Data Mining Approach to Explore the Contributing Factors to Fatal Wrong-Way Crashes by Local and Non-Local Drivers

Abstract

Despite significant research efforts into wrong-way driving crashes, the fatality rate in the United States remains persistently high year after year. However, few studies have concentrated on how the driver’s familiarity with the road affects wrong-way driving. This study aims to examine if there is a difference in contributing factors to fatal wrong-way driving crashes by local and non-local drivers by utilizing Fatality Analysis Reporting System (FARS) data from 2016 to 2020. Descriptive statistics were first used to give insight into the data, and then the association rule mining method was applied to help uncover the hidden connections between contributing factors to wrong-way driving crashes for both local and non-local drivers. The findings indicated that several factors, including intoxicated drivers, an urban environment, and late-night hours from 12 A.M. to 6 A.M., play a significant role in causing local wrong-way driving crashes. On the other hand, non-lighted conditions in a rural setting significantly contributed to fatal wrong-way driving crashes by non-local drivers.

1. Introduction

A driver’s familiarity with road geometric design and conditions is crucial in enabling them to make better decisions and avoid potential crashes; this is because being aware of the road layout and its condition allows drivers to navigate safely and respond appropriately to various situations.

Wrong-way driving (WWD) is characterized by the movement in the opposite direction of traffic flow [1]. The resultant crashes represent a significant concern for traffic safety due to the higher severity compared to other types of crashes. The nature of WWD crashes, primarily involving head-on collisions or opposite-direction sideswipes, contributes to the increased severity of their consequences [2].

The occurrence of WWD crashes is relatively low, yet their fatality rate is 1.34, meaning that in each fatal wrong-way driving crash, on average, 1.34 fatalities occur. It surpasses that of all other crashes with 1.10, and this higher fatality rate results in 24 additional fatalities per 100 fatal crashes in comparison to the general occurrence of fatal crashes [3]. The elevated fatality rate observed in WWD crashes necessitates an investigation of the underlying factors contributing to such incidents.

A study used the distance between drivers’ homes and crash locations as a proxy to indicate the familiarity or unfamiliarity of drivers. This study focused on single and two-vehicle crashes [4]. However, a few studies have examined the impact of drivers’ familiarity with the road on wrong-way crashes. In this regard, WWD crashes are divided into two different categories: local drivers and non-local drivers. Considering the presumption that local drivers possess familiarity with the road, making a mistake in finding the correct direction is not expected. Therefore, the motivation behind local wrong-way drivers, as well as the factors behind non-local drivers, must be examined.

The definition of locality for crashes is a controversial topic. Some researchers have proposed using state borders as a criterion, but this is not ideal, as the average vehicle trip length in the United States was 12 miles in 2017 [5], and the state drivers, on average, travel approximately 45 miles [3]. According to [3], one way to determine the locality status of WWD crashes is by considering the borders of the counties.

In this research, drivers involved in a crash within the county where they reside are classified as local drivers, whereas those who experienced a crash outside their residential county (either within the same state or in a different state) are categorized as non-local drivers.

After categorizing WWD crashes into local and non-local categories using the driver’s ZIP code, association rule mining was employed to identify the factors contributing to the fatalities of wrong-way drivers. The results of this study are expected to assist Departments of Transportation and transportation agencies in implementing countermeasures specific to the needs of different users.

The Fatality Analysis Reporting System (FARS) is an excellent resource for conducting safety research due to the extensive details it contains. It is a thorough national database overseen by the National Highway Traffic Safety Administration (NHTSA). FARS records fatal motor vehicle accidents occurring on public U.S. roads, offering detailed insights into each incident, including accident circumstances, involved vehicles and drivers, environmental factors, and timing. However, it is important to note that FARS exclusively contains police-reported fatal crashes, which means it cannot be used to draw conclusions about injury-only crashes or those resulting solely in property damage. This extensive dataset plays a vital role in analyzing trends and formulating road safety enhancement strategies.

2. Literature Review

The primary research areas concerning WWD crashes encompass three main aspects: identifying (i) high-risk crash locations, (ii) factors contributing to crashes, and (iii) effective countermeasures [3].

Extensive investigations in previous studies revealed that significant factors leading to WWD crashes are impaired drivers, nighttime driving, urban areas, early morning hours, male drivers, older drivers, weekends, dark-but-not-lighted conditions, the geometric design of exit/entrance roads and crossroads, the placement of WRONG WAY signs, pavement arrows, and high annual average daily traffic (AADT) on entrance ramps [1,3,6,7,8,9].

