Riding Risk: Factors Shaping Helmet Use Among Two-Wheeled Electric Vehicle Riders in Fuzhou, China

Abstract

With the rapid increase in the number of two-wheeled electric vehicles, the number of accidents related to them has also greatly increased. However, despite facing a huge threat from accidents, the helmet, an efficient and legally required protection for riders, is not popular with Chinese two-wheeled electric vehicles riders. To study the factors affecting helmet use for these riders, this paper conducted an observational study to collect helmet use data for 16,207 two-wheeled electric vehicle riders in Fuzhou, China. With these data, this paper built a multivariate logistic regression model to study the main effects of various factors on helmet use, and analyze the interaction effects of these factors. Results showed that, on the one hand, area, weather, temperature, controller, separated non-motor-vehicle lanes, time, rider’s age, and type of vehicle had significant effects on helmet use and the interaction between these factors is significant, especially the interaction between weather, temperature and other factors. On the other hand, level of service, gender and whether the riders are food delivery workers have no significant impact on helmet use, but show significant interaction effects with other factors.

1. Introduction

The widespread use of two-wheeled electric vehicles (TWEVs) has become a global phenomenon, with the total number exceeding 300 million worldwide [1]. This surge is attributed to TWEVs’ affordability, convenience, and environmental benefits. However, the increase in TWEV usage has been accompanied by a corresponding rise in traffic-related accidents, significantly impacting public safety [2]. Globally, traffic crashes involving TWEVs are a major concern, with TWEV riders facing disproportionately higher risks compared to users of other vehicle types [3]. Electric bicycle accidents are more severe than traditional bicycle accidents due to higher speeds and greater momentum. This heightened risk presents new challenges for road safety, requiring updated measures to protect riders and pedestrians alike [4].

Effective interventions, including red-light cameras, wearing helmets, seat belts, and addressing safety incidents, can reduce road traffic accidents by 25–40% [5]. Helmet use, a critical factor in reducing the severity of injuries, has been identified as a primary means of mitigating these risk [6]. Studies have consistently demonstrated that helmets reduce the risk of severe head injuries by 69% and fatalities by 42% in the event of a crash [7,8]. Despite this evidence, helmet usage among TWEV riders remains suboptimal. However, progress in helmet use is not satisfactory, although many countries have enacted legislation to force riders to wear helmets when riding [9]. A large number of studies have shown that riding without wearing a helmet is still common, which is proved by observational studies, like that of Felix et al. [10] in Myanmar, Várheyi et al. [11] and Rusli R et al. [12] in Malaysia, and Xuequn [13] in China. Due to a significant number of motorcyclists not wearing helmets, ASEAN countries have implemented stricter enforcement of mandatory helmet laws for two-wheeler riders [14].

Understanding the reasons for this reluctance is essential to improving safety. Research indicates that helmet use is influenced by both subjective and objective factors. Subjective factors include individual attributes, such as education level, personal attitudes, and risk-taking behaviors [15,16,17]. However, these factors are challenging to address through short-term interventions. Consequently, researchers have turned to objective factors, such as secondary tasks during riding [18], law enforcement intensity [19], and demographic characteristics [13]. Despite this shift, there remains a lack of clarity regarding how these objective factors interact to influence helmet use, particularly among TWEV riders, whose behaviors may differ significantly from those of motorcyclists.

A review of existing studies reveals significant methodological gaps. Data collection often relies on direct observation or questionnaires, with analysis typically conducted using logistic or probit models to identify the relationships between variables and helmet use [20,21,22]. However, most studies have focused on isolated factors, neglecting the potential interaction effects that could provide deeper insights into rider behaviors. For example, while enforcement efforts are known to increase compliance, their interaction with other factors, such as rider demographics or traffic density, has not been adequately explored. Studies have shown, that for every 1% increase in police patrol time, the frequency of accidents during the same period decreases by 0.15% [23].

