Influence of Built Environment on Micromobility–Pedestrian Accidents

Abstract

The use of micromobility (MM), a form of sustainable urban mobility which has expected effects such as reducing traffic congestion and greenhouse gases, has been rapidly increasing across the world. However, this growth has resulted in a considerable number of MM-related accidents. Most previous studies have explored MM user injuries to improve the safety of MM users, but the threat to pedestrians by MM is not yet fully understood. Therefore, this study aims to identify built environment factors which contribute to MM–pedestrian accidents by using MM–pedestrian crash data in Seoul, Korea from 2020 to 2021. Setting the spatial unit of analysis as a hexagonal grid with an apothem of 150 m, we developed the SZINB (spatial zero-inflated negative binomial) models for the accidents, controlling spatial autocorrelation, zero-inflated, and overdispersion. The model results showed that road intersections, sidewalks, and subway entrances have significant impacts on MM–pedestrian accidents. Thus, it should be suggested that safety measures for both MM and pedestrians are reducing MM speed limits in intersections, preventing MM use on sidewalks through modified sidewalk designs, and installing MM stations near subway stations.

1. Introduction

Micromobility (MM, including e-scooters, e-bikes, hoverboards, segways, etc.) are gaining popularity around the world. For example, due to advantages such as the reduction in carbon emissions through electric energy, improvement in public transportation by acting as a first–last mile solution, and easy accessibility, the use of MM in the United States has increased rapidly with the introduction of shared e-scooter services in September 2017 [1]. Similarly, in Korea, MM has become an attractive means of transportation [2,3].

However, with the increase in the popularity and use of MM, MM-related accidents have also increased rapidly. In the United States, 8016 e-scooter-related crashes were identified in 2017, which increased to 14,641 in 2018 [4]. In Korea, 216 MM-related crashes occurred in 2017, which doubled to 449 in 2018 [3]. This upward trend has been continuing to this day, and in the case of Korea, such accidents have been increasing significantly. Referring to the Traffic Accident Analysis System (TAAS) in Korea, 876 MM accidents were identified to have occurred in 2019, 1525 cases in 2020, and 2842 cases in 2021. Most of the accidents occurred in Seoul, the capital city of Korea.

In this situation, many countries have been taking preventive measures against MM-related accidents. Most of these measures (limits on speed, age eligibility for MM use, wearing a helmet, where to ride, etc.) are aimed at the safety of MM users. As a measure for the safety of pedestrians, use on sidewalks has been banned. For example, e-scooters have been banned from pavements in France, and in Germany and United Kingdom, using MM on sidewalks is illegal [5]. Similarly, in Korea, the government has implemented policies for MM users such as requirement of a driver’s license, prohibition of MM operation under the influence of alcohol, prohibition of driving without helmets, and prohibition of driving on sidewalks. In spite of these efforts, MM-related accidents have increased, threatening the safety of pedestrians who are relatively vulnerable road users.

MM-related accidents can be divided into two categories: those where the MM user is the offender and those where the MM user is the victim. From 2019 to 2021 in Seoul, it was reported that there were 996 accidents in which the MM user was the offender and 491 accidents in which the MM user was the victim. That is, the former was almost twice as many accidents as the latter. This indicates that the introduction of MM threatened the safety of existing road users. Noting that most of the accidents were MM–pedestrian crashes in which the MM user was the offender and the pedestrian was the victim (41.5%), it should be recognized that MM can be a threat to the safety of pedestrians who are relatively vulnerable road users. Therefore, we aim to improve the safety of both MM users and pedestrians by identifying factors to affect MM–pedestrian crashes.

2. Literature Review

A number of recent studies have attempted to improve the safety of MM. Fang [6] investigated the situation and severity of injured micromobility users. He analyzed micromobility accident data from 2014 to 2019 from NEISS, which provides visitor records to emergency rooms for about 100 hospitals in the United States. It was found that most of the micromobility accidents were collisions with vehicles and occurred on the streets. Yang et al. [7] analyzed crash reports and news of 40 U.S. states where shared e-scooter services operated, from 2017 to 2019. It was reported that accidents mainly occurred at intersections, and accidents did not occur on bicycle roads. In addition, they showed that hitting pedestrians was the major type of accident that happened on sidewalks. These studies show that it is important to explore how built-environment factors are related to MM accidents.