Various methods are utilized to identify crash contributing factors, including descriptive statistics, logistic regression, multiple correspondence analysis (MCA), machine learning, and data mining approaches like association rule mining. Subsequently, considerable effort is dedicated to developing traditional countermeasures and introducing new technologies. Recently suggested countermeasures to address the phenomenon of WWD involve lowered and enhanced signs, reflective pavement marking, additional signs, LED attached on WRONG WAY sign borders, rectangular flashing beacon (RFB)-equipped signs, rectangular rapid flash beacon (RRFB), and two RFB WW signs [3,10,11].

In spite of all the efforts made, there is still a significant prevalence of WWD crashes [3]. Previous studies on WWD crashes have primarily focused on the crash location, contributing factors, and countermeasures.

As drivers become more familiar with certain roads, they are more prone to violating minor traffic rules and engaging in riskier driving behaviors, demonstrating a decrease in attentiveness on those familiar roads [4], but assessing the impact of a driver’s local or non-local status in WWD crash incidents has been infrequently conducted [3].

The definition of the locality and non-locality of a driver has been a long-standing topic of debate among researchers [3]. In the United States, drivers from outside the state are referred to as non-local drivers [3,12,13,14,15]. Based on the previous study, utilizing the county border as a parameter proves to be a successful method in determining a driver’s locality status [3].

Association rule mining is a potent tool in machine learning for discovering connections between various factors because it identifies groups of items that frequently co-occur [16]. This is why researchers aim to uncover extensive patterns with the association rule mining method since numerous factors can lead to WWD crashes [16]. Various recent research has utilized association rule mining to identify meaningful correlations concerning various safety issues [16,17,18,19,20].

Association rule mining (ARM) is a practical approach in data mining that uncovers connections between variables within large databases [21]. Agrawal [22] initially introduced it in 1993 to analyze transactional databases [21]. Descriptive statistics may not be enough to establish complex relationships among variables. At the same time, parametric models rely on assumptions about the distribution of independent and dependent variables that may not be accurate. Consequently, the use of ARM can be helpful [16]. The utilization of ARM has resulted in notable advancements in various fields, including transportation safety, and various researchers in this field are utilizing it [21,23,24].

Various applications of the ARM method in transportation safety research have been implemented. The study referenced as [23], presents an ARM framework to examine key factors associated with accident severity in datasets of run-off-the-road crashes. ARM is utilized on the dataset to identify rules linked to crash severity. Another study referenced as [25] analyzed crash data from expressways using association rule mining to investigate the characteristics and factors influencing these incidents. It aims to improve traffic regulation enforcement and enhance transportation infrastructure, thereby boosting overall road safety.

In our study, we employ association rule mining to reveal the underlying factors responsible for WWD incidents. The analysis focuses on both local and non-local drivers, aiming to compare them and determine if there are any differences between the two groups.

3. Materials and Methods

3.1. Association Rule Mining

Association rule mining is a technique utilized to discover relationships between variables in big datasets.

Consider A = {a_1, a_2, …, a_n}, which includes all potential contributing factors to fatal wrong-way driving incidents, and T = {t_1, t_2, …, t_m}, where t_i represents the set of factors that contributed to incident i. Since we put all needed factors in set A, each member in T is a subset of items in A.

A rule is represented as C→D, where C and D are subsets of A and have no common elements (C ∩ D = ϕ).

C is referred to as the antecedent (left-hand side or LHS), and D is referred to as the consequent (right-hand side or RHS) [21]. Later, specific criteria will be introduced to evaluate the relevance and strength of a rule, such as C→D, with respect to the set T [26].

It is important to highlight that these rules represent relationships among factors, suggesting interdependencies rather than direct causation.

Association rules use three fundamental criteria: support, confidence, and lift [27].

The support denotes the fraction of all incidents in the database that contain a specific item [21].

Support(C) = Count(C)/M

Support(D) = Count(D)/M

Support(C⇒D) = Count (C ∩ D)/M

where M is the total number of members in the set T and Count (C) represents the number of appearances of item set C in the set T.

The accuracy of a rule is evaluated using confidence, which indicates the likelihood of incident D occurring given the occurrence of incident C [26]. Confidence is calculated using the following equation:

Confidence(C⇒D) = Support(C⇒D)/Support(C)

The lift is utilized to gauge the correlation between C and D, and as the correlation degree increases, the influence of C on D becomes stronger [25].

A lift value greater than 1 shows a positive correlation, while a value less than 1 indicates a negative correlation, and a lift value close to 1 suggests that C is independent of the likelihood of D [25].

