Analysis of factors affecting injury severity of shared electric bike riders

Zhang Xiaolong Huang Jianling,2 Bian Yang Zhao Xiaohua Li Jia

(1Faculty of Architecture,Civil and Transportation Engineering,Beijing University of Technology,Beijing 100124,China)(2Beijing Intelligent Transportation Development Center,Beijing 100073,China)

Abstract：This study investigated the potential factors affecting the injury severity of shared electric bike (e-bike) riders and analyzed potential heterogeneity using a dataset comprising of 1 343 shared e-bike insurance accidents recorded by a shared e-bike company as the research object.The injury severity was categorized into two levels: not injured and injured.Twelve independent variables were selected based on six aspects involving attributes of shared e-bike rider,vehicle,road,environment,time,and accident.The effects of different factors on the injury severity of shared e-bike riders were assessed using the random parameter logit model with heterogeneity in means.Results indicate that the variable “other traffic participants at fault” in the accident scenarios featured a random parameter that adhered to a normal distribution and exhibited mean heterogeneity.This increased the likelihood of injury among shared e-bike riders.However,the probability of injury decreased when the scenario involved both the variable “other traffic participants at fault” and component damage.The variables female,intact road surface,dry road pavement,nighttime,single-vehicle accidents,and both at-fault accidents could increase the injury probability among shared electric bike riders to varying degrees.The findings of this research provide a theoretical basis for the development of traffic safety strategies targeted at shared electric bike riders.

Key words：traffic engineering; injury severity; shared electric bike riders; random parameter model; heterogeneity in mean

Under China’s rapid economic development,rapid industrialization progress,and continuous urbanization,cities continue to expand.Consequently,road network traffic has increased,and traffic travel distances continue to rise.Owing to their speed,convenience,and cost-effectiveness,electric bikes (e-bikes) have become an essential means of urban transportation for short distances[1].In recent years,China’s e-bike market has experienced rapid growth,with the social ownership of e-bikes reaching nearly 300 million[2].However,the issue of e-bike traffic safety has become more pronounced.In 2019,e-bike-related casualties accounted for about 70% of the nonmotor vehicle losses in national road traffic accidents[3].

In recent years,the rapid advancement of Internet technology has led to an emerging service industry represented by shared e-bikes.By 2019,the number of shared e-bikes had exceeded 1×106[4].Moreover,shared e-bikes have introduced certain safety concerns owing to poor management and insufficient fundamental investment.In addition,the simple registration procedure of the platform[5]and the inappropriate age limit for riders[6]have resulted in a low entry barrier for shared e-bikes,thereby posing a significant safety risk.Moreover,some riders lack proper driving skills and legal awareness[7],contributing to repeated traffic violations such as wrong-way riding,thereby increasing the risk of traffic accidents.Research indicates a rising trend in traffic accident injuries involving shared e-bikes[8].Therefore,investigating the traffic safety challenges posed by shared e-bikes to propose reasonable improvement suggestions will help alleviate the pressure on traffic safety management and assist in safety prevention and control.

In recent years,numerous scholars have performed extensive research on the safety risks and accidents associated with private e-bikes.Hertach et al.[8]obtained e-bike accident data through questionnaires and used a binomial logit model to analyze the influence of various factors such as individual,vehicle,road,and environmental attributes on the accident injuries of e-bike riders.Their findings revealed that accident injuries are more prevalent among women,older adults,riders traveling at speeds up to 45 km/h,and those who perceive their physical condition as inferior to those of their peers.Panwinkler et al.[9]extracted e-bike accident data based on accident text and used an ordered probability model for research.They found that excessive speed,drinking alcohol,and downhill riding can cause serious injury to e-bike riders.Dong et al.[10]analyzed the factors affecting accident injury based on accident video data involving e-bike collisions with cars using binomial logit and multiple logit models and compared the prediction accuracies of the two models.