To this end, Heydari et al. [8] identified the relationship between the neighborhood environment and the frequency of e-scooter accidents. They analyzed crash data from STATS19 in the United Kingdom, from January 2020 to June 2021. It was found that the frequency of e-scooter accidents increased as the walking and cycling levels increased. On the other hand, there was a negative relationship between the proportion of greenspace area and e-scooter accidents. Additionally, it was shown that the frequency of e-scooter accidents increased as the number of schools increased. Azimian and Jiao [9] analyzed the relationship between e-scooter accidents and socio-economic, transportation, and built environment factors. E-scooter accidents were collected from local news and Patch, an information platform, and the number of e-scooter accidents was counted by census areas. Analyzing these data, Azimian and Jiao [9] showed e-scooter accidents’ tendency to decrease in census areas with a high percentage of sidewalks. Additionally, the land use entropy index had a positive effect on the accidents, and it was found that e-scooter accidents tended to decrease in census areas with a high proportion of public transportation users. As such, most previous studies have attempted to improve MM user safety by analyzing injury accidents of MM users, leading to the possibility of an insufficient understanding regarding the impact of MM on pedestrians who are relatively vulnerable road users.

Gehrke et al. [10] explored the observed safety risks that e-scooter users may face. They investigated the relationship between e-scooter trips and the number of accidents related to vulnerable road users such as pedestrians, cyclists, skateboarders, and motorcyclists. They also identified the network and environmental factors affecting them. The travel data came from Lime e-scooter trip records from 1 April to 15 November 2019, and the accident data of vulnerable road users were collected from 2015 to 2019, in Brooklyn. Then, these data were aggregated to 300 m grid cells. It was found that accidents decreased in areas with a wide-open space. On the contrary, they showed that the more bus stops per square mile, the more vulnerable road users were to accidents. Oh and Kim [2] investigated the difference between personal mobility (PM, including e-scooters, e-bikes, hoverboards, segways, etc.)–vehicle accidents and PM–pedestrian accidents in Korea. PM accident data were collected from TAAS, from 2017 to 2019. Most of the PM–pedestrian accidents occurred while walking on sidewalks or crossing roads. In addition, the highest proportion of PM–pedestrian crashes occurred during walking on sidewalks in roads with both bicycle roads and sidewalks.

Additionally, for a better understanding of MM–pedestrian crashes, we reviewed the literature related to pedestrian crash analysis. Ukkusuri et al. [11] analyzed the relationship between pedestrian accidents and built environment factors in New York City. The data for pedestrian accidents were collected from the New York City Department of Transportation (NYCDOT), from 2002 to 2006, and were aggregated into census areas. They developed several negative binomial models for pedestrian accidents: a negative binomial regression model, negative binomial regression with heterogeneity in dispersion parameter model, and a zero-inflated negative binomial regression model. The model results showed that the greater industrial and commercial land use, subway stations, and transit ridership, the higher the pedestrian accidents. Lee et al. [12] investigated how the built environment affects the risk of elderly pedestrian accidents. Elderly pedestrian crashes were collected from TAAS, from 2015 to 2017, in Korea. They aggregated the data by dividing Seoul into 1419 “daeguyeog” which matches with the census block group. Negative binomial regression analysis was performed to investigate the determinants of pedestrian crashes. They found that the four-way intersections were positively related to elderly pedestrian accidents, and residence of multi-family housing was negatively correlated with elderly pedestrian crashes. In addition, land use and public transportation factors were found to be significantly associated with elderly pedestrian accidents.

According to the literature review, most previous studies focused on improving the safety of MM users, despite the fact that MM is also threatening the safety of pedestrians who are relatively vulnerable road users. In addition, previous studies of MM–pedestrian accidents had several limitations. First, Gehrke et al. [10] analyzed the relationship between e-scooter trips and vulnerable road user crashes with sophisticated spatial analysis units instead of using actual MM–pedestrian accidents. Thus, since this is not an analysis of the actual MM–pedestrian crash accident, there could be limitations. Second, Oh and Kim [2] conducted only a descriptive statistical analysis of PM–pedestrian accidents, and accident severity analysis was performed by focusing on PM user injury severity. This means that there is a lack of understanding of the MM–pedestrian accidents in terms of pedestrians. Therefore, considering the limitations of previous studies, we analyze the impact of transportation, land use, and built-environment factors on MM–pedestrian accidents by developing econometric models such as negative binomial regression models. This study could improve our understanding of MM and pedestrian safety and help develop relevant safety policies.