The lift metric is calculated as the ratio of the confidence of the rule C→D to the support of D [26]. And it can be calculated using the equation below:

Lift(C⇒D) = Confidence(C⇒D)/Support(D)

In this study, the categorical data of crashes was converted to Boolean attributes using the one-hot encoding method.

The Apriori, FP-Growth (Frequent Pattern Growth), and Eclat algorithms are most widely used in association rule mining. We utilize the Apriori algorithm for this research due to its adaptability and efficiency.

The Apriori algorithm operates by identifying frequent item sets within the dataset and subsequently generating rules based on these item sets. It adopts a bottom-up approach, progressively increasing the size of the item sets until no frequent item sets can be identified [28]. The algorithm uses support, lift, and confidence measures to assess the strength and significance of the generated rules.

When exploring data with ARM, numerous association rules can be discovered. To assess the validity of an association rule, commonly used parameters include confidence, support, and lift. Consequently, it is necessary to establish minimum thresholds for each of these parameters [26]. Only association rules that surpass all thresholds should be reported.

Previous studies on traffic safety have commonly established minimum thresholds for support and confidence in the ranges of 0.01–0.1 and 0.6–0.7, respectively [26].

3.2. Variable Selection for Analysis Using the Random Forest Method

Choosing the right variables is a key process for pinpointing the significant factors to be incorporated into the ARM model. Including an excessive number of variables can lead to the generation of trivial patterns caused by noise, which complicates the interpretation of the outcomes [29].

The random forest method is extensively utilized in traffic safety research to identify the most significant predictors based on mean decrease accuracy.

In our article, we employed the random forest method for feature engineering due to its robust ability to handle high-dimensional data and its effectiveness in ranking predictor importance based on mean decrease accuracy. The random forest method was employed to prioritize the significance of attributes initially identified based on prior studies aiming to narrow down the variables considered for the ARM analysis. We focused on variables widely recognized in previous studies as significant factors in WWD crashes to thoroughly explore their influence within our study’s scope.

The principles of random forest involve constructing multiple decision trees during training and outputting the mode of the classes for classification tasks.

3.3. Overview of Materials and Methods

Our study employs a two-step methodological approach to analyze fatal wrong-way driving incidents using the FARS dataset. Initially, we utilize the random forest method for feature engineering, taking advantage of its ability to handle high-dimensional data and rank the importance of attributes. This step helps identify and focus on the most significant variables for further analysis.

Subsequently, we apply association rule mining techniques using the Apriori algorithm to discover relationships between these key variables in the dataset. Through this combined approach, we identify important rules based on the factors contributing to fatal wrong-way driving crashes, which can inform the development of effective countermeasures.

4. Data Collection

Fatality Analysis Reporting System (FARS) data used for this study are from 2016 to 2020. FARS is a nationwide database for fatal crashes, overseen by the National Highway Traffic Safety Administration (NHTSA). It records motor vehicle crashes on public roads that lead to at least one fatality [30].

Similar to the study [31], to extract Wrong-Way Driving (WWD) data from the FARS database files, the following attributes were used:

• Roadway Function Class: This field includes attribute codes for Rural Principal Arterial Interstate, Rural Principal Arterial Other, Urban Principal Arterial Interstate, Urban Principal Arterial Other Freeways or Expressways, and Urban Other Principal Arterial.
• Trafficway Flow: This field encompasses attribute codes for Divided Highway—Median Strip (Without Traffic Barrier), Divided Highway—Median Strip (With Traffic Barrier), One-way Trafficway, and Entrance/Exit Ramp.
• Sequence of Events: This field includes the attribute code for Cross Median/Centerline crashes.
• Violations Charged: This field includes attribute codes for Driving Wrong Way on One-way Road and Driving on Left—Wrong Side of Road.
• Driver-Related Factors: This field includes attributes for Driving Wrong Way on One-way Traffic and Driving on Wrong Side of Road (Intentional or Unintentional).

Incidents regarding these criteria are classified as WWD crashes. These incidents may involve single-vehicle crashes or multiple vehicles, but our dataset specifically includes only those drivers identified as WWD.

Finally, data from 1875 crashes in 50 states and one district were used for this study.

4.1. Analysis to Define Local and Non-Local Drivers

First, based on the drivers’ ZIP code in the FARS data, the state and the county of the drivers’ residences were found. In this regard, the “requests” library in Python was used to communicate with the OpenStreetMap Nominatim API to retrieve state and county information based on the provided ZIP codes. A driver is classified as a county driver if the crash and home locations are within the same county’s boundary. If the crash location is outside the driver’s home county but still within the same state, they are referred to as state drivers. Lastly, drivers from outside the drivers’ home state’s boundary are labeled as out-of-state drivers [3]. Here, a driver is considered a local driver if the crash location is in the same county as their home county; otherwise, they are considered a non-local driver.

