Investigation on Hazardous Material Truck Involved Fatal Crashes Using Cluster Correspondence Analysis

Abstract

Although hazardous material (HAZMAT) truck-involved crashes are uncommon compared to other types of traffic crashes, these crashes pose considerable threats to the public, property, and environment due to the unique feature of low probability with high consequences. Using ten-year (2010–2019) crash data from the Fatality Analysis Reporting System (FARS) database, this study applies cluster correspondence analysis to identify the underlying patterns and the associations between the risk factors for HAZMAT-truck-involved fatal crashes. A low-dimensional space projects the categorical variables (including the crash, road, driver, vehicle, and environmental characteristics) into different clusters based on the optimal clustering validation criterion. This study reveals that fatal HAZMAT-truck-involved crashes are highly distinguishable concerning collision types (angle and front-to-front crashes, single-vehicle crashes, and front-to-end crashes) and roadway geometric variables, such as two-way undivided roadways, curve alignments, and high-speed (65 mph or more) urban interstate highways. Driver behavior (distraction, asleep or fatigue, and physical impairment), lighting conditions (dark–lighted and dark–not lighted), and adverse weather are also interrelated. The findings from this study will help HAZMAT carriers, transportation management authorities, and policymakers develop potential targeted countermeasures for HAZMAT-truck-involved crash reduction and safety improvement.

1. Introduction

Due to the rapid and broad development of urbanization and industrialization, the demand for hazardous materials (HAZMAT) has increased significantly. According to the 2017 Economic Census: Transportation report, HAZMAT shipments were approximately 3 billion tons in the United States in 2017. Trucks as a single mode accounted for nearly 65% of all HAZMAT shipments in value and over 60% of all HAZMAT shipments in tonnage [1]. HAZMAT-truck-involved crashes are uncommon compared to general traffic crashes, and the frequency is relatively smaller. However, these crashes tend to cause more fatalities and severe injuries. In the United States, 5005 heavy trucks were involved in fatal crashes in 2019, with 120 (2.4%) carrying HAZMAT [2]. The US Department of Transportation compiled a review of HAZMAT-related incidents during the ten-year period (2012–2021), which showed that HAZMAT highway transportation accounted for 100% of the total number of fatalities (83) and 74.02% of the total number of injuries (1732) [3].

Because of the unique physical and chemical properties, particularly the higher possibility of spillage, fire, and explosion during or after an incident, HAZMAT-truck-involved crashes are frequently referred to as “low probability with high consequences” [4]. According to a guidance issued by Federal Emergency Management Agency in 2019, these crashes may cause harm to people, the environment, critical infrastructure, and property. Their potential for harm exists regardless of whether hazardous materials are released by accident, fire, or weather-related event [5]. Hazardous material incidents impact various community stakeholders. First responders, first receivers, transportation personnel, local residents, and students are all at risk of negative health effects from these substances. Understanding where, when, and why crashes happen is critical to reducing the number of crashes and improving HAZMAT road transportation safety.

Many research efforts aim to identify the risk factors and the relationships among variables that influence HAZMAT-truck-involved crashes. Previous studies have proven that traditional statistical regression models can effectively establish the associations between different explanatory variables and crash occurrence or severity. However, these parameter-oriented methods often assume that the crash data form a specific distribution. It is difficult to reveal the heterogeneous relationships in the crash data. More and more researchers apply data mining and machine learning techniques to investigate the risk factors and their impacts on crashes. Due to the greater flexibility, and comparable or superior performance, these innovative methods could assist in comprehensive crash analysis and decision-making. Nevertheless, little research has utilized these approaches for HAZMAT-truck-involved crash analysis. Thus, there is a need for research efforts with additional resources and newer approaches and techniques in HAZMAT safety analysis. By focusing on the following objectives, this study aims to contribute to safer HAZMAT road transportation.

• Analyze the fatal HAZMAT-truck-involved crash characteristics and the risk factors;
• Investigate the crash patterns and the collective associations between risk factors and fatal HAZMAT-truck-involved crashes;
• Apply machine learning techniques for HAZMAT crash pattern recognition.

Using ten-year (2010–2019) crash data from the Fatality Analysis Reporting System (FARS) database, this study applies cluster correspondence analysis to identify the underlying patterns and the associations between the risk factors for HAZMAT-truck-involved fatal crashes, intending to present the collective associations of risk factors that have yet to be explored or hidden in the whole crash dataset. This approach integrates cluster analysis and correspondence analysis by grouping the crash data into different clusters. A low-dimensional space based on the optimal scaling values projects the categorical variables (including the crash, road, driver, vehicle, and environmental characteristics) in a multivariate dataset. The findings from this study will help HAZMAT carriers, transportation management authorities, and policymakers develop potential targeted countermeasures for HAZMAT-truck-involved crash reduction and safety improvement.