3. Data

MM–pedestrian accident data used in this study were collected from TAAS, which collects its data from accident records by local police agencies and insurance companies in Korea. We used 342 MM–pedestrian accidents in Seoul from 2020 to 2021 in consideration of the impact of pre- and post-COVID-19. Additionally, since bicycle accidents are collected separately in TAAS, they are not included in the MM accident data we used. The TAAS database provides accident information such as accident type (accident with motor vehicle, accident with pedestrian, etc.), offender vs. victim, gender, age, injury severity, time, season, weather, region, etc. However, TAAS does not provide accurate geographical location coordinates for each accident. Instead, TAAS provides location information on the accident site through GIS maps. Therefore, we used Geocoding (Qgis) to map locations to the most accurate coordinates. Based on the geocoded location, the number of MM–pedestrian accidents is considered as a dependent variable and transportation, land use, and the built-environment attributes are considered as explanatory variables.

In this study, explanatory variables were mainly employed from previous studies. First, the number of intersections [10,11,12] where MM users more often interact with pedestrians was collected from the National Spatial Data Infrastructure Portal (NSDIP) in Korea. Next, sidewalks and bicycle roads [2,9] were considered. At first, MM was prohibited from use on sidewalks and bicycle roads by law, but on 10 December 2020, MM was permitted to use bicycle roads in Korea. Regardless of the law, MM users tend to use sidewalks instead of roads when bicycle roads are not available [2]. Generally, MM is known to provide first-last mile connections for public transit [10]. As such, public transit facilities are expected to have high MM usage as well as high pedestrian volumes. Thus, location data of subway station entrances and bus stops were selected and were obtained from Forest Big Data Exchange Platform (FBDEP) and Seoul Open Data Plaza (SODP). In addition, land use information such as residential, commercial, and green areas was collected from NSDIP, often used in previous studies [2,8,9,10,11,12]. Furthermore, as the land use mix index is shown to have a negative effect on MM trips, we calculated the land use mix index [13]. Land use mix index has a value from 0 to 1, where 1 means “mixed land use” and 0 means “single land use”. Finally, since the hilly terrain with slopes is related to the use of MM [2,6], slope data from NSDIP were also selected. The slope data are provided as a GIS shape file with six scales (0~2%, 2~7%, 7~15%, 15~30%, 30~60%, and 60~100%). Considering the characteristics of the data, the slopes were converted to median values (1%, 4.5%, 11%, 22.5%, 45%, and 80%). Next, we calculated the average slope per grid using the area ratio of each slope as a weight.

The previous studies have used census blocks as the spatial analysis unit, but such blocks have limitations in reflecting spatial characteristics because shapes and sizes of such blocks are irregular. In addition, due to difference in distance between adjacent corner squares and adjacent side squares, the square grid has limits in reflecting spatial characteristics in relation to adjacent squares. In this respect, the hexagonal grid has the advantage that each adjacent hexagon is the same distance from each other. Therefore, in this study, we divided Seoul into a hexagonal grid with an apothem of 150 m and aggregated the dependent and explanatory variables at the grid basis. Additionally, grids of mountain areas and river areas where MM is not available were excluded. Finally, 5792 hexagonal grids were used for this study, as can be seen in Figure 1.

Descriptive statistics of MM–pedestrian accidents and explanatory variables at a hexagonal grid level are provided in

Table 1. Descriptive statistics.

Variable	Obs	Mean	Std. Dev.	Min	Max
The number of MM–pedestrian accidents	5792	0.059	0.278	0	4
The number of intersections	5792	1.158	1.169	0	7
Average slope (%)	5792	10.498	11.858	0	80
Proportion of sidewalk area (%)	5792	2.534	2.229	0	16.599
Average sidewalk width (m)	5792	3.276	1.856	0	13.21
Proportion of bicycle road area (%)	5792	0.278	0.473	0	4.121
The number of subway entrances	5792	0.433	1.828	0	32
The number of bus stops	5792	1.861	2.005	0	13
Ratio of residential area	5792	0.679	0.394	0	1
Ratio of commercial area	5792	0.057	0.176	0	1
Ratio of green spaces area	5792	0.208	0.348	0	1
Land use mix index	5792	0.154	0.204	0	0.900

. In terms of MM–pedestrian accidents, the average number of accidents per grid is 0.059. On average, there was one intersection per grid, the sidewalk area was 2.5% and the sidewalk width was 3.3 m. Bicycle roads accounted for a small proportion (0.28%) on average. There were 0.4 subway entrances and 1.9 bus stops per grid on average. The average land use mix index is 0.154.