4.2. Descriptive Statistics

• Variable Selection

In this paper, the primary area of investigation focuses on various attributes associated with WWD crashes based on previous studies [32,33,34].

In our variable selection process, we prioritized factors that have been extensively documented in previous studies as significant contributors to WWD crashes. This approach aimed to ensure a thorough exploration of key variables influencing WWD incidents within the scope of our study.

These include factors such as weather conditions, time of the crash, season, land use, lighting conditions, age of the driver, driver condition, drug usage, gender, road horizontal alignment, road profile, surface condition, and the fatalities of the wrong-way driver.

4.3. Crash Contributing Factor Analysis

The contributing factors to the crash, along with their categories and distribution considering three different driver groups (out of state, in the same county, in the same state), are presented in detail in

Table 1. Crash factors for different categories of wrong-way drivers.

Variable	Categories	Out of State	County	State
Weather	Clear	62%	68%	70%
	Cloudy	14%	16%	15%
	Other (Not Reported, Fog, Smog, Smoke)	16%	11%	9%
	Rain	8%	5%	6%
Days of the Week	Weekday	60%	59%	59%
	Weekend	40%	41%	41%
Crash Time	12 A.M.–6 A.M.	52%	48%	49%
	12 P.M.–6 P.M.	10%	12%	9%
	6 A.M.–12 P.M.	8%	6%	8%
	6 P.M.–12 A.M.	30%	29%	28%
Season	Fall	28%	27%	26%
	Spring	22%	21%	24%
	Summer	27%	26%	28%
	Winter	23%	26%	22%
Land Use	Rural	48%	25%	43%
	Urban	52%	75%	57%
Lighting Condition	Dark-Lighted	23%	36%	28%
	Dark-Not Lighted	58%	43%	55%
	Daylight	17%	17%	15%
	Other	2%	4%	2%
Age	Middle Adult (25 to 45)	48%	46%	48%
	Middle Aged (46 to 65)	19%	18%	18%
	Older (65+)	18%	18%	17%
	Young (Less Than 25)	15%	18%	16%
Driver Condition	No (Alcohol Not Involved)	22%	20%	22%
	Not Reported	11%	10%	11%
	Reported as Unknown	14%	9%	11%
	Unknown (Police Reported)	8%	8%	9%
	Yes (Alcohol Involved)	45%	53%	47%
Drug Usage	No (Drugs Not Involved)	36%	38%	34%
	Not Reported	26%	25%	25%
	Reported as Unknown	14%	10%	11%
	Unknown	10%	8%	9%
	Yes (Drugs Involved)	14%	19%	21%
Gender	Female	25%	28%	30%
	Male	74%	72%	70%
Road Horizontal Alignment	Curve-Left	4%	5%	5%
	Curve-Right	3%	5%	5%
	Curve-Unknown Direction	4%	2%	1%
	Curve Left	3%	4%	2%
	Curve Right	2%	4%	2%
	Not Reported	1%	2%	1%
	Straight	83%	79%	84%
Road Profile	Downhill	4%	6%	4%
	Grade, Unknown Slope	11%	10%	10%
	Hillcrest	2%	3%	4%
	Level	74%	70%	72%
	Not Reported	4%	7%	5%
	Sag (Bottom)	0%	0%	1%
	Uphill	5%	4%	4%
Surface Condition	Dry	85%	88%	87%
	Wet	14%	11%	13%
Is WWD Driver Dead?	Alive	27%	25%	27%
	Dead	73%	75%	73%

.

Based on Table 1, rainy weather has a significant influence on out-of-state drivers. Rain can hinder visibility, especially during nighttime, potentially posing challenges for drivers who are unfamiliar with the area, which is consistent with the findings of [3]. Also, most of the WWD crashes happened in clear weather, which aligns with the findings of [34].

It is also important to note that more than half of the alcohol-involved wrong-way crashes were local drivers, higher than the “out of state” and “same state” categories.

Another important note is that the accuracy of drug usage data in the FARS dataset is questionable, as many states do not consistently conduct or report drug testing results. Therefore, these data likely reflect suspected drug involvement rather than confirmed cases.

5. Results

5.1. Variable Selection Results for Analysis

Initially, using the random forest method, we sorted the variables based on their importance.

These 11 attributes are land use (urban/rural), lighting condition, trafficway description, police reported alcohol involvement, police reported drug involvement, horizontal alignment, age, weekday, hour, driver’s gender and weather.