In response to these gaps, this study conducted a detailed observational analysis of TWEV riders in Fuzhou, China. A total of 16,206 TWEV riders were observed across eight locations, varying by traffic conditions, enforcement levels, and demographic composition. Using a multivariate logistic regression model, we analyzed the effects of factors such as gender, age, traffic density, and enforcement presence on helmet use, along with the interactions between these variables. This approach aims to provide policymakers and professionals with actionable insights into the factors driving helmet use and to support the development of targeted interventions. Ultimately, the findings are expected to enhance the effectiveness of helmet regulations and improve TWEV safety management.

2. Literature Review

2.1. Review of Research on TWEV Safety

In the field of traffic safety, research related to TWEV safety focuses mainly on the investigation of riding behavior and the analysis of crash data. Wang et al. conducted a questionnaire survey of TWEV riders in Guilin City and used structural equation modeling to analyze the relationship between safety knowledge, safety attitudes, and risk perception and risky driving behaviors, and used multiple regression modeling to test the effect of safety knowledge on various types of risky driving behaviors. The results showed that safety knowledge was significantly associated with risky driving behaviors in Chinese TWEV riders [24]. Chai et al. compared the behaviors of Chinese e-bike riders and bicycle riders at unsignalized intersections. They found that e-bike riders were less sensitive to inclement weather and intersection risk, but more sensitive to infrastructure quality [25]. Qian et al. developed a two-wheeled vehicle rider behavior questionnaire and conducted a survey in the city of Xi’an, finding that the characteristics of the abnormal behaviors of TWEV riders were more similar to those of motorcycle riders. They both showed a higher frequency of abnormal riding behavior compared to bicycle riders [26]. Gioldasis et al. conducted a face-to-face road survey of TWEV riders to explore the relationship between risk factors and user attributes and travel behavior [27].

In addition to surveys of TWEV riding behaviors, a number of scholars have analyzed crash data. Shah et al. used crash data from the Tennessee Integrated Transportation Analytics Network (TITAN) database from April 2018 to April 2020 and analyzed this using the Pedestrian and Bicycle Crash Analysis Tool (PBCAT), which showed that every 10 e-scooter and bicycle motorized vehicle collisions resulted in injury or death to a TWEV rider or bicyclist. In addition, there were statistically significant differences between TWEV and bicycle collisions in terms of spatial and temporal distribution, demographics, lighting conditions, and distance from home [28]. Sadeghi et al. combined the XGBoost algorithm with a stochastic parametric binary logit model and used likelihood ratio tests to validate the temporal instability of motorized vehicle collisions in the UK over different periods of the COVID epidemic [29]. Gao et al. categorized EV crashes into two-vehicle crashes and single-vehicle crashes, and used a random parameter approach with mean and variance heterogeneity to examine the factors affecting injury severity [30]. In addition, some studies have explored the factors affecting the frequency of TWEV crashes from a macroscopic perspective, combining crash data and environmental characteristics using crash counting models [31,32].

2.2. Review of Research on Helmets

Numerous studies have shown that failure to properly use seat belts and helmets is one of the leading causes of road traffic accidents [33,34]. Wearing a helmet can be effective in reducing the risk of head injuries [35]. Thompson et al. investigated the effectiveness of bicycle helmets in preventing and protecting against head injuries. The results showed that bicycle helmets, regardless of type, can provide protection for cyclists of all ages involved in collisions, including those involving motor vehicles [36]. The World Health Organization (WHO) produced the World Report on Road Traffic Injury Prevention [37]. It was found that proper wearing of motorcycle helmets reduced the risk of traffic fatalities by 40% and the risk of severe head injuries by 70%. Moftakhar et al. conducted a retrospective multicenter study of all consecutive patients admitted to hospitals for TWEV-related injuries (riders and non-riders) between 1 May 2018 and 30 September 2019, and the results showed that 17.7% had serious head injuries. Therefore, proper helmet wear may reduce the chance of injury in TWEV accidents [38]. Hamzani used a retrospective cross-sectional study design. All patients, including those referred to the emergency departments of tertiary care centers for oral and maxillofacial injuries involving TWEVs or P-scooters, from 2014–2020 were investigated. Results showed that patients who used helmets also had lower rates of fractures (18.2%) and alveolar injuries (23.7%) than those who did not (68.8% and 37.3%, respectively) [39].

