An Exploratory Computational Piecewise Approach to Characterizing and Analyzing Traffic Accident Data

w w w . i j m r e t . o r g Page 26
International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 1 ǁ June 2016.
An Exploratory Computational Piecewise Approach to Characterizing and
Analyzing Traffic Accident Data.
Farhan Saleem, Eric Asa1
and Joseph Membah.
Associate Professor, North Dakota State University, Department of Construction Management and Engineering,
Dept. 2475, Box 6050, Fargo, ND, 58108-6050
Abstract : In spite of the many advancementsinsafety technology, roadway design and engineering as well as
several policy initiativesaimed at addressingtraffic crashes (and it concomitant injuries and fatalities); it
continuesto saddle humanity and present significant health hazards and threats to the socio-economic wellbeing
of the inhabitants of this earth. Even though federal and state transportation engineers, policy makers, planners
and researchers have spent large sums of money and effort on this complex and ubiquitous problem, traffic
crashes is one of the top causes of fatal and non-fatal injuries in the United States. In this work a piecewise
approach was used to perform an exploratory,systemic characterization and analysis of the six-year (2008-
2013) traffic crash data from North Dakota to discover the nature, peculiarities and trends in the data.Several
important features and trends in the data were discovered and the outcome of this paper can be further used for
engineering design, planning and policy analysis. The research approach could be duplicated in any other state
to enhance its societal benefits. Heterogeneity and uncertainty will be fully addressed in future work.
Key words: Traffic accidents; Safety; Exploratory analysis; Piecewise.
Introduction
From an engineering standpoint the most
challenging problem in highway safety is the
persistence of traffic crashes in spite of several
improvements in engineering design and planning,
technological advances and policy initiatives to halt
this unfortunate phenomenon (Mannering et al.
2016). Traffic accidents are a major source of
property damage, injuries and deaths in the United
States and across the globe.Traffic fatalities is
considered the leading cause of deaths in the first
thirty years of life and the second highest leading
cause of deaths before age 75 in the U.S (Conner
and Smith 2014; CDC, 2010). Although several
policies and preventive measures have been
implemented in the United States over the years, an
average of about 35,000 people lose their lives to
traffic accidents every year. According to the
National Safety Council38,300 people lost their
lives to traffic accidents and another 4.4 million
(1.37% of the entire US population)were injured on
this country’sroads in 2015. It was the largest one-
year percentage increase in traffic fatalities in the
United States in half a century.In North Dakota,
traffic accidents accounted for 135 fatalities and
5,289 personal injuries in 2014 (NDDOT, 2014). In
North Dakota, a traffic accident occurred every 33
minutes; one person was injured from a traffic-
related accident every 1.7 hours; and one person
died every 2.7 days from traffic accidents in 2014
(NDDOT, 2014). It is quite evident that traffic
accidents and the concomitant fatalities and injuries
remain a major health problem and present a
significant challenge to the socio-economic well-
being of the inhabitants of the United
States.Generally, ttraffic safety analyses are
employed to evaluate potential safety issues and
identify opportunities for improving safety. In order
to identify the challenges and address the traffic
accident problem, several methods have been
employed to model and analyze traffic accident data
over the years.The predominant approach in traffic
crash research is the use of generalized linear
models (GLMs) to characterize crash data, develop
crash modification factors,analyze the relationship
between traffic accidents and different covariates,
predict values, identify hot spots and screen
variables (Geedipally et al. 2012; Wu, Chen et al.
2014; Shirazi et al. 2016).Other benefits of traffic
safety analyses are the identification of low
cost/high-impact improvement options;promotion of
safety conscious planning, design, and
implementation culture;provision of data and
information to facilitate good decision making in
future; economic benefits from reduced accidents;
liability cost reduction due to safer roads and others
(Mahmud et al, 2015). The two commonly used
methods are negative binomial and Poisson
regression models. Flavors of the Poisson-based
modelsare Poisson lognormal conditional-
autoregressive (Wang and Kockelman 2013), full
Bayesian multivariate Poisson lognormal (El-
Basyouny et al. 2014), multivariate Poisson-
lognormal regression (Ma et al. 2008), Poisson-
Weibull (Cheng et al. 2013), Poisson-gamma
mixture (Mothafar et al. 2016), geographically-
weighted Poisson (Hadaayeghi et al. 2010), diagonal
inflated bivariate Poisson regression (Lao et al.