For initial feature selection, we considered variables that significantly contributed to the model’s accuracy. The model’s accuracy was validated using cross-validation techniques, ensuring reliability. Variables with low importance scores were excluded to enhance model efficiency and interpretability without compromising performance.

Based on the results of this method, as shown in

Table 2. Results of random forest method for variable selection.

Variable Name	Ranking
RUR_URBNAME.x	1
LGT_CONDNAME	2
VTRAFWAYNAME	3
VALIGNNAME	4
DRINKINGNAME	5
WEATHERNAME	6
DRUGSNAME	7
AGE	8
HOUR	9
WEEKDAY	10
SEXNAME	11

, we decided to remove the variable “SEXNAME” since it had the lowest mean decrease accuracy, significantly lower than the “WEEKDAY” variable, which was second from the bottom.

5.2. Association Rule Mining Results

Setting appropriate minimum support and confidence values is crucial in association rule mining because a low support threshold can result in many unimportant rules, while a high support value may miss significant relationships between attributes. Finding the right balance is crucial for obtaining meaningful and interesting results [35]. Also, a lift amount higher than 1 is considered here to show that the presence of the antecedent (the item or items on the left-hand side of the rule) positively impacts the consequent (the item on the right-hand side of the rule).

Here, for the analysis, 11 attributes are considered. Land use (urban/rural), lighting condition, weather condition, trafficway description, police reported alcohol involvement, police reported drug involvement, horizontal alignment, time of crash, weekday, age, and obviously, the locality status of the drivers.

5.2.1. Rules for Local Wrong-Way Driver Crashes

Rules were filtered using a minimum lift amount of 1.4, minimum support of 0.1, and a confidence threshold of 0.6. These amounts have been determined based on previous traffic safety research that often defines minimum thresholds for support and confidence, typically ranging between 0.01 to 0.1 and 0.6 to 0.7, respectively. The rules were obtained based on these criteria and then arranged in ascending order of their confidence values.

Table 3. Rules for local wrong-way driver crashes.

Rules (Antecedent → Consequent)		Support	Confidence	Lift
1	Injury severity: Fatal Injury (K), Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban	0.11	0.95	1.48
2	Trafficway Description: Two-Way, Divided, Positive Median Barrier, Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban	0.11	0.94	1.47
3	Weather condition: Clear, Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban	0.11	0.94	1.47
4	Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban	0.16	0.94	1.46
5	Locality status: Local, Lighting condition: Dark-Lighted, Horizontal alignment: Straight → Setting: Urban	0.12	0.93	1.46
6	Drinking Name: Yes (Alcohol Involved), Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban	0.10	0.93	1.46
7	Hour: 12 A.M.–6 A.M., Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban	0.11	0.93	1.45
8	Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban	0.12	0.92	1.44
9	Trafficway Description: Two-Way, Divided, Positive Median Barrier, Hour: 12 A.M.–6 A.M., Horizontal alignment: Straight, Locality status: Local → Setting: Urban	0.11	0.90	1.41
10	Hour: 12 A.M.–6 A.M., Setting: Urban, Locality status: Local → Trafficway Description: Two-Way, Divided, Positive Median Barrier	0.14	0.76	1.43
11	Hour: 12 A.M.–6 A.M., Horizontal alignment: Straight, Setting: Urban, Locality status: Local → Trafficway Description: Two-Way, Divided, Positive Median Barrier	0.11	0.76	1.43
12	Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban, Horizontal alignment: Straight	0.12	0.73	1.44
13	Trafficway Description: Two-Way, Divided, Positive Median Barrier, Drinking Name: Yes (Alcohol Involved), Locality status: Local → Hour: 12 A.M.–6 A.M.	0.11	0.70	1.41
14	Lighting condition: Dark-Lighted, Locality status: Local → Setting: Urban	0.12	0.70	1.51
15	Lighting condition: Dark-Lighted, Locality status: Local → Hour: 12 A.M.–6 A.M.	0.12	0.70	1.40
16	Lighting condition: Dark-Lighted, Locality status: Local → Trafficway Description: Two-Way, Divided, Positive Median Barrier, Setting: Urban	0.11	0.69	1.71
17	Lighting condition: Dark-Lighted, Locality status: Local → Hour: 12 A.M.–6 A.M., Setting: Urban	0.11	0.65	1.84
18	Lighting condition: Dark-Lighted, Locality status: Local → Weather condition: Clear, Setting: Urban	0.11	0.63	1.43
19	Hour: 12 A.M.–6 A.M., Horizontal alignment: Straight, Locality status: Local → Trafficway Description: Two-Way, Divided, Positive Median Barrier, Setting: Urban	0.11	0.63	1.57
20	Lighting condition: Dark-Lighted, Locality status: Local → Drinking Name: Yes (Alcohol Involved), Setting: Urban	0.10	0.62	1.91
21	Hour: 12 A.M.–6 A.M., Locality status: Local → Trafficway Description: Two-Way, Divided, Positive Median Barrier, Setting: Urban	0.14	0.62	1.55

presents the rules for local wrong-way driver crashes, while

Table 4. Rules for non-local wrong-way driver crashes.

