Analysis of Environmental Factors on Intersection Accidents

Abstract

In recent years, the number of fatal traffic accidents has been on the rise in Taiwan, with a total of 2865 fatalities in 2019, representing a 3% increase from the previous year, and indicating an urgent need for the improvement of Taiwan’s traffic conditions. This research focuses on the three-way intersections on provincial roads. In Taiwan, such three-way intersections account for more than 70% of all intersections on provincial roads. From 2018 to 2019, there were 41,137 accidents at three-way intersections, accounting for about 50% of the traffic accidents in Taiwan. Relevant research in Taiwan has been mostly focused on driver characteristics and other factors were less addressed. In this study, we looked into the environmental factors, conducted analysis and the results were proposed for future roadway safety improvements. This study uses a regression model for analysis, in which the dependent variable is the number of traffic accidents at each intersection. However, for 68% of the three-way intersections surveyed, the number of traffic accidents recorded during the observation period was zero. Therefore, this study uses zero-inflated models to construct a model to identify important variables that affect the severity of traffic accidents at three-way intersections. The research results show that two types of factors are related to the risk of traffic accidents at three-way intersections. The first type includes the width of the provincial road, the width of the shoulder, the width of the dividing line and the number of lanes, while the second type relates to the presence of convenience stores, gas stations, supermarkets, and other attractions, such as public retail markets at the intersection.

1. Introduction

Accident data analysis has long been used as a reference basis for road and vehicle design and the formulation of safety policy. Many accident-related issues have been discussed in the international community, such as the identification of accident-prone locations and analysis of accident factors. However, the frequency of traffic accidents has continued to rise, requiring continued scrutiny of traffic safety-related issues.

Previous research on traffic accident factors in Taiwan has focused on the impact of driver characteristics on accidents. However, in addition to driver characteristics, accidents are caused by additional factors, including vehicle, road conditions and environmental conditions. Therefore, improvement strategies that also incorporate an understanding of road and environmental factors will be more effective in reducing the risk of traffic accidents. This research analyzes accident data to identify key road and environmental factors that affect the frequency of traffic accidents as the basis for developing strategies to improve traffic safety.

Previous studies of accident frequency mostly used the Poisson regression model (Ma, 2009 [1]; Ye et al., 2013 [2]), duration mode (Chung, 2010 [3]; Jovanovic et al., 2011 [4]), multivariate models (Castro et al., 2012 [5]; Caliendo et al., 2013 [6]), neural networks (Abdelwahab, 2001 [7]), Bayesian neural networks (Xie et al., 2007 [8]), and vector machine models (Yu and Abdel-Aty, 2013 [9]). Analysis of accident severity relied on logit and probit models (Moudon et al., 2011 [10]; Santolino et al., 2012 [11]), ordered logit and probit models (Mergia et al., 2013 [12]; Ye, 2014 [13]); and artificial neural networks (Delen et al., 2006 [14]; Chimba, 2009 [15]).

This study considers the random and scarce data characteristics of traffic accidents, based on a review of the relevant literature (Miaou, 1994 [16]; Qin et al., 2004 [17]; Lord et al., 2007 [18]; Malyshkina and Mannering, 2010 [19]). When too many data points are recorded as zero (i.e., Y-type intersections with zero accidents), one can use the zero-inflated model to analyze and incorporate the characteristics of the probability distribution between Poisson and the negative binomial, and then construct zero-inflated Poisson (ZIP) and zero-inflated negative binomial models (ZINB). Model analysis can identify road condition and environmental factors that affect the incidence of traffic accidents to propose corresponding improvement suggestions.

2. Background and Literature Review

2.1. Research Scope and Objectives

Provincial roads are the subject of study and they constitute the backbone of the road network in Taiwan. Some of them are intercity roadways, some are urban corridors and some are rural roads, as shown in Figure 1.

The data from the National Land Surveying and Mapping Center in Taiwan indicates the following: there are 29,208 intersections in provincial roads in Taiwan and 70% of them are three-way intersections. About 50% of traffic accidents happened in these three-way intersections, from 2018 to 2019, and this is the motivation of this study.

The number of yearly traffic accidents for these three-way intersections in 2018 and 2019 are listed in

Table 1. Statistical summary of accidents at intersections.

