavioral control and takeover behavior (r = 0.494) particularly highlights the dual role of PBC as both a direct determinant of behavior and an indirect influence through intention, as theorized by [37]. This dual-pathway effect advances our understanding of how control perceptions influence complex behaviors involving human–machine interaction.

5.3. Practical Implications

The findings from this study provide a foundation for developing a comprehensive safety enhancement framework that can be directly applied to automated driving system design, driver training, and policy development to significantly improve safety during takeover transitions. By simultaneously targeting drivers’ cognitive processes (attitudes), social influences (subjective norms), and skill components (perceived behavioral control), a more robust takeover safety assurance system can be constructed.

Given the significant effect of attitude on takeover intention (β = 0.324, p < 0.001), automobile manufacturers and relevant institutions should strengthen drivers’ awareness of the importance of timely takeover through education and publicity, cultivating positive takeover attitudes. Cultivating positive attitudes toward takeover may contribute to improved takeover performance and safety, though further research is needed to confirm these effects. Emphasizing the safety value and utility of takeover behavior can increase drivers’ willingness to take over when necessary, thereby reducing safety incidents caused by hesitation. Marketing and communication strategies for automated vehicles should highlight not only the convenience of automation but also the critical role of human intervention in ensuring safety during system limitations.

Considering the dual effects of perceived behavioral control on takeover intention (β = 0.302, p < 0.001) and takeover behavior (β = 0.296, p < 0.001), driver takeover skill training should be strengthened to improve drivers’ takeover confidence and actual ability. Simulator studies demonstrate that drivers who undergo systematic takeover training exhibit more stable lane-keeping capability (32% less deviation), reduced oversteering (45% reduction), and smoother braking patterns (29% fewer sudden braking events) during emergency takeovers—metrics directly associated with collision avoidance capabilities. This study suggests developing targeted takeover training modules, such as simulated takeover scenario training, to help drivers accumulate takeover experience and enhance their sense of control. This is particularly important for drivers with less experience, who demonstrated lower perceived behavioral control in our analysis.

The effect of subjective norm on takeover intention (β = 0.215, p < 0.001) indicates that creating a social atmosphere supporting safe takeover is also important. By establishing clear takeover norms and social expectations, especially for drivers with low driving experience and infrequent automated system users who rely more heavily on normative influences, the formation of safe takeover behavior can be promoted. This could involve developing industry standards for takeover protocols that emphasize safety benchmarks and incorporating these standards into driver education programs. Safety campaigns highlighting responsible takeover practices have been shown to improve compliance rates by 27% among novice automated system users.

Based on the moderating effects of individual characteristics, we recommend adopting differentiated design strategies for drivers with different characteristics and implementing complementary takeover safety monitoring systems:

For drivers with low driving experience (≤6 years) and infrequent automated system users (several times a month or less), human–machine interfaces should incorporate more normative guidance elements, such as explicit instructions and recommendations during takeover requests, leveraging these drivers’ greater responsiveness to subjective norms. Additionally, systems should monitor these drivers’ takeover performance in real-time, providing immediate corrective feedback when unsafe takeover patterns (such as oversteering or sudden braking) are detected, which has been shown to reduce takeover-related accident risks by 23%.

For drivers with high driving experience (>6 years) and frequent users (several times a week or more), system designs should focus on enhancing their sense of control and positive attitudes toward takeover. This might include providing more detailed system status information and customizable takeover settings to accommodate their stronger reliance on perceived behavioral control and attitudes. Furthermore, adaptive warning systems can be developed to dynamically adjust takeover request lead times based on driver distraction levels and environmental complexity, an approach demonstrated to increase successful safe takeover rates by 17%.

These differentiated approaches would optimize human–machine interaction design to meet the needs of different drivers, enhance safety during takeover transitions, and provide theoretical support and practical pathways for future system design. Implementation of these safety-focused improvements based on TPB constructs could significantly reduce the estimated 38% of disengagements that currently occur in challenging traffic environments, particularly addressing the 22% that escalate into near-collision events due to insufficient or delayed driver responses.

5.4. Limitations and Future Directions

This study adopts a cross-sectional survey design, which cannot establish causal relationships between variables. The relationships identified between TPB constructs and takeover behavior represent correlational patterns rather than definitive causal mechanisms. Future research should adopt longitudinal designs or experimental methods to more accurately reveal the causal relationships between TPB constructs and takeover behavior, tracking how these relationships evolve over time as drivers gain experience with automated systems. The measurement of takeover behavior relies on self-report data, which may involve subjective bias and recall errors. While self-reports provide valuable insights into drivers’ perceptions, they may not accurately reflect actual behavior under real-world conditions. Future research can combine objective measurement methods such as driving simulators or real vehicle tests to more accurately assess drivers’ actual takeover performance, potentially revealing discrepancies between perceived and actual takeover abilities.

The present sample was collected from an urban area in a highly developed province, which may limit the generalizability of the findings to other regions or populations. Future research should recruit participants from diverse geographic and socioeconomic backgrounds to enhance external validity. Our model explained 48.7% of the variance in takeover intention and 58.3% of the variance in takeover behavior, suggesting that other factors beyond the traditional TPB constructs may influence takeover behavior. Future research could expand the theoretical framework by incorporating additional constructs such as trust in automation, risk perception, situation awareness, and mental workload to further enhance the model’s explanatory power. This study only considered two moderating variables, driving experience and automated driving system use frequency. While these variables demonstrated significant moderating effects, other individual difference factors may also influence the relationships between TPB constructs and takeover behavior. Future research should explore the effects of additional individual difference factors, such as personality traits, cognitive abilities, age, and technology acceptance tendencies, to develop a more comprehensive understanding of the factors that shape takeover behavior.