Rules (Antecedent → Consequent)		Support	Confidence	Lift
1	Locality status: Non-Local, Setting: Rural, Horizontal alignment: Straight → Lighting condition: Dark-Not Lighted	0.11	0.74	1.49
2	Locality status: Non-Local, Hour: 6 P.M.–12 A.M., Horizontal alignment: Straight → Lighting condition: Dark-Not Lighted	0.11	0.73	1.47
3	Locality status: Non-Local, Setting: Rural → Lighting condition: Dark-Not Lighted	0.12	0.73	1.47
4	Locality status: Non-Local, Injury severity: Fatal Injury (K), Setting: Rural, Horizontal alignment: Straight → Lighting condition: Dark-Not Lighted	0.12	0.72	1.46
5	Locality status: Non-Local, Age: Middle Adult (25 to 45), Setting: Urban → Hour: 12 A.M.–6 A.M.	0.11	0.72	1.46
6	Locality status: Non-Local, Injury severity: Fatal Injury (K), Setting: Rural → Lighting condition: Dark-Not Lighted	0.14	0.72	1.45
7	Locality status: Non-Local, Hour: 6 P.M.–12 A.M. → Lighting condition: Dark-Not Lighted	0.12	0.71	1.44
8	Locality status: Non-Local, Setting: Rural, Horizontal alignment: Straight → Lighting condition: Dark-Not Lighted	0.15	0.71	1.44
9	Locality status: Non-Local, Setting: Rural → Lighting condition: Dark-Not Lighted	0.17	0.71	1.43
10	Locality status: Non-Local, Weather condition: Clear, Setting: Rural → Lighting condition: Dark-Not Lighted	0.11	0.71	1.42
11	Locality status: Non-Local, Days of the week: Weekday, Setting: Rural → Lighting condition: Dark-Not Lighted	0.10	0.70	1.41
12	Trafficway Description: Two-Way, Divided, Unprotected Median, Setting: Rural → Locality status: Non-Local, Horizontal alignment: Straight	0.11	0.63	1.41
13	Locality status: Non-Local, Injury severity: Fatal Injury (K), Setting: Rural → Lighting condition: Dark-Not Lighted, Horizontal alignment: Straight	0.12	0.63	1.52
14	Locality status: Non-Local, Hour: 6 P.M.–12 A.M. → Lighting condition: Dark-Not Lighted, Horizontal alignment: Straight	0.11	0.63	1.52
15	Locality status: Non-Local, Setting: Rural → Lighting condition: Dark-Not Lighted, Horizontal alignment: Straight	0.11	0.62	1.51
16	Locality status: Non-Local, Setting: Rural → Lighting condition: Dark-Not Lighted, Horizontal alignment: Straight	0.15	0.61	1.49
17	Locality status: Non-Local, Lighting condition: Dark-Not Lighted, Injury severity: Fatal Injury (K), Horizontal alignment: Straight → Setting: Rural	0.12	0.60	1.68
18	Locality status: Non-Local, Lighting condition: Dark-Not Lighted, Injury severity: Fatal Injury (K) → Setting: Rural	0.14	0.60	1.67

is dedicated to non-local wrong-way driver crashes.

According to the rules obtained for local drivers, the urban setting was present in 19 different rules, encompassing all of them except for rules 13 and 15, which shows its significant effect on the occurrence of WWD fatal crashes for local drivers.

The hours from 12 A.M. to 6 A.M. is a frequent time that appeared in multiple rules (rules 7, 9, 10, 11, 13, 15, 17, 19, 21), highlighting its significant role in contributing to local wrong-way drivers.

Rules 6 and 20 show the association between local intoxicated drivers and urban settings. Also, rule number 20 has the highest amount of lift among other rules; this suggests a strong association between well-lit urban environments and local drivers driving under the influence. Rule 13 associates intoxicated local wrong-way drivers with hours between 12 A.M. to 6 A.M.

