Understanding Electric Bikers’ Red-Light Running Behavior: Predictive Utility of Theory of Planned Behavior vs Prototype Willingness Model

To date, electric bikers’ (e-bikers’) red-light running (RLR) behavior is often viewed as one of the main contributors to e-bikerelated accidents, especially for traffic scenarios with high e-bike ridership. In this paper, we aim to understand e-bikers’ RLR behavior based on structural equation modeling. Specifically, the predictive utility of the theory of planned behavior (TPB), prototype willingness model (PWM), and their combined form, TPB-PWMmodel, is used to analyze e-bikers’ RLR behavior, and a comparison analysis is made. +e analyses of the three modeling approaches are based on the survey data collected from two online questionnaires, where more than 1,035 participant-completed questionnaires are received. +e main findings in this paper are as follows: (i) Both PWM and TPB-PWM models could work better in characterizing e-bikers’ RLR behavior than the TPB model. +e former two models explain more than 80% (81.3% and 81.4%, respectively) of the variance in e-bikers’ RLR behavior, which is remarkably higher than that of the TPB model (only 74.3%). (ii) It is also revealed that RLR willingness contributes more on influencing the RLR behavior than RLR intention, which implies that such behavior is dominated by social reactive decisionmaking rather than the reasoned one. (iii) Among the examined psychological factors, attitude, perceived behavioral control, past behavior, prototype perceptions (favorability and similarity), RLR intention, and RLR willingness were the crucial predictors of e-bikers’ RLR behavior. Our findings also support designing of more effective behavior-change interventions to better target e-bikers’ RLR behavior by considering the influence of the identified psychological factors.


Introduction
In recent decades, as a green, cost-effective, and easy-tocarry transport means, electric bikes (e-bikes) have been adopted and promoted in an increasing number of countries such as Switzerland, Norway, the Netherlands, and China [1][2][3][4]. However, the use of e-bikes has been causing increasing e-bike-involved traffic accidents [1,5,6]. For example, in China, around 56,200 traffic accidents were caused by e-bikers, resulting in 8,431 fatalities, 63,400 injuries, and direct property losses of 111 million-yuan (equivalent to 16.42 million dollars based on the 2017 average closing exchange rate) between 2013 and 2017 [7]. According to the statistics of accidents related to two-wheeled vehicles (e.g., e-bikes, regular bicycles, and e-scooters) in typical Chinese cities (including Beijing, Changchun, Ningbo, Foshan, and Weihai), e-bike-involved accidents accounted for 34.8% of the total accidents from July 2011 to June 2016; of the e-bikeinvolved accidents, the proportions of minor injuries, serious injuries, and fatalities to e-bikers were 70.0%, 12.6%, and 10.6%, respectively [8]. Some studies reported that the speed and weight of e-bikes were both higher than those of regular bicycles, thereby leading to more injuries and fatalities [9]. In addition, previous findings showed that traffic violating behaviors, especially red-light running (RLR) behavior at signalized intersections, partially contributed to e-bikers-related accidents and fatalities [10][11][12][13][14][15][16]. us, studies on e-bikers' RLR behavior are needed, especially for countries with high e-bike ownership and e-bike-involved accidents.
Previous studies mainly focused on the influence of the demographics and traffic environment on e-bikers' RLR behavior. Moreover, in the studies of psychological factors associated with violating behavior, e-biker-related studies are less than the pedestrian-related and driver-related studies. A recent study has employed a social cognitive theory, the theory of planned behavior (TPB), to recognize the psychological determinants of e-bikers' intention to violate a red light [17]. Nevertheless, some evidence presented that the utility of the TPB framework is insufficient in predicting the unconscious risk-taking behaviors [18][19][20], while the prototype willingness model (PWM) might be a superior social cognitive theory in predicting such behaviors than the TPB framework [21]. Herein, the main goal of our work is to investigate the predictive utility of TPB, PWM, and their combined form, TPB-PWM, in e-bikers' RLR behavior and explore the psychological mechanism behind this behavior.

