Investigating Risky Behaviors and Safety Countermeasures for E-Bike Riders in China: A Traffic Conflict Analysis Approach

Abstract

In recent years, e-bikes have rapidly gained popularity in China. However, riders frequently engage in aberrant behaviors, posing significant traffic safety concerns. Field observation combined with traffic conflict techniques offer an effective approach for identifying risky riding behaviors that significantly affect traffic safety. This study aims to address two major limitations in existing research that can lead to estimation biases: the unsystematic and incomplete inclusion of potential risky riding behaviors, and the insufficient consideration of unobserved heterogeneity in conflict data. Data on 437 e-bike–motor vehicle conflicts were collected at four signalized intersections in Hefei, covering 21 variables including illegal, negligent, and error-prone riding behaviors, as well as sociodemographic factors. Appropriate conflict risk indicators were selected for straight-line and angle conflicts, respectively. A random parameters binary logit model with heterogeneity in means and variances (RPBL-HMV) was developed and compared against binary logistic and mixed logit models. The results indicate that the RPBL-HMV model provides a significantly better goodness-of-fit than the other two models. Six factors with fixed parameters are positively associated with high-risk conflicts, while two factors exhibit random parameters—one of which decreases in mean when riders fail to slow down before turning. The identified risky behaviors and the corresponding targeted countermeasures offer practical insights for regulating unsafe e-bike riding and improving intersection safety.

1. Introduction

Electric bicycles (e-bikes) are lightweight, two-wheeled vehicles powered fully or partially by rechargeable batteries [1]. Unlike conventional bikes, e-bikes provide electric motor assistance, which reduces pedaling effort while enhancing riding efficiency and comfort. Additional advantages include flexibility, convenience, affordability, and environmental friendliness. Given the increasing severity of urban traffic congestion and environmental pollution, e-bikes have consequently become an essential mode of daily commuting for urban residents worldwide [2]. In 2024, the global e-bike market was valued at approximately US$35 billion and is projected to reach US$62.25 billion by 2030, with the Asia-Pacific region accounting for the largest market share, approximately 63% [3].

Over the past decade, the rapid expansion of China’s sharing economy and food delivery industry has driven the emergence of new business models and industries centered on e-bikes. For example, shared e-bikes have become a popular option for short-distance urban travel, while e-bike delivery has evolved into a critical logistics method for instant retail. As a result, both the demand for and production of e-bikes in China have experienced substantial growth. By 2025, the total number of e-bikes in use in China reached 380 million, representing a 72.7% increase from approximately 220 million units in 2015 [4]. Supporting this trend, domestic production of e-bikes in the first quarter of 2025 totaled roughly 11 million units, marking a 25% year-on-year increase compared to the same period in 2024 [5].

Despite offering tremendous convenience for daily commuting, e-bikes have given rise to increasingly serious road traffic safety issues in China due to their poor stability, high speeds, lack of reliable safety protection devices, and the fact that riders are not required to hold a driver’s license or undergo safety training—combined with generally weak riding skills and low safety awareness [6]. According to a study by the Songguo Think Tank [7], the research arm of Songguo Travel, a leading shared e-bike company in China, e-bikes were involved in up to 75% of crashes involving non-motorized vehicles. In 2023, e-bike riders were responsible for 35,566 crashes in China, resulting in 4698 deaths, 41,009 injuries, and direct property losses totaling 71.21 million Chinese yuan (approximately 9.8 million US dollars) [8]. Compared with 2018, these figures increased by 50.0%, 39.4%, 50.4%, and 39.9%, respectively.

E-bike crashes primarily involve collisions with motor vehicles, accounting for 74% of incidents [9]. In these crashes, e-bike riders sustain severe injury, disability, and fatality rates of 37.5%, 17.5%, and 12.0%, respectively [10]. These high casualty rates stem from the substantial differences in mass and speed between e-bikes and motor vehicles, compared with collisions involving fixed objects, other e-bikes, conventional bicycles, or pedestrians [11].

A large number of studies have demonstrated that aberrant riding behavior is a key determinant of both the occurrence and severity of e-bike crashes [12,13,14]. The most prevalent aberrant riding behaviors include running red lights, riding in motor vehicle lanes, traveling against the flow of traffic, speeding, and illegally carrying passengers [15,16,17,18]. Identifying aberrant riding behaviors that are genuinely risky—that is, those that significantly contribute to safety-critical events between e-bikes and motor vehicles (e.g., crashes or conflicts)—has therefore become a major concern for both traffic safety researchers and policymakers in recent years.

Based on data collection approaches, research on e-bike riders’ risky behaviors can be categorized into four main approaches: self-report questionnaire surveys, crash data analysis, naturalistic cycling studies, and field observation studies [12,14,19]. Among these, questionnaire surveys are the most widely adopted approach due to their straightforward implementation and low cost. Typically, researchers gather information on e-bike riders’ riding behaviors, safety attitudes, risk perceptions, demographic information, and traffic safety outcomes, such as experiences of crashes or traffic conflicts. Structural equation modeling and discrete outcome models are then employed to identify significant factors affecting traffic safety [20,21,22]. Despite their advantages, questionnaire surveys rely heavily on self-reported data, whose validity may be compromised by recall bias and social desirability effects, potentially resulting in biased estimates of risk factors [18,23].

Another prevalent methodology is crash data analysis. This approach involves collecting road crash data over extended periods and applying statistical and machine learning models to examine how riders’ pre-crash behaviors influence crash severity. Nevertheless, crash data—predominantly derived from police records—have several limitations, including underreporting (as property-damage-only crashes are often unreported), incomplete and unreliable pre-crash behavior data (usually based on e-bike riders’ self-reports), and the lengthy time required to accumulate sufficient crash data [24,25].

The limitations of the two aforementioned methods can be addressed through naturalistic cycling studies and field observation studies. The former equips e-bikes with cameras and sensors to record riders’ natural behaviors, whereas the latter employs unmanned aerial vehicles and roadside camera systems for the same purpose [26,27]. Statistical and machine learning methods are then applied to identify risky riding behaviors associated with critical events, such as traffic conflicts or hard braking incidents [28]. Both approaches feature large sample sizes, short data collection cycles, and high ecological validity [29,30]. However, naturalistic cycling studies are often constrained by the limited field of view of on-bike cameras, which typically capture only forward-facing scenery. This constraint impedes the accurate calculation of traffic conflict indicators, making conflict identification heavily dependent on the subjective judgment of video annotators [31,32]. In contrast, field observation studies using unmanned aerial vehicles can accurately capture the motion trajectories of all interacting traffic participants from video recordings. Such data facilitate the precise computation of objective conflict indicators, thereby offering substantial advantages for accurate quantitative analysis and interpretation [33,34].

Research on e-bike riders’ risky riding behaviors based on field observations has made notable progress [35,36,37,38]. However, two critical issues remain unresolved in this area. First, the selection of potential risky riding behaviors (i.e., aberrant riding behaviors) in existing studies is often subjective and fragmented, lacking a comprehensive and systematic integration of illegal, non-compliant, negligent, and error-prone behaviors into the analytical framework. This shortcoming may lead to the omission or bias of significant risky riding behaviors. Second, although various models—such as binary logit, multinomial logit, and mixed logit—have been employed for data analysis, previous studies have not accounted for potentially more complex forms of heterogeneity, such as heterogeneity in the means and variances of random parameters.