However, despite the protective advantages of helmets, helmet use among e-bike riders is relatively low [19]. Some riders have low willingness to wear helmets because of their heaviness and inconvenience [40]. Yuan et al. found that helmet wearing among e-bike riders was influenced by area of residence, age, gender, marital status, and education level [41]. Yang et al. developed an integrated theoretical model combining the Theory of Planned Behavior (TPB) and the Health Belief Model (HBM) to test the influence of factors among college students on the willingness of TWEV riders to wear helmets, and the results showed that law enforcement emerged as the most influential factor, highlighting the key role of enforcing regulations and raising awareness [42]. Ma et al. used logistic regression to study the impact of helmet wearing regulations on TWEV riders, and the results showed that mandatory regulation was more effective than encouraging regulation in reducing head injuries [43].

This paper takes Fuzhou city as an example and analyzes the factors affecting the influence of helmet wearing on TWEV riders through video data. It provides a theoretical basis for the safety management of electric vehicles in Fuzhou city.

3. Materials and Methods

3.1. Site Characteristics

Four intersections and four road sections in Fuzhou, China were selected for investigation. The distribution of observation sites is shown in Figure 1, and their characteristics are shown in

Table 1. Observation site characteristics.

Name	Mark in the Figure	Type	Area	Separate Non-Motor Vehicle Lane
Jiangfu Road-Bincheng Avenue intersection	triangle-1	intersection	suburb	nothing
Jianping Road-Qishan Avenue intersection	triangle-2	intersection	suburb	yes
Wusi Road-Hualin Road intersection	triangle-3	intersection	urban	yes
Guping road-Hudong Road intersection	triangle-4	intersection	urban	nothing
Bincheng Avenue section	circle-1	road section	suburb	nothing
Qishan Avenue section	circle-2	road section	suburb	yes
Wusi Road section	circle-3	road section	urban	yes
Guping road section	circle-4	road section	urban	nothing

. The choice of sites was based on the data provided by the traffic police and the distribution of TWEV traffic volume. The observation was carried out through monitoring by traffic police. The view is shown in Figure 2.

3.2. Data Collection Procedure

The observational data came from the surveillance video of the observation site from 11 March to 10 April 2022, from which a sunny day and a rainy day were randomly selected as the survey dates. The survey time was from 7:30 am to 8:00 am (morning peak hours), from 12:00 to 12:30 pm (off peak hours) and from 17:30 pm to 17:30 pm (evening peak hours). Each site provided a total of 3 h of surveillance video, and the total amount of video material observed was 24 h.

All observation data were coded by 15 master’s degree students majoring in traffic and transportation. The data were collected in a pre-set table. All coders had been trained in different types of TWEV identification, rider characteristics identification and data quality control. The coding helmet use data was reviewed by three experts in the field of traffic engineering and traffic safety to ensure data quality.

4. Data Analysis

4.1. Descriptive Statistics

A total of 16,206 TWEV riders were observed, including 5375 riders not wearing helmets and 10,831 riders wearing helmets. The observed helmet use data at each site is shown in Figure 3. The number on the right side of the bar chart refers to the number of riders observed, and the x-axis as a percentage refers to the proportion of riders without helmets compared to all riders.