The aforementioned research indicates that scholars have achieved specific advancements in the investigation of personal e-bike traffic accident injuries.However,the research on shared e-bike traffic accident injuries remains inadequate.Given the numerous differences between shared and private e-bike riders across various facets,including mental and physical health and riding behavior characteristics[6-7],it remains uncertain whether the outcomes of accident causation analysis and the corresponding improvement measures derived from the private e-bike riding group can be applied to the shared e-bike riding group.Consequently,for the emerging demographic of shared e-bike riders,reanalyzing the mechanism behind the impact of traffic accidents and formulating targeted measures to enhance traffic safety are imperative.Moreover,although numerous studies have modeled e-bike traffic accident injuries and analyzed causal relations,several unresolved or overlooked issues persist,such as the heterogeneity of e-bike accident injury data.The discrete choice model utilized in the aforementioned research,particularly the conventional logit regression model,cannot effectively identify unobserved heterogeneity in the traffic accident data[11].Heterogeneity arises when specific variables influencing accident severity are omitted from the model or when correlations exist between variables in the model and omitted variables.Studies have demonstrated that variables such as the physiological characteristics of drivers[12],traffic violations[13],and road lighting[14]are affected by other variables in the model during traffic accident analysis.This results in biased model outcomes and an inability to accurately analyze potential interaction relations among factors affecting e-bike traffic accident injuries.

Several scholars have utilized the random parameter logit (RP-logit) model to address heterogeneity in traffic accident data.Chen et al.[15]employed the RP-logit model to assess injury severity in single-car and multicar accidents involving trucks on rural roads.Kim et al.[16]investigated the effect of driver age and gender on injury severity in single-vehicle collision accidents using the RP-logit model.The RP-logit model allows explanatory variables to be spontaneous; however,individual heterogeneity in the random parameters can still exist.To address this issue,Mannering et al.[17]introduced a method that relates the mean and variance of random parameters with other explanatory variables in the model when analyzing highway traffic accidents to determine individual heterogeneity.Yang et al.[18]employed the RP-logit model with heterogeneity in means and variances (RP-logit-HMV) to explore the heterogeneity of factors affecting the severity of accidents involving nonhelmeted motorcycle riders and various vehicle types.The abovementioned study thoroughly investigates the effect of heterogeneity on the modeling and analysis of motor vehicle traffic accidents,introducing novel perspectives to traffic accident analysis.Therefore,incorporating the RP-logit or RP-logit-HMV models proves advantageous in addressing heterogeneity in shared e-bike injury data and investigating potential interactions among variables.

In this research,data on shared e-bike insurance accidents from a shared e-bike enterprise during the period of May to September 2020 were obtained.Combining information such as shared e-bike riders’,accident,road,and weather attributes,the RP-logit-HMV,RP-logit,and binomial logit models were used for fitting and comparative analysis.Furthermore,the optimal model was used to analyze the injury severity of shared e-bike riders and provide a theoretical basis for developing traffic safety strategies to protect shared e-bike riders.

1 Data Preparation

The dataset used in this research was comprised of two components.1) 2 667 shared e-bike accident data spanning from May to September 2020,sourced from an insurance accident management platform of a shared e-bike company.The platform documents accident specifics,including location,time,causation,and detailed information about the accident,along with the name,gender,and date of birth of the rider for each insured accident.After deleting the missing values in the original data,1 343 accidents were retained as the foundational dataset for this research.2) Historical weather data,including weather conditions,temperature,wind direction,and other relevant indicators,were retrieved from a query website[19].We aligned the weather data with the time and location of each accident to establish an accurate correlation.

The traffic accidents in the insurance accident database of the shared e-bike company are divided into three categories—not injured,injured,and fatal accidents—based on the injury severity of shared e-bike riders.Among the 1 343 accidents obtained,395 were not injured accidents,945 were injured accidents,and three were fatal accidents.Considering that fatal accidents account for only a small proportion of the sample,they were combined with injured into one level,namely injured,for subsequent modeling analysis.