For the MM–pedestrian accident distribution in Figure 2. It should be noted that the number of grids with zero was 5498 (94.92%), whereas the number of cells more than zero was 294 (5.08%). This indicates the presence of excessive zero-value grids in the MM–pedestrian accident data. Furthermore, when analyzing aggregated data in geographic analysis units, spatial correlation may exist. Therefore, we need to utilize an appropriate analysis method that can identify the presence of excessive zero-value grids and the spatial correlation of the data.

4. Methods

In general, since traffic accident data have the characteristics of count data which randomly occurs, the count model has been widely used in previous studies. The dependent variable used in this study is the number of MM–pedestrian accidents, so it is suitable for the count model. The count model is classified into a Poisson regression model and a negative binomial regression model (NB). The Poisson model assumes that the dependent variable y follows the Poisson distribution function of Equation (1), where μ is the mean of $Y$ [9,10,11,14,15,16,17].

$$\mathrm{Pr}\left(Y=y_i\right)=\frac{e^{-\mu }{\mu }^{y_i}}{y_i!}$$

(1)

The major characteristics of the Poisson distribution model is that the mean and variance are equal. However, the problem of overdispersion, instances in which the variance is larger than the mean, frequently occurs in reality. This causes bias regarding the standard error of the regression coefficient. To improve this problem, the negative binomial distribution of Equation (2) is generally applied by mixing Poisson distribution and Gamma distribution [9,10,11,14,15,16,17]. Here, $\mathsf{\Gamma }()$ is a Gamma function, and α is a parameter indicating overdispersion.

$$\mathrm{Pr}\left(Y=y_i\right)=\frac{\mathsf{\Gamma }\left(y_i+\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\alpha $}\right.\right)}{\mathsf{\Gamma }\left(y_i+1\right)\mathsf{\Gamma }\left(\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\alpha $}\right.\right)}{(\frac{1}{1+\alpha \mu })}^{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\alpha $}\right.}{(\frac{\alpha \mu }{1+\alpha \mu })}^{y_i}$$

(2)

We noted that the dependent variable (number of the MM–pedestrian accidents) has excessive zero-values. This is not fit when using the count model. To solve this problem, zero-inflated count models have been proposed [18]. This model uses both binary and count processes. The binary process uses logit or probit, and the count process uses Poisson or NB. In the zero-inflated count model, the probability distribution of the dependent variable y is divided into always 0 group and 1 group (Poisson or NB). This can be written as Equation (3) [9,14,15,16,17,18]. In this regard, the inflation (binary process) model interprets belonging to always 0 group as an “occurrence of event”.

$$\mathrm{Pr}\left(Y=y_i\right)=\{\begin{array}{l}{P}_i+\left(1-P_i\right)f\left(y_i\right), y_i=0\\ \left(1-P_i\right)\frac{{\mu }_i^{y_i}f\left(y_i\right)}{y_i!}, y_i>0\end{array}$$

(3)

Zero-inflated count models can be classified into zero-inflated Poisson regression model (ZIP) and zero-inflated negative binomial regression model (ZINB). In Equation (3), the zero-inflated count model becomes ZIP if $f\left(y_i\right)$ is a Poisson model, and the Zero-inflated count model becomes ZINB if it is a negative binomial model. $f\left(y_i\right)$ used in a ZIP model can be written as Equation (4), and $f\left(y_i\right)$ used in a ZINB model can be written as Equation (5) [9,14,15,16,17,18].

$$\mathrm{ZIP}: f\left(y_i\right)=e^{-{\mu }_i}$$

(4)

$$\mathrm{ZINB}: f\left(y_i\right)=\frac{\mathsf{\Gamma }\left(y_i+\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\alpha $}\right.\right)}{\mathsf{\Gamma }\left(y_i+1\right)\mathsf{\Gamma }\left(\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\alpha $}\right.\right)}{(\frac{1}{1+\alpha {\mu }_i})}^{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\alpha $}\right.}{(\frac{\alpha {\mu }_i}{1+\alpha {\mu }_i})}^{y_i}$$

(5)