To address these two issues, this study first conducts a comprehensive review of e-bike-related laws and regulations in China and Anhui Province. It then systematically identifies aberrant riding behaviors of e-bike riders and employs unmanned aerial vehicles and roadside cameras to collect data on both these behaviors and e-bike–motor vehicle traffic conflicts at four typical signalized intersections in Hefei, China. Subsequently, appropriate conflict risk indicators are selected according to the distinct characteristics of straight-line and angle conflicts, enabling the classification of traffic conflicts into high- and low-risk categories. Finally, three statistical models—a binary logistic model, a mixed binary logit model, and a random parameters binary logit model with heterogeneity in means and variances (RPBL-HMV model)—are developed. The optimal model is determined based on goodness-of-fit tests, and it is employed to quantitatively analyze the effects of significant e-bike riding behaviors on conflict risk. Based on the results, targeted traffic safety improvement measures are proposed.

The remainder of this paper is organized as follows. Section 2 reviews previous studies on risky e-bike riding behaviors. Section 3 details the data collection process for aberrant e-bike riding behaviors and the kinematic characteristics of e-bike–motor vehicle traffic conflicts. In Section 4, conflict risk indicators are determined and the statistical modeling approach is described. Section 5 compares the performance of different statistical models. Section 6 discusses significant risky riding behaviors and proposes targeted safety measures, while Section 7 concludes the paper.

2. Literature Review

2.1. Self-Report Questionnaire Survey

Depending on the research objectives, studies on risky e-bike riding behavior based on questionnaire surveys primarily fall into two categories. The first category investigates the influence of riding behaviors and socio-demographic factors on the occurrence of safety-critical events (crashes or conflicts) to identify significant predictors [39]. The second focuses on the mechanisms through which socio-demographic, psychological, and cognitive- or capability-related factors affect aberrant riding behaviors [12]. Socio-demographic factors are included in both categories because they help precisely identify high-risk rider groups, thereby facilitating the development of more targeted safety interventions for risky riding behaviors.

Studies in the first category primarily collect data on daily riding behaviors, demographic characteristics, and experiences of safety critical events via questionnaires. After conducting reliability and validity tests, discrete outcome models are typically employed for analysis. For example, Haustein and Møller surveyed 685 e-bike riders in Denmark to examine factors influencing involvement in safety-critical events [40]. Binary logistic regression revealed that individuals who frequently rode at higher speeds and accelerations, and those reporting greater excitement about riding e-bikes, were more likely to engage in unsafe riding behaviors, thereby increasing the likelihood of such events. Fyhri et al. applied the same modeling approach to compare risk factors between e-bikes and conventional bikes [41]. They found that e-bike use significantly increased crash probability relative to conventional bike use, primarily due to higher risks among female e-bike riders relative to female conventional bike riders and male riders of both types. However, other studies have suggested that male riders are more likely to be involved in crashes than females [11,42]. This inconsistency may be attributed to a lack of consideration of interactions between gender and other observed or unobserved factors in these studies. Zhang et al., Qian et al., and Nayar et al. incorporated a broader range of aberrant riding behaviors into their models [16,17,20]. Their results indicated that behaviors such as running red lights, drinking and riding, carrying children or adults as passengers, riding in motor vehicle lanes, riding against traffic, frequently braking abruptly to avoid obstacles, and having a history of crashes significantly increased the probability of e-bike crashes.

In the second category of studies, after completing reliability and validity testing of the questionnaires, structural equation modeling (SEM) is typically employed to explore the causal relationships among demographic factors, key psychological factors (e.g., safety attitudes and risk perception), other psychological factors (e.g., riding confidence and personality traits), cognitive- and capability-related factors (e.g., safety knowledge, educational level, safety skills, and perceptual–motor skills), and aberrant riding behaviors [6,43,44]. Discrete outcome models are subsequently used to quantitatively assess the influence of significant factors on aberrant riding behaviors [45,46]. Wang et al. and Liu & Chen specifically examined psychological factors influencing mobile phone use while riding e-bikes, identifying past phone use, mobile phone addiction, and attitudes toward phone use as the strongest predictors [45,47]. More recent studies have extensively investigated factors influencing helmet-wearing behavior. For instance, Gebru et al., Yang et al. and Guo et al. identified perceived safety, perceived behavioral control, attitudes, penalties, subjective norms, safety awareness, and helmet service levels as key determinants of helmet use [46,48,49]. Beyond these single behaviors, numerous scholars have also applied the same theoretical framework to simultaneously examine factors influencing multiple types of aberrant riding behaviors—such as aggression, negligence and errors, rule violations, and leading behaviors [6,18,43,50,51].

The theoretical framework underlying questionnaire surveys supports an in-depth analysis of risky riding behaviors and their associated motivational factors. However, data obtained from questionnaires largely rely on self-reports, the accuracy of which is susceptible to social desirability and recall bias [31]. Social desirability is the tendency of respondents to modify their answers to align with social norms, public expectations, or moral standards, rather than to truthfully reflect their actual attitudes or behaviors [52]. Recall bias refers to systematic errors in data that arise when research relies on respondents’ recollection of past events or behaviors (e.g., aberrant riding behaviors or traffic conflicts), as respondents may provide inaccurate or selectively remembered information [23]. These inaccuracies can introduce bias into the estimated relationships among variables derived from questionnaire-based methods [40].

2.2. Crash Data Analysis

Research on risky riding behaviors of e-bikes based on crash data analyses typically involve collecting and processing road traffic crash reports over extended periods from departments such as traffic police, hospitals, and transportation bureaus [25,53]. Data extracted from these reports are then used to develop discrete outcome models or machine learning models that examine the relationship between riding behaviors, socio-demographic factors, and the severity or occurrence of e-bike crashes [54,55,56]. Discrete outcome models typically perform quantitative analyses using metrics such as odds ratios, pseudo-elasticities, and marginal effects, whereas machine learning models employ techniques such as SHAP (Shapley Additive Explanations), ALE (Accumulated Local Effects), and PFI (Permutation Feature Importance) for quantitative interpretation [57].

Hu et al. collected 205 injury records from e-bike and conventional bike crashes in Hefei, obtained from the People’s Liberation Army No. 105 Hospital [58]. Using a binary logistic regression model, they found that traffic rule violations and rider age above 45 were significantly associated with higher crash severity. Wang et al. examined factors influencing crash severity through a generalized ordered logit model based on 3814 e-bike–motor vehicle crash records from the Guilin Municipal Public Security Bureau Traffic Accident Processing Center [59]. Their results indicated that riders aged 55 years and older were more likely to sustain severe or fatal injuries compared with younger riders. Moreover, crash severity increased when e-bike riders ran red lights, turned left or right, or crossed mid-block. Zhong et al. conducted a similar analysis, reporting that behaviors such as chasing and playing, answering phone calls while riding, continuing to ride after e-bike malfunctions, not wearing helmets, and riding against traffic significantly increased injury severity among e-bike riders [60]. Wang et al. analyzed 1760 e-bike crashes over a three-year period in Hunan Province [61]. Riders were initially classified using the Chi-squared Automatic Interaction Detection (CHAID) method, followed by binary logistic regression to identify key risk factors influencing crash severity across different rider groups. The results indicated that the absence of improper behaviors decreased the likelihood of severe or fatal injuries by 3.4%. For non-professional riders under 55 years old, failing to dismount and push the e-bike while crossing the road was associated with higher crash severity. Wei et al. found that LightGBM outperformed logistic regression, KNN, SVM, DT, and XGBoost models in predicting the severity of e-bike–motor vehicle crashes [62]. SHAP analysis revealed that male riders and those with lower saddle heights were more likely to experience greater injury severity.