The transportation system encompasses a multitude of factors,including traffic participants,vehicles,roads,and the environment.Any problem within these elements could potentially result in traffic safety issues and cause accidents.Therefore,considering the effect of various factors on accidents,12 variables were selected based on shared e-bike rider,vehicle,road,and environmental attributes as candidate explanatory variables.The coding and descriptive statistics for each explanatory variable are presented in Tab.1.

Tab.1 Descriptive statistics for variables

2 Method

2.1 Random parameter logit model with heterogeneity in means and variances

The utility function for the injury severity of shared e-bike riders is shown as follows:

Sin=βiXin+εin

(1)

whereSinrepresents a severity function determining the probability of injury severity categoryi(not injured,injured) in crashn;Xinrepresents the vector of explanatory variables;βirepresents a vector of estimable parameters;εinrepresents the random error.

Allowing heterogeneity in the means and variances of random parameters entails reformulating Eq.(1) so thatβinbecomes a vector of estimable parameters that varies across observations.

βin=βi+δinZin+σinexp(ωinWin)vin

(2)

whereβirepresents a mean parameter estimate in all accidents;Zinrepresents a vector of attributes that represents heterogeneity in the means;δinrepresents a corresponding vector of estimable parameters;Winrepresents a vector of attributes that captures heterogeneity in the standard deviationσinwith parameter vectorωin;vinrepresents a disturbance term (a term with a random distribution that captures unobserved heterogeneity among accidents).

The probability of injury severity categoryiattributable to the crashn,Pn(i),is expressed by allowing the vectorβinto obtain a continuous density function,implying that Prob(βn=β)=f(β|φ),

(3)

wheref(β|φ) represents the density function ofβwithφreferring to the vector of parameters (mean and variance) of that density function,and other terms are as previously defined.

2.2 Marginal effect

The RP-logit-HMV cannot quantify the magnitude of the effect of each explanatory variable on the probability of the injury severity outcome; therefore,the marginal effect is used to evaluate the effect of a unit increase in each explanatory variable on the probability of the injury severity outcome.The marginal formula is shown as follows:

(4)

2.3 Binomial logit model

A binomial logit model is used as a comparison model herein.The logit regression equation can be expressed as follows:

(5)

wherePrepresents the probability when the injury severity is classified as injured;Xrepresents the vector of explanatory variables,βrepresents a vector of estimable parameters;β0represents a constant.

When injured,the injury severity probability for a given value of the vector of explanatory variablesXcan be theoretically calculated as follows:

(6)

3 Parameter Identification and Results

3.1 Parameter identification

In this study,the NLogit software and the Monte Carlo method were used to establish the RP-logit model with heterogeneity in means and variances,with a significance level of 0.1.The solution process was as follows:

1) As no closed-form solution exists for the coefficient solution of the RP-logit-HMV,only the simulation solution can be used.The simulation solution involves random and Halton sequence sampling.The Halton sequence sampling exhibits high efficiency and uniform distribution of sampling points.Consequently,the paper adopted Halton sequence sampling,with a sampling frequency of 200[11].

2) Before model estimation,the specification of the probability density function for the parameters was essential.Probability density function forms of parameters exhibit uniform distribution,normal distribution,and log-normal distribution.The normal distribution is superior to other probability density functions[20].Therefore,it was assumed that all explanatory variables to be estimated were random parameters,and the parameters to be evaluated were assumed to follow a normal distribution in the simulation.The simulation results revealed that the opposing side of the accident fault exhibited the attributes of random parameters,and the remaining explanatory variables were fixed parameters.

3) The RP-logit-HMV was established along with the RP-logit and binomial logit models using the NLogit software.The parameter estimation results are shown in Tab.2.Notably,no variance heterogeneity existed in the model; thus,RP-logit-HMV eventually degenerates into the RP-logit model with heterogeneity in means (RP-logit-HM).The comparison of the three models is presented in Tab.3.The log-likelihood value of the RP-logit-HM was-658.161,with an Akaike information criterion (AIC) value of 1 340.300,a Bayesian information criterion (BIC) value of 1 402.800,anR2value of 0.293,and an adjustedR2value of 0.287.The results reveal that the RP-logit-HM exhibits a better fit and interpretation.