2. Literature Review

Many past studies have identified and investigated the risk factors influencing the frequency and severity of HAZMAT-truck-involved crashes. The commonly used methods for determining the relationship between explanatory variables and crashes are exploratory data analysis with statistical description [6,7,8] and regression modeling [9,10,11,12]. Shen et al. [8] conducted a statistical distribution study of 708 HAZMAT road transportation crashes in China from 2004 to 2011. The most common collision types were rollover, run-off-road, and rear-end collisions. Freeways, early morning (4:00–6:00 a.m.) and midday (10:00 a.m.–12:00 p.m.) hours, vehicle-related defects, and human-related errors were all major influencing factors for HAZMAT tanker crashes. Uddin and Huynh [9] investigated factors affecting crash severity involving HAZMAT large trucks through fixed- and random-parameters ordered probit models. The model results showed that male occupants, dark–unlighted and dark–lighted conditions, crashes in rural areas, and weekday crashes are associated with a higher probability of severe injuries. Xing et al. [10] collected 1721 HAZMAT crashes in China from 2014 to 2017. They developed a random-parameters ordered probit model to quantify the influence of risk factors on HAZMAT crash severity, accounting for the unobserved heterogeneity in the crash data. HAZMAT category (compressed gas, explosive, and poison), roadway characteristics (tunnel, slope, county road, and dry road surface), collision types (rear-end and multi-vehicle crashes), driver characteristics (misoperation, fatigue, and speeding), and environmental characteristics (winter and dark lighting conditions) were associated with increased probability of severe injury crashes. Song et al. [11] developed an ordered logistic model for factor analysis of HAZMAT road transportation crashes. According to their findings, the six most important factors affecting HAZMAT crash severity were: traffic volume, driver fatigue or asleep, number of lanes, speeding, adverse weather conditions, and lighting conditions. Similarly, a study conducted by Ma et al. [12] also applied the ordered logistic model to investigate the relationship between risk factors and HAZMAT crashes. The model estimation results revealed that violations, unsafe driving behaviors, vehicle defects, vehicle types, weather, lighting, seasonal distribution, and road grade are closely related to more severe crashes. In addition, there were many other commonly used regression models in large-truck-involved crash studies, such as ordered probit [13], multinomial logit [14,15], mixed logit [16,17,18,19], hierarchical Bayesian random intercept [20], Bayesian binary logit [21], and Bayesian model averaging-based logistic regression [22].

Traditional statistical models can successfully identify the correlations between HAZMAT crashes and independent variables. However, they usually require certain assumptions in the crash data, and it is challenging to discover the underlying crash patterns and the interaction among various variables. Data mining and machine learning approaches for truck-related crash data processing and analysis, including classification trees [23], taxicab correspondence analysis [24], association rules mining [25], and Bayesian networks [26,27,28], have become increasingly popular in recent years. For example, Das et al. [24] utilized taxicab correspondence analysis to uncover the complex interactions between multiple risk factors and large truck-involved fatal crashes. The study identified five clusters displaying the association patterns of different variables, such as intersection types, posted speed limits, two-lane undivided roadways, collision types, number of vehicles, driver impairment, and weather conditions. Using crash data from 2008 to 2017, Hong et al. [25] performed association rules mining to investigate the risk factors contributing to HAZMAT vehicle-related crashes on expressways. Male drivers, daytime, mainline segments, single-vehicle crashes, and clear weather conditions were highly associated with HAZMAT vehicle-related crashes. Zhao et al. [26] utilized the Bayesian networks approach to prioritize the risk factors that impact HAZMAT road transportation crashes. According to their findings, the most important attributes were packing and loading of HAZMAT, vehicle and facility-related factors, and human factors. Similarly, Ma et al. [27] also utilized Bayesian networks to explore the most probable factors or the combination of factors leading to the crash—a recent HAZMAT crash severity prediction study [28] combined random forest and Bayesian networks. Random forest ranked the importance of risk factors, and Bayesian networks developed the probabilistic inference. The results showed that driver characteristics (driver age less than 25, fatigue, distraction, and violation), roadway characteristics (posted speed limits over 66 mph, intersections, ramps, and bridges), crash characteristics (fire/explosion/spill, midnight and early morning hours, and head-on crashes), vehicle characteristics (more than four vehicles), and environmental characteristics (poor lighting conditions and adverse weather) were all related to fatal and severe injury HAZMAT crashes.

Table 1. Summary of HAZMAT road transportation crash studies.

Reference	Data Source	Method	Variables
[6]	Beginning of the 20th century to 2004, Major Hazard Incidents Data Service	Exploratory data analysis and statistical description	Human, mechanical, and external characteristics
[7]	2000–2007, Journal of Safety and Environment	Exploratory data analysis and statistical description	Driver, crash, road, vehicle, management, and weather characteristics
[8]	2004–2011, China Chemical Safety Association	Exploratory data analysis and statistical description	Driver, roadway class, crash, temporal variables, vehicle, and weather characteristics
[9]	2005–2011, Highway Safety Information System, California	Fixed- and random- parameters ordered probit	Occupant, crash, vehicle, roadway, environmental, and temporal characteristics
[10]	2014–2017, China Chemical Safety Association	Random-parameters ordered probit	Driver, crash, vehicle, roadway, and environmental characteristics
[11]	Five-year period, Highway Safety Information System	Ordered logit	Driver, roadway, and environmental characteristics
[12]	2018–2019, Ministry of Emergency Management	Ordered logit	Driver, vehicle, roadway, and environmental characteristics
[25]	2008–2017, Korea Expressway Corporation	Association rules mining	Driver, crash, vehicle, roadway, environmental, and temporal characteristics
[26]	2005–2009	Bayesian network	Human, vehicle, mechanical, and external characteristics
[27]	2015–2016, State Work Accident Briefing System and Chemical Accidents Information Network	Bayesian network	Driver, crash, vehicle, roadway, and environmental characteristics
[28]	2013–2017, Highway Safety Information System	Random forest and Bayesian network	Driver, crash, vehicle, roadway, and environmental characteristics

summarizes HAZMAT road transportation crash studies, including data sources, methods, and considered variables. Although data mining and machine learning approaches can provide comparable or superior performance on crash modeling, such applications in HAZMAT crash risk factor analysis were not a significant focus in the literature. Thus, there still is a need for research efforts with additional resources and innovative methods. This study applies cluster correspondence analysis to explore the complex HAZMAT-truck-involved fatal crash data, reveal the underlying crash patterns, and investigate the relationships between variables (including the crash, road, driver, vehicle, and environmental characteristics).

3. Materials and Methods

3.1. Crash Data

This study collected ten-year (2010–2019) fatal crash data involving HAZMAT trucks from the FARS database. FARS is a nationwide census providing National Highway Traffic Safety Administration (NHTSA), Congress, and the American public yearly data on fatal injuries in motor vehicle traffic crashes. This study used the unique crash identification number to retrieve and match data from the various datasets (crash data, vehicle data, and person data). The research conducted a thorough data cleaning process to address incorrect (such as speed limit recorded as 200 mph) and missing values in the final dataset development. The fatalities recorded in the source data include all that occurred for any reason at an accident involving HAZMAT trucks. There were 1322 HAZMAT-truck-involved fatal crashes after the data cleaning process. Figure 1 shows the distribution of these crashes in the US by different states from 2010 to 2019.