Number of Accidents Which Resulted in Injuries or Fatalities	Number of Intersections (2018)	%	Number of Intersections (2019)	%
0 Accidents	13,555	69.23%	13,063	66.72%
1~10 Accidents	5705	29.14%	6191	31.62%
11~20 Accidents	239	1.22%	250	1.28%
21~30 Accidents	48	0.25%	53	0.27%
31~40 Accidents	19	0.10%	15	0.08%
41~50 Accidents	8	0.04%	6	0.03%
Over 51 Accidents	6	0.03%	2	0.01%
Total	19,580		19,580

. Most of them have a value of zero, meaning no traffic accident happened on them in 2018 or 2019.

2.2. Variables of Accident-Prone Location

This section summarizes the important influencing variables used in the previous literature and individually explains each one, with respect to factors related to road conditions and the environment, seen below:

• Traffic:
  • Road width: road width mainly reflects the relationship between flow and capacity. Past studies have shown that the degree of road risk will decrease with decreased road width (Greibe, P, 2003 [20]; Abhishek Das, 2011 [21]).
  • Speed limit: the higher the speed limit, the faster the velocity of steady traffic, which increases accident severity. Research has shown that the road speed limit is related to accident frequency. The higher the speed limit, the higher the frequency of accidents (Xuesong Wang, 2006 [22]; Abhishek Das, 2011 [21]).
  • Central divider: studies have shown that the installation of a central divider will increase the severity of accidents (Greibe, P, 2003 [20]; Xuesong Wang, 2006 [22]; Abhishek Das, 2011 [21]).
  • Sidewalk: studies have shown that accident severity is also increased by the presence of sidewalks (Greibe, P, 2003 [20]; Anderson, 2009 [23]; Abhishek Das, 2011 [21]).
  • Traffic flow: most studies invest in the analysis of traffic volume (ADT or AADT) factors and find that it is correlated with the severity of accidents and the frequency of accidents. The greater the traffic flow, the higher the frequency and severity of accidents (Greibe, P, 2003 [20]; Xuesong Wang, 2006 [22]; John C. Milton, 2008 [24]; Abhishek Das, 2011 [21]; Sandra Vieira Gomes, 2012 [25]; Chunjiao Dong, 2014 [26]; Cantillo, 2015 [27]).
• Geographic:
  • Stations: relevant studies have shown that accidents are more likely to occur around stations (Greibe, P, 2003 [20]; Anderson, 2009 [23]).
  • Schools: relevant research results show that accident-prone areas in metropolitan areas typically depend on the surrounding environment and land use. For example, accidents often occur near schools (Anderson, 2009 [23]).
  • Business district: relevant research results show that commercial use areas have relatively high accident frequencies, and the types and functions of buildings along roads also have an impact on the risk of accidents (Greibe, P, 2003 [20]; Anderson, 2009 [23]).

3. Methodology

Many previous studies on the causes of traffic accidents have used analysis methods such as regression analysis, multivariate analysis, and neuroanalysis. However, if too many data points in the data sample are recorded as zero it will cause greater variation in the count data, along with over-dispersion in the zero-inflated model.

Shankar et al. [28] empirically explored the applicability of the zero-inflation model for road section accident frequency analysis. Although some data showed zero values, these data still have the potential to reflect accident risk. The Poisson regression and negative binomial analysis models cannot be used for observation, so their research attempted analysis using the zero-inflated model. The research results found that the distribution of zero-inflated Poisson (ZIP) and zero-inflated negative binomial (ZINB) are in line with reality. The accident pattern is suitable for exploring the causal relationship between related variables and the possible accident risk of data points recording zero accidents.

Based on the above, this study adopts two research methods (ZIP and ZINB) to construct the following analysis model.

3.1. Zero-Inflated Poisson, (ZIP)

The zero-inflated Poisson regression model contains two regressions, the logistic regression (LR) and the Poisson regression (PR). The logistic regression explains whether the sample’s characteristics will affect the probability of the sample becoming a “zero” value, as follows:

$${\mathsf{\omega }}_i=\frac{{\pi }_i}{1+{\pi }_i}$$

(1)

where ${\pi }_i$ is the logit function of the data sample, and ${\mathsf{\omega }}_i$ is the probability that this data sample is zero.

The logit function is calculated as follows:

$${\pi }_i=exp\left[{\beta }_0+{{\displaystyle \sum }}_{n=1}^k{\beta }_n{X}_{ni}\right]$$

(2)

where ${\beta }_0$ is model intercept obtained by the Logit regression, ${\beta }_n$ is the logistic regression coefficient obtained by the correction of the nth characteristic variable, and $X_{ni}$ is the nth characteristic variable of the ith data sample. This study collected the surrounding environment and traffic information for three-way intersections on provincial roads in 2018 and 2019, as seen in Table 2.

