Determinants of switching behavior to wear helmets when riding e-bikes, a two-step SEM-ANFIS approach

1.   Introduction

E-bikes, electric-powered two-wheeled vehicles, have received considerable attention during the past decade due to environmental protection, labor-saving and low-cost advantages. In China, E-bikes have become one of the most popular travel modes ^[1], and the total number of e-bikes had been close to 300 million in 2021 ^[2]. However, accidents related to e-bikes took up a large proportion of all traffic accidents. For example, more than 50% of the total road traffic accidents were related to e-bike accidents in two Chinese provinces (Jiangsu and Zhejiang) in 2019 ^[3]. In accidents for e-bike riders, head injury is the leading cause of injury and death ^[4,5]. Previous research revealed that wearing helmets can effectively ensure the head safety of e-bike riders and reduce 51% of head injury risk ^[6]. A high helmet-wearing rate seems crucial to protect travel safety for e-bike riders ^[7].

The helmet-wearing rate is the aggregate reflection of e-bike riders' individual wearing behavior. The literature related to riders' intention and behavior for helmet use focuses on bicycles and motorcycles, which contributes to scientific helmet promotion plans for policymakers to improve helmet-wearing rates and protect riders' travel safety ^[8]. In China, e-bikes have high social recognition, and ownership continues to increase ^[9]. E-bikes have a different risk of accident fatalities than motorcycles and bicycles ^[10]. In addition, existing studies have found a discrepancy in riding violations between different riding modes. For instance, the results from zhou et al. ^[11] showed that e-bike riders are more likely to break the law than bicycle riders, and Truong et al. ^[12] found that motorcycle riders were more than twice as likely to use mobile phones while riding as e-bike riders. Riding without wearing a helmet belongs to violating traffic laws, especially in countries like China that have enacted helmet regulations. Different travel modes may lead to a discrepancy in helmet use behavior and influencing factors for riders ^[13]. Therefore, the current research results on bicycles and motorcycles' helmet-wearing behaviors may not be applicable in China for e-bikes. To our knowledge, only zhou et al. ^[14] analyzed helmet-wearing behavior from the perspective of objective variables, such as weather, travel time and travel purpose after the Chinese helmet regulation was enacted. Apart from that, Wang et al. ^[15] focused on the delivery riders' e-bike helmet use with the survey data in 2019, but the helmet regulation had not yet begun to spread at that time and the delivery riders' wearing behavior may not behave the same as regular riders for occupational safety. In general, there is a lot of research work on the wearing behavior of e-bike helmets to explore.

The Traffic Administration Bureau of China launched a nationwide safety campaign named "One Helmet, One Belt" on April 21, 2020, to increase helmet-wearing ^[16]. Currently, 31 Chinese provinces have issued more than 70 local regulations (as shown in Figure 1) to require residents to wear helmets when riding e-bikes. The legislation imposes mandatory constraints on people's behavior and is an effective way for the government to regulate public behavior ^[17], but the legislation might not be sufficient. For example, the helmet-wearing rate reached 93.39% after Qingyuan, one city in Guangdong Province, issued a helmet-wearing rule in November 2020 but quickly dropped to 56.40% after only three months ^[18]. The refined interventions are essential to cooperate with the enforcement activities of helmet regulations ^[19], which needs to identify influencing factors and explore further relationships between factors and helmet-wearing intention, such as nonlinear relationships ^[20]. Compared with linear models, such as linear regression and the Structural Equation Model(SEM)in existing studies for helmet-wearing behavior ^[14,21,22], the Adaptive Network-based Fuzzy Inference System (ANFIS) is a very powerful approach to dealing with the complex, nonlinear relationship between input and output factors ^[23] and rank the factors' importance ^[24]. At present, helmet regulations are being promoted nationwide in China, and the intention of switching from not wearing to wearing helmets when riding e-bikes involves hundreds of millions of people's daily travel safety. However, less research has focused on helmet-wearing behavior and the switching intention of e-bike riders in China. To fill in the gaps and formulate refined interventions to effectively improve the wearing rate of helmets, we proposed the following research questions (RQs) and tried to solve them.