All rules, except for rules 9, 10, 11, 13, 19, and 21, take into account the “Dark-Lighted” condition. The absence of dark conditions in any rules indicates that local drivers’ familiarity aided them in navigating the correct direction under low-light conditions.

Based on rules 5, 9, 11, 12, and 19, roads without curves are associated with wrong-way driving crashes, which aligns with [16].

Rules 8 and 14 for local drivers associate wrong-way drivers with urban settings, which is consistent with the findings of a previous study [3].

5.2.2. Rules for Non-Local Wrong-Way Driver Crashes

Rules were filtered using a minimum lift amount of 1.4, minimum support of 0.1, and a confidence threshold of 0.6. These amounts have been determined through trial and error to achieve the best outcomes. The rules were obtained using these criteria and then arranged in ascending order of their confidence value.

According to the rules obtained for non-local drivers, the rural setting was present in 14 different rules, encompassing all of them except for rules 2, 5, 7, and 14, which shows its significant effect on the occurrence of WWD fatal crashes for non-local drivers.

All rules, except for rules 5 and 12, contain the “Dark-Not Lighted” condition for non-local drivers. This condition suggests that driving in dark areas can lead unfamiliar drivers to go the wrong way.

Rule 12 highlights that an unprotected median in rural areas is associated with the absence of a horizontal curve and non-local wrong-way drivers.

Rules 4, 6, 13, 17, and 18 feature various combinations of rural settings, poorly lit conditions, and the absence of horizontal alignment, all of which contain fatal injuries for non-local wrong-way drivers. These rules have been repeated five times with different scenarios of these factors.

6. Discussion

6.1. Difference between Contributing Factors to Local and Non-Local Wrong-Way Drivers

Through the use of descriptive statistics and ARM, the study has identified the contributing factors to local and non-local wrong-way drivers. Now, based on the findings, the differences between contributing factors for these two groups can be determined.

Urban areas see a higher likelihood of local drivers engaging in WWD crashes, whereas, in rural areas, non-local drivers are more inclined to go in the wrong direction.

More than half of WWD crashes involving local drivers are attributed to alcohol, and the generated rules consistently linked alcohol involvement to multiple instances of local wrong-way drivers; on the other hand, alcohol involvement was not present in the high-confidence rules for non-local drivers.

Lighting condition is another contributing factor to wrong-way crashes. Non-local wrong-way drivers are often associated with dark and poorly lit conditions, while this specific condition did not appear in high-confidence rules for local drivers.

To delve deeper into these findings, it is essential to consider the underlying reasons for these differences between local and non-local drivers and compare them with existing literature.

Previous studies, such as those by [30], have also highlighted the role of alcohol in WWD crashes among local drivers, supporting the notion that local drivers are more likely to drive under the influence due to a perceived familiarity with the roads.

In contrast, non-local drivers, often navigating unfamiliar territories, might be more cautious but still vulnerable to errors in poorly lit conditions, as highlighted by [31]. This difference can be attributed to the different driving behaviors and risk perceptions between local and non-local drivers. Moreover, the difference in lighting conditions as a contributing factor suggests that non-local drivers rely more heavily on visual cues, which are compromised in dark conditions, leading to a higher incidence of WWD crashes. These findings align with studies by [16], which also found a significant correlation between poor lighting and non-local driver errors.

6.2. Countermeasures Recommended Based on the Study Findings

While different countermeasures might be designed for various categories of drivers (local and non-local), it is challenging to target the exact intended driver group. However, these countermeasures can be implemented in areas where the intended driver group is more likely to benefit, reducing the risk of wrong-way driving.

The transportation community widely accepts the Safe System Approach as a highly effective method for reducing risks within our vast and intricate transportation network. This approach involves constructing multiple layers of protection to proactively prevent collisions and mitigate the harm to individuals affected by crashes. It represents a thorough strategy that serves as a framework to enhance safety across all settings and all different users [36]. Considering the Safe System Approach, we explore various countermeasures aimed at addressing risks in both the roadway environment and drivers’ behavior that can lead to WWD crashes. These countermeasures align with the Safe System Approach principles by establishing multiple layers of protection and addressing the various factors that contribute to WWD incidents.

The study findings highlight the critical need to implement different countermeasures to address WWD issues among local and non-local drivers. It is evident that different strategies should be employed for each group to target their specific risk factors and challenges on the roads. Previous studies did not determine whether the WWD crash contributing factors vary among different driver groups; consequently, there has been no assessment of whether a particular countermeasure is specifically efficient for local or non-local drivers [3].

Policy-level countermeasures are recommended in this context, and further investigation is necessary for both local and non-local wrong-way drivers to recommend specific ones.