In addition to examining how e-bike riding behaviors influence crash severity, several studies have explored how such behaviors relate to rider characteristics—such as individual risk, crash responsibility, and behavioral traits—using crash data. For instance, Wang et al. applied quasi-induced exposure theory to classify riders into high-risk and low-risk groups based on violation records and crash histories [56]. They developed four tree-based machine learning models—Gradient Boosting Decision Tree (GBDT), XGBoost, AdaBoost, and Random Forest (RF)—to predict these risk categories, finding that GBDT achieved the highest predictive accuracy. The results also showed that riders with more frequent traffic violations (e.g., running red lights, aggressive driving, speeding, and drunk driving) were more likely to belong to the high-risk group, whereas those with accumulated fines and demerit points were more often classified as low-risk. Wang et al. analyzed 412 fatal e-bike crashes in Taixing City, China, from 2012 to 2014 using a mixed logit model to identify factors influencing fault allocation [1]. Their findings indicated that riders crossing at mid-block locations were significantly more likely to be solely at fault, and those carrying passengers were more likely to bear full responsibility. Guo et al. collected data on 276 e-bike crashes from the Ningbo Police Department and surveyed 862 non-crash riders via questionnaires and telephone interviews [63]. Employing a bivariate binary probit model, they jointly analyzed factors influencing crash occurrence and license plate usage, revealing that gender, age, aggressive riding, and law compliance were significantly associated with both outcomes.

Despite the valuable progress made in the aforementioned studies, two critical issues warrant further attention. First, official bike crash reports—covering both conventional bikes and e-bikes—suffer from substantial underreporting. A COST 1101 survey conducted across 17 EU member states found that, on average, only about 10% of bike crashes are reported to the police, with Germany having the highest reporting rate at merely 35% [25]. Reporting rates for minor bike injuries may be even lower; according to the European Road Safety Observatory, they were only 21% in the UK and 4% in the Netherlands [64]. Second, although crash records accurately capture vehicle, road, and environmental factors, they lack completeness and accuracy regarding detailed pre-crash cycling behaviors [24,65]. This limitation largely arises from the reliance on cyclists’ self-reports for behavioral information in crash records, which are often incomplete or biased due to riders’ concerns about legal liability [56,61]. Consequently, the reliability of analyses concerning risky e-bike riding behaviors based solely on crash records remains questionable [41,66,67].

2.3. Naturalistic Cycling Study

Naturalistic cycling studies involve a certain number of e-bike riders cycling under real-world conditions. Cameras, GPS/speed sensors, inertial measurement units, and wheel brake pressure sensors are installed on the e-bikes to collect data on riding behaviors and traffic conflicts. These data are then analyzed to identify factors affecting traffic safety [68,69,70].

In a study by Schleinitz et al., 90 participants rode their daily commuter bikes—either conventional bicycles, pedelecs, or S-pedelecs—for four weeks [26]. Each bike was equipped with two cameras and a speed sensor: one camera recorded the road ahead, and the other captured the rider’s upper body. Among 8000 encounters with red lights, 16.3% involved red-light running, with no significant difference observed among riders of the three bike types. Dozza et al. collected two weeks of riding data from 12 Swedish e-bike riders [29]. In addition to cameras, GPS, and inertial measurement units, an electronic button mounted on the handlebars of the e-bikes allowed riders to time-stamp critical events (crashes and conflicts) that occurred during their rides. By combining interviews, button presses, and video analyses, 88 critical events and 176 randomly selected baseline events were identified. It was found that approaching intersections or encountering parked motor vehicles in bike lanes increase the likelihood of critical events. Moreover, comparison with conventional bike data showed that e-bikes exhibited higher average speeds, consistent with the findings of Huertas-Leyva et al. and Petzoldt et al. [19,31,32]. In contrast to the video acquisition methods employed in the aforementioned studies, Wang et al. analyzed 1201 e-bike–motor vehicle crash videos obtained from in-vehicle dashcams [53]. Using a mixed logit model, they found that rider violations and the use of sunshades were both associated with greater injury severity.

Data collected from naturalistic cycling studies are characterized by large sample sizes, short data collection periods, and high ecological validity, effectively addressing the limitations of self-reported questionnaire surveys and crash data analyses [19,68]. However, constrained by the forward-facing field of view of e-bike cameras, such studies rely on video annotators or riders to subjectively identify traffic conflicts. This approach reduces analytical precision and precludes the classification of conflict risk levels, thereby hindering accurate and quantitative assessments of risky riding behaviors [26,29].

2.4. Field Observation Study

Unlike naturalistic cycling studies, field observation studies typically employ hidden roadside cameras to monitor e-bike riders’ behaviors and use drones or cameras mounted on elevated roadside structures to capture the motion data of all traffic participants [33,34]. As a result, these studies can accurately compute various traffic conflict indicators, facilitating precise identification and classification of conflicts [28,36].

Numerous field observation studies have examined red- and yellow-light running behaviors among e-bike riders at signalized intersections, employing various regression models such as binary logistic, PCA-based binary logistic, binary mixed logit, and multinomial mixed logit models. The findings indicate that factors including riders’ visual search during the waiting period, waiting beyond the stop line, higher average e-bike speeds, riding alone, fewer waiting e-bikes, and the presence of red-light–running e-bikes significantly increase the likelihood of red-light violations [27,35,71,72,73,74]. For yellow-light violations, critical predictors include the e-bike’s approach speed to the intersection at the onset of the yellow phase, the number of acceleration rate changes, and the distance to the stop line [34,75].

Beyond these studies, Zhang et al. classified e-bike–heavy vehicle conflict risks into three levels using traffic conflict indicators [33]. Results from a multinomial logit model showed that aggressive e-bike behaviors—such as rushing or sudden acceleration—significantly shortened heavy-vehicle drivers’ reaction times and increased the likelihood of high-risk conflicts. Similarly, Miao et al. applied a random forest model with SMOTE sampling to analyze factors influencing traffic conflict risk on shared slow lanes [28]. Their results indicated that illegal lane occupation (e.g., stopping mid-lane, weaving) and counterflow riding by e-bikes were strongly associated with higher probabilities of high-risk conflicts.