RQ1: What factors can significantly influence the intention of switching to helmet-wearing?

RQ2: Among the influencing factors, which factors are more important?

RQ3: Under the combined effect of different important factors, will the intention of e-bike riders to wear helmets change?

This study aims to clarify the factors affecting helmet-wearing switching intention and analyze the factor's importance and relationship to provide targeted suggestions for helmet policy promotion. According to the research questions and purpose, we set three objectives: 1) Identifying the factors that affect the switching intention of helmet-wearing; 2) Ranking the factors' importance; 3) Providing targeted suggestions to improve helmet-wearing rates. This research introduces the Push-Pull-Mooring (PPM) framework into the field of helmet-wearing behavior for the first time and constructs a theoretical model of the helmet-wearing switching intention of e-bikes. The Partial Least Square Structural Equation Model (PLS-SEM) and the ANFIS were combined to reveal the possible linear and nonlinear relationship between the influencing factors and the dependent variable. A two-step SEM-ANFIS method was constructed to rank the importance of psychological variables while verifying the significance of path relationships. We obtained a more refined path relationship by analyzing the changes in helmet-wearing switching intention under the combined effect of different important factors. Overall, the PPM model we constructed is a theoretical framework for the e-bike riders' helmet-wearing behavior, and the two-step SEM-ANFIS method results could provide targeted suggestions for improving helmet-wearing rates.

The remainder is organized as follows: the next section provides a brief review of the literature on the construction of the PPM model and hypothesis. Section 3 describes the method of data collection, measurement structure and two-step SEM-ANFIS applied for the analysis. Section 4 presents the results of both PLS-SEM and ANFIS analysis. The theoretical and practical implications, limitations and further directions are given in Section 5. The final Section 6 consists of the conclusion of this research.

2.   Literature review and theoretical model

The PPM model originated in the field of human migration ^[25] and has been widely applied to explain switching behavior in many fields, such as environmental protection, healthcare services and online applications ^[26,27,28]. Based on the Theory of Planned Behavior (TPB) and Protection Motivation Theory (PMT), this study constructs a PPM model for helmet-wearing switching behavior of e-bikes by expanding the characteristic variables, such as perceived legal norms, habits and conformity tendency. The PPM model was deemed useful for understanding switching behavior because of its powerful explanation. We analyzed the influencing factors and path relationships of helmet-wearing switching intention from three perspectives: push, mooring and pull. The switching intention (SI) was defined as the behavior intention to switch from not wearing to wearing a helmet. The PPM model for helmet-wearing switching intention and hypotheses were constructed in Figure 2.

2.1.   Push effects

In the PPM model, the push effect refers to the positive factor that motivates people to leave the original place ^[29]. This study uses the perceived vulnerability and perceived severity in the PMT, and extended perceptual legal normative factors as the push factors, analyzing the positive impact of the disadvantages of not wearing a helmet on the switching intention.

Perceived vulnerability (PVU) refers to an e-bike rider's estimate of the likelihood of injury in a traffic accident. Perceived severity (PSE) refers to the severity of injury without a helmet. Generally, the perceived risk of injury and perceived injury severity directly affect an individual's safe behavior ^[30]. Due to concerns about accidents and serious injuries, e-bike riders may wear helmets to protect their heads. Therefore, we incorporated PVU and PSE into the push effects to examine their switching intention of helmet use.

Perceived legal norm (PLN) is the level at which road users perceive legal consequences for violating traffic rules ^[31]. Paschall et al. ^[32] found that legal norms correlated significantly with young people's intention to drink. Traffic police's punishment for non-helmeted cyclists may prompt them to switch to wearing helmets, so we expand PLN as one of the push factors.

Based on the analysis of PVU, PSE and PLN on the helmet-wearing intention of e-bike riders, we propose the following three hypotheses:

H1: Perceived Vulnerability has a positive impact on switching intention;

H2: Perceived Severity has a positive impact on switching intention;

H3: Perceived Legal Norm has a positive impact on switching intention.