4.2. Multivariate Logistic Regression Model

The multivariate logit model is applicable in cases where the result variable is a category variable (with or without helmet), and multiple explained variables are classified variables or sequence variables. Therefore, a multivariate logistic regression model was built to analyze the relationship between helmet use and other factors. The formula for the model is shown in Equation (1) below:

$$logit\left[\mathrm{P}\left(\mathrm{y}=j\left|X\right.\right)\right]=\log \left({\displaystyle \frac{\mathrm{P}\left(\mathrm{y}=j\left|X\right.\right)}{1-\mathrm{P}\left(\mathrm{y}=j\left|X\right.\right))}}\right)={\alpha }_j-x_i^{\prime }{\beta }_j,$$

(1)

where $j$ is the level of the result variable and i is the number of variables to be explained,$\mathrm{i}ϵ\left[1,I\right],jϵ\left[1,J\right]$. y is the result variable. $x_i^{\prime }$ is the change value of the i-th explained variable. ${\beta }_i$ is the slope of the i-th explained variable. ${\alpha }_j$ is the constant term of the result variable corresponding to the j-th level.

When y is the j-th level, the formula can be written as Equation (2):

$$P_j=\mathrm{P}\left(\mathrm{y}=j\left|X\right.\right)={\displaystyle \frac{1}{1+\exp \left[-\left({\alpha }_j-x_i^{\prime }{\beta }_j\right)\right]}}$$

(2)

In a logistic model, a certain group of explained variables is selected as the baseline to estimate the impact of other explained variables on the result variables. The impact is generally reflected in the odds ratio. Assuming that a selected baseline is the first group, then the impact of other explained variables on the result variable relative to the baseline can be written as follows in Equation (3):

$${\displaystyle \frac{P_j}{P_1}}={\displaystyle \frac{\mathrm{P}\left(\mathrm{y}=j\left|X\right.\right)}{\mathrm{P}\left(\mathrm{y}=1\left|X\right.\right)}}=e^{x_i^{\prime }{\beta }_j},j=2,\dots ,\mathrm{J}$$

(3)

The effect of the change of $l$-th explained variables on the odds ratio can be written as $e^{\Delta x_{il}{\beta }_{jl}}$, where, ${\beta }_{jl}$ is the $l$-th element in the $j$-th group coefficient vector ${\beta }_j$. The equation shows that, when the other explained variables remain unchanged, for each additional unit of $x_{il}$, i.e., $\Delta x_{il}=1$, the odds ratio of selecting the $j$-th groups relative to baseline is $e^{{\beta }_{jl}}$. It can be seen that if, and only if, ${\beta }_{jl}$ > 0, the odds ratio of the $j$-th group relative to the baseline is greater than 1, and the increase in $x_{il}$ will increase the odds ratio of selecting the $j$-th group relative to the baseline.

4.3. Data Preparation

In this study, helmet use was the result variable. The other factors recorded during observation were explained variables. Finally, 16,206 rows of table data were obtained.

Table 2. Result variables, explained variables and their codes.

Result Variable	Explained Variable	Data Type	Classification	Code
Helmet use	Area	Category	Urban, Suburban	1, 0
	Time	Category	Morning Peak hours, Off Peak hours, Evening Peak hours	0, 1, 2
	Level of service (LOS)	Sequence	1, 2, 3, 4 Levels	1, 2, 3, 4
	Weather	Category	Rainy, Sunny	1, 0
	Existence of controller (abbreviated as EC)	Category	Yes, No	1, 0
	Average daily temperature (abbreviated as ADT)	Category	≥20 °C, <20 °C	1, 0
	Whether the rider is a food delivery worker (abbreviated as FD)	Category	Yes, No	1, 0
	Separate non-vehicle lane (abbreviated as SNL)	Category	Yes, No	1, 0
	Gender	Category	Male, Female	1, 0
	Age	Category	≥35 Years old, <35 Years old	1, 0
	Vehicle type	Category	Electric Moped, Electric Bicycle	1, 0

shows the result variable, explained variables and their codes. STATA 17.0 MP software was used to build the logistic model, with mlogit command.