Based on the availability and relevance to explain the fatal crash occurrence, this study selected 26 risk factors, including crash, road, driver, vehicle, and environmental characteristics, as explanatory variables.

Table 2. Descriptive statistics of HAZMAT-truck-involved fatal crashes.

Variables	Count	%	Variables	Count	%
Driver characteristics			Road characteristics
Driver age (DrA)			Trafficway Description (TrW)
<25	24	1.82%	Two-way, not divided	686	51.89%
25–35	183	13.84%	Two-way, divided, unprotected median	289	21.86%
36–45	312	23.60%	Two-way, divided, positive median barrier	261	19.74%
46–55	430	32.53%	Two-way, not divided with a continuous left-turn lane	37	2.80%
56–65	312	23.60%	Entrance/exit ramp	24	1.82%
>65	61	4.61%	Others	25	1.89%
Driver gender (DrG)			Speed limit (SpL)
Female	27	2.04%	<25 mph	56	4.24%
Male	1295	97.96%	30–45 mph	206	15.58%
Driver violation (DrV)			50–65 mph	790	59.76%
No	1264	95.61%	>65 mph	270	20.42%
Yes	58	4.39%
Commercial motor vehicle license status (CDL)			Roadway alignment (Alg)
Valid	1290	97.58%	Straight	1028	77.76%
Not Valid	32	2.42%	Curve	248	18.76%
Previous recorded crashes (PrA)			Others	46	3.48%
No	1112	84.11%	Roadway grade (Grd)
Yes	210	15.89%	Level	896	67.78%
Previous recorded suspensions, revocations, and withdrawals (PrvSs)			Grade	374	28.29%
No	1280	96.82%	Others	52	3.93%
Yes	42	3.18%	Roadway surface condition (SrC)
Previous speeding convictions (PrvSp)			Dry	1071	81.01%
No	1106	83.66%	Non-dry	251	18.99%
Yes	216	16.34%	Setting (Stt)
Previous other moving violation convictions (PrO)			Rural	890	67.32%
No	1040	78.67%	Urban	432	32.68%
Yes	282	21.33%	Roadway function class (RdF)
Distraction (Dst)			Interstate	335	25.34%
No	956	72.32%	Arterial/collector	936	70.80%
Yes	75	5.67%	Local	51	3.86%
Others	291	22.01%	Intersection type (InT)
Impairment (Imp)			Segment	1050	79.43%
None/apparently normal	998	75.49%	Four-way intersection	177	13.39%
Asleep, fatigue or physical impairment	53	4.01%	T-intersection	84	6.35%
Others	271	20.50%	Others	11	0.83%
Crash characteristics			Environmental characteristics
Crash hour (Hor)			Lighting condition (LgC)
12:00–5:59 a.m.	296	22.39%	Daylight	783	59.23%
6:00–11:59 a.m.	405	30.64%	Dark–lighted	131	9.91%
12:00–5:59 p.m.	406	30.71%	Dark–not lighted	348	26.32%
6:00–11:59 p.m.	215	16.26%	Dawn/dusk	60	4.54%
Manner of collision (MnC)			Weather (Wth)
Not collision with motor vehicle	408	30.86%	Clear	909	68.75%
Angle	377	28.52%	Cloudy	212	16.04%
Front-to-front	200	15.13%	Rain/snow	136	10.29%
Front-to-rear	227	17.17%	Others	65	4.92%
Sideswipe	87	6.58%	Visual obstruction (VsO)
Others	23	1.74%	No	1259	95.23%
Fire (Fir)			Yes	63	4.77%
No	1129	85.40%
Yes	193	14.60%
Vehicle characteristics
Hazardous material class (HzC)			Number of total vehicles (VhT)
Flammable/combustible liquid	727	54.99%	Single	314	23.75%
Gases	191	14.45%	Two	765	57.87%
Corrosive	94	7.11%	Multi	243	18.38%
Explosives	29	2.19%
Others	281	21.26%

summarizes the descriptive statistics of the HAZMAT-truck-involved fatal crashes. The majority of fatal crashes involving HAZMAT trucks occurred on two-way undivided roadways with speed limits between 50–65 mph on segments of arterials or collectors. Notably, fatal crashes occurring in the dark–lighted and dark–not lighted conditions accounted for over 36% of the whole dataset. In addition, more HAZMAT-truck-involved fatal crashes occurred in rural areas (67.32%). Regarding driver characteristics, this study grouped the driver’s age into six levels. Drivers aged 46–55 accounted for the highest proportion in this fatal crash dataset. Some variables form high skewness distribution. For example, approximately 98% of drivers involved in the crashes are male. The dataset also contains driver status records, such as previously recorded crashes, previously recorded suspensions, revocations, withdrawals, previous speeding convictions, and previous other moving violation convictions.

3.2. Cluster Correspondence Analysis

Parametric and non-parametric models are two primary crash data analysis methods used to investigate the associations between risk factors and crash occurrence. In a parametric model, individual variables can explain the impact of analyzed risk factors on either crash probability or injury severity levels. Random parameter models are one example of the parametric technique to address data heterogeneity. However, these parametric modeling techniques require distributional hypotheses, which may be difficult to capture the variations for the observations with shared unobserved heterogeneity [29]. Given large and complex datasets such as crash data, which contains a lot of categorical variables with high dimensions from input data, there is a need for advanced methods to handle nonlinear relationships.