However, these models (Poisson, NB, ZIP, and ZINB) have limitations in considering spatial correlation. To improve this problem, a method controlling spatial autocorrelation has been proposed [19]. In general, spatial autocorrelation can be calculated using a variable (MM–pedestrian accident in this study) between a spatial unit and another adjacent spatial unit. By adding this to the independent variable in the model estimation process, spatial correlation can be controlled. $Mora{n}^{\prime }s I$ is the widely used statistic to measure spatial correlation and can be written as Equation (6) [20]. Where $\overline{Y}$ is the average $Y_i$ and ${\omega }_{ij}$ is a weight representing the relationship between regions $i$ and $j$.

$$I=\frac{n}{{{\displaystyle \sum }}_i{{\displaystyle \sum }}_j{\omega }_{ij}}\times \frac{{{\displaystyle \sum }}_i{{\displaystyle \sum }}_j{\omega }_{ij}\left(Y_i-\overline{Y}\right)\left(Y_j-\overline{Y}\right)}{{{\displaystyle \sum }}_i{(Y_i-\overline{Y})}^2}$$

(6)

Therefore, we first tested $Mora{n}^{\prime }s I$ to diagnose if there is a spatial correlation, then determined the appropriate model among the spatial zero-inflated Poisson (SZIP), spatial zero-inflated negative binomial (SZINB), ZIP or ZINB models.

5. Model Results

Figure 3 shows the results of $Mora{n}^{\prime }s I$ for MM–pedestrian accidents. It turns out that $Mora{n}^{\prime }s I$ was 0.197 (significant at $p-value=0.1$). This means that spatial correlation exists. In addition, it is found that the high–high area has a lot of MM usage and floating population [2,3].

We tried to identify transportation, land use, and built-environment factors to affect MM–pedestrian accidents using SZIP or SZINB, since spatial correlation exists. Then, we selected SZINB because overdispersion exists ($\alpha =0.24$, ($chibar2\left(01\right)=2.44$, $Pr\ge chibar2=0.059$)). Finally, we employed the SZINB model because both spatial correlation and overdispersion exist.

Furthermore, goodness-of-fit measures were compared with other models such as NB and ZINB: AIC (Akaike information criterion), BIC (Bayesian information criterion), and log-likelihood. It is interpreted that the lower the AIC and BIC, and the higher the value of log-likelihood, the better model.

Table 2. Analysis results of the NB, ZINB, and SZINB models.

Variable	NB		ZINB		SZINB
	β	IRR	β	IRR	β	IRR
Count model
Constant	−4.164 ***	0.016 ***	−3.275 ***	0.038 ***	−2.736 ***	0.065 ***
The number of intersections	0.035	1.036	−0.278 ***	0.757 ***	−0.235 ***	0.791 ***
Average slope (%)	−0.011	0.990	0.003	1.003	0.008	1.008
Proportion of sidewalk area (%)	0.066 *	1.068 *	0.092 *	1.096 *	0.084 *	1.088 *
Average sidewalk width (m)	0.233 ***	1.262 ***	0.281 ***	1.325 ***	0.130 **	1.139 **
Proportion of bicycle road area (%)	−0.100	0.905	−0.393 **	0.675 **	−0.042	0.959
The number of subway entrances	0.079 ***	1.082 ***	0.039 **	1.040 **	0.036 *	1.037 *
The number of bus stops	0.040	1.041	0.089 **	1.093 **	0.010	1.010
Ratio of residential area	0.353	1.423	0.207	1.230	−0.011	0.989
Ratio of commercial area	0.187	1.206	−1.002 *	0.367 *	−0.976 *	0.377 *
Ratio of green spaces area	−2.666 ***	0.070 ***	−2.252 ***	0.105 ***	−2.609 ***	0.074 ***
Land use mix index	0.389	1.475	0.218	1.243	0.016	1.016
Spatial autocorrelation					1.813 ***	6.130 ***
Inflation model
Constant			−0.473	−0.473	0.706	0.706
The number of intersections			−1.380 ***	−1.380 ***	−1.322 ***	−1.322 ***
Average slope (%)			0.032	0.032	0.039 *	0.039 *
Proportion of sidewalk area (%)			0.145	0.145	0.088	0.088
Average sidewalk width (m)			0.249	0.249	0.066	0.066
Proportion of bicycle road area (%)			−1.484 **	−1.484 **	−0.664	−0.664
The number of subway entrances			−0.285 *	−0.285 *	−0.474	−0.474
The number of bus stops			0.251 **	0.251 **	0.081	0.081
Ratio of residential area			0.315	0.315	−0.003	−0.003
Ratio of commercial area			−36.835 ***	−36.835 ***	−26.496 **	−26.496 **
Ratio of green spaces area			−1.606	−1.606	−2.954	−2.954
Land use mix index			5.582 ***	5.582 ***	4.859 **	4.859 **
N	5792		5792		5792
α	1.858 ***		0.809 ***		0.240 *
Log likelihood	−1172.604		−1136.688		−1079.267
AIC	2371.208		2323.376		2210.535
BIC	2457.843		2489.982		2383.805