E-Bikers' RLR Behavior.
So far, most existing evidence has revolved around the red-light infringement behavior of cyclists. Specifically, investigations have been performed on the influence of demographics [11][12][13][14][15], psychological factors [11][12][13], and other risk behaviors (e.g., unhelmeted riding, carrying passengers, using a phone, and listening to music) [14,16,22,23] on cyclists' RLR behavior. However, due to the differences in weight, speed, and acceleration of e-bikes and regular bicycles, there are some discrepancies in the riders' psychology. Some studies discovered that e-bikers have a stronger intention to violate a red light than regular cyclists [13,[24][25][26][27][28]. Among these studies, Yang et al. [27] proposed that the prominent contributors to RLR behavior of cyclists and e-bikers were individual features (i.e., gender and rider type), psychological factors (i.e., conformity tendency), and traffic factors (i.e., waiting position and traffic volume). Investigations have also been performed on the impact of infrastructure conditions on the RLR behavior of cyclists and e-bikers, such as sunshades [26], carriageway, and intersection types [28]. However, apart from the above studies, much less literature has been published specifically on e-bikers' RLR behavior. Recent research conducted by Yang et al. [17] identified the psychological factors (i.e., attitude, perceived behavioral control, moral norm, and selfidentify) which significantly influenced the e-bikers' RLR intention based on the TPB framework. However, their study focused on e-bikers' RLR intention, but not on their behavior. Also, it is relatively undesirable to employ the extended TPB framework they proposed to predict e-bikers' RLR intention.
us, other more efficient models were applied to obtain an in-depth investigation of e-bikers' RLR intention and behavior in our study.

eoretical Framework
1.2.1. TPB Framework. TPB has been the most prevalent framework used to identify the social cognitive constructs that influence traffic violations of drivers, pedestrians, and e-bikers. In reviews of drivers' violations, the TPB framework has successfully explained the violation intentions, including high-speeding [29,30], drink-driving [31][32][33], mobile phone use while driving [34,35], and seatbelt nonuse [36,37]. In terms of pedestrians' violations, the TPB framework has been employed to understand various pedestrian behavior intentions, such as walking while intoxicated [38], red-light infringement [39,40], and distracted crossing [41]. In addition to drivers' and pedestrians' violations, Yang et al. [17] also determined the utility of the TPB framework in explaining e-bikers' RLR intention.
As mentioned above, TPB has been supported as a framework for explaining various traffic violations. Nevertheless, it is still unclear whether these behaviors are totally decided by volition. Sheeran et al. [18] pointed out that the TPB framework ignores the spontaneous or heuristic processes, and its hypothesis that a particular behavior is reasoned and premeditated could not match real situations. Also, some studies indicated that the utility of the TPB framework is insufficient in predicting the unconscious risktaking behaviors [19,20]. For example, even though e-bikers have a negative attitude towards e-bikers' RLR behavior, they might also have the willingness to violate a red light when there is a traversable space-time gap in the motor vehicle traffic flows.
In this investigation, an extension of the TPB framework ( Figure 1) was applied. According to Ajzen's studies [42,43], RLR intention (RI) reflects an e-biker's readiness to violate a red light. Attitude (AT) refers to an e-biker's positive/favorable or negative/unfavorable perception of RLR behavior. Subjective norm (SN) represents an e-biker's perceived view of whether social referents (e.g., family members and friends) approve or disapprove of RLR behavior. Perceived behavioral control (PBC) reflects an e-biker's perceived view of their ability to violate a red light. In light of previous studies, past behavior was also incorporated into the original TPB framework [29,44,45]. Past behavior (PB) represents an e-biker's violating behavior or illegal riding habits in the past and can indicate his/her habit strength [43,44]. A range of previous research studies has discussed the explanatory power of the extension of the TPB framework with the addition of past behavior, suggesting that past behavior emerged as a key contributor to behavioral intentions and behaviors in particular correlation with perceived behavioral control [19,29,43,44]. In the present model framework, attitude, subjective norm, perceived behavioral control, and past behavior have direct effects on RLR intention and indirect effects on RLR behavior via RLR intention. Also, perceived behavioral control directly influences RLR intention and behavior while indirectly influences RLR intention and behavior through past behavior.