Correspondence analysis extends principal component analysis suited to exploring relationships among categorical data. This multivariate statistical tool utilizes contingency tables for dimension reduction and data visualization. A low-dimensional space displays the associations between the column elements and row elements in the data matrix, and the positions of the row and column points are consistent with their associations in the table. Many researchers have applied different forms of correspondence analysis in transportation safety studies, such as taxicab correspondence analysis [24,30], multiple correspondence analysis [31,32,33,34,35,36], and joint correspondence analysis [37,38]. As one of the unsupervised machine learning approaches, correspondence analysis does not require any distribution hypothesis among data. A previous study has identified its ability to efficiently reduce dimensionality and compile results into easy-to-read plots for in-depth crash analysis [32]. However, distinguishing different clusters depends on subjective judgment, which relies on the categorical variable feature approximation results in a low-dimensional space [36]. In recent years, some transportation safety studies used cluster correspondence analysis due to its ability to discover the underlying cluster structures and reduce multicollinearity among data [24,39,40,41,42]. This approach has outperformed other earlier approaches that attempt to apply dimension reduction and clustering either sequentially or concurrently in terms of cluster convergence [42].

In general, cluster analysis is an exploratory approach that enables one to classify a set of multivariate observations into groups by maximizing their similarity within each group. Correspondence analysis is a process of identifying, quantifying, separating, and plotting associations among the characteristics and relationships among the various levels. It complements the cluster analysis in terms of dimension reduction and data visualization. By combining cluster analysis and correspondence analysis, this approach groups the data points into different clusters based on the profiles of the categorical variables with optimal scaling values. It depicts the clusters and variables to a low-dimensional approximation. Below presents a brief description of cluster correspondence analysis. Readers can refer to van de Velden et al. [43,44] for the complete theoretical concept.

Suppose there are n number of individuals (for example, HAZMAT-truck-involved fatal crashes) with $\mathrm{p}$ categorical variables (for example, driver, road, vehicle, crash, and environmental characteristics) in the dataset. A super indicator matrix Z with $\mathrm{n} \text{×} \mathrm{Q}$ (where $\mathrm{Q}={\sum }_{\mathrm{j}=1}^{\mathrm{p}}{\mathrm{q}}_{\mathrm{j}}$, $\mathrm{j}=1, 2, \cdots , \mathrm{p}$) represents this dataset. Each variable can also have ${\mathrm{q}}_{\mathrm{j}}$ modalities. Cluster membership can be expressed as an indicator matrix Z_K, $\mathrm{K}$ is the number of clusters. A table-cross-tabulation cluster memberships with categorical variables can be expressed as F = Z′_KZ, where Z_K is the $\mathrm{n} \text{×} \mathrm{K}$ indicator matrix describing cluster membership. This approach automatically finds the optimal scaling values for rows (clusters) and columns (categories) so that the between-cluster variance is the maximum. That is, it optimally separates the clusters concerning the distributions over the categorical variables. Similarly and simultaneously, it optimally separates the categories with differing distributions over the clusters. The optimal cluster allocation Z_K can be calculated as [44]:

$$\max {\mathsf{\varphi }}_{\mathrm{clusca}}\left({\mathrm{Z}}_{\mathrm{K}}{,\mathrm{B}}^{*}\right)=\frac{1}{\mathrm{p}}{\mathrm{traceB}}^{*}{{}^{\prime }\mathrm{D}}_{\mathrm{Z}}^{-\frac{1}{2}}{{\mathrm{Z}}^{\prime }\mathrm{M}\mathrm{Z}}_{\mathrm{K}}{\mathrm{D}}_{\mathrm{Z}}^{-1}{{\mathrm{Z}}^{\prime }}_{\mathrm{K}}{\mathrm{MZD}}_{\mathrm{Z}}^{-\frac{1}{2}}{\mathrm{B}}^{*}$$

(1)

The objective function of cluster correspondence analysis is shown below [44]:

$$\min \mathsf{\varphi }{{}^{\prime }}_{\mathrm{clusca}}\left({\mathrm{Z}}_{\mathrm{K}},\mathrm{G}\right){\Vert \sqrt{\frac{\mathrm{n}}{\mathrm{p}}}{\mathrm{MZD}}_{\mathrm{Z}}^{-\frac{1}{2}}{\mathrm{B}}^{*}{ - \mathrm{Z}}_{\mathrm{K}}\mathrm{G}\Vert }^2$$

(2)

where, $\mathrm{Z}$ is the $\mathrm{n} \text{×} \mathrm{Q}$, matrix, ${\mathrm{Z}}_{\mathrm{j}}$ is an ${\mathrm{n} \text{×} \mathrm{q}}_{\mathrm{j}}$ indicator matrix for the j-th categorical variable, ${\mathrm{Z}}_{\mathrm{j}}{1}_{{\mathrm{q}}_{\mathrm{j}}}{=1}_{\mathrm{n}}$, $1_{\mathrm{q}}$ denotes a q dimensional vector of ones; ${\mathrm{Z}}_{\mathrm{K}}$ is the $\mathrm{n} \text{×} \mathrm{K}$ cluster membership indicator matrix; $\mathrm{G}$ is the cluster centroid matrix; $\mathrm{B}$ is the column coordinate matrix of rank $\mathrm{k}$, $\mathrm{k}$ denotes the dimensionality of the optimally scaled quantifications; ${\mathrm{B}}^{*}=\frac{1}{\sqrt{\mathrm{np}}}{\mathrm{D}}_{\mathrm{Z}}^{\frac{1}{2}}\mathrm{B}$; ${\mathrm{M}=\mathrm{I}}_{\mathrm{n}}{ - 1}_{\mathrm{n}}{1^{\prime }}_{\mathrm{n}}/\mathrm{n}$; ${\mathrm{D}}_{\mathrm{Z}}$ is the diagonal matrix, ${\mathrm{D}}_{\mathrm{Z}}{1}_{\mathrm{Q}}{={\mathrm{Z}}^{\prime }1}_{\mathrm{n}}$.