Despite their valuable contributions, previous field observation studies exhibit two major limitations. First, the selection of aberrant riding behaviors in existing studies is often subjective and fragmented, lacking systematic integration. This shortcoming may lead to the omission or bias of significant e-bike risky riding behaviors identified in the studies. Therefore, it is essential to incorporate illegal, non-compliant, negligent, and error-prone e-bike riding behaviors into the analytical framework in a more systematic and comprehensive manner. Second, although various statistical and machine learning models have been applied, insufficient consideration of unobserved heterogeneity in traffic conflict data has constrained model fit and predictive accuracy, highlighting the need for further methodological refinement.

3. Data Collection and Processing

To obtain data of aberrant riding behaviors and e-bike-motor-vehicle conflicts, field observations with roadside cameras and unmanned aerial vehicle were employed in this study [24,33,34,76].

3.1. Site Selection

The observation sites were selected in Hefei, the capital of Anhui Province in eastern China, which has a permanent population exceeding 10 million. Intersections were chosen for field research since they serve as hubs and nodes within urban road networks, and are focal points for traffic conflicts and crashes. Site selection was based on the following criteria [27,36]: First, the intersections had to be four-legged and signal-controlled. Second, they needed to be representative of the geometric characteristics, traffic facilities, and traffic conditions commonly observed in Hefei. Third, sufficient volumes of both e-bikes and motor vehicles had to be present to ensure adequate collection of conflict data.

Based on these criteria, four signal-controlled intersections were selected as research sites: Wuhu Road–Tongcheng Road, Huancheng Road–Tongcheng Road, Lujiang Road–Tongcheng Road, and Hongxing Road–Tongcheng Road. All the four intersections are located in the downtown area. The durations of red lights at the four intersections range from 25 to 107 s, green lights from 22 to 56 s, and yellow lights are uniformly set at 3 s. To ensure site diversity, some intersection approaches and departures featured physical separation facilities for motor vehicles and non-motor vehicles, while others accommodated mixed traffic. The characteristics of the selected sites are summarized in

Table 1. Characteristics of the observed intersections.

Site	Intersection	Signal Type	Number of Lanes	Number of Signal Phases	Cycle Length (s)
1	Wuhu Rd. & Tongcheng Rd.	Countdown	6 & 3	3	151
2	Huancheng Rd. & Tongcheng Rd.	Countdown	2 & 3	2	66
3	Lujiang Rd. & Tongcheng Rd.	Countdown	2 & 3	2	70
4	Hongxing Rd. & Tongcheng Rd.	Countdown	2 & 3	2	72

.

3.2. Data Collection

The data collection area within each intersection was defined as the zone enclosed by the inner edge of the crosswalk, its extension lines, and the curb [77], as illustrated in Figure 1. A DJI Inspire 2 drone equipped with a Zenmuse X4S gimbal camera was deployed to record the movement trajectories of e-bikes and motor vehicles. The drone hovered approximately 60 m above ground level, capturing 4K video footage (see Figure 1). In addition, four concealed cameras were installed at the corners of each intersection—behind trees, within green belts, or hidden by parked cars—to unobtrusively observe e-bike riders’ behaviors without influencing their natural riding patterns [74], as shown in Figure 2. Before recording, all cameras were synchronized with the drone to ensure temporal consistency across video sources. Field observations were conducted under clear weather conditions during the evening rush hour, when commuter and delivery e-bike traffic volumes were both relatively high. Video data were collected between 17:00 and 18:30 on one weekday in May 2021 for each intersection, yielding a total of six hours of footage. It should be noted that the evening rush hours were selected for data collection because most rare categories within the independent variables—such as “rider is a courier,” “red-light running,” and “speeding”—exhibit higher occurrence rates during peak periods than during off-peak times. Sufficient sample sizes for these rare categories help mitigate Type II errors and enhance the statistical model’s ability to identify genuine effects [78].

The video analysis and modeling software Tracker was used to process footage captured by the drone, as shown in Figure 3. First, a Cartesian coordinate system was established based on a reference point within the video frame. The length and width of pedestrian crosswalks were then used for calibration to create a mapping relationship between the video coordinate system and the real-world coordinate system. Finally, target e-bikes and motor vehicles were tracked manually or automatically at a frame rate of 30 frames per second. This process yielded motion-state data for road users, including trajectories, velocities, and acceleration/deceleration profiles.

A comprehensive and systematic collection of aberrant behaviors among e-bike riders is essential for yielding unbiased and exhaustive analysis results regarding risky behaviors. In this study, aberrant e-bike riding behaviors are classified into four categories: illegal, non-compliant, negligent, and error-prone behaviors. The first two categories were identified through an in-depth review of the Road Traffic Safety Law of the People’s Republic of China, the Implementing Regulations of the Road Traffic Safety Law, and the Anhui Province E-Bike Management Regulations. The latter two categories, representing risky behaviors arising from insufficient safety awareness or overconfidence, were adapted from previous research [42,43,63]. The 21 explanatory variables, encompassing aberrant e-bike riding behaviors and riders’ sociodemographic information, are presented in

Table 2. Explanatory variables and their coding values.

Category	Variable	Coding Value
Sociodemographic information	Gender	Male: 0; female: 1
	Courier a	No: 0; Yes: 1
	E-bike type	Bicycle-style: 0; Scooter-style: 1
Illegal, non-compliant behavior	Riding in the wrong direction	No: 0; Yes: 1
	Occupying motor vehicle lanes
	Running red light
	Running yellow light
	Speeding b
	E-bike turning without yielding to straight-through motor vehicles
	Failure to turn left from the right side of the intersection’s center
	Stop beyond the stopping line
	Failure to slow down before turning
	Reckless riding c
	Using a mobile phone while riding
	Not wearing a helmet
	Failure to display an e-bike license plate
	Weaving through parked non-motorized vehicles ahead
	Assembled, modified, or retrofitted e-bike
Negligent and error-prone behavior	Talking to the passenger or other riders while riding	No: 0; Yes: 1
	Entering intersections without slowing down
	Failing to maintain a safe lateral distance d

Note: ^a. Whether a rider is a delivery courier is determined by the presence of delivery platform logos on their uniform or helmet. ^b. Speeding is defined as exceeding 15 km per hour. ^c. Reckless riding includes towing or being towed by non-motorized vehicles, riding with hands off the handlebars, holding objects in hands while riding, riding side-by-side while leaning on each other, chasing one another, or weaving during riding. ^d. Failing to maintain a safe lateral distance refers to riding an e-bike with a lateral distance of less than 1.5 m from motor vehicles.

.

It should be noted that other abnormal riding behaviors specified in the aforementioned laws and regulations—such as riders under 16 years of age, passengers under 12 or over 16 years of age, and overloaded e-bikes—were difficult to identify from the video footage and were therefore excluded from this study. Moreover, since field observations were conducted under clear weather conditions with good visibility, the behavior of “failing to turn on lights during low visibility” was also not considered.

4. Methodology

4.1. Selection of Conflict Risk Indicators

Traffic conflict risk indicators should be continuous measures capable of distinguishing normal traffic interactions from conflicts and classifying the latter into different risk levels. Specifically, indicator selection should consider two analytical perspectives. First, the indicators should be capable of capturing risk in scenarios both with and without a collision course [79]. Second, they should be able to assess the risks faced by road users both before they reach the conflict point and after one has passed it [80].