2.2.   Mooring effects

Mooring effects contain negative factors that may hinder the switching intention ^[25]. This study applied habit (HAB) and perceived inconvenience (PI) as mooring factors in the PPM model. We analyzed the negative impact of the inconvenience caused by the helmet and the long-term habit of not wearing behavior on helmet-wearing switching intention.

Habit is an unintentional response that occurs when an individual engages in an activity over a long period ^[33]. Lai et al. ^[28] incorporated Habit into the PPM model and found that habit negatively impacts the switching intention of middle-aged and elderly patients to adopt cloud medical services. In this study, the habit was defined as the habit of riding an e-bike without a helmet.

Perceived inconvenience is defined as the extent to which an individual feels inconvenient and uncomfortable ^[30]. Barbarossa and Pelsmacker ^[34] used structural equation models to reveal that consumer perceived inconvenience is a negative and significant factor in green purchase intention. This research defined Perceived Inconvenience as the degree of inconvenience the riders feel caused by helmet-wearing and involved perceived inconvenience as a negative factor to explain e-bike riders' switching intention.

Based on the above application and results on habit and perceived inconvenience in behavior switching research, we propose the following hypotheses:

H4: Habit negatively influences switching intention;

H5: Perceived Inconvenience negatively influences switching intention.

2.3.   Pull effects

The pull effects are positive factors facilitating migrants to their destination ^[25]. Five psychological variables, namely conformity tendency (CT), response efficiency (RE), attitude (ATT), subjective norm (SN) and perceived behavior control (PBC), were included in the pull effects. We analyzed five psychological variables' pull effect on switching intention.

Conformity stands for the tendency to follow others' behavior ^[35]. Zhou et al. ^[36] found that conformity tendency is a significant factor influencing pedestrians' intention to cross the street. In addition, e-bike riders with a higher conformity tendency are more likely to violate traffic rules ^[37]. Whether or not other e-bike riders wear a helmet could impact the individual intention and behavior, so we apply conformity tendency as one of the pull factors influencing the switching intention.

Response efficiency refers to the perceived effectiveness of risk-protective behaviors and is a sub-constituent of the coping evaluation in PMT ^[38]. The study by Chamroonsawasdi et al. ^[39] found that response efficiency significantly and positively affected dietary activity to prevent diabetes. Since wearing a helmet protects the e-bike rider's head safety and reduces the chance of injury, we defined response efficiency as the perceived protective effectiveness of helmets when riding an e-bike and included it in the pull effects.

Subjective norm (SN), perceived behavior control (PBC) and Attitude (ATT) are all factors from the TPB, which is currently one of the most widely used models in behavioral research ^[40]. Previous research demonstrates that SN, PBC and ATT in TPB positively affect human intention ^[40]. In this study, subjective norm refers to the influence of relatives and friends on the individual's helmet-wearing switching intention, perceived behavior control refers to the individual's ability to wear a helmet, and attitude is defined as a positive attitude towards wearing a helmet.

The above five psychological variables were included in the pull effects to analyze their influence on helmet-wearing switching intention, and the following hypotheses were proposed:

H6: Conformity tendency positively impacts switching intention;

H7: Response efficiency positively impacts switching intention;

H8: Subjective norm positively impacts switching intention;

H9: Perceived behavior control positively impacts switching intention;

H10: Attitude positively impacts switching intention.

3.   Methods

According to the constructed theoretical PPM model, we designed the variable measurement scale and questionnaire. After obtaining the data through the survey, the mathematical analysis method and the two-step SEM-ANFIS approach were used to test the model and obtain the output value to rank the factors' importance. Moreover, we analyzed the nonlinear relationship and combined effects. The analysis process is shown in Figure 3.

3.1.   Measurement

This study designed a survey using a questionnaire to investigate helmet-wearing for e-bike riders. The questionnaire includes three sections. The first section introduces the research purpose and investigation background. The respondents' socio-economic characteristics and trip attributes were counted in the second section. The final section focuses on the impact of psychological factors on respondents' helmet-wearing intention. We adopted measurement items for eleven constructs from previously validated research, and these items were answered according to a seven-point Likert scale ranging from one strongly disagree to seven strongly agree. Table 1 shows the eleven constructs with items and their source.