Iterating between correspondence analysis of the contingency matrix ${\mathrm{F}={\mathrm{Z}}^{\prime }}_{\mathrm{K}}\mathrm{Z}$ and applying K-means cluster analysis to the reduced space coordinates obtained using the correspondence analysis category quantifications until convergence can obtain the optimal category quantifications (i.e., column coordinates) and an optimal cluster allocation. The cluster correspondence analysis approach can analyze two-way and multi-way tables that contain relationships between the rows and columns from datasets with various nominal variables. Interpreting the distributions over the variables in the different clusters is straightforward. Crash data is too heterogeneous, making it challenging to identify specific patterns. With its dimension reduction and visualization characteristics, the cluster correspondence analysis approach performs well in representing crash scenarios, analyzing high-dimensional categorical crash data, and recognizing the underlying patterns behind crash data.

This study used an open-source R package, “clustrd”, to perform the cluster correspondence analysis in R programming [45,46]. The clustering validation criterion used in the research is the Calinski–Harabasz index. This index (also known as the variance ratio criterion) determines how similar an object is to its own cluster (cohesion) when compared to other clusters (separation) [47]. This index uses the distance between the data points in a cluster and its cluster centroid to estimate cohesion. In contrast, it uses the distance between the cluster centroids and the global centroid to estimate separation. A higher index value suggests that the clusters are well separated.

4. Results and Discussion

4.1. Cluster Overview

To select the number of clusters, this study randomly conducted K-means clustering running multiple times to obtain the optimal values of the criterion. The study determined the final cluster number as four with the optimal criterion value of 4.3396 for the HAZMAT-truck-involved fatal crash dataset. Thus, the study grouped the HAZMAT-truck-involved fatal crashes into four clusters accordingly.

Table 3. Key cluster measurements.

Cluster	Cluster Centroid		Size	Percentage	Sum of Squares
	Dimension 1	Dimension 2
C1	−0.0191	−0.0047	561	42.44%	0.0492
C2	0.0229	−0.0116	402	30.41%	0.0323
C3	0.0071	0.0226	324	24.51%	0.0447
C4	−0.0099	−0.0354	35	2.64%	0.0324

summarizes the key measurements of the clusters, including cluster centroid, number and percentage of crashes, and the sum of squares. Cluster 1 includes over 40% of the crash data. Cluster 2 and Cluster 3 account for 30.41% and 24.51% of the crash data, respectively. Cluster 4 contains the smallest number of crashes, only 2.64% of the crash data.

Figure 2 illustrates the projection of crash data points and individual variable categories onto a two-dimensional biplot. It provides the approximations and visualization of their associations. This study groups the crash data points into four clusters with the colors red (C1), green (C2), blue (C3), and purple (C4), respectively. Different shapes represent the cluster centroids with the shapes circle (C1), triangle (C2), square (C3), and plus (C4), respectively. The small black circles represent the variable categories with their names on the side. This biplot visualizes all variables’ locations and associations in a two-dimensional space. However, it is difficult to observe individual variable categories within specific clusters. For better visualization and interpretation, four separate bar plots in Figure 3 display the variable contributions in the format of standardized residuals.

The bar plots in Figure 3 show the top 30 variable categories with the highest standardized residuals (positive or negative in absolute values) in each cluster. It helps to identify variable categories with the most deviation from the independence condition. A positive residual indicates that the variable category has an above-average frequency within the cluster. A negative residual indicates that the variable category has a below-average frequency. A higher residual value indicates a stronger association and a more significant contribution to cluster segmentation. The purpose of this research is to identify HAZMAT-truck-involved fatal crash patterns. Thus, the following sections discuss only variable categories with positive standardized residuals.

4.1.1. Cluster 1

There are several variable categories with positive standardized residuals in this cluster: angle crashes, two vehicles involved crashes, two-way not divided roadways, four-way intersections, front-to-front crashes, arterials or collectors, daylight, two-way not divided roadways with a continuous left-turn lane, speed limits between 30 to 45 mph, T-intersections, the crash hour between 12:00 p.m. to 5:59 p.m., normal driver conditions, and no distraction. Cluster 1 indicates an association of fatal crashes involving HAZMAT trucks and another vehicle at intersections on two-way, undivided roadways during daytime.

4.1.2. Cluster 2

This cluster shows a group of variable categories closely associated with each other (measured by the positive standardized residuals on the right side in Figure 3b). These variable categories include single-vehicle crashes, not collision with motor vehicle crashes, two-way divided roadways with unprotected median, roadway functional class–local, roadway segment, driver age under 25, dark–not lighted conditions, roadway with grade, cloudy weather conditions, driver age between 25 to 35, curve alignment, manner of collision–others, the crash hour between 6:00 p.m. to 11:59 p.m., the driver with previously recorded suspensions, revocations, and withdrawals, and adverse weather conditions (rain or snow). Cluster 2 indicates an association of fatal crashes involving single HAZMAT trucks on local roadways with relatively younger drivers under dark lighting conditions.

4.1.3. Cluster 3

This cluster contains several variable categories with positive standardized residuals: interstate highways, two-way divided roadways with a positive median barrier, speed limits greater than 65 mph, front-to-rear crashes, the crash hour between 12:00 a.m. to 5:59 a.m., multi-vehicle crashes, fire, dark–lighted conditions, abnormal driver conditions–asleep, fatigue, physical impairment, or others, entrance/exit ramps, urban areas, distraction, roadway segment, HAZMAT class–explosives, other distraction, and female drivers. Cluster 3 indicates an association of fatal crashes involving HAZMAT trucks and multiple vehicles at urban interstate highway segments and entrance/exit ramps with relatively high speed limits, under dark lighting conditions during midnight and early morning. Results show that driver behavior, such as distraction, asleep or fatigue, and physical impairment, are closely associated with fatal crashes under the abovementioned conditions. Female drivers also have associations with fatal crashes.

4.1.4. Cluster 4

This cluster includes the following variable categories with positive standardized residuals: traffic way–others, speed limits less than 25 mph, roadway grade–others, roadway alignment–others, dark–lighted conditions, roadway surface condition–non-dry, four-way intersections, angle crashes, roadway functional class–local, urban areas, the driver with previously recorded crashes, the crash hour between 12:00 a.m. to 5:59 a.m., and the crash hour between 6:00 p.m. to 11:59 p.m. This cluster contains the least information, covering many other values of different variable categories.