*** p < 0.01, ** p < 0.05, * p < 0.1

shows the estimation results of the NB, ZINB, and SZINB models. It can be seen that the SZINB model has the lowest value of AIC and BIC, and the highest value of log-likelihood. Thus, we focused on evaluating and interpreting the results of SZINB.

6. Discussions

The SZINB model results showed that the number of intersections, proportion of sidewalk area, average sidewalk width, number of subway entrances, ratio of commercial area, and ratio of green spaces area significantly affected the number of MM–pedestrian accidents at $p-value=0.1$ in the count model. Slightly differing from the count model, the inflation model results showed that the number of intersections, average slope, ratio of commercial area, and land use mix index significantly affected the number of MM–pedestrian accidents at $p-value=0.1$.

For the count model in SZINB, areas with a large number of intersections tended to have a low frequency of MM–pedestrian accidents. On the other hand, in the inflation model, areas with a large number of intersections were found to have a higher probability of accidents. The inflation model result was consistent with those reported by previous studies [10,11,12]. In contrast, the result of the count model was found to be different from the results of the previous literature. The intersection is typically a space where MM users and pedestrians interact, so conflicts may occur between MM users and pedestrians. On the other hand, because both MM users and pedestrians decrease their speed at intersections, a situation where both parties can avoid collision could be created. This result shows that the frequency of MM–pedestrian accidents tends to be low in intersections where accidents occur. This can be seen as a result of MM users and pedestrians paying attention at intersections where accidents occur. Based on these results, it is important to separate pedestrians and MM users in intersections. In addition, it is necessary to improve intersections so that the speed of MM users can be sufficiently reduced. For example, a recommended measure to reduce speed is by using speed bumps on roads approaching the intersections and by using rumble strips to warn MM users. Additionally, visual elements such as traffic signs are also good methods to give warnings to MM users. However, these facilities should not cause inconvenience and risk to MM users.

The average slope showed a positive coefficient in the inflation model. This means that areas with a high average slope tend to decrease the probability of accidents. This result is consistent with the results of previous studies, which show that fewer accidents occur in hilly areas [6]. It makes sense that areas with high slopes affect the speed of MM users, as the speed of MM decreases on an uphill slope and decreases due to careful driving of MM users on a downhill slope.

For the sidewalk variable, it was found that the proportion of sidewalk area and the average sidewalk width tend to increase the frequency of MM–pedestrian accidents in the count model. This result is different from the result reported by Azimian and Jiao [9], but consistent with the result reported by Oh and Kim [2]. However, the result by Azimian and Jiao [9] is different from this study because their study analyzed the injury accidents of e-scooter users. Accordingly, because the sidewalk is a relatively safer space for e-scooter users than the road, it is expected that the number of e-scooter user injuries decrease. However, since this study focused on MM–pedestrian accidents, the results should be interpreted from a different perspective. MM users tend to use sidewalks when sidewalks are wide and available [2]. Therefore, accidents in areas with a high proportion and large width of sidewalks are likely to occur due to increased MM user use on such sidewalks. In Korea, using on sidewalks is illegal, but such regulations are not well-enforced. Thus, strict enforcement of the law should be implemented to solve this problem. Moreover, there is a need for sidewalk design that prevents MM use on sidewalks. For example, traffic calming is a method to provide a safe road environment for pedestrians by reducing the speed and traffic of vehicles. If such a method is introduced to sidewalks, reducing the speed and traffic of MM, accidents can be reduced. With this in mind, public agencies should establish long-term plans and funding for sustained enforcement of such policies in heavily used MM areas.