3.2.   Data collection and sample

We conducted a pre-survey online and collected 112 questionnaires, and the formal questionnaire was modified and improved according to the pre-survey results. The formal survey period is from June 24 to June 25, 2021, and was carried out by investigators in shopping malls and other public places in Zhenjiang, Jiangsu. A total of 48 investigators were involved in the survey, and they were trained to select respondents randomly to ensure sampling randomness. We collected 387 questionnaires with the interviewee informed and excluded questionnaires with filling logical errors or choosing all the same options. Furthermore, we removed the respondents who wore helmets before the promulgation of the law because we aimed to switch intentions. Finally, 322 valid questionnaires constituted our final sample with an effective rate of 83.2%.

Table 2 shows the demographic distribution and travel characteristics of the respondents. The males were 50.93%, and the females were 49.07%, which was almost equal. The average age of the respondents was 33.7 years. The most occupation is service workers (34.79%), and more than 70% of respondents earned less than 6000 yuan per month. Although the samples in this study are younger and have low-level income compared to the statistics of Chinese people ^[47], they fit with the crowd characteristics of e-bike riders ^[48]. A total of 40.37% had ridden e-bikes for more than four days per week and 14.91% every day. However, most of the respondents (67.41%) wore helmets less than three times per week after enacted legislation. We investigated the Participants' experiences of being punished for not wearing a helmet and found that 19.88% were stopped and fined by police. Therefore, our sample could be considered representative of Chinese e-bike riders.

3.3.   SEM-ANFIS

Structural equation modeling (SEM) has been widely used to assess human behavior and examines the relationships among latent variables as a linear method ^[49]. Literature suggests two methods of SEM. First, Covariance-based SEM (CB-SEM) determines the relationships based on minimizing the difference between theoretical and estimated covariance matrices ^[50]. Second, Partial Least Squares-SEM (PLS-SEM) estimates the relationships by maximizing the explained variance of the endogenous variables ^[51]. Considering that PLS-SEM is better suited for examining exploratory models ^[52], requires no assumptions of homogeneity and normality in the data, and deals with small sample sizes, unlike CB-SEM ^[53], we use PLS-SEM to test the relationship in the PPM model.

The PLS-SEM approach is usually explained by the measurement model and the structural model. The measurement model exhibits the relationship of observed variables with their respective constructs, and the structural model exhibits the interrelationships between the constructs by determining the path coefficients. The measurement model is expressed by flowing reflective measurement (Eq (1)) and formative measurement (Eq (2)) ^[54]:

where, ${x}_{i}$ is observed variable; ${\pi }_{i}$ is the loading corresponding to the observed variable; $\xi$ is the latent variable related to observed variables; ${\varepsilon }_{i}$ is the error term; ${w}_{i}$ is the weight corresponding to observed variables; $\delta$ is the residual term.

The structural model is expressed as follows:

where ${\xi }_{j}$ is jth latent variable with ith of latent variables; $\beta$ is the latent variables' regression coefficient term; ${v}_{j}$ is an error term related to ${\xi }_{j}$.

Jang proposed an adaptive network based fuzzy inference system (ANFIS), also known as the adaptive neural fuzzy reasoning system, in 1993 ^[55]. ANFIS is a new type of fuzzy reasoning system structure that organically combines fuzzy logic and neuronal network, which is a multi-layer feedforward network using a hybrid algorithm of backpropagation algorithm and least squares method to adjust the premise parameters and conclusion parameters, and finally generates If-Then rules ^[55]. ANFIS model has been applied in empirical behavioral research to predict and reveal the influencing factors and their importance to the willingness to accept autonomous vehicles and the successful development of hotels ^[20,24]. Considering that the relationship between variables might be possibly nonlinear and the SEM focuses on the linear relationship mainly, we use the ANFIS to overcome the drawback.

For simplicity, a singleton Sugeno fuzzy model with two inputs, "x1" and "x2", and one output Y has the following If-Then rules:

Rule1:If "x₁" is "A₁" and "x₂" is "B₁" then "z₁ = p₁ x₁ + q1 x₂ + r₁"

Rule2:If "x₁" is "A₂" and "x₂" is "B₂" then "z₂ = p₂ x₁ + q₂ x₂ + r₂"

where "x₁" and "x₂" are non-fuzzy inputs, A_i (or B_i) are fuzzy sets, z_i is the output value; p_i, q_i and r_i are parameters of the output function.