4.2. HAZMAT Crash Characteristics Analysis

The cluster-based analysis identifies the patterns of the risk factors associated with HAZMAT-truck-involved fatal crashes. The standardized residuals measure the significance of variable categories within the clusters.

Table 4. Proportion distributions of the variables by clusters.

Variables		Number of Crashes	Cluster 1 (561)	Cluster 2 (402)	Cluster 3 (324)	Cluster 4 (35)
Driver age (DrA)	<25	24	16.67%	58.33%	25.00%	0.00%
	25–35	183	35.52%	37.70%	22.95%	3.83%
	36–45	312	45.19%	32.69%	18.91%	3.21%
	46–55	430	44.64%	27.91%	25.12%	2.33%
	56–65	312	40.39%	27.56%	29.81%	2.24%
	>65	61	47.54%	24.59%	26.23%	1.64%
Driver gender (DrG)	Female	27	18.52%	25.93%	51.85%	3.70%
	Male	1295	42.93%	30.50%	23.94%	2.63%
Driver violation (DrV)	No	1264	42.09%	30.85%	24.53%	2.53%
	Yes	58	50.00%	20.69%	24.14%	5.17%
Commercial motor vehicle license status (CDL)	Valid	1290	42.79%	30.31%	24.26%	2.64%
	Not Valid	32	28.13%	34.37%	34.37%	3.13%
Previous recorded crashes (PrA)	No	1112	43.07%	30.31%	24.46%	2.16%
	Yes	210	39.05%	30.95%	24.76%	5.24%
Previous recorded suspensions, revocations, and withdrawals (PrvSs)	No	1280	42.57%	30.08%	24.69%	2.66%
	Yes	42	38.10%	40.47%	19.05%	2.38%
Previous speeding convictions (PrvSp)	No	1106	43.13%	30.11%	24.23%	2.53%
	Yes	216	38.89%	31.94%	25.93%	3.24%
Previous other moving violation convictions (PrO)	No	1040	44.52%	30.48%	22.31%	2.69%
	Yes	282	34.76%	30.14%	32.62%	2.48%
Distraction (Dst)	No	956	49.27%	29.18%	19.14%	2.41%
	Yes	75	13.33%	33.33%	48.01%	5.33%
	Others	291	27.49%	33.68%	36.08%	2.75%
Impairment (Imp)	None/Apparently Normal	998	49.90%	30.26%	17.33%	2.51%
	Asleep, Fatigue or Physical Impairment	53	3.77%	26.42%	67.92%	1.89%
	Others	271	22.51%	31.73%	42.44%	3.32%
Trafficway Description (TrW)	Two-Way, Not Divided	686	63.85%	30.17%	4.81%	1.17%
	Two-Way, Divided, Unprotected Median	289	28.03%	42.90%	28.72%	0.35%
	Two-Way, Divided, Positive Median Barrier	261	2.68%	24.52%	72.80%	0.00%
	Two-Way, Not Divided With a Continuous Left-Turn Lane	37	94.59%	5.41%	0.00%	0.00%
	Entrance/Exit Ramp	24	0.00%	20.83%	75.00%	4.17%
	Others	25	0.00%	0.00%	0.00%	100.00%
Speed limit (SpL)	<25 mph	56	14.29%	19.64%	14.29%	51.78%
	30–45 mph	206	63.11%	30.58%	4.85%	1.46%
	50–65 mph	790	48.99%	31.77%	18.86%	0.38%
	>65 mph	270	13.33%	28.52%	58.15%	0.00%
Roadway alignment (Alg)	Straight	1028	44.66%	30.54%	23.44%	1.36%
	Curve	248	34.68%	35.89%	29.03%	0.40%
	Others	46	28.26%	4.35%	23.91%	43.48%
Roadway grade (Grd)	Level	896	45.98%	29.35%	23.33%	1.34%
	Grade	374	35.83%	36.10%	28.07%	0.00%
	Others	52	26.92%	9.62%	19.23%	44.23%
Roadway surface condition (SrC)	Dry	1071	43.23%	31.19%	24.09%	1.49%
	Non-Dry	251	39.05%	27.09%	26.29%	7.57%
Setting (Stt)	Rural	890	46.52%	30.22%	21.69%	1.57%
	Urban	432	30.32%	30.79%	34.03%	4.86%
Roadway function class (RdF)	Interstate	335	0.30%	22.69%	76.41%	0.60%
	Arterial/Collector	936	58.02%	32.05%	6.94%	2.99%
	Local	51	33.33%	50.99%	5.88%	9.80%
Intersection type (InT)	Segment	1050	30.76%	34.67%	33.05%	1.52%
	Four-Way Intersection	177	81.36%	10.17%	0.00%	8.47%
	T-Intersection	84	73.81%	21.43%	0.00%	4.76%
	Others	11	72.73%	18.18%	9.09%	0.00%
Crash hour (Hor)	12:00–5:59 a.m.	296	18.24%	27.03%	50.00%	4.73%
	6:00–11:59 a.m.	405	52.60%	31.11%	15.06%	1.23%
	12:00–5:59 p.m.	406	55.42%	29.80%	13.30%	1.48%
	6:00–11:59 p.m.	215	32.09%	34.89%	28.37%	4.65%
Manner of collision (MnC)	Not Collision with Motor Vehicle	408	13.24%	48.53%	35.78%	2.45%
	Angle	377	80.64%	10.34%	3.18%	5.84%
	Front-to-Front	200	77.50%	20.50%	2.00%	0.00%
	Front-to-Rear	227	24.23%	34.80%	40.09%	0.88%
	Sideswipe	87	42.53%	35.63%	18.39%	3.45%
	Others	23	0.00%	43.48%	56.52%	0.00%
Fire (Fir)	No	1129	45.79%	31.18%	20.55%	2.48%
	Yes	193	22.80%	25.91%	47.66%	3.63%
Number of total vehicles (VhT)	Single	314	3.18%	54.46%	39.81%	2.55%
	Two	765	62.74%	21.70%	12.55%	3.01%
	Multi	243	29.22%	28.40%	40.73%	1.65%
Hazardous material class (HzC)	Flammable/Combustible Liquid	727	43.06%	30.67%	22.83%	3.44%
	Gases	191	52.88%	29.32%	15.18%	2.62%
	Corrosive	94	39.36%	26.60%	34.04%	0.00%
	Explosives	29	37.93%	20.69%	41.38%	0.00%
	Others	281	35.23%	32.74%	30.25%	1.78%
Lighting condition (LgC)	Daylight	783	55.04%	29.63%	14.18%	1.15%
	Dark–Lighted	131	23.66%	25.19%	40.46%	10.69%
	Dark–Not Lighted	348	22.13%	41.66%	33.62%	2.59%
	Dawn/Dusk	60	36.67%	33.33%	25.00%	5.00%
Weather (Wth)	Clear	909	43.90%	29.15%	23.87%	3.08%
	Cloudy	212	35.85%	42.92%	18.87%	2.36%
	Rain/Snow	136	29.41%	36.76%	33.09%	0.74%
	Others	65	32.31%	32.31%	33.84%	1.54%
Visual obstruction (VsO)	No	1259	43.05%	30.50%	23.83%	2.62%
	Yes	63	30.16%	28.57%	38.10%	3.17%