PWM Framework.
In fact, e-bikers' crossing behavior is a spontaneous risk-taking behavior that generally requires reactive decisions on how to act in response to changing traffic environmental demands (e.g., traffic signals, traffic volume, and other e-bikers' crossing behaviors) at signalized intersections. us, other decision-making models might be considered in governing RLR behavior. e PWM framework, which focuses on both social reactive decision-making and reasoned one, might contribute to a better understanding of RLR behavior than TPB framework, which is concentrated on reasoned decision-making only. e PWM framework has been successfully applied in examinations of drivers' and pedestrians' violations, such as drink-driving [46], high-speeding [47,48], texting while driving [48], and pedestrians' violations [21]. As far as we are concerned, no previous research studies have employed the PWM framework in the study on e-bikers' violations. In this work, the PWM framework was adapted from the study of Gibbons et al. [49] (Figure 2). Among the PWM-based variables, prototype perceptions refer to the images of e-bikers (e.g., age and gender) who engage in RLR behavior, consisting of prototype similarity (PS) and prototype favorability (PF) [21,49]. RLR willingness (RW) denotes the e-biker's willingness to violate a red light when such an opportunity is provided [21,49]. In the reasoned decision-making pathway, attitude, subjective norm, and past behavior exert direct effects on RLR intention and indirect effects on RLR behavior via RLR intention, while past behavior directly influences RLR behavior. In the social reactive decisionmaking pathway, attitude, subjective norm, prototype similarity, prototype favorability, and past behavior have direct effects on RLR willingness, and RLR willingness directly affects RLR intention and behavior.

Integrating TPB and PWM Frameworks.
A few studies have applied the integrative model which incorporates TPB and PWM frameworks (hereafter, referred to as "TPB-PWM") to investigate a particular behavior in other domains, such as health-protective/risk behavior [50] and organ donation communication [51]. In recent research, Demir et al. [21] compared the utility of TPB, PWM, and TPB-PWM frameworks in explaining the pedestrians' violations, among which PWM and TPB-PWM frameworks had a better predictive power than TPB framework. In light of the original PWM proposed by Gibbons et al. [49], the past behavior was also a crucial predictor and could improve the PWM's performance. However, the past behavior was not included in Demir et al.'s [21] study, which might have negatively influenced the representativeness of the analysis results.
us, past behavior was incorporated into three decision-making frameworks in this work. e TPB-PWM framework in this study refers to an extended PWM with the addition of perceived behavioral control, which has a direct effect on RLR intention and behavior ( Figure 3).

Paper Position and Contribution.
Little empirical evidence has been reported to support the utility of TPB and PWM frameworks in understanding e-bikers' RLR behavior. In this study, we compared the utility of TPB, PWM, and TPB-PWM frameworks in understanding such behavior, by using the TPB framework as a baseline comparator. is study aimed to test our hypothesis of e-bikers' RLR behavior as being more social reactive (using the PWM/TPB-PWM framework) rather than rational (using the TPB framework). Another aim of this work was to investigate the social cognitive determinants that influence e-bikers' RLR behavior.

2.1.
Procedures. An experiment was designed using selfreported questionnaires. In the survey, participants who often used e-bikes (on average at least once a day) were asked to answer a wide range of questions related to each variable of the proposed model frameworks in a hypothetical situation. e situation assumed was described as follows: "You are riding an e-bike to work/school or somewhere on time. When you reach a signalized intersection, the traffic light turns red. In this situation, please answer the following questions." Before the formal survey in this study, we performed a pilot survey among a small total number of 50 participants using online questionnaires. e purpose of this pilot survey was to ensure that the questionnaire items were easy to understand. Also, the pilot survey was helpful to evaluate the completion time of the survey. Cronbach's α correlation analysis and principal component analysis (PCA) were  Journal of Advanced Transportation employed to examine the reliability and validity of the items, and those items that did not pass the examinations were removed. e finalized questionnaire was designed based on the participants' feedback and the test results. e formal survey was carried out in Shanghai, China, from May to July 2018. e participants were recruited from universities, communities, shopping malls, and office districts. ey were required to complete two online questionnaires distributed by an online survey platform (http://www.wjx.cn) one month apart. To be specific, the participants were required to complete the first (Time 1) questionnaire, including all items of variables, except for RLR behavior (Table 1). One month later, the participants should complete the second (Time 2) questionnaire to measure the subsequent RLR behavior. Finally, the participants were rewarded with an amount of about $1 for completing the survey.