For angle candidate traffic events (e.g., sideswipe and side-impact), before road users reach the conflict point, T₂ is an ideal indicator because it can measure risk in scenarios both with and without a collision course [81,82]. T₂ represents the time it would take for the second road user to reach the potential collision point, assuming unchanged speeds and intended trajectories [79]. Once one road user has passed the collision point, PET (Post-Encroachment Time) is widely recognized as an effective measure for evaluating conflict risk [67,83]. PET is defined as the time interval between the departure of one road user and the arrival of another at a potential collision point or area [84].

Regarding straight-line candidate traffic events (rear-end, head-on), both rear-end and head-on events are on collision courses, making the consideration of crossing course scenarios unnecessary. In addition, the conflict point in straight-line events is mobile in nature, meaning that both road users would never reach it unless an actual crash occurs. As a result, only the scenario before the two road users reach the conflict point needs to be analyzed. Extensive studies have demonstrated the effectiveness of TTC (Time-to-Collision) in such situations [2,85,86,87,88].

In summary, this study adopts TTC as the conflict risk indicator for straight-line candidate traffic events. For angle candidate traffic events, T₂ and PET are simultaneously used as risk indicators. To determine the thresholds for these indicators, a rigorous approach would involve examining (i) the correlation between the number of motor vehicle–e-bike conflicts of a given type (e.g., straight-line or angle conflicts, including both low- and high-risk conflicts) and the number of motor vehicle–e-bike crashes of the same type, and (ii) the correlation between the number of high-risk motor vehicle–e-bike conflicts of that type and the number of motor vehicle–e-bike crashes of the same type, across four intersections under varying threshold values of the risk indicators. The optimal thresholds for distinguishing normal traffic interactions, low-risk conflicts, and high-risk conflicts are identified when the highest correlation coefficients are achieved, respectively [89,90]. However, due to confidentiality restrictions imposed by the Hefei Traffic Police Department on traffic accident data, historical records of e-bike–motor vehicle accidents at the four intersections were not available.

To the best of the authors’ knowledge, no previous studies on e-bike–motor vehicle conflicts have applied the aforementioned methodology to determine indicator thresholds. Therefore, this study draws on existing research on TTC and PET thresholds from the widely applied Swedish Traffic Conflict Technique (Swedish TCT) and the Dutch Objective TCT for Operation and Research (DOCTOR) [66,67,91], as well as other well-recognized studies [92,93]. Specifically, low and high-risk conflicts are identified when 1.5 s < TTC ≤ 3 s, TTC ≤ 1.5 s, respectively. The PET thresholds for identifying low- and high-risk conflicts are 3 s and 1 s, respectively. Given the similarity in their definitions, thresholds of T₂ are referenced from TTC for events classification. It should be noted that, for angle candidate traffic events, if T₂ and PET yield inconsistent risk levels, the higher risk level is adopted. The aforementioned studies comprehensively considered motor vehicle–motor vehicle, motor vehicle–bicycle, and motor vehicle–pedestrian conflicts when determining the threshold values of the indicators. Although these values may not be optimal for e-bike–motor vehicle conflicts, they exhibit reasonable general applicability.

4.2. Random Parameters Binary Logit Model with Heterogeneity in Means and Variances

After dividing traffic conflict risks into high-risk and low-risk categories, it is essential to select an appropriate analytical model to examine the relationship between explanatory variables and risk levels to identify significant risky behaviors. Previous studies investigating risky e-bike riding behaviors based on field observations have primarily employed binary logit, multinomial logit, or mixed logit models [27,33,72,74]. However, these approaches often fail to adequately capture the complex heterogeneity inherent in conflict data. To address this limitation, the present study employs a RPBL-HMV model for the analysis [94,95,96].

For the i-th e-bike–motor vehicle traffic conflict, let p_i denote the probability that the conflict is high-risk, and 1 − p_i the probability that it is low-risk. The conventional binary logistic model can be represented as follows:

$$y ~ Binomial(p_i,n)$$

(1)

$$logit (p_i)=\log ({\displaystyle \frac{p_i}{1-p_i}})={\beta }_0+{\displaystyle \sum _{k=1}^K{\beta }_k{x}_{ik}}$$

(2)

where ${\beta }_0$ is the model constant. and ${\beta }_k$ represents the parameter estimate corresponding to the k-th (k = 1, 2, …, K) explanatory variable $x_{ik}$.

From Equation (2), it follows that:

$$p_i={\displaystyle \frac{EXP(\mathit{\beta }{\mathit{X}}_i)}{1+EXP(\mathit{\beta }{\mathit{X}}_i)}}$$

(3)

The conventional binary logistic model does not account for unobserved heterogeneity in conflict data. The mixed binary logit model overcomes this limitation by allowing some or all parameter estimates to vary across observations. Accordingly, Equation (3) can be rewritten as follows:

$$p_i={\displaystyle \int \frac{EXP(\mathit{\beta }{\mathit{X}}_i)}{1+EXP(\mathit{\beta }{\mathit{X}}_i)}f(\mathit{\beta }\left|\mathit{\phi }\right.)}d\mathit{\beta }$$

(4)

where $f(\mathit{\beta }\left|\mathit{\phi }\right.)$ is the density function of $\mathit{\beta }$, and $\mathit{\phi }$ denotes the vector of parameters characterizing this density function.

The mixed binary logit model assumes that the means and variances of the random parameters are fixed. To further account for unobserved heterogeneity, a RPBL-HMV model is developed in this study, which is defined as:

$${\beta }_k=\beta +{\mathit{\theta }}_k{\mathit{Z}}_k+{\delta }_{k}EXP({\mathit{\omega }}_k{\mathit{W}}_k){\xi }_k$$

(5)

where $\beta$ is the mean parameter estimate across all conflicts; ${\mathit{Z}}_k$ is a vector of explanatory variables that capture heterogeneity in the mean; ${\mathit{\theta }}_k$ denotes the vector of estimable parameters associated with ${\mathit{Z}}_k$; ${\mathit{W}}_k$ is a vector of explanatory variables that captures heterogeneity in the standard deviation ${\delta }_k$ with corresponding parameter vector ${\mathit{\omega }}_k$, and ${\xi }_k$ represents a disturbance term. It should be noted that the RPBL-HMV model reduces to a mixed logit model when ${\mathit{\theta }}_k$ and ${\mathit{\omega }}_k$ are not significantly different from zero, and further reduces to a binary logistic model when ${\delta }_k$ is set to zero.

Since the RPBL-HMV model incorporates random parameters, the effects of explanatory variables on risk vary across observations [97]. Therefore, the marginal effect is employed to measure the influence of a one-unit change in an explanatory variable on the dependent variable. For indicator (dummy) variables, the calculation is given as follows [96]:

$$\overline{M_k}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N[P(x_{ik}=1)-P(x_{ik}=0)]}$$

(6)

where $\overline{M_k}$ represents the marginal effect of the k-th explanatory variable, and N denotes the number of conflicts.

5. Results

5.1. Descriptive Statistics of Conflict Data

Over six hours of field investigation, a total of 437 traffic conflicts between e-bikes and motor vehicles were observed. According to the conflict risk classification criteria outlined in Section 4.1, 162 incidents (37.1%) were classified as low-risk conflicts, whereas 275 incidents (62.9%) were classified as high-risk conflicts.