The typical ANFIS model structure is shown in Figure 4, and each of the five layers for the ANFIS model is described below ^[56].

Layer 1: layer 1 is an adaptive unit with a function defined by:

where x and y are inputs, ${O}_{1, i}$ represents the output of node i in layer l, and ${A}_{i}$ and ${B}_{i-3}$ are membership functions for inputs.

Layer 2: Every node in layer 2 is a fixed node labeled Π, which multiplies the incoming inputs from layer 1 (Eq (2)):

Each node output in this layer is called the firing strength of a rule, which means the weight degree of the if-then rules.

Layer 3: Every node in this layer is a fixed node labeled N. The ratio of the ith rules' firing strength is calculated in the ith node as a sum of all rules' firing strengths and normalizes the firing strengths from layer 2.

Layer 4: The nodes in this layer are adaptive nodes comprising a linear combination of functions.

where $\stackrel{-}{{w}_{i}}$ is a normalized firing strength output from layer 3, and ${f}_{i}$ is the consequent parameter.

Layer 5: The single node in this layer is a fixed node labeled and overall output, summing the output of layer 4.

Based on the PPM theoretical model and path hypothesis, a mathematical analysis method of two-stag SEM-ANFIS was constructed to analyze the linear and nonlinear relationship in factors and dependents by using PLS-SEM and ANFIS. We used PLS-SEM to test and obtain significant factors affecting switching intention and included these significant factors in the ANFIS model for prediction as input variables. According to the importance ranking of the factors from the predicted value of ANFIS, we obtained a more refined path relationship by analyzing the changes in switching intention under the combined influence of different important factors in the three types of factors of push, mooring and pull.

4.   Results

4.1.   Data quality assessment

Data quality is a prerequisite for constructing models, including reliability tests and validity tests. Reliability was decided by Cronbach's Alpha and composite reliability (CR). The formula for Cronbach's alpha is:

where the ${\sigma }_{x}^{2}$ is the total test score variance, n is the number of variables' items, and ${\sigma }_{i}^{2}$ is the item variance.

The CR value reflects whether all items in each variable could explain the variable consistently, and the formula stands for the total amount of true score variance relative to the total score variance ^[57]:

where ${\lambda }_{i}$ is the ith indicator's standardized loading, and $\theta$ is of indicator's error item.

As shown in Table 3, both Cronbach's Alpha and CR are higher than the standard values of 0.7 and 0.6. Scale validity is determined by convergence validity and discriminant validity. Convergence validity is based on two conditions: item loading (λ) must exceed 0.7 ^[58], and the average variance extracted (AVE) for each construct must exceed 0.5 to indicate good convergence validity ^[59]. As shown in Table 4, the square root value of AVE for each variable is greater than the correlation value with other variables, indicating good discriminant validity ^[59]. The results of the data quality assessment showed that the data had good reliability and validity. The AVE is computed using the following formula:

where ${\lambda }_{i}$ is the ith indicator's standardized loading, and $\theta$ is of indicator's error item.

4.2.   PLS-SEM results

4.2.1.   Structural model quality

The PLS-SEM model was built using the SmartPLS Software (Version 3.0), The model R² and goodness of fit (GoF) were selected as the evaluation indicators of model quality, and the results are shown in Table 5. R² reflects the explanatory strength of the model, and the evaluation criteria are 0.19 (low), 0.33 (moderate) and 0.67 (high) ^[58]. The results show that the model constructed in this institute has a reasonable explanatory strength, reaching 44.2%. GoF is used to test the validity of model prediction, the evaluation criteria are 0.1 (low), 0.25 (moderate) and 0.36 (high) ^[55], and the model test results show that the GoF value is 0.381, which has a high model prediction effectiveness. Overall, the PPM model constructed by this institute is of good quality. The GoF is computed using the following formula:

where the AVE is calculated by Eq (12), ${R}^{2}$ is the model ${R}^{2}$.