Tab.2 Parameter estimation of the three logit models

Tab.3 Measure-of-fit of the three logit models

3.2 Results

Comparing the results of the RP-logit-HM with the RP-logit and binomial logit models revealed that the influence of factors such as female,damaged road surface,wet road pavement,nighttime,both at fault,vehicle-pedestrian accident,and vehicle-to-vehicle accident on the injury severity of shared e-bike riders were in the same direction (see Tab.2).In contrast,in the case of other scenarios such as those involving different traffic participants at fault and component damage,the RP-logit-HM identified the stochastic and heterogeneous attributes of these variables,thereby mitigating potential biases in model parameter estimation and inference.Consequently,only the parameter outcomes of the RP-logit-HM model are discussed in subsequent sections of the paper.

3.2.1 Random parameter

According to Tab.2,the variable “other traffic participants at fault” within accident scenarios followed a normal distribution of random parameters.As shown in Fig.1,the variable “other traffic participants at fault” followed a normal distribution characterized by a mean of 2.028 and a standard deviation of 2.175.This indicates that about 82.380% of shared e-bike riders in accidents were more susceptible to injuries when other traffic participants were at fault.In contrast,about 17.620% of shared e-bike riders were less prone to injury if the accident was attributed to other traffic participants.

Fig.1 Distribution of random parameters

3.2.2 Heterogeneity in the mean of the random parameter

Tab.2 shows that the mean of the random parameter for the variable “other traffic participants at fault” was influenced by vehicle status,specifically component damage.The mean value of the random parameter for the variable “other traffic participants at fault” exhibited a negative correlation with the presence of component damage in the vehicle.As shown in Fig.2,the mean value of the parameter for accidents involving other traffic participants at fault was 2.028 and decreased to 0.064 when the accident involved component damage.The result suggests that while the likelihood of injury among shared e-bike riders increased when other traffic participants were at fault,the presence of component damage mitigated this escalation.One possible explanation is that the shared e-bike platform promptly immobilized and prohibited the use of vehicles after a user reported vehicle component damage.Consequently,this measure curtailed accidents arising from vehicle component damage,thereby reducing the likelihood of rider injury.

Fig.2 Distribution of random parameters with heterogeneity in means for other traffic participants at fault

3.2.3 Fixed parameter

According to the model calculation results (see Tab.3),several factors exhibited significant effects on the injury severity of shared e-bike riders.Hence,in this section,the factors influencing injuries sustained by shared e-bike riders,in conjunction with marginal effects (see Tab.4),were analyzed.

Tab.4 Marginal effect of various parameters in the RP-logit-HM

1) Shared e-bike riders’ attributes.The gender of shared e-bike riders significantly influenced injury severity.Following accidents,female shared-e-bike riders displayed a higher likelihood of injury than males,with the probability of female injury escalating by 2.200%.This trend is attributable to multiple factors,such as the greater weight of shared e-bikes,the potential difficulties female riders face in maneuvering these bikes,and the less ability of female riders to control and react to accidents compared with males[21].The above factors led to a higher probability of injury for female shared e-bike riders than males after an accident.

2) Road attributes.The condition of the road surface significantly influenced the severity of accident injuries.The likelihood of shared e-bike riders sustaining injury under damaged road surface conditions was 0.900% lower than that under intact road surface conditions.Moreover,the state of the road pavement significantly affected injury severity,and the likelihood of shared e-bike riders sustaining injuries under wet road pavement conditions was 0.700% lower than that under dry pavement conditions.Interestingly,previous research yielded inconsistent results concerning the effects of road surface and pavement on accidents.Wang[1]found that road pavement did not significantly influence accident severity.In contrast,Hertach et al.[8]identified damaged roads and difficulties maintaining balance owing to wet road surfaces as significant accident contributors.Although riding under conditions of damaged road surfaces and wet road pavement entails elevated risk,our research indicated that accidents might not necessarily occur,attributable to a shift in the psychological approach of shared e-bike riders when confronted with unfavorable road conditions.Riders often adopt anticipatory measures,exhibiting defensive driving behaviors such as heightened focus and cautious operation to mitigate the risks associated with poor road conditions.