Participants.
In this survey, a total number of 1,147 participants completed the Time 1 and Time 2 questionnaires. After excluding 112 invalid questionnaires (e.g., inconsistent scoring logic, unusually short completion time, and abnormal score), we obtained a complete and valid sample of N � 1,035 participants. e mean age of the final sample of participants was 34.6 (SD � 9.92), 68.0% of which were female (N � 704); 89.2% used e-bikes at least twice a day on average (N � 923), and 74.3% had more than two years of e-bike riding experience (N � 769).

Measures.
A 29-item questionnaire was designed to obtain data of participants for this work. e TPB-based items for each variable (i.e., attitude, subjective norm, perceived behavioral control, and RLR intention) and PWM-based items for each variable (i.e., past behavior, prototype perceptions, and RLR willingness) are presented in Table 1. Each item was measured using a 7-point response format. For the variables, we adopted the confirmatory factors analysis (CFA) and Cronbach's α correlation analysis to examine the validity and internal consistency of the items, thereby evaluating the reliability of the survey data. CFA was conducted by PCA. PCA identified that the value of      (Table 1). Apart from these items, participants' demographic data (e.g., gender, age, use frequency, and years of riding experience) were also collected.  Table 2 that the participants regarded themselves slightly dissimilar to the prototypical RLR e-bikers (prototype similarity, mean � 3.80) and rated the prototypical RLR e-bikers as moderately negative (prototype favorability, mean � 3.74). Additionally, the participants had a weak intention to violate a red light (RLR intention, mean � 4.06) but a relatively strong willingness to violate a red light whenever an opportunity was provided (RLR willingness, mean � 4.60). Finally, the participants reported that they violated a red light to a moderate extent during one month after completing the Time 1 questionnaire (RLR behavior, mean � 4.20). Pearson's bivariate correlations were calculated to evaluate the association between each variable (Table 2). e correlations established suggested that all study variables were significantly associated with each other. In particular, RLR intention, RLR willingness, and RLR behavior positively and significantly correlated with attitude, perceived behavioral control, past behavior, prototype similarity, and prototype favorability, while negatively and significantly with subjective norm. RLR behavior was positively and significantly related to RLR intention and willingness. e correlation analysis results support the efficacy of TPB, PWM, and TPB-PWM frameworks in explaining e-bikers' RLR behavior.

Results of the Structural Equation Modeling
. In this section, we have presented the process and results of the path analysis conducted using structural equation modeling in Amos 24. e path models for e-bikers' RLR behavior was evaluated based on TPB, PWM, and TPB-PWM frameworks. Degree-of-fit of the proposed model frameworks was evaluated and modified by multiple indexes, including chisquare divided by degrees of freedom (CMIN/df ), rootmean-square error of approximation (RMSEA), goodness-of-fit index (GFI), normal fit index (NFI), incremental fit index (IFI), comparative fit index (CFI), and Tucker-Lewis index (TLI) [52,53]. Subsequently, we employed the path analysis to estimate the significance of the direct, indirect, and total effects of the variables in each model and assess the predictive utility of the proposed models. e path analysis results of each model are presented below. e degree-of-fit indexes for three models are summarized in Table 3, while  Table 4 displays the direct, indirect, and total effects of the independent variables on RLR intention, willingness, and behavior in three models. Table 3 showed that the CMIN/df was more than 3 in the initial TPB model, indicating the degreeof-fit of this model to our data was inadequate (CMIN/ df � 3.191, RMESA � 0.046, GFI � 0.953, NFI � 0.969, IFI � 0.979, CFI � 0.979, and TLI � 0.975). Based on the suggested modification indices in Amos 24, the initial model path was revised by adding a path from the subjective norm to past behavior. e modified TPB model fitted well to the data (CMIN/df � 2.925, RMESA � 0.043, GFI � 0.958, NFI � 0.972, IFI � 0.981, CFI � 0.981, and TLI � 0.978). e path analysis results revealed that the attitude, subjective norm, and perceived behavioral control were the crucial determiners of RLR intention. RLR behavior was also significantly predicted by RLR intention, perceived behavioral control, and past behavior. e modified TPB model explained 80.4% of the variance of RLR intention and 73.6% of RLR behavior (Figure 4).