The Poisson regression interprets the probability of accidents caused by the characteristics of the data sample. The calculated expected value is combined with the probability of occurrence of “zero” given by the Logit regression, to obtain the overall ZIP result, as follows:

Assuming accident occurrences Y = ($y_1$, $y_2$, …, $y_n$) are mutually independent, the Poisson probability density function is Equation (3).

$$g\left(y_i\right)=\frac{exp\left(-{\lambda }_i\right){\lambda }_i{}^{y_i}}{y_i!}$$

(3)

The probability density function of the ZIP overall mode, below, is:

$$P\left(y_i\right)=\left\{\begin{array}{ll}{\mathsf{\omega }}_i+\left(1-{\mathsf{\omega }}_i\right)\times exp\left(-{\lambda }_i\right),& y_i=0\\ \left(1-{\mathsf{\omega }}_i\right)\frac{exp\left(-{\lambda }_i\right){\lambda }_i^k}{k!},& y_i>0\end{array}\right.$$

(4)

where ${\lambda }_i$ is the expected value obtained from the Poisson regression, and $k$ in this study is the accident occurrence in the data sample Y = ($y_1$, $y_2$, $y_3$, …, $y_n)$.

3.2. Zero-Inflated Negative Binomial (ZINB)

If the data suffers from over-dispersion, the above-mentioned ZIP regression model is changed to the ZINB regression model, that is, the original Poisson model distribution is changed to a negative binomial distribution, and the probability density function as Equation (5), shown below:

$$P\left(y_i\right)=\left\{\begin{array}{c}{\mathsf{\omega }}_i+\left(1-{\mathsf{\omega }}_i\right){\left(\frac{{\alpha }^{-1}}{{\alpha }^{-1}+{\lambda }_i}\right)}^{{\alpha }^{-1}}, y_i=0\\ \left(1-{\mathsf{\omega }}_i\right)\frac{\Gamma \left({\alpha }^{-1}+y_i\right)}{\Gamma \left({\alpha }^{-1}\right)\Gamma (y_i+1)}{\left(\frac{{\alpha }^{-1}}{{\alpha }^{-1}+{\lambda }_i}\right)}^{{\alpha }^{-1}}{\left(\frac{{\lambda }_i}{{\alpha }^{-1}+{\lambda }_i}\right)}^{y_i},\hspace{1em} y_i>0\end{array}\right.$$

(5)

where $y_{\mathrm{i}}$ is the observed value; $x_i$ is the matrix of related factors that affect the probability of the overall event, and ${\mathsf{\omega }}_i$ is the probability value that reflects the special zero value generation mechanism.

4. Empirical Study

4.1. Data Description

This study collected data on accidents causing death and injury in Taiwan in 2018 and 2019, using Geographic Information Systems to conduct spatial analysis of such accidents at three-way intersections on Taiwan’s provincial roads, as shown in Figure 2.

4.1.1. Response Variable in Model

Accounting for road conditions and variables related to the surrounding environment for each intersection, the number of accidents at each intersection is calculated separately for 2018 and 2019, with a dataset consisting of a total of 39,160 records.

4.1.2. Independent Variables in Model

This study uses GIS to identify traffic information for three-way intersections on provincial roads in 2018 and 2019, using the characteristics of the surrounding environment as independent variables, like those summarized in

Table 2. Variables used in the analysis.