4.2.2.   Hypothesis testing results

Figure 5 shows the hypothesis testing results of the PPM model. The results showed that perceived legal norm (β = 0.234, p < 0.001), perceived severity (β = 0.23, p < 0.01), perceived vulnerability (β = 0.192, p < 0.01), conformity tendency (β = 0.241, p < 0.05), subjective norm (β = 0.15, p < 0.05) and attitude (β = 0.198, p < 0.05) had a positive and significant impact on switching intention, and perceived inconvenience (β = -0.117, p < 0.001) and habit (β = -0.113, p < 0.001) had a significantly negative impact on switching intention, while response efficiency (β = 0.01, p > 0.05) and perceptual behavior control (β = 0.059, p > 0.05) had no significant impact on switching intention. See Table 6 for details.

4.3.   ANFIS results

The ANFIS model in this research was conducted in MATLAB Version R2018b with the Neuro-Fuzzy Designer toolbox (The MathWorks, Natick, MA, USA). Eight variables significantly impact switching intention in the PLS-SEM model test results were extracted as input variables for ANFIS. The total samples were divided into a training set (50%, 161 samples), a checking set (20%, 65 samples) and a testing set (30%, 96 samples). Gaussian membership functions (Gaussian MFs) were used to convert the 7-level Likert scale into three fuzzy levels of low, medium and high, and set the training number was 200 epochs. We obtained the predicted values, and the Root Mean Square Error (RMSE) between the observed and predicted values is 0.7321, within an acceptable range.

4.3.1.   Importance ranking

By analyzing the nonlinear relationship between the input variable and the predicted value of switching intention, the importance of each variable is obtained according to the change of the switching intention brought by the input variables (as shown in Figure 6). We found that the perceived legal norm, perceived inconvenience and conformity tendency are the most important factors in push, mooring and pull effects.

4.3.2.   Combined effect

Furthermore, we analyzed the changes in switching intention under the combined effect of two factors in the push, mooring and pull effects in order to more accurately analyze the influence relationship of various types of factors on switching intention (as shown in Figure 7).

5.   Discussion

5.1.   Theoretical implications

This research attempts to use the PPM model to explain e-bike riders' switching intention from the push, mooring and pull aspects. The model fitness test states that the conceptual model has strong explementary power on e-bike riders' switching intention on wearing helmets. The results indicate the adaptability of this proposed PPM model, therefore the researchers could use this theoretical framework to analyze the helmet-wearing behavior with different research purposes or other travel behavior switching. Furthermore, the ANFIS results show that the most important influencing factors in the three effects of push, mooring and pull are the perceived legal norm, the perceived inconvenience and the conformity tendency. We analyzed the change of switching intention under the combination of the most important factors and residual factors in each effect to explore the relationship between influencing factors and helmet-wearing switching intention in a deeper way. The two-step SEM-ANFIS approach provides an analysis framework to find the significant factors, rank their importance and explore the combined effects. Meanwhile, the analysis framework above is a theoretical guidance to help the practitioners implement work. First, constructing the theoretical model and involving possible influencing factors. Second, testing the significance by SEM and choosing the significant factors for inputs. Third, inputting the significant factors into ANFIS and obtaining the predicted output. Finally, analyzing the nonlinear relationship and combined effects, and drawing a conclusion.

5.2.   Practical implications

5.2.1.   Push factors

We found that perceived legal norm was significantly correlated with the switching intention. This conclusion is consistent with ^[60], which assertion that traffic accident legislation could make drivers actively or consciously avoid traffic violations. Inversely, this result is different from ^[61], who reached the opposite conclusion. Two main reasons may contribute to mixed views of perceived legal norm between this study and the others. First, Jiangsu province early started implementing regulations on e-bikes, issuing 50 RMB penalties (equal to around 7 ＄) for no-wearing helmet behavior. Most Chinese e-bike riders belong to economically disadvantaged subpopulations ^[37]. They might value 50 RMB more than others. Second, local traffic authorities have carried out activities and increased police force to investigate and punish this behavior after the "One Helmet One Belt" event. E-bike riders feel the increase of police on the road and are likely to wear helmets

The most important influencing factor in the push factors was the perceived legal norm. Figure 7 shows that under the influence of lower perceived vulnerability, with the rise of perceptual legal norms, the improvement of switching intention is not obvious, indicating that people with low perceived vulnerability have limited deterrence of legislation. The main reason might be that they do not think an accident will happen to ride an e-bike for fluke mentality. Therefore, policymakers should pay more attention to publicizing the possibility of a traffic accident to e-bike riders, such as providing the annual number and the e-bike-related accident rate.