Table 3. Summary statistics of explanatory variables.

Category	Variable	Coding Value	Count	Percentage	Mean	Std. Dev
Sociodemographic information	Gender (X1)	0	132	30.21%	0.30	0.460
		1	305	69.79%
	Courier (X2)	0	381	87.19%	0.13	0.335
		1	56	12.81%
	E-bike type (X3)	0	99	22.65%	0.77	0.419
		1	338	77.35%
Illegal, non-compliant behavior	Riding in the wrong direction (X4)	0	350	80.09%	0.20	0.100
		1	87	19.91%
	Occupying motor vehicle lane (X5)	0	340	77.80%	0.22	0.416
		1	97	22.20%
	Running red light (X6)	0	315	72.08%	0.28	0.449
		1	122	27.92%
	Running yellow light (X7)	0	418	95.65%	0.04	0.204
		1	19	4.35%
	Speeding (X8)	0	197	45.10%	0.55	0.498
		1	240	54.90%
	E-bike turning without yielding to straight-through motor vehicles (X9)	0	349	79.86%	0.20	0.401
		1	88	20.14%
	Failure to turn left from the right side of the intersection’s center (X10)	0	349	79.86%	0.20	0.401
		1	88	20.14%
	Stop beyond the stopping line (X11)	0	359	82.15%	0.18	0.383
		1	78	17.85%
	Failure to slow down before turning (X12)	0	372	85.13%	0.15	0.356
		1	65	14.87%
	Reckless riding (X13)	0	422	96.57%	0.03	0.182
		1	15	3.43%
	Using a mobile phone while riding (X14)	0	417	97.25%	0.04	0.188
		1	12	2.75%
	Not wearing a helmet (X15)	0	125	28.60%	0.71	0.452
		1	312	71.40%
	Failure to display an e-bike license plate	0	437	100%	0	0
		1	0	0%
	Weaving through parked non-motorized vehicles ahead	0	437	100%	0	0
		1	0	0%
	Assembled, modified, or retrofitted e-bike (X16)	0	302	69.11%	0.31	0.463
		1	135	30.89%
Negligent and error-prone behaviors	Talking to the passenger or other riders while riding (X17)	0	432	98.86%	0.01	0.106
		1	5	1.14%
	Entering intersections without slowing down (X18)	0	338	77.35%	0.23	0.419
		1	99	22.65%
	Failing to maintain a safe lateral distance (X19)	0	349	79.86%	0.20	0.401
		1	88	20.14%

summarizes the statistics for all explanatory variables.

As shown in Table 3, the values of “failure to display an e-bike license plate” and “weaving through parked non-motorized vehicles ahead” have values of zero across all observations. This lack of variation means these variables do not provide any discriminative information for the RPBL-HMV model. In such cases, the standard errors of their coefficient estimates tend toward infinity, causing convergence issues during model estimation [98]. Consequently, these two explanatory variables were excluded from the analysis prior to model estimation.

5.2. Estimation Results of RPBL-HMV Model

To eliminate multicollinearity among the independent variables, Spearman’s correlation test was first conducted to identify pairs of explanatory variables with absolute correlation coefficients exceeding 0.7 [99]. Next, the Variance Inflation Factor (VIF) was calculated for each variable, and the variable with the higher VIF was excluded from highly correlated pairs [51]. Finally, variables with a VIF greater than 5 were also removed to further mitigate multicollinearity [34].

Figure 4 illustrates the results of the Spearman correlation test. It can be observed that the absolute values of the correlation coefficients are all below 0.7, indicating that no pairs of variables exhibit strong correlation. Also, the maximum VIF among all explanatory variables is 1.53, suggesting no multicollinearity issues. Consequently, all 21 explanatory variables were retained for inclusion in the RPBL-HMV model.

Model estimation was carried out using simulation-based maximum likelihood estimation with 1000 Halton draws, implemented in the NLOGIT 5 statistical software package [100]. Furthermore, several distributional forms—including uniform, gamma, Weibull, triangular, and normal distributions—were examined for estimating the random parameters [95]. If the estimated standard deviation of a parameter is significantly different from zero at the 95% confidence level, the parameter is considered random; otherwise, it is regarded as fixed. The statistical results showed that the normal distribution provided the maximum Log-likelihood value and the minimum AIC value; therefore, it was adopted for the final model. Table 2 shows the estimated results of the RPBL-HMV model.

As shown in

Table 4. Parameter estimation results of the RPBL-HMV model.

Variable	Coefficient	T-Statistic	p-Value	Marginal Effect
Gender	−0.418 **	−2.25	0.025	−0.049
Courier	0.999 ***	2.83	<0.01	0.117
Running red light	0.457 **	2.19	0.029	0.053
Failure to slow down before turning	1.451 **	2.47	0.014	0.169
Entering intersections without slowing down (X18)	0.478 **	2.18	0.029	0.056
Failing to maintain a safe lateral distance (X19)	1.395 ***	3.84	<0.01	0.163
Random parameters
Occupying motor vehicle lane	0.867 ***	2.96	<0.01	0.101
Standard deviation	0.803 **	2.35	0.019	-
E-bike turning without yielding to straight-through motor vehicles	1.343 **	2.53	0.011	0.157
Standard deviation	2.553 ***	3.72	<0.01	-
Heterogeneity in the random parameter’s mean
Occupying motor vehicle lane: Failure to slow down before turning	−2.629 ***	−3.13	<0.01	-
Goodness-of-fit measures
Number of observations	437
Log-likelihood at constant	−288.128
Log-likelihood at convergence	−229.395
McFadden ${\rho }^2$	0.204
AIC	488.8
BIC	550.00

Note: **, and *** denote statistical significance at the 0.05, and 0.01 levels, respectively.

, there are eight significant explanatory variables. Among them, the regression coefficients of six explanatory variables are fixed, indicating that their effects on conflict risk do not vary across observations. The coefficients of two explanatory variables—“occupying motor vehicle lane” and “e-bike turning without yielding to straight-through motor vehicles”—exhibit randomness. Notably, the mean coefficient of “occupying motor vehicle lane” varies with “failure to slow down before turning”, indicating an interaction effect between these two variables. In addition, no variables were found to significantly explain variance heterogeneity.

5.3. Comparison of Models

To evaluate model performance, the goodness-of-fit of the RPBL-HMV model was first compared with that of the binary logistic model and the mixed binary logit model (see

Table 5. Goodness-of-fit of the competing models.

Model	Binary Logistic	Mixed Binary Logit	RPBL-HMV
Number of observations	437	437	437
Number of parameters	7	7	11
Log-likelihood at constant	−288.128	−288.128	−288.128
Log-likelihood at convergence	−241.349	−237.214	−229.395
McFadden ${\rho }^2$	0.162	0.177	0.204
AIC	496.7	496.4	488.8
BIC	525.3	541.3	550.00

). As shown in the table, the RPBL-HMV model achieves the highest Log-likelihood at convergence and McFadden’s ${\rho }^2$, as well as the lowest AIC among the three models, demonstrating superior goodness-of-fit. However, it is worth noting that the RPBL-HMV model also exhibits the highest BIC value, consistent with findings reported by Chen et al. and Waseem et al. [101,102]. This is primarily due to the greatest number of parameters included in the RPBL-HMV model, with BIC imposing a stronger penalty for model complexity than AIC. Nevertheless, the increased number of parameters allows the model to better capture the complex characteristics of the data, thereby enhancing interpretability. Overall, the RPBL-HMV model outperforms the other two models.