Label	Variable	Type	Descriptive Statistics	Researcher
$y_1$	Accident	Continuous	The number of accident-related fatalities and injuries, for three-way intersections on provincial roads in 2018 and 2019.	-
$x_1$	Station	Binary	(1) Station near intersection: 2638 (6.7%) (2) No station near intersection: 36,522 (93.3%)	[20,23]
$x_2$	School	Binary	(1) School near intersection: 13,310 (36.5%) (2) No school near intersection: 24,850 (63.5%)	[23]
$x_3$	Hospital	Binary	(1) Hospital near intersection: 4158 (10.6%) (2) No hospital near intersection: 35,002 (89.4%)	-
$x_4$	Market	Binary	(1) Market near intersection: 8088 (20.7%) (2) No market near hospital: 31,072 (79.3%)	-
$x_5$	Gas Station	Binary	(1) Gas station near intersection: 12,796 (32.7%) (2) No gas station near intersection: 26,364 (67.3%)	-
$x_6$	Department store	Binary	(1) Department store near intersection: 536 (1.4%) (2) No department store near intersection: 38,624 (98.6%)	-
$x_7$	Convenience store	Binary	(1) Convenience store near intersection: 5384 (13.7%) (2) No convenience store near intersection: 33,776 (86.3%)	-
$x_8$	Parking lot	Binary	(1) Parking lot near intersection: 2554 (6.5%) (2) No parking lot near intersection: 36,606 (93.5%)	-
$x_9$	Speed limit	Classification	(1) Under 40 km/h: 2194 (5.6%) (2) 40~49 km/h: 3210 (8.2%) (3) 50~59 km/h: 10,356 (26.5%) (4) 60~69 km/h: 13,528 (34.6%) (5) 70~79 km/h: 9272 (23.7%) (6) Above 80 km/h: 600 (1.53%)	[21,22]
$x_{10}$	Road width(unit: m)	Continuous	Minimum: 2 Maximum: 88 Mean: 18.74 Standard deviation: 8.74	[20,21]
$x_{11}$	Main lane width(unit: m)	Continuous	Minimum: 2 Maximum: 34 Mean: 7 Standard deviation: 3.11	-
$x_{12}$	Slow vehicle lane width(unit: m)	Continuous	Minimum: 0 Maximum: 9.5 Mean: 1 Standard deviation: 1.55	-
$x_{13}$	Central divider(unit: m)	Continuous	Minimum: 0 Maximum: 2.28 Mean: 0.6 Standard deviation: 1.05	[20,21,22]
$x_{14}$	Sidewalk width(unit: m)	Continuous	Minimum: 0 Maximum: 7 Mean: 0.9 Standard deviation: 1.76	[20,21,23]
$x_{15}$	Traffic Flow	Continuous	Minimum: 2.41 Maximum: 4 Mean: 5.22 Standard deviation: 0.42	[20,21,22] [24,25,26,27]

Note: “Near” indicates the POI is within 300 m in radius of the intersection.

.

The data of environmental factors and traffic are from MOI Open Data. Traffic Flow is AADT (Annual Average Daily Traffic) and the values are taken the natural logarithm.

4.2. Analytic Model

This study used a regression analysis model to explore the factors affecting accident incidence. The accident data is statistical data with non-negative integer values; thus, data analysis was performed using the Poisson regression and negative binomial regression. However, an excessive number of data points have a value of zero, which may lead to poor adaptation of ZIP and ZINB. Therefore, the Poisson and negative binomial regressions are combined with the zero-inflated model, to produce ZIP and ZINB, to identify significant influencing factors that affect the incidence of accidents at three-way intersections on provincial roads.

The results of the ZIP model (

Table 3. ZIP Model Calibration Results.

First Segment: Intersections with No Accident Casualties (Zero-Inflated, Logit)
${\mathit{X}}_{\mathit{i}}$		Estimate	Std.Error	Pr(>|z|)	${\mathit{X}}_{\mathit{i}}$		Estimate	Std.Error	Pr(>|z|)
Intercept term		5.581	0.180	<0.000	Road width		−0.052	0.002	<0.000
School	No	--	--	--	Traffic flow		−0.706	0.041	<0.000
	Yes	−0.264	0.030	<0.000	Speed limit	Under 30 km/h	--	--	--
Shopping mall	No	--	--	--		40 km/h	−0.298	0.112	0.008
	Yes	−0.639	0.036	<0.000		50 km/h	−0.683	0.098	<0.000
Gas station	No	--	--	--		60 km/h	−0.587	0.097	<0.000
	Yes	−0.345	0.029	<0.000		70 km/h	−0.517	0.100	<0.000
Department store	No	--	--	--		Above 80 km/h	0.435	0.168	0.010
	Yes	−0.306	0.113	0.007	Fast lane width		−0.013	0.005	0.010
Convenience store	No	--	--	--	Slow lane width		0.027	0.009	<0.000
	Yes	−0.748	0.039	<0.000	Sidewalk width		−0.052	0.009	<0.000
Parking lot	No	--	--	--
	Yes	−0.272	0.053	<0.000
Second Segment: Intersections with Accident Casualties (Poisson)
${\mathit{X}}_{\mathit{i}}$		Estimate	Std.Error	Pr(>|z|)	${\mathit{X}}_{\mathit{i}}$		Estimate	Std.Error	Pr(>|z|)
Intercept item		−1.918	0.086	<0.000	Road width		0.021	0.001	<0.000
Station	No	--	--	--	Traffic flow		0.388	0.017	<0.000
	Yes	0.034	0.015	0.024	Speed limit	Under 30 km/h	--	--	--
Hospital	No	--	--	--		40 km/h	0.400	0.051	<0.000
	Yes	0.134	0.013	<0.000		50 km/h	0.316	0.042	<0.000
Shopping mall	No	--	--	--		60 km/h	0.220	0.043	<0.000
	Yes	0.257	0.013	<0.000		70 km/h	0.078	0.044	0.075
Gas station	No	--	--	--		Above 80 km/h	−0.149	0.080	0.062
	Yes	0.110	0.011	<0.000	Fast lane width		0.028	0.002	<0.000
Department store	No	--	--	--	Slow lane width		0.009	0.003	0.002
	Yes	0.127	0.025	<0.000	Central divider width		0.011	0.005	0.022
Convenience store	No	--	--	--	Sidewalk width		0.021	0.003	<0.000
	Yes	0.196	0.012	<0.000
Parking lot	No	--	--	--
	Yes	0.215	0.014	<0.000