3) Environmental attributes.The likelihood of causing injury to shared e-bike riders was higher during nighttime than during the daytime,with the probability of injury for those traveling at night increasing by 1.500%.This phenomenon is attributable to reduced rider visibility and the accelerated speeds of shared e-bikes during nighttime.These factors lead to diminished perception and reaction time for accident participants to evade danger or mitigate personal harm[22].

4) Accident attributes.Both accident fault and accident form displayed significant effects on injury severity.Relative to single-vehicle accidents,incidents involving vehicle-pedestrian and vehicle-to-vehicle collisions were 4.500% and 14.200% less likely to cause injury to shared e-bike riders,signifying that single-vehicle accidents were more prone to causing injuries.This outcome diverged from findings in prior research[23].Riders might be unfamiliar with the structural mechanics of shared e-bikes,potentially leading to improper or excessively abrupt braking and unsuitable or excessive speeds[8].Inexperienced riding and a lack of awareness regarding traffic safety among shared e-bike riders could also contribute to these disparities[24].Traffic management must prioritize the safety of shared e-bike riders.When both parties are at fault in an accident,the likelihood of injury for shared e-bike riders increases by 1.200% compared with accidents where only the rider is responsible.This is attributable to the reduced reaction time for shared e-bike riders in situations involving mutual fault,thereby elevating the risk of injury[1].

The results of the paper could provide a basis for improving the traffic safety of shared e-bike riders: 1) The mean heterogeneity of “other traffic participants at fault” and component damage reveals that while component damage can mitigate the degree of injury escalation,regular vehicle inspections and maintenance by shared e-bike enterprises are essential to enhance vehicle safety performance.2) To address road surface,pavement,and lighting factors,governmental initiatives should focus on enhancing road traffic facilities,optimizing riding environments,incorporating road lighting facilities,and reinforcing road maintenance efforts.3) Addressing gender discrepancies,accident faults,and accident forms necessitates increased traffic safety advocacy by traffic management authorities.Moreover,the awareness of law-abiding behavior should be enhanced,with specific attention to training programs for women and other vulnerable groups to enhance their protection awareness.

4 Conclusions

1) The variable “other traffic participants at fault” exhibited characteristics of a random parameter with a normal distribution and demonstrated mean value heterogeneity within accident fault scenarios.Specifically,when the vehicle had component damage,accidents in which the fault lay with other traffic participants resulted in an increased likelihood of injury to shared e-bike riders.However,the presence of component damage could mitigate the extent of this increase.

2) The model estimation and marginal effect outcomes highlight that various factors—such as female riders,intact road surfaces,dry road pavements,nighttime incidents,single-vehicle accidents,and accidents involving both parties at fault—contribute to varying degrees in elevating the probability of injury among shared electric bike riders.

3) This study elucidated the factors influencing injury severity among shared e-bike riders.According to insurance accident data,potential data heterogeneity was addressed.The insights provided serve as valuable references for shaping policies aimed at enhancing the safety of shared e-bike traffic systems.Nevertheless,certain limitations exist within this study.The insurance accident records for shared e-bikes cannot cover all pertinent accident-related information.Future research could more comprehensively analyze factors impacting the injury severity of shared e-bike riders by obtaining more accurate and comprehensive shared e-bike traffic accident data.Moreover,according to the extensive data from shared e-bike trajectories,an all-encompassing analysis of safety,incorporating risky riding behaviors,and proposing corresponding enhancement measures could better elucidate the current state of shared e-bike safety.