5. Results

5.1. Multicollinearity Diagnostics

Multicollinearity diagnostics are used to examine whether there are correlations among the influencing factors in the electric bicycle rider helmet use model. Severe multicollinearity can increase the standard deviation and variance of model parameters, reduce the significance level of variables, and even cause the model to fail to converge, leading to ineffective predictions. Therefore, this study employs the Variance Inflation Factor (VIF) method and the correlation coefficient matrix method to diagnose multicollinearity among the 11 factor variables included in the model, and Pearson correlation analysis is used to analyze the correlation between variables.

Based on the diagnostic results of the VIF and the correlation coefficient matrix method, it can be concluded that there is no severe multicollinearity among the 11 factor variables in the model (

Table 3. Results of multicollinearity diagnostics for the independent variables.

Variable	VIF	1/VIF
Los	1.67	0.60
ADT	1.67	0.60
SNL	1.12	0.89
Age	1.12	0.89
Vehicle type	1.11	0.90
Weather	1.08	0.92
EC	1.08	0.93
Time	1.06	0.94
Gender	1.06	0.95
FD	1.04	0.96

Mean VIF: 1.20.

). Specifically, the VIF values are all below 5 and within a reasonable range, indicating that the correlations among the factors are low and the variables are relatively independent. Therefore, the parameter estimates of the model are reliable, and the standard deviations or variances will not be inflated due to multicollinearity.

In the correlation coefficient matrix, the correlations between variables do not show significant multicollinearity issues, and the highest correlation coefficients between variable combinations are still within an acceptable range, with no near-perfect correlations close to 1 or −1 (Figure 4). This further indicates that the factor variables in the model can function independently, ensuring the stability of the model.

Overall, the multicollinearity diagnostics indicate that the correlations between selected variables do not significantly affect the regression results, and the model’s parameter estimation and predictive performance remain unaffected by multicollinearity. Therefore, the model can be used for further analysis and prediction.

5.2. Analysis of Multivariate Logistic Regression Model Results

After 8 iterations, the model converges (log likelihood = −8444.285). The model prob > χ²= 0.000, pseudo R² = 0.2147, which indicates that the model fits better than the model containing only constant terms, but the ability of independent variable x to explain the total variation is low. The results of the multivariate logistic model are shown in

Table 4. Results of main effect analysis.

Explained Variable		Odds Ratio	Possibility of Helmet Use	Standard Error	z	P > |z|	95% Confidence Interval
							Lower Limit	Upper Limit
LOS	1	Baseline
	2	2.798	179.8%	2.630	1.090	0.274	0.443	17.658
	3	0.398	−60.2%	0.360	−1.020	0.309	0.067	2.349
	4	0.115	−88.5%	0.171	−1.450	0.146	0.006	2.126
Area	Suburb	Baseline
	Urban	0.007	−99.3%	0.002	−14.460	0.000	0.003	0.013
Weather	Sunny	Baseline
	Rainy	0.062	−93.8%	0.026	−6.570	0.000	0.027	0.142
ADT	Below 20 °C	Baseline
	Above 20 °C	0.000	−100.0%	0.000	−6.630	0.000	0.000	0.001
EC	No	Baseline
	Yes	4.460	346.0%	2.898	2.300	0.021	1.248	15.940
FD	No	Baseline
	Yes	3.466	246.6%	3.645	1.180	0.237	0.441	27.229
SNL	No	Baseline
	Yes	0.204	−79.6%	0.141	−2.290	0.022	0.052	0.794
Time	Morning peak hours	Baseline
	Off peak hours	0.036	−96.4%	0.060	−1.980	0.047	0.001	0.962
	Evening peak hours	0.025	−97.5%	0.030	−3.110	0.002	0.002	0.258
Gender	Female	Baseline
	Male	1.316	31.6%	0.326	1.110	0.269	0.809	2.139
Age	≤35 years old	1 baseline
	>35 years old	2.922	192.2%	0.751	4.170	0.000	1.766	4.837
Vehicle type	Electric bicycle	Baseline
	Electric moped	0.326	−67.4%	0.097	−3.760	0.000	0.182	0.585

. Among them, the odds ratio refers to the probability of using helmets when all other explained variables are controlled at the same level. Taking area as an example, riders in urban areas are more likely to use helmets than those in suburbs by 99.3%. After the main effect analysis, the interaction analysis was carried out. The following

Table 5. Result of the interaction analysis.