) and the ZINB model (

Table 4. ZINB Model Calibration Results.

First Segment: Intersections with No Accident Casualties (Zero-Inflated, Logit)
${\mathit{X}}_{\mathit{i}}$		Estimate	Std.Error	Pr(>|z|)	${\mathit{X}}_{\mathit{i}}$		Estimate	Std.Error	Pr(>|z|)
Intercept term		6.097	0.416	<0.000	Road width		−0.068	0.009	<0.000
School	No	--	--	--	Traffic flow		−1.026	0.105	<0.000
	Yes	−0.447	0.091	<0.000	Speed limit	Under 30 km/h	--	--	--
Shopping mall	No	--	--	--		40 km/h	−0.088	0.223	0.694
	Yes	−2.191	0.363	<0.000		50 km/h	−0.033	0.207	0.872
Gas station	No	--	--	--		60 km/h	−0.456	0.203	0.025
	Yes	−0.902	0.108	<0.000		70 km/h	−0.536	0.220	0.015
Department store	No	--	--	--		Above 80 km/h	1.692	0.346	<0.000
	Yes	−3.555	1.239	0.004	Fast lane width		−0.053	0.021	0.011
Convenience store	No	--	--	--	Sidewalk width		−0.392	0.062	<0.000
	Yes	−0.668	0.219	0.002
Second Segment: Intersections with Accident Casualties (Negative Binomial)
${\mathit{X}}_{\mathit{i}}$		Estimate	Std.Error	Pr(>|z|)	${\mathit{X}}_{\mathit{i}}$		Estimate	Std.Error	Pr(>|z|)
Intercept item		−3.689	0.167	<0.000	Road width		0.041	0.002	<0.000
Station	No	--	--	--	Traffic flow		0.475	0.032	<0.000
	Yes	0.112	0.035	0.001	Speed limit	Under 30 km/h	--	--	--
School	No	--	--	--		40 km/h	0.632	0.110	<0.000
	Yes	0.054	0.026	0.040		50 km/h	0.641	0.090	<0.000
Hospital	No	--	--	--		60 km/h	0.335	0.091	<0.000
	Yes	0.145	0.029	<0.000		70 km/h	0.196	0.092	0.034
Shopping mall	No	--	--	--		Above 80 km/h	0.053	0.155	0.731
	Yes	0.399	0.028	<0.000	Fast lane width		0.034	0.004	<0.000
Gas station	No	--	--	--	Sidewalk width		0.038	0.006	<0.000
	Yes	0.142	0.024	<0.000
Convenience store	No	--	--	--
	Yes	0.360	0.027	<0.000
Parking lot	No	--	--	--
	Yes	0.264	0.035	<0.000

) show that the set of variables that significantly affect death- and injury-causing accidents at three-way intersections are largely the same, and the coefficient directions are also the same. The first part of the model accounts for no fatalities or injuries within the intersection area (Logit model), outlined below:

• Most of the significant variables are the same for both models, and they all show negative effects, indicating that there are environmental features for intersections that can be altered to reduce the probability of death or injury.
• The presence of facilities in which people congregate together (e.g., schools) and commercial developments (e.g., shopping malls, convenience stores), significantly increases the risk of accidents resulting in fatalities and injuries.