illustrates the variable distribution in each cluster in a heatmap format. Risk factors highlighted in dark blue indicate overrepresentation. This table effectively illustrates the “between clusters” proportions of variable categories. The combination of variable categories identified in the clusters helps to discover the distinct and interpretable patterns for fatal crashes involving HAZMAT trucks.

• HAZMAT-truck-involved fatal crashes at intersections on two-way undivided roadways overrepresent in Cluster 1. Intersection crashes (including four-way intersections and T-intersections) account for nearly 20% of the whole dataset, as shown in Table 2, and it is over 70% in Cluster 1. Similarly, 51.89% of fatal HAZMAT truck-related crashes occurred on two-way undivided roadways in the whole dataset. The variable accounts for 63.85% of the crashes in Cluster 1. This cluster represents 40% of the crash data, indicating HAZMAT-truck-involved fatal crashes under certain conditions. The increased conflicting points and interfering factors, including vision field, pedestrian crossing, and traffic lights at intersections, could explain the higher probability of fatalities once a crash occurs. This finding is consistent with previous studies [20,22,23,27,28,48].
• Intersection crashes, especially angle (80.64%) and front-to-front (77.50%) crashes on arterials or collectors, are prominently associated with fatalities, as shown in Cluster 1. Previous studies have shown that head-on and angle crashes are more harmful than other collision types at such locations, as these crashes commonly result in fatalities and severe injuries [13,28].
• Curve alignments also have higher associations with HAZMAT-truck-involved fatal crashes. The variable accounts for nearly 20% of the whole dataset, as shown in Table 2, and it is 34.68% and 35.89% in Cluster 1 and Cluster 2, respectively. The combined intersections and curves challenge drivers because of their unique design and functions, increasing the crash risk of vehicles running off-road or rollovers, leading to fatalities and severe injuries. This finding is consistent with previous studies [10,14,15,16,20,21,25,49,50].
• The result shows that driver status is a critical risk factor in HAZMAT-truck-involved fatal crashes in Cluster 3—specifically, multiple vehicles at urban interstate highway segments with relatively high speed limits under dark lighting conditions. Distraction, asleep or fatigue, and physical impairment could impair driving ability and cause significant cognitive inadequacies, increasing the possibility of fatal and severe injury crashes. The result complies with the previous studies [10,11,12,13,14,16,20,27,28,47,51]. Commercial truck drivers are more commonly subjected to longer travel time and driving distances than passenger vehicle drivers, bringing them a higher safety risk for distraction and fatigue [52,53]. The driver behavior-related crash pattern identified in this study emphasizes the need for safety training, education program, and advanced driver supervision (i.e., a driver monitoring system with fatigue and distracted driving detection/alert function) for HAZMAT truck drivers. Crash prevention training is one of the most effective ways of promoting the application of safe HAZMAT road transportation practices and procedures [54].
• The dark–lit condition is associated with interstate highways, predominantly on two-way divided roadways with a positive median barrier in urban areas (40.46% in Cluster 3). On the contrary, the dark–not lighted condition is associated with a lower functional class in rural areas (41.66% in Cluster 2). Poor visibility caused by adverse lighting conditions during nighttime is a critical risk factor for HAZMAT-truck-involved fatal crashes. It adheres to past research [9,11,12,15,16,17,20,24,28,47,55]. According to a Florida study, installed lighting positively affects the reduction in crashes for all crash types and severity levels by 37% [56].
• This study identifies the association between HAZMAT-truck-involved fatal crashes and adverse weather conditions. Adverse weather conditions could pose challenges to drivers due to the reduced road friction coefficient and relatively poor visibility. The presence of driver inattention, fatigue, physical impairment (Cluster 2), dark lighting conditions (dark–lighted and dark–not lighted), or special road locations such as entrance/exit ramps (Cluster 3) could intensify the challenges.
• Front-to-end crashes on high-speed (65 mph or more) urban interstate highways and freeways overrepresent in Cluster 3. Higher speed limits are associated with HAZMAT-truck-involved fatal crashes. An explanation for these rear-end crashes on high-speed interstate highways and freeways is failure to keep a safe speed and distance for maneuver adjustment. In addition, multiple vehicle crashes at such locations also increase the potential risk of HAZMAT release and explosion. Cluster 3 illustrates the overrepresented fire (47.66%), as shown in Table 4.
• The presence of overrepresented other information identified in Cluster 4 implies a higher possibility of hit-and-run situations, combined with angle crashes, dark–lighted conditions, speed limits less than 25 mph, and local roadways. This finding is similar to a previous study [57]. However, there are only 35 crashes in this group, which is challenging to have a conclusive result.
• Additionally, some factors with small observations in the whole dataset may not be apparent to have a chance to show their impact. However, these factors show different proportion distributions in different clusters, further explaining their correspondence with fatal HAZMAT crashes. For example, Cluster 2 illustrates the overrepresented factor of drivers with previously recorded suspensions, revocations, and withdrawals (40.47%) and drivers aged under 25 (58.33%). The combination of the factors may be rare but strongly associated with fatality given an accident, which was hidden in the whole data descriptive statistics.