Explained Variable		Odds	Possibility of Using Helmet	Standard Error	z	P > |z|	95% Confidence Interval
							Lower Limit	Upper Limit
Area*Weather	Area = Suburb * Weather = Sunny	Baseline
	Area = Urban * Weather = Rainy	14.227	1322.7%	4.419	8.550	0.000	7.740	26.151
Area*FD	Area = Suburb * FD = No	Baseline
	Area = Urban * FD = Yes	0.032	−96.8%	0.039	−2.790	0.005	0.003	0.360
Area*Gender	Area = Suburb * Gender = Female	Baseline
	Area = Urban * Gender = Male	1.483	48.3%	0.269	2.170	0.030	1.040	2.117
Weather*EC	Weather = Sunny * EC = No	Baseline
	Weather = Rainy * EC = Yes	0.608	−39.2%	0.072	−4.190	0.000	0.482	0.767
Weather*ADT	Weather = Sunny * ADT = <20 °C	Baseline
	Weather = Rainy * ADT = ≥20 °C	0.001	−99.9%	0.001	−5.350	0.000	0.000	0.010
Weather*FD	Weather = Sunny * FD = No	Baseline
	Weather = Rainy * FD = Yes	3.957	295.7%	1.424	3.820	0.000	1.954	8.012
Weather*SNL	Weather = Sunny * SNL = No	Baseline
	Weather = Rainy * SNL = Yes	1.405	40.5%	0.147	3.260	0.001	1.145	1.725
Weather*Gender	Weather = Sunny * Gender = Female	Baseline
	Weather = Rainy * Gender = Male	1.356	35.6%	0.121	3.400	0.001	1.138	1.615
Weather*Vehicle type	Weather = Sunny * Vehicle type = Electric bicycle	Baseline
	Weather = Rainy * Vehicle type = Electric moped	1.355	35.5%	0.135	3.060	0.002	1.115	1.646
ADT*EC	ADT = <20 °C * EC = No	Baseline
	ADT = ≥20 °C * EC = Yes	0.002	−99.8%	0.002	−7.490	0.000	0.000	0.010
ADT*FD	ADT = <20 °C * FD = No	Baseline
	ADT = ≥20 °C * FD = Yes	0.179	−82.1%	0.139	−2.220	0.027	0.039	0.819
ADT*Age	ADT = <20 °C * Age = ≥35 Years old	Baseline
	ADT = ≥20 °C * Age = <35 Years old	1.746	74.6%	0.342	2.850	0.004	1.189	2.564
ADT*Vehicle type	ADT = <20 °C * Vehicle type = Electric bicycle	Baseline
	ADT = ≥20 °C * Vehicle type = Electric moped	2.843	184.3%	0.765	3.880	0.000	1.677	4.819
EC*FD	EC = No * FD = No	Baseline
	EC = Yes * FD = Yes	3.468	246.8%	1.566	2.750	0.006	1.431	8.404
EC*SNL	EC = No * SNL = No	Baseline
	EC = Yes * SNL = Yes	0.399	−60.1%	0.051	−7.160	0.000	0.310	0.513
EC*Gender	EC = No * Gender = Female	Baseline
	EC = Yes * Gender = Male	1.711	71.1%	0.176	5.230	0.000	1.399	2.093
EC*Age	EC = No * Age = ≥35 Years old	Baseline
	EC = Yes * Age = <35 Years old	0.738	−26.2%	0.088	−2.550	0.011	0.584	0.932
FD*Age	FD = No * Age = ≥35 Years old	Baseline
	FD = Yes * Age = <35 Years old	0.374	−62.6%	0.131	−2.810	0.005	0.189	0.742
SNL*Age	SNL = No * Age = ≥35 Years old	Baseline
	SNL = Yes * Age = <35 Years old	0.764	−23.6%	0.082	−2.520	0.012	0.620	0.942
Gender*Age	Gender = Female * Age = ≥35 Years old	Baseline
	Gender = Male * Age = <35 Years old	1.242	24.2%	0.111	2.430	0.015	1.043	1.479