The second part of the model accounts for the number of casualties at an intersection (Poisson or Negative binomial model), described below:

• Most of the significant variables for the two models are the same and both show positive effects, indicating there are certain characteristics of intersections that increase the number of accident-related fatalities and injuries.
• Most of the significant variables have positive effects and are related to commercial development or the gathering of crowds, such as schools, hospitals, department stores, convenience stores, etc.
• Variables related to the presence of vehicle entrances and exits (e.g., gas stations and parking lots) increase disruption to traffic flows and thus have a significant impact on the number of accidents.
• There is an absolute positive relationship between traffic volume and the number of accidents, related to the concept of traffic exposure. Therefore, the greater the traffic volume at an intersection, the higher the risk of accidents. In the two regression models, the coefficient values of the traffic flow variables are larger than other variables, so traffic flow is clearly sensitive to accident severity.
• In addition, the greater the road width, the higher the accident risk, possibly because wider roads correlate with higher vehicle speeds, while allowing different vehicle types (e.g., trucks and buses) to operate in the same driving space, and traffic flow at intersections is also more complicated, leading to a higher accident risk.

4.3. Model Evaluation

The following is based on the proposed zero-inflated Poisson and zero-inflated negative binomial models. Different index values are used to judge and compare the suitability of each model, including the log-likelihood, AIC (Akaike information criterion) and BIC (Bayesian information criterion) of each model. Smaller AIC and BIC index values mean that the selected model fits the original data well (Lord and Mannering, 2010 [29]). The results are shown in

Table 5. Comprehensive comparison of model calibration results.

	Zero-Inflated Poisson	Zero-Inflated Negative Binomial
Sample size	39,160	39,160
log-likelihood	−52,099.83	−42,538.91
AIC	104,271.7	85,143.82
BIC	104,580.4	85,426.82

. The AIC and BIC values of the zero-inflated negative binomial model are smaller than those of the zero-inflated Poisson model and, thus, fit better. Most accident data feature discrete problems and the Poisson regression model generally has limitations for making assumptions in the data analysis. Due to these limitations, the negative binomial model has gradually become more widely used because it can handle over-dispersed data, where the average value is much greater than the variance. The negative binomial distribution of the zero-inflated negative binomial model is a better fit for the data analyzed in this research.

4.4. Discussion

This research conducted analysis using zero-inflated Poisson (ZIP) and zero-inflated negative binomial (ZINB) and found important variables that affect the incidence of intersection accidents, to facilitate subsequent research on intersection safety improvement strategies. According to the comparison results of different models, the negative binomial distribution is better suited to the data, and the zero-expansion negative binomial model is a better fit than the zero-expansion Poisson.

Traffic flow is a key influencing variable, and the incidence of accidents at three-way intersections increase with traffic volume. However, in terms of actual strategy implementation, traffic volume cannot be directly reduced, leaving only time or spatial diversion measures, or reducing the number of intersection conflict points to reduce the incidence of accidents.

Accident frequency is positively correlated to the presence of commercial districts or places where people gather, such as schools, hospitals, shopping malls, department stores, and convenience stores, where cars stop to drop off or pick up passengers, along with pedestrian traffic volume that can easily increase traffic merging and other points of conflict. In addition to enhancing the enforcement and prohibition of moving violations, it is also important to increase the protections for crossing pedestrians and to enhance measures to prevent speeding and reckless passing.

Analysis results show that parking lots and gas stations are both important variables for accident risk. When building new parking lots or gas stations, attention should be taken to ensure entrances and exits are located away from intersections. At the same time, improved signage, potentially using lights or sound alerts, should warn through-drivers of potential merging vehicles to prevent accidents. In addition, in terms of strategy, strong law enforcement is needed to prevent vehicles exiting gas stations or parking lots from failing to merge properly, thus reducing the accident risk.

The findings show that wider lane widths increase accident risk. Narrower driving lanes would limit maneuverable space for cars, prompting drivers to reduce speed and, therefore, reduce accident risk.

5. Concluding Remarks

The results of this research correspond well with prior findings, such as the significance and directionality of significant variables, comparing the results with related studies in other countries. The present findings can serve as a useful reference to improve intersection design and safety. Recommendations include reducing vehicle speed and applying traffic management methods to segregate vehicles in time and space; thus, reducing accident risk.

Furthermore, points of interest, such as schools and commercial areas near an intersection, usually lead to a higher risk of traffic accident incidence. In order to lower the risk and improve safety, the location of future POI, the speed limit of intersections with POI nearby, etc., should be taken into consideration. The effectiveness of safety measures can be analyzed in future studies based on the findings of this study.

Follow-up research could consider the incorporation of artificial intelligence research methods, such as ANN, CART, etc., to analyze factors contributing to traffic accidents, raising the possibility of predicting accidents at intersections. In addition, the current research focuses exclusively on three-way intersections, but future research could consider four- or more way intersections.