To further analyze whether the RPBL-HMV model significantly outperforms the other models, a likelihood ratio test was conducted, as shown in

Table 6. Results of the likelihood ratio test.

Likelihood Ratio Test	Binary Logistic vs. RPBL-HMV	Mixed Binary Logit Vs. RPBL-HMV
Degrees of freedom	4	4
Level of significance	0.05	0.05
Critical ${\chi }^2$	9.488	9.488
Computed ${\chi }^2$	23.91	15.64

.

It can be seen from Table 6 that, for both pairs of models, the computed χ² values are greater than the corresponding critical χ² values at the 0.05 significance level, indicating that the RPBL-HMV model performs significantly better than the other two models in terms of goodness-of-fit.

Regarding the classification performance of the three models, widely used evaluation indicators include accuracy, precision, recall, the F1-score, and the area under the receiver operating characteristic curve (AUC-ROC) [46]. Among these indicators, the F1-score has the advantage of jointly considering precision and recall in a balanced manner. The AUC-ROC evaluates a model’s overall classification performance across all possible classification thresholds, thereby allowing for effective handling of class imbalance [28]. Therefore, this study adopts the F1-score and AUC-ROC as evaluation indicators for model classification performance.

Table 7. F1-scores of the three models.

Indicator	Binary Logistic	Mixed Binary Logit	RPBL-HMV
F1-score	0.777	0.772	0.777

presents the F1-scores of the three models, while Figure 5 illustrates their corresponding ROC curves.

As shown in Table 7, the F1-score values of the RPBL-HMV model and the binary logistic model are identical and slightly higher than that of the RPBL model. This indicates that, when the classification threshold separating high-risk and low-risk conflicts is set to 0.5, both the RPBL-HMV and binary logistic models achieve better classification performance than the RPBL model. When evaluating the models’ ability to discriminate between positive and negative samples across all possible classification thresholds, the RPBL-HMV model achieves the highest AUC-ROC value among the three models (see Figure 5). These findings indicate that, while maintaining a comparable balance between precision and recall at the default threshold, the RPBL-HMV model offers the best overall discriminative performance under varying classification thresholds.

6. Discussions

6.1. Factors with Random Parameters

As shown in Table 4, the factor “occupying motor vehicle lane” produces random parameters. The normal distribution plot of this factor’s random parameter (see Figure 6) indicates that 86.12% of violators are associated with an increased probability of high-risk conflicts. This is because e-bikes are relatively small in size, making it difficult for drivers to detect them promptly when they enter motor vehicle lanes. The combination of drivers’ long reaction times and e-bikes’ limited braking capabilities significantly elevates the risk of traffic collisions. In the remaining 13.88% of cases, this risky behavior decreases the likelihood of high-risk conflicts, primarily because e-bike riders tend to compensate for safety by riding more cautiously in complex traffic conditions [100]. Overall, the marginal effect results indicate that occupying motor vehicle lanes increases the probability of high-risk conflicts by 10.1% compared with riding in non-motor vehicle lanes, which is consistent with the findings of Miao et al. and Ye [28,50].

Table 4 also reveals that the mean of the random parameter for this factor is heterogeneous and negatively correlated with “failure to slow down before turning”. That is, the latter factor weakens the effect of the former on high-risk conflicts. Although this finding may appear counterintuitive at first glance, a closer analysis provides an explanation: none of the intersections investigated in this study are equipped with dedicated signal phases for non-motorized vehicles, i.e., e-bikes follow the same turning phases as motor vehicles. When an e-bike occupies a motor vehicle lane and fails to slow down before turning, its speed becomes more similar to that of surrounding motor vehicles. This reduced speed differential in turn diminishes the positive effect of “occupying motor vehicle lane” on high-risk conflicts.

The key to improvement measures targeting “occupying motor vehicle lane” lies in establishing dedicated waiting areas and travel paths for non-motorized vehicles across the entire intersection area. Specifically, at intersection approaches, physical separation barriers (e.g., green belts, railings) are suggested to be added between motor vehicle lanes and bike lanes [36,60]. Regarding the interior of intersections, traffic channelization should be enhanced by applying conspicuous colored pavement (e.g., red or green) to non-motorized lanes, providing clear visual cues to both e-bike riders and motor vehicle drivers. These measures can effectively guide the movement of e-bikes, minimizing their interaction and interference with motor vehicles.

Another factor generating random parameter is “e-bike turning without yielding to straight-through motor vehicles.” As shown in Figure 7, this random parameter follows a normal distribution with a mean of 1.343 and a standard deviation of 2.553. Cases exhibiting positive and negative effects on high-risk conflicts account for 70.05% and 29.95%, respectively, and this factor increases the overall probability of high-risk conflicts by 15.7%. This result can be explained as follows: at intersections without dedicated left-turn phases, especially during peak hours, left-turning e-bike riders often take advantage of the gaps between straight-through vehicles to cross the intersection due to competitive tendencies, thereby increasing the likelihood of high-risk conflicts [1,103]. This unsafe behavior can be mitigated by enforcing a two-stage crossing strategy for left-turning e-bikes [104]. This strategy is expected to significantly reduce both the number of conflict points and the level of conflict risk between e-bikes and motor vehicles. Key traffic engineering measures include: (i) painting large counterclockwise directional arrows on non-motorized lanes within the intersection; (ii) installing “Left Turn: Two-Stage Crossing” signs upstream of the non-motorized stop lines; and (iii) establishing clearly marked two-stage crossing waiting areas for e-bikes at all four corners of the intersection.

6.2. Factors with Fixed Parameters

Table 4 indicates that the parameter for “gender” is fixed, with a marginal effect of −4.9% for females, suggesting that female e-bike riders are less likely to be involved in high-risk conflicts. This finding is consistent with those reported by Savitsky et al., Guo et al. and Wang et al. [6,11,63]. A primary explanation is that female riders generally exhibit higher levels of compliance with traffic regulations and adopt more cautious riding styles, whereas male riders tend to display greater risk tolerance and more adventurous riding tendencies [1,34]. Although some studies have reached the opposite conclusion—namely, that female e-bike riders are associated with higher risk than males [41,59]—it is important to recognize that these studies differ substantially from ours in terms of research methodology (e.g., reliance on questionnaire surveys) or outcome measures (e.g., focusing on crash severity rather than conflict risk). To address this issue, it is recommended to install electronic surveillance systems at intersections to capture illegal and non-compliant riding behaviors. In addition, personal traffic credit records should be established to systematically manage and penalize such violations.