With respect to public transit factors, the number of bus stops had no significance. On the other hand, differing from the results by Azimian and Jiao [9], the number of subway entrances tended to increase the number of accidents in the count model. This contrast is due to the difference between their analysis of MM user injuries and our analysis of MM–pedestrian accidents. In general, the subway station is a place with a large amount of MM usage and pedestrians. It is expected that there will be frequent conflicts between MM users and pedestrians at the subway station. Thus, an area with a large number of subway entrances is likely to have a larger number of accidents because there may be many conflicting points between MM users and pedestrians. Spatial separation between MMs and pedestrians is essential in such areas. A highly recommended measure to facilitate such separation is the MM station. However, many currently installed MM stations are located in front of subway entrances, so there are persisting conflicts with pedestrians. In order to solve such problems, the installation of MM stations in consideration of pedestrians and MM movement can be useful. Accordingly, public agencies need to design and fund MM stations considering spacing required for safe MM–pedestrian interaction.

For land use variables, the ratio of commercial area tended to decrease the number of MM–pedestrian accidents in the count model and tends to increase the probability of accidents in the inflation model. The result of the inflation model is consistent with the previous studies [9,11,12]. MM usage is positively related to regions with high employment rates [10,13], so this may increase the probability of accidents. On the contrary, the ratio of commercial land use was found to decrease the frequency of accidents in the count model. This is related to the period of data. The MM–pedestrian accident data used in this study are from 2020 to 2021 when the impact of COVID-19 existed. Globally, COVID-19 has brought many changes to our society. Many companies implemented working from home, and government policies in most countries restricted personal travel. As a result, there was a decrease in personal trips for shopping and business purposes [21]. Taking into account that the impact of COVID-19 in Korea grew during the period of analysis, the ratio of commercial land use can be interpreted as reducing the frequency of accidents.

In addition, the ratio of green space areas tended to decrease the number of MM–pedestrian accidents in the count model. This is consistent with the result reported by Heydari et al. [8]. In Seoul, Hangang Park is a representative green space. The public agencies approved the use of MM in Hangang Park on 10 December 2020. Due to its direct relation to increased MM use, such a policy could be expected to bring an increase in accidents. However, Hangang Park was well-installed with bicycle roads (length = 78.0 km), separating pedestrians and MM users. Moreover, it is illegal to use MM in most other green spaces in Seoul, Korea. Therefore, it can be inferred that green spaces have a negative correlation with the number of MM–pedestrian accidents.

The land use mix index showed a tendency to decrease the probability of an accident in the inflation model. This is related to the negative correlation between land use mix and e-scooter usage reported by Bai and Jiao [13]. According to their analysis, balanced land use reduces e-scooter usage. Thus, due to areas with mixed land use having less MM usage, the land use mix index can be seen as having negative correlation with probability of accidents. Furthermore, this means that single land use areas have a higher probability of MM–pedestrian accidents than mixed land use areas. Considering that internal trips mainly occur in mixed land use areas and pass-through trips mainly occur in single land use areas, government departments should focus on single land use areas when implementing policies and should identify and mange routes with high pass-through and MM usage.

7. Conclusions

The spread of MM around the world has also brought a rapid increase in MM related accidents. Recognizing the potential threat of MM to relatively vulnerable pedestrians, this study aimed to understand MM accidents not only from the perspective of MM users, but also from the perspective of pedestrians. To provide an improved understanding of MM–pedestrian accidents, we selected Seoul in Korea as our study area to identify transportation, land use, and built environmental factors contributing to MM–pedestrian accidents. Then, in consideration of data attributes and spatial correlation, we proposed the SZINB model advanced from the NB and the ZINB models, confirming that the SZINB model estimated data better and was the most suitable model for this study.

As a main result of the SZINB model, it was found that road intersections, sidewalks, subway entrances, and the land use mix index were significantly correlated with MM–pedestrian accidents. Based on these results, we proposed recommended measures such as speed bumps, rumble strips, traffic signs, traffic calming, and MM stations.

This study could provide a better understanding of MM–pedestrian accidents, as previous studies focused on the safety of MM users, but we focused on the safety of pedestrians threatened by the introduction of MM. Additionally, this study could contribute academically by proposing a methodology that considers zero excess and spatial autocorrelation, which are the limitations of the previous studies.

However, this study has some limitations. First, there are various sharing MM service providers in Seoul. In this regard, the number of sharing MM service companies and devices operating in the region were unavailable. Second, we did not identify individual-level factors such as injury severity, age, and gender. Therefore, there is a need to address such factors in future work to provide better insights.