PWM Model.
e initial PWM model did not obtain a good fit to our data due to the value of more than 3 for the CMIN/df (CMIN/df � 3.246, RMESA � 0.047, GFI � 0.940, NFI � 0.965, IFI � 0.975, CFI � 0.975, TLI � 0.971). us, the initial model path was improved based on the suggested modification indices. We removed the path from subjective norm to RLR willingness and added the paths from prototype favorability and similarity to RLR intention, as well as the path from prototype favorability to RLR behavior. e modified PWM model obtained adequate degree-of-fit to the data (CMIN/df � 2.627, RMESA � 0.040, GFI � 0.951, NFI � 0.972, IFI � 0.982, CFI � 0.982, and TLI � 0.979) and explained 76.6%, 77.4%, and 81.3% of the variance of RLR intention, willingness, and behavior, respectively. It is noteworthy that RLR willingness had a greater impact on RLR behavior than RLR intention (β � 0.395 vs β � 0.173) ( Figure 5).  (Figure 6).     Tables 3  and 4, the modified TPB, PWM, and TPB-PWM models provided adequate degree-of-fit to the data, but the modified PWM and TPB-PWM models had a more considerable utility in predicting e-bikers' RLR behavior than the modified TPB model. at is, the PWM and TPB-PWM frameworks provided a more complete account of e-bikers' RLR behavior than the TPB framework.    [38][39][40][41]. However, these studies focused on predicting pedestrians' behavioral intention but not on their behavior. Demir et al. [21] confirmed the path from intention to behavior for pedestrians' violations. Currently, insufficient evidence exists on the intention-behavior path for e-bikers' violations. All the variables from the TPB framework had a statistically significant total effect on RLR intention and behavior. Of these variables, perceived behavioral control had the highest total effect (the sum of the direct and indirect effects) on RLR intention, which can be supported across a few studies on pedestrians' violation intention and drinkdriving intention [41,45]. Also, the total effect of perceived behavioral control on e-bikers' RLR behavior was the strongest. In accordance with earlier related findings on pedestrians' violations [21], perceived behavioral control also played an important role in predicting pedestrians' violating behavior. However, in contrast to the results obtained by Demir et al. [21], we established that the subjective norm was statistically significantly associated with RLR intention and behavior. is result was in line with previous research studies which revealed that subjective norm contributed a moderate effect on pedestrians' violation intentions [39,40,54]. Among the extended TPB framework, past behavior as an extended variable contributed a weak effect on RLR intention, which is inconsistent with the related studies on the behavioral intentions of risky riding [30] and drink-driving [31,45]. However, past behavior was found to be a second-crucial determiner of RLR behavior and added an additional 7.2% to the explained variance for RLR behavior compared to the original TPB model. is result was in good agreement with a previous study on drivers' speeding behavior [55], which reported that the additional past behavior could enhance the explained variance for speeding behavior by an additional 4%. Our findings highlighted the fact that past behavior, with the incorporation into the original TPB framework, improved the predictive power. To some extent, e-bikers' past behavior partially represented their behavioral habits, while the habitual, familiar behavior has been recognized by Ajzen [42] that plays an important role in predicting behavior.