6.2.1. Recommendations for Local Drivers

The study identifies intoxication as a significant factor contributing to WWD crashes, particularly among local drivers in urban areas. The alarming finding that more than half of local wrong-way drivers were drunk necessitates targeted countermeasures to address this specific issue. Preventive measures aimed at self-correction for intoxicated drivers are of the highest importance [3]. Installing specialized road elements, such as low-mounted WRONG WAY signs at interchanges, can prove effective in catching the attention of impaired drivers [3]. By employing these kinds of measures in areas with higher WWD incidents, local drivers can be deterred from entering one-way lanes in the wrong direction, significantly reducing the risk of crashes and enhancing overall road safety.

One proactive approach for preventing intoxicated drivers regarding the safe system approach are the Ignition Interlock Devices (IIDs). Ignition Interlock Devices (IIDs) are breathalyzer instruments installed in vehicles. They necessitate the driver to blow into them before initiating the engine. If the device detects alcohol levels exceeding a predetermined threshold (typically aligned with legal blood alcohol concentration limits), it prohibits the engine from starting.

Additionally, the following countermeasures can be implemented by installing detector systems that can accurately detect and alert authorities to instances of wrong-way driving. These systems can trigger responses for wrong-way drivers and notify law enforcement and transportation agencies [37].

6.2.2. Recommendations for Non-Local Drivers

For non-local drivers, the primary issue lies in their unfamiliarity with road geometry, making them more prone to going in the wrong direction in dark areas; therefore, specific countermeasure packages tailored to the needs of non-local drivers should be considered, aiming to provide clearer navigation guidance that helps prevent wrong-way incidents. By enhancing the road signage and implementing raised pavement markings and wrong-way arrows, non-local drivers can be better guided to travel in the correct direction of rural areas, reducing the risk of wrong-way driving occurrences [1,3].

Additionally, the following countermeasures can be implemented:

• Geometric Improvements: Adjusting the design of ramps and intersections to prevent unintentional entries from the wrong direction. This includes increasing separation between entrance and exit ramps, using raised medians to restrict certain turns, and reducing the curvature of potential entry paths to discourage incorrect maneuvers [37].
• Blank-Out Signs: Signs that activate messages only during wrong-way driving events, improving their visibility to drivers unfamiliar with the area [37].
• Signs with Flashing Enhancements: WRONG WAY signs can be equipped with flashing beacons or LED borders to improve nighttime visibility. These enhancements are activated only when a wrong-way driving event is detected [37].
• Supplemental Signs: Additional DO NOT ENTER and WRONG WAY signs can be installed to reinforce the message to wrong-way drivers. Using larger signs can enhance visibility and effectiveness [37].
• Pavement Marking Arrows: Arrows painted on freeway exit ramps and in through lanes near intersections on divided highways can help deter wrong-way movements. These markings are relatively low-cost treatments that provide additional guidance to drivers within their primary field of vision [37].

7. Conclusions

This study aims to identify the differences between factors contributing to the fatality of wrong-way drivers, both local and non-local, based on the FARS data from 2016 to 2020. Only a few research studies have explored whether the locality status of drivers impacts WWD crash occurrences [3]. In this paper, WWD crashes are categorized into two groups: local and non-local. Various attributes are considered, and descriptive statistics and association rule mining are employed. Finally, distinct rules were discovered for each group of drivers (local and non-local).

According to the descriptive statistics, the states of Mississippi, Virginia, Indiana, Maryland, Georgia, Iowa, Oklahoma, Missouri, and Arkansas experienced a significant number of WWD crashes by non-local drivers, and the percentage for non-local drivers was over 60 percent of all WWD crashes. This finding necessitates further investigation to determine the underlying factors contributing to these occurrences. Also, a noticeable increase in the percentage of non-local drivers during the early summer is observed, possibly due to long-distance travel. Additionally, there is a higher frequency of fatal WWD crashes involving local drivers between November and February.

The rules underscore the increased danger that intoxicated local drivers in urban areas encounter, with a 93 percent confidence level in rule 6. Moreover, the period from midnight to 6 A.M. corresponds to when local drivers are most susceptible to entering the wrong direction on the roads.

Moreover, the vulnerability of non-local drivers in dark conditions to wrong-way driving incidents highlights the significance of addressing their reduced ability to navigate unfamiliar areas, which emphasizes the need for further investigation into the impact of factors such as inadequate signage or pavement markings that may contribute to the driver’s inability to correctly identify directions.

The present study possesses certain limitations that warrant attention in future research. For instance, one possible approach for future research could involve expanding the number of attributes to uncover additional underlying rules.