4.3. Implications of Study Findings

Research on crash safety analysis ultimately aims to provide a scientific basis for crash and injury reduction by developing state of the art in modeling, management, and policymaking. This study applies an advanced categorical data mining technique, cluster correspondence analysis, to identify the critical clusters that describe the patterns of influencing factors associated with fatal crashes involving HAZMAT trucks. The factors examined in the current study for various scenarios could help transportation authorities build data-driven interventions and countermeasures. Considering the study scope, the following recommendations are based on the results, focusing on the planning level without involving the project level.

• Considering the high percentage of HAZMAT-truck-related fatal crashes at intersections on rural two-way undivided roadways, strategies suggested by Federal Highway Administration (FHWA) could benefit safety improvement. For example, improving driver awareness of intersection approach(es) by providing enhanced signing, delineation, and supplementary messages with the systemic application of multiple low-cost countermeasures may help to reduce fatal and injury crashes by 10–27% for both signalized and stop-controlled intersections [58,59].
• To reduce intersection fatal crashes, especially angle and front-to-front crashes involving HAZMAT trucks, dedicated left- and right-turn lanes for physical separation between turning traffic that is slowing or stopped and adjacent through traffic at approaches to intersections may serve as one of the potential countermeasures. Offset turn lanes can provide added safety benefits for a reduction in fatal and injury crashes (36%) and total crashes (14–26%) [60].
• HAZMAT shipment carriers and stakeholders can utilize the findings of driver-behavior-related factors to develop safety training materials and education programs.
• Enhancing roadway lighting and installing a roadside weather-responsive warning system could reduce HAZMAT-truck-related fatal crashes during nighttime and adverse weather conditions.
• The real-time advisory speed limit on high-speed urban interstate highways and freeways can mitigate corresponding rear-end crash risks.

One of the most important factors in risk mitigation is the site of crashes involving HAZMAT trucks. If the locations are close to businesses and residential areas, it will increase the possibility of catastrophic effects in the event of a leak or explosion. Based on the cluster correspondence analysis results, roadway conditions, such as intersections on two-way undivided roadways, curve alignments, and high-speed urban interstate highways, are strongly associated with HAZMAT-truck-involved fatal crashes. As part of the permitting process, jurisdictions and policymakers may set guidelines for an interactive mechanism allowing HAZMAT shipment carriers, transportation management authorities, and stakeholders to communicate with local first responders and create emergency plans. This ensures that first responders are informed of and ready to address the new potential risks under different roadway conditions.

5. Conclusions

HAZMAT-truck-involved crashes pose considerable threats to the public, property, and environment. To reduce crashes, particularly fatal crashes, it is critical to identify the crash patterns and risk factors associated with the crash occurrence. This study uniquely contributes to the research of HAZMAT-truck-involved crashes in several aspects. By collecting ten-year (2010–2019) fatal crash data from the FARS database, this study applies a comparatively new categorical data analysis approach combining cluster and correspondence analyses. It depicts the clusters and variable categories to a low-dimensional approximation through a biplot. Furthermore, the estimation of top positive residuals demonstrates the correlations between various variables, highlighting the most prevalent fatal crash scenarios involving HAZMAT trucks. Instead of a subjective method, the cluster correspondence analysis approach uses automated data segmentation following the dimensional reduction in categorical data.

A crash is the complicated and interconnected result of risk factors, including driver, road, vehicle, and the environment. The four clusters for fatal HAZMAT-truck-involved crashes provide interesting insights. Through interactions with associated factors, this study identifies the underlying patterns and the high-risk scenarios where fatal HAZMAT-truck-involved crashes are more likely to occur. This study reveals that fatal HAZMAT-truck-involved crashes are highly distinguishable concerning collision types (angle and front-to-front crashes, single-vehicle crashes, and front-to-end crashes) and roadway geometric variables, such as two-way undivided roadways, curve alignments, and high-speed (65 mph or more) urban interstate highways. In addition to the factors mentioned above, driver behavior (distraction, asleep or fatigue, and physical impairment), lighting conditions (dark–lighted and dark–not lighted), and weather are also interrelated. The study also provides “between clusters” proportions by the variable categories, which help identify important risk factors with small observation and the combination of variables strongly associated with HAZMAT-truck-involved fatal crashes that have yet to be explored or hidden in the whole crash dataset.

The associations of risk factors identified in this study provide implications for potential preventive measures to improve HAZMAT road transportation safety. Understanding the crash mechanism and recognizing why, where, and how HAZMAT-truck-involved fatal crashes occurred is important. HAZMAT shipment carriers, transportation agencies, and stakeholders can use the results to develop countermeasures that prioritize safety goals and objectives in the emphasized areas of targeted crash patterns. Several critical tasks need further consideration: developing a data-driven summary of the causes of crashes and primary risk factors, creating an information guide for the risk factors to enhance managerial and operational awareness and expertise of preventive measures, and enhancing knowledge of effective technologies, safety culture, and prevention practices and procedures to improve HAZMAT road transportation safety.

This study has several limitations. The research identifies risk factors from the FARS database’s ten-year (2010–2019) fatal crash data. Based on the data availability, the researchers limit the scope of this study to fatal HAZMAT-truck-involved crashes. With more injury crashes and no injury crash data, the crash patterns for different severity levels could shed more light on HAZMAT-truck-involved crash characteristic analysis and HAZMAT road transportation safety improvement. Future studies should explore and examine the integration of spatial, temporal, and additional data (such as HAZMAT-crash-related environmental damage and the population along the HAZMAT transportation route).