Predictive Utility of PWM and TPB-PWM.
Compared to the modified TPB framework, the modified PWM and TPB-PWM frameworks explained an additional 7.7% and 7.8% of the variance in RLR behavior. e modified TPB-PWM framework also explained an additional 1.6% of the variance in RLR intention. Interestingly, our findings showed that RLR intention and willingness were crucial predictors of RLR behavior, and the total effect of willingness on RLR behavior was prominently stronger than that of intention. Elliott et al. [47] and Demir et al. [21]  ey also revealed that behavioral willingness contributed more substantially to violating behavior than behavioral intention. us, we concluded that e-bikers might decide on whether to violate a red light through the social reactive pathway rather than the reasoned pathway. is result was also consistent with the conclusions of Gibbons et al. [49], which underlined that the social reactive decision-making was more suitable for predicting risk-taking behavior. Accordingly, as expected, the results revealed that social reactive decision-making is more crucial than reasoned decision-making in determining e-bikers' RLR behavior.
is finding is paralleled to the dynamic nature of the riding task, which suggests that social reactive decision-making needs to respond to changing traffic environmental factors. e model estimation results revealed that the attitude still significantly influenced RLR intention as the TPB model, and furthermore, had a significant total effect on RLR behavior via a mediator variable "RLR willingness." However, the influence of subjective norm on RLR intention and behavior was weakened, and it was no longer a significant variable in the TPB-PWM model. A plausible reason is that the approval of one's family members or friends on RLR behavior has little influence on his/her RLR intention and behavior, which is supported by the results of several investigations on violation intentions of drivers, e-bikers, and pedestrians [17,21,29,32,40]. For instance, Zhou et al. [40] suggested that the behavior of pedestrian's family members or friends exerts a stronger influence on the pedestrian's intention to violate crossing than their approval of such behavior.
Prototype perceptions (favorability and similarity) played a significant role in predicting RLR intention, willingness, and behavior, except for a special case; that is, the effect of prototype favorability on RLR intention was not significant in the TPB-PWM model, which was caused by the addition of perceived behavioral control. ese results revealed that e-bikers who perceived people of their age and gender that regularly violate a red light as favorable and similar had a higher intention/willingness to commit RLR behavior. Although the original PWM framework has no relationship between prototype perceptions and intention, recent studies on pedestrian violations also found that prototype perceptions significantly contributed to intention [47]. More specifically, in the modified PWM model, prototype perceptions had a more significant influence on RLR intention than RLR willingness. Since perceived behavioral control was integrated into the modified TPB-PWM model and significantly influenced RLR intention, prototype perceptions contributed to RLR willingness more considerably.
Perceived behavioral control, as one of the TPB-based variables, still exerted a strong influence on RLR intention, whereas it had a weak effect on RLR behavior since it did not directly affect RLR behavior. In the modified PWM, past behavior was always a crucial predictor of RLR intention, willingness, and behavior, as also established by Gibbons et al. [20,49]. In the modified TPB-PWM model, although the effect of past behavior on RLR intention was no more significant due to the addition of perceived behavioral control, past behavior still substantially contributed to RLR willingness and behavior. ese findings revealed that e-bikers' perceived ability to violate a red light (perceived behavioral control) was more likely to influence behavioral intention, whereas behavioral habits or experiences (past behavior) had a greater impact on their behavioral willingness. In summary, e-bikers' behavioral habits or experiences substantially influence RLR behavior through social reactive and reasoned pathways. erefore, bias and deviations in the final results might have been caused in Demir et al.'s [21] study without considering the impact of past behavior.

Implications of Safety Interventions.
e outputs obtained from this study could support the application of safety interventions and thereby further discourage the e-bikers' RLR behavior. e attitude, perceived behavioral control, past behavior, prototype perceptions, RLR intention, and RLR willingness represent particularly good intervention targets since they were the crucial predictors of e-bikers' RLR behavior.
Among these identified psychological factors, the perceived behavioral control was a crucial contributor to e-bikers' RLR intention. is result revealed that if e-bikers perceived the difficulty of violating a red light, they had less intention to commit it. us, reducing e-bikers' perceived control might decrease their RLR behavior. For example, radio-frequency identification (RFID) can be integrated into e-bike license plates, combined with traffic cameras, to automatically monitor and capture e-bikes' RLR behavior at signalized intersections. Meanwhile, learning from the management system of drivers' red-light infringement behavior, more stringent penalties (e.g., higher fines and demerit point system) can be employed to limit the e-bikers' RLR behavior. e RLR willingness was especially crucial for RLR behavior. e aforementioned countermeasures would also reduce the opportunity for RLR behavior, thereby weakening the willingness to violate a red light.
Since attitude, prototype perceptions, and past behavior were found to be pronounced predictors of both intention and willingness to engage in RLR behavior, preventative safety interventions are needed to focus on addressing this issue. erefore, future advertisements and education related to traffic safety could consider creating the images of negative and undesirable e-bikers who violate a red light and enhance the negative and unfavorable perception of the most typical violators. For instance, riding simulation could allow e-bikers to experience traffic crashes or near-crashes as the results of their RLR behavior at signalized intersections, which might make e-bikers directly perceive the negative attributes of RLR behavior. It is expected that identifying violating e-bikers with negative portrayal would reduce e-bikers' RLR behavior. is expectation is supported by the findings of Demir et al. [56], which indicated that identifying speeders with negative portrayal on campus could reduce the incidence of speeding behavior. Also, school-based education can be an important approach to changing or correcting individual attitudes and prototype perceptions, especially correcting younger people's attitudes and prototype perceptions [40]. In essence, past behavior can indicate individual behavioral habits [42]. erefore, the proposed advertisements and education, riding simulation, and automated traffic rule enforcement might help e-bikers develop good riding habits. Furthermore, if e-bikers are required to participant in e-bike riding training and pass some riding tests, their illegal riding habits might be corrected.