In addition, the use of a median split for subgroup analyses is inherently limited. We recommend that future studies apply the Johnson–Neyman technique or continuous moderation analysis to provide more nuanced insights. Another limitation of this study is that all available data were used to estimate the structural equation model. Due to sample size constraints, we were unable to conduct split-sample validation or cross-validation. Future studies should employ larger datasets to allow for training/testing splits or cross-validation to assess model robustness. Finally, this study did not distinguish between different types of automated driving systems and disengagement contexts, while different systems and contexts may affect the relationships between TPB constructs. The current analysis treated automated driving systems as a general category, potentially masking important variations across system types and disengagement scenarios. Future research should focus on specific automated driving systems and disengagement contexts (such as planned disengagement vs. emergency disengagement) to explore the predictive power of TPB constructs in different scenarios, allowing for more targeted interventions and design recommendations.

6. Conclusions

This study constructed a prediction model for driver takeover behavior after automated driving system disengagement based on the TPB and validated its effectiveness through a questionnaire survey of 385 drivers with automated driving system experience and structural equation modeling analysis.

First, drivers’ attitudes toward takeover behavior, subjective norms, and perceived behavioral control all had significant positive effects on takeover intention, collectively explaining 48.7% of its variance. Among these factors, attitude had the strongest influence (β = 0.324, p < 0.001), followed by perceived behavioral control (β = 0.302, p < 0.001) and subjective norms (β = 0.215, p < 0.001). This indicates that cultivating positive evaluations of takeover behavior’s safety and necessity is crucial for forming takeover intentions.

Second, takeover intention (β = 0.534, p < 0.001) and perceived behavioral control (β = 0.296, p < 0.001) both demonstrated significant positive predictive effects on actual takeover behavior, explaining 58.3% of its variance. This result confirms TPB’s core hypothesis that behavioral intention is the most direct antecedent of behavior, while also indicating that in automated driving takeover contexts, drivers’ self-assessment of their capabilities directly influences their takeover performance.

Third, individual driver characteristics showed significant moderating effects on TPB construct relationships. In the high driving experience group (>6 years, n = 130), perceived behavioral control had stronger effects on takeover intention (βhigh = 0.358 vs. βlow = 0.249, p < 0.05) and takeover behavior (βhigh = 0.343 vs. βlow = 0.253, p < 0.05), while in the low driving experience group (≤6 years, n = 255), subjective norms had a stronger influence on takeover intention (βlow = 0.265 vs. βhigh = 0.172, p < 0.05). Similarly, for frequent users of automated driving systems (several times a week or more, n = 243), attitude (βhigh = 0.375 vs. βlow = 0.279, p < 0.05) and perceived behavioral control (βhigh = 0.347 vs. βlow = 0.258, p < 0.05) had stronger effects on takeover intention, while for infrequent users (several times a month or less, n = 142), subjective norms had a more pronounced influence (βlow = 0.287 vs. βhigh = 0.156, p < 0.01).

These findings expand the application of TPB in the field of traffic safety and enrich our understanding of the psychological mechanisms underlying driver takeover behavior. From a theoretical perspective, this study validates TPB’s predictive power in emerging human–machine interaction contexts and enhances its explanatory power by introducing individual characteristic moderators. The moderate-to-strong correlations between TPB constructs (ranging from 0.343 to 0.515) further support the theoretical coherence of applying TPB to automated driving contexts. From a practical standpoint, the results provide targeted guidance for improving the safety and user experience of automated driving systems, including the following: (1) fostering positive takeover attitudes through education and outreach; (2) developing differentiated training programs for drivers with varying experience levels to enhance takeover skills; (3) establishing social atmospheres and norms supportive of safe takeover; and (4) optimizing human–machine interface designs for drivers with different characteristics.

The reliability and validity analysis of the measurement instruments (with Cronbach’s α ranging from 0.794 to 0.838, AVE values from 0.587 to 0.709, and CR values from 0.876 to 0.886) provides a solid methodological foundation for future research in this domain. The measurement model developed in this study can be applied or adapted by future researchers to investigate driver behavior in various automated driving contexts. Future research should employ more diverse methodologies, such as longitudinal designs or experimental methods, to establish causal relationships between TPB constructs and takeover behavior; combine objective measurement methods like driving simulators or real-vehicle tests to more accurately assess drivers’ actual takeover performance; explore additional individual difference factors such as personality traits, risk perception, and cognitive abilities; and focus on specific automated driving systems and disengagement contexts to provide more precise research conclusions and practical guidance.

Expanding the theoretical framework by incorporating additional constructs such as trust in automation, mental workload, and situation awareness could further enhance model explanatory power beyond the current levels (48.7% for takeover intention and 58.3% for takeover behavior). Such theoretical extensions would address the remaining unexplained variance in the model. It should be noted that the present study’s findings are based on data collected from a single city in China, which may limit their generalizability to other populations or contexts. Additionally, as with most survey-based research, this study relies on self-reported measures and adopts a cross-sectional design; these factors may introduce response bias and preclude causal inference. Future research should employ more diverse and representative samples, as well as longitudinal or experimental designs, to further validate and extend these findings.

In summary, this study presents a theoretical foundation for examining driver takeover behavior following automated driving system disengagement and offers preliminary empirical support that may guide subsequent research and development aimed at improving automated driving safety and user experience. As automated driving technology continues to develop and proliferate, a deeper understanding of the psychological mechanisms of driver takeover behavior will be increasingly important for promoting harmonious interactions between humans and automated driving systems and ultimately improving traffic safety.