lists all statistically significant (p < 0.05) interaction factors without singular values.

6. Discussion

The main effect analysis shows that the possibility of helmet use for riders in urban areas is higher than that of riders in suburban areas, and riders are more reluctant to wear helmets on rainy days. This finding aligns with previous studies, which indicate that stricter law enforcement and better infrastructure in urban areas contribute to higher helmet use rates [44,45,46]. However, our study further reveals that regulatory differences regarding TWEV modifications play a crucial role. An interaction is found, increasing the possibility of helmet use to 1322.7% relative to the baseline group of Area = Suburb * Weather = Sunny. Apart from the stricter enforcement of helmet use laws in urban areas, this enormous increase may be due to the different regulations regarding TWEV modification in suburbs and urban areas [47]. In urban areas, installing rain canopies on TWEVs is prohibited, making riders more likely to wear helmets on rainy days. However, in suburban areas, regulations on such modifications are more relaxed, leading riders to forgo helmets in favor of installing rain shelters for protection. Compared to similar studies that focus solely on law enforcement, our findings offer a novel perspective by highlighting the indirect effects of vehicle modifications on helmet use

The rise of ADT also reduces the probability of riders wearing helmets. This contradicts previous research suggesting that higher traffic density leads to increased safety awareness among riders [48]. Interaction analysis proves that this interacts with age. When the ADT rises, older riders are more willing to wear helmets. The reason may be that older riders are more safety conscious, which is also verified by the main effect analysis of age variables. This highlights that helmet use behavior is not only dependent on external traffic conditions but also significantly influenced by demographic factors, an aspect that previous models may have overlooked [44,49].

The existence of traffic controllers (EC) greatly increases the probability of helmet use. On the contrary, separated non-motor vehicle lanes (SNLs) do not improve the use of helmets but increase the probability of riders not wearing helmets. This contrasts with studies, which suggest that dedicated lanes improve overall traffic safety [50,51]. In interaction analysis, even if traffic controllers exist, the separated non-motor vehicle lane still reduces the possibility of helmet use. The source of this problem may be that riders overestimate the protective effect of lane separation, thus reducing helmet use, seen as implementation of an “over-protection” measure [52]. Unlike previous research that primarily focuses on the benefits of lane separation, our study reveals a potential unintended consequence—riders’ false sense of security leading to lower helmet use.

Time of day affects helmet use. Compared with the morning peak hours, the probability of riders wearing helmets during off peak hours and evening peak hours is reduced. The reasons for this situation may be very complicated, involving traffic flow, the characteristics of riders, the existence of controllers, etc. Our findings support previous research indicating that law enforcement intensity varies throughout the day [53], but we further highlight that this variation directly influences riders’ compliance behavior. Most likely, stricter law enforcement during morning peak hours may lead to this situation.

It is noteworthy that, in the main effect analysis, gender does not have a significant impact on helmet use (p > 0.05). However, regarding interaction, that between gender and age shows a significant impact. Over 35-years-old male riders are more likely to wear helmets, which may reflect the different attitudes of male and female riders of different ages towards helmet use. This nuanced finding extends previous studies that focus solely on gender differences by demonstrating the necessity of considering interaction effects between demographic variables [49,54]. At the same time, although being a food delivery worker has no influence on helmet use, in interactive analysis, the probability of the over 35-years-old food delivery worker wearing the helmet is less than that of under-35 years-old riders. This may indicate that work-related urgency or risk perception differences contribute to helmet use behavior, an area that warrants further study.