Limitations and Future Research.
is study employed the social cognitive theories, including the TPB, PWM, and TPB-PWM frameworks, in understanding e-bikers' RLR behavior. Our work extended upon previous studies in e-bikers' RLR behavior which was investigated by using the TPB frameworks. Furthermore, the application of the PWM and TPB-PWM frameworks allowed for an in-depth understanding of the psychological factors influencing e-bikers' RLR behavior. Despite this, our work still had some limitations that need attention and discussion. First, the study sample included e-bikers that were predominantly female and with a large age variance. Previous studies showed that men or young people are more prone to take the risk when driving or riding; also their likelihood to get involved in traffic accidents is greater [11-15, 57, 58]. Hence, the gender and age composition of the study sample might have influenced the results of the proposed model frameworks. More comprehensive studies need to be conducted to elucidate the effects of gender and age differences in the future. Second, the findings of this study might have been affected by the bias inherent in self-report data. Specifically, such bias might have led to an underestimation of past behavior and subsequent RLR behavior due to social desirability and recall bias. Although this work was conducted by the use of voluntary and anonymous questionnaires to reduce social desirability bias, recall bias was still an issue that needed to be addressed. ird, one aim of this study is to predict potential RLR behavior, which requires high accuracy for measuring subsequent RLR behavior, whereas we measured this variable only by self-reports. erefore, the absence of an objective measurement of subsequent RLR behavior is also a major limitation of this work. In future studies, we can use some objective measurements such as an e-bike data acquisition system to obtain the frequency of subsequent RLR behavior in a given period (one month or more) after completing the Time 1 questionnaire. e fourth limitation is that this study assumes that the use of e-bikes of participants is an equal amount, whereas it varies significantly between participants based on the collected metric of exposure (e.g., use frequency and years of riding experience). As the exposure might be correlated to the psychological variables, the study results would be altered. erefore, the influence of the exposure factors on the psychological variables and behavior should be considered in future work.

Conclusions
e findings of our research suggest that the TPB and PWM are the promising theoretical frameworks for predicting e-bikers' intention or willingness to engage in RLR behavior, as well as for understanding the effects of psychological variables on RLR behavior. Furthermore, the PWM and TPB-PWM frameworks were found to be superior to the TPB framework in predicting e-bikers' RLR behavior. e present findings validate our hypothesis that e-bikers' RLR behavior is governed by both reasoned and social reactive decision-making, and the influence of the latter is stronger. In the PWM model, attitude, past behavior, and prototype similarity were the crucial predictors of both RLR intention and willingness. Prototype favorability had a relatively weak influence on RLR intention and willingness while exerted a pronounced direct effect on RLR behavior. In the TPB-PWM model, with the addition of perceived behavioral control, the impact of past behavior on RLR intention was reduced by the substantial influence of perceived behavioral control on RLR intention, but past behavior was still a crucial factor in predicting RLR behavior.
ese results imply the significance of specifically changing the e-bikers' attitude, perceived behavioral control, past behavior, and prototype perceptions towards RLR behavior. us, the design of more effective behavior-change interventions is needed, such as advertisements and education, training programs (e.g., riding simulation, riding training, and riding tests), and automated traffic rule enforcement.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper.

Authors' Contributions
T. Tang, H. Wang, and X. Zhou were responsible for study conception and design. H. Gong and H. Wang were involved in data collection. T. Tang, H. Wang, and H. Gong carried out analysis and interpretation of results. T. Tang and X. Zhou prepared the draft manuscript. All authors reviewed the results and approved the final version of the manuscript.