Compared with non-professional e-bike riders, delivery couriers are more likely to be involved in high-risk conflicts. This elevated risk stems from the multitasking demands associated with courier riding—such as navigation, order confirmation, and communication—as well as aggressive riding behaviors driven by time pressure and overconfidence in riding skills [30,33,37]. Regulatory authorities are advised to encourage delivery companies to revise their performance evaluation and incentive systems, placing greater emphasis on “safe riding” than on on-time delivery rates. For couriers who frequently engage in risky behaviors, simulation-based experiential safety training is recommended to provide riders with a direct understanding of the hazards associated with such behaviors (e.g., loss of vehicle control, insufficient braking distance).

E-bike riders running red lights also represent a significant risky behavior, increasing the probability of high-risk conflicts by 5.3%. Observations at the four intersections in this study reveal that most red-light violations occur during the early and late stages of the red light phase, a finding consistent with those reported by Yang et al. and Wu et al. [35,71]. Red-light violations in the early stage may be attributed to some e-bike riders’ risk-taking and impulsive tendencies, as they attempt to cross the intersection quickly to save time. As waiting times increase, another group of riders may lose patience and eventually choose to violate the signal during the late stage of the red-light phase. Both scenarios make it difficult for motor vehicle drivers to react or brake in time, thereby increasing the likelihood of high-risk conflicts. Safety promotion measures can reference the discussion under “gender” and are therefore not further elaborated here.

“Entering intersections without slowing down” and “failure to slow down before turning” are two speed-related factors that both exhibit a significant positive association with higher conflict risk. As e-bike speed increases, riders have less time to anticipate and respond to dynamic changes in the traffic environment, thereby elevating the likelihood of conflicts [31,40]. Previous studies suggest that, from the perspectives of risk perception and safety attitudes, e-bike riders who are more accustomed to violating traffic regulations are more likely to engage in non-deceleration behavior, whereas riders with stronger safety awareness are less inclined to exhibit such speed-related risky actions [6,56]. Regarding countermeasures, it is important to note that China’s latest national standard stipulates that, starting from 1 September 2025, all newly manufactured e-bikes must be equipped with communication modules, and newly manufactured e-bikes for commercial use must also be equipped with BeiDou Navigation Satellite System modules. For newly manufactured e-bikes intended for ordinary household use, consumers may choose whether to retain the BeiDou module. It is recommended that the government guide all newly manufactured e-bikes to be equipped with BeiDou modules, enabling the transmission of real-time information such as location and speed to traffic safety management authorities via communication modules. In addition, facial recognition technology should be integrated into electronic surveillance systems at intersections. As e-bikes compliant with the new national standards become increasingly prevalent on the roads, traffic safety management authorities would be able to promptly contact e-bike riders when excessive speeds are detected at intersections, thereby enabling remote safety education and warnings.

The last significant factor is “failing to maintain a safe lateral distance”, which is positively associated with conflict risk. This is an interesting finding that has rarely been reported in previous studies. The underlying explanation is as follows: when an e-bike travels alongside a motor vehicle (e.g., when both are turning left or proceeding straight in the same direction), the driver may be blinded by the A- or B-pillars of the vehicle, preventing them from noticing the rider on the right. In such situations, if the e-bike fails to maintain a safe lateral distance, even slight directional deviations by either party can substantially elevate the risk level of conflict. This risk can be further amplified when both the motor vehicle and e-bike turn right simultaneously due to the inner wheel difference [33,54]. To mitigate this risk, it is recommended that, in addition to applying colored anti-skid pavement within intersections to highlight non-motorized vehicle paths, reflective flexible posts—which are more cost-effective and space-efficient compared with rigid separation facilities—be installed to help maintain safe lateral distances between motor vehicles and e-bikes during right-turn maneuvers.

7. Conclusions and Limitations

To address the limitations of existing traffic conflict-based studies on e-bike rider behavior—specifically, the incomplete inclusion of potential risky behaviors and the insufficient consideration of unobserved heterogeneity in conflict data—this study collected data on 21 explanatory variables, encompassing illegal, non-compliant, negligent, and error-prone e-bike riding behaviors, as well as sociodemographic information, from four typical signal-controlled intersections in Hefei City, Anhui Province. Appropriate conflict risk indicators were selected by analyzing the characteristics of straight-line and angle conflicts, respectively. To account for the unobserved heterogeneity in means and variances of random parameters, an RPBL-HMV model was developed and compared with binary logistic and mixed binary logit models. Finally, in-depth analyses and discussions of significant risky e-bike riding behaviors were conducted based on the optimal model. The main findings are as follows:

• The RPBL-HMV model demonstrates significantly better goodness-of-fit than both the binary logistic model and the mixed binary logit model.
• Six factors with fixed parameters are positively associated with the likelihood of high-risk conflicts, including male, courier, running red light, failure to slow down before turning, entering intersections without slowing down, and failing to maintain a safe lateral distance.
• “Failing to maintain a safe lateral distance” is a rarely reported but important factor. When an e-bike travels alongside a motor vehicle, it can easily fall into the driver’s blind spot. Under such conditions, even minor directional deviations by either party can substantially increase conflict risk.
• The factors “occupying motor vehicle lanes” and “e-bike turning without yielding to straight-through motor vehicles” generate random parameters, and their marginal effects indicate a positive influence on the probability of high-risk conflicts.
• “Failure to slow down before turning” decreases the mean of the random parameter for the factor “occupying motor vehicle lane”. This can be attributed to the fact that when e-bikes share the same turning phases as motor vehicles under the condition of occupying motor vehicle lanes, riders who do not decelerate reduce the speed differential between themselves and motor vehicles. This, in turn, lowers the likelihood of high-risk conflicts compared with cases where riders slow down before turning.
• No variables were found to significantly explain variance heterogeneity. This suggests that the impacts of the two factors with random parameters on high-risk conflicts vary among different e-bike riders. However, the degree of dispersion in these impacts remains relatively stable across different scenarios.
• The findings of this study offer important insights into the behavioral mechanisms underlying risky riding among e-bike users. Safety improvement recommendations are proposed from the perspectives of engineering design, traffic management, and behavioral guidance, offering practical implications for regulating unsafe e-bike behaviors and enhancing intersection safety.

Although the findings of this research demonstrate some insightful relationships between traffic conflict risk and risky riding behaviors of e-bike riders, several limitations still need to be addressed in future work. First, this study analyzed six hours of video footage from four intersections during the weekday evening peak period (5:00 PM–6:30 PM). It should be noted that the identified significant risk factors are applicable only to weekday evening rush hours. As time periods vary, these factors are not expected to remain stable, owing to changes in the exposure proportions of different categories within the independent variables. Future research is recommended to investigate how season, day type, and time of day influence cyclists’ risky riding behaviors. Second, since aberrant driving behaviors of motor vehicle drivers could not be observed, they were not included as explanatory variables. This limitation could be addressed by integrating in-vehicle driver monitoring systems with vehicle-to-infrastructure (V2I) technologies. Lastly, beyond the three models employed in this paper, other statistical models that account for unobserved heterogeneity—such as the random thresholds random parameters hierarchical ordered probit model and the correlated random parameters logit model—could be employed for comparative analysis to uncover additional insights. In addition, integrating machine learning methods, particularly ensemble learning algorithms, with statistical models to accurately characterize and quantitatively analyze unobserved heterogeneity (e.g., spatio-temporal correlations) warrants further exploration.