5.2.2.   Mooring factors

The influence of two factors on the switching intention is both negative in the mooring factors. The significant result on perceived inconvenience is consistent with ^[34], and the ANFIS results showed that perceived inconvenience is more important than habit as a hindrance factor. The combined effect results of these two factors further prove that people with lower perceived inconvenience and weaker habits of not wearing helmets have a higher intention to wear helmets. Therefore, helmet manufacturers should focus on the portability of e-bike helmets. Besides, anti-theft and storage are also important factors, so e-bike manufacturers should add storage and anti-theft systems to helmets.

5.2.3.   Pull factors

Conformity tendency statistically impacts e-bike riders' switching intention positively, which is consistent with previous studies exploring pedestrians' violating crossing behavior ^[61]. Based on the combined effect results, we found that conformity tendency has a significant two-sidedness. Under low subjective norms and attitudes, with the rise of conformity tendency, the switching intention shows a trend of first rising and then declining. The main reason might be that the individual's conformity will lead to a vacillating behavior intention. The switching intention rises when the conformity tendency rises from a low to a medium level. When the conformity tendency rises from a medium level to a high level, the population with low subjective norms and attitudes is easily affected by riders who do not wear helmets, so their own helmet-wearing intention declines. Low attitudes and subjective norms will lead to blind conformable intention not to wear a helmet.

5.3.   Limitations and future direction

This research provided exciting findings while also having several limitations. First, since the switching behavior is rather complex, many other factors, such as satisfaction and moral norms, might also influence helmet-wearing behavior. Second, the road travel environment, weather, helmet type and travel time might influence the helmet-wearing behavior intention. Hence, future research could extend these factors and explore the influence from the combined effects of subjective and objective factors. Third, the data was only collected with a structured interview in one city. Data collected from more samples, more channels and cities with different legislation situations could be considered to improve the persuasion of the results, especially for the ANFIS model. Furthermore, the punishment from traffic police, the direct effect of helmet regulations, is not involved in this research. The participants could be divided into different groups by multi-group analysis based on their experience of punishment to analyze the heterogeneity and identify the effects of law enforcement.

6.   Conclusions

We constructed a PPM model and used the two-step SEM-ANFIS method to analyze the questionnaire survey data, verify the path significance of push, mooring and pull variables that affect the helmet-wearing switching intention for e-bike riders in China. After ranking the importance of variables and analyzing the changes in switching intention under the interaction of different types of factors, the following conclusions are obtained:

The most important variables in push, mooring and pull are the perceived legal norm, perceived inconvenience and conformity tendency. On the basis of maintaining law enforcement actions, helmet storage facilities and devices could be installed in parking places or e-bikes so as to improve e-bike riders' willingness to wear helmets, further influence more riders to wear helmets, and form a good travel atmosphere for wearing helmets.

Under the low perceived vulnerability, the increase in switching intention brought by the rise of the perceived legal norm is not obvious, indicating that for people with low perceived vulnerability, the impact of traffic police' deterrence will be reduced under the fluke mentality. Awareness of the vulnerability of accidents for e-bike riders is needed to ensure the effectiveness of helmet legislation.

For people with low attitudes and subjective norms, it is easy to blindly follow the crowd and ride without wearing helmets. It is necessary to improve the attitude and personal safety awareness through media publicity, safety lectures, etc., so as to avoid e-bike riders' blind conformism and follow the trend without wearing helmets, and improve the helmet-wearing rate.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (grant number 71871107).

Conflict of interest

The authors declare there is no conflict of interest.