Types of TWEV have a significant influence on helmet use. Compared with electric bicycle riders, electric moped riders are more reluctant to wear helmets. This aligns with previous research suggesting that riders of larger, more powerful vehicles often exhibit greater risk-taking behavior [55,56]. However, in interactive analysis, when the control temperature is different, the possibility of electric moped riders wearing helmets is higher than that of electric bicycle riders. Further experiments are needed to check and verify this conclusion. This temperature-related factor has been rarely explored in prior studies, suggesting that external environmental conditions could play a more critical role than previously assumed. Further experiments are needed to check and verify this conclusion.

From Table 3 and Table 4, it seems that different LOS has no influence on helmet use. All the p values for this are greater than 0.05 in the main effect analysis. This result is inconsistent with previous studies on traffic density and risky behaviors [57], which often report a correlation between congestion levels and safety compliance. The reason for this phenomenon may be that helmet use is not closely related to aggressive riding behavior, which is obstructed by other vehicles in high-density traffic flow. Our findings suggest that helmet use decisions may be less about risk-taking tendencies and more about external regulatory and environmental factors, distinguishing our study from previous behavioral-focused research.

Despite the strong explanatory power of the multinomial Logit model, this study has several limitations. First, while the model provides valuable insights into the influence of various factors on helmet use, it assumes that the relationships between variables are linear and independent. This may oversimplify real-world behavior, as helmet use decisions are influenced by complex and dynamic factors that may require more flexible models, such as mixed-effects models or machine learning approaches. Second, the study relies on observational data collected at specific locations in Fuzhou, which may limit the generalizability of the findings to other cities or regions with different regulatory and cultural contexts. Third, some unobserved variables, such as individual risk perception, attitudes toward law enforcement, and socioeconomic factors, were not included in the model, potentially affecting the robustness of the conclusions. Future research should consider incorporating more comprehensive behavioral data and exploring alternative modeling techniques to enhance predictive accuracy and generalizability.

7. Conclusions

This study conducted an observational analysis to collect helmet use data from 16,206 two-wheeled electric vehicle (TWEV) riders in Fuzhou, China. A multivariate logistic regression model was developed to investigate the influence of various factors, including level of service (LOS), area type (urban/suburban), average daily traffic (ADT), weather, etc., on helmet use. The model demonstrated high goodness of fit, effectively explaining how these factors impact helmet-wearing behavior.

The findings revealed that several factors, such as area type, weather conditions, average daily temperature, the presence of traffic controllers, separated non-motor vehicle lanes, time of day, rider age, and vehicle type, significantly influenced helmet use. Moreover, the interactions among these factors were notable, with weather and average daily temperature showing particularly strong interaction effects. In contrast, LOS, gender, and whether the rider was a food delivery worker or not did not have significant direct effects on helmet use but exhibited significant interaction effects with other factors.

A key observation was that separated non-motor vehicle lanes might create a false sense of security, leading to lower helmet use among riders. Additionally, the analysis highlighted discrepancies between the conclusions drawn from interaction effect analysis and those from main effect analysis. These findings align with previous research but provide additional insights into the behavioral dynamics of helmet use, emphasizing the role of environmental and regulatory factors.

Future research should focus on several key areas to build upon these findings. First, alternative modeling approaches, such as mixed-effects models or machine learning methods, should be explored to better capture the complexity of helmet use behaviors. Second, additional behavioral variables, including riders’ individual risk perceptions, attitudes toward law enforcement, and socioeconomic factors, should be integrated to enhance the robustness of the analysis. Third, expanding the study to different cities or regions with varying regulatory environments will help assess the generalizability of the conclusions. Finally, controlled experiments could be conducted to further validate the interaction effects identified in this study, ensuring a more precise understanding of the factors influencing helmet use.