2011), Poisson regression(Powers at al.2010), non-
canonical Poisson regression (Polus and Cohen
2012) and other Poisson-based models. Examples of
the recent applications of the negative binomial

models to traffic accident data are finite mixture of
negative binomial (Zou et al. 2014), negative
binomial-Lindley (Lord and Geedipally
2011),semiparametric negative binomial generalized
linear (Sherazi et al. 2016), finite mixture of
negative binomial with varying weight parameter
(Zou et al. 2013), negative binomial-generalized
exponential (Vangala et al. 2015), random
parameters negative binomial panel (Coruhet al.
2015), negative binomial and a generalized
estimating equation (Mohammadi et al. 2014),
random effect negative binomial (Chin and
Quddus2003) and other models. In view of the
limitations of the negative binomial and Poisson
models, machine learning, Bayesian and other
classes of models have recently attracted
considerable attention. Examples are
quantileregression (Wu, Gao et al. 2014), time series
(Quddus 2008), empirical Bayes approach (Gkritza
et al 2014), Bayesian networks (de Ona et al. 2013),
latent class cluster (de Ona et al. 2013; Cerwick et
al. 2014),spatial analysis (Pirdavani et al. 2014),
spatial regression (Rhee et al. 2016), mixed logit
(Cerwick et al. 2014; Wu et al. 2014), multinomial
logit (Ye and Lord 2014), ordered probit (Ye and
Lord 2014),Tobit regression (Anastasopoulos 2008),
discrete generalized Pareto distribution (Prieto et al.
2014), multinomial logistic regression (Bham et al.
2012), fuzzy logic (Effati et al. 2015), data mining –
classification trees and rules discovery (Montella et
al. 2011; Pakgohar et al. 2010), clustering
(Anderson 2009), support vector machines (Li et al
2012), artificial neural networks (Abdelwahab and
Abdel-Aty, 2001) and other models.
Wu et al. (2014)and Effati et al. (2015) argued that
fully parametric (as well as mixed and random)
models, lack robustness, cannot adequately handle
heterogeneity and could result in biased estimation
of crash incidence and severity as homogeneity and
independence are assumed during modeling.
Researchers have noted that traffic crash data could
(1) contain many zero values and thus be highly
dispersed (Lord and Geedipally 2011; Lord and
Geedipally 2012; Shirazi et al. 2016); (2) exhibit
high variance (low amount of information) or low
variance (high amount of information); (3) be
heterogeneous – process and statistical (Rhee et al
2016);(4) complex (Mannering et al. 2016 and
Delorme and Lassarre 2014); and (5) uncertain
(Rhee et al. 2016).Even though methodological
advances like random parameters count models
(Barua et al. 2016), random parameters Tobit
models (Anastaopoulos et al. 2012), random
parameters generalized count model (Bhat et al.
2014), latent-class (finite mixture) model (Yasmin et
al. 2014), Markov switching count model
(Malyshkina and Mannering 2010), and
bivariate/multivariate models with random
parameters (Barua et al 2015, 2016) have been
developed to address the unobserved heterogeneity;
addressing heterogeneity is still a significant
challenge as the disadvantages of these approaches
outweigh their advantages (Zou et al. 2013; Vangala
et al. 2015; Mannering et al. 2016). In general;
traffic crash data were collected from different
scales (cities, counties, states), roads (urban and
rural); road characteristics/geometries (intersections,
curves, type of surface and others) and others and
are aggregated when used in modeling. The
quantitative models were applied to the traffic
accident data without prior exploratory analysis to
discover the trends; peculiarities; and the effects of
the various factors on traffic crashes (Rhee et al
2016). In view of the inherent complexity and
uncertainty in traffic crash data, an understanding of
the relationships among the various factors and
traffic crashes could be a useful source of
information for transportation engineers (Rhee et al.
2016; Mannering et al 2016).Traffic accidents are
driven by several discrete variables and research to
understand the peculiarities of the variables and how
they relate to crashes is needed (Quddus 2008; de
Ona 2013; Zou et al. 2014). Kim et al. (2006)
indicated in their work that there are differences in
the relationship between different crash types and
the explanatory variables and using crash type
model could enhance understanding of traffic
crashes and the effectiveness of countermeasures.
According to Zou et al. (2014) and Mannering and
Bhat (2014), sub-groups of data are homogeneous as
compared to aggregates of data from different sites,
which are heterogeneous. Segregating the data into
different subgroups could improve modeling and
analysis. In thisresearch an exploratory piecewise
approach was used to characterize the traffic
accident data types to discover their peculiarities,
nature and trends.
Research Approach
This research is aimed atunderstanding the nature,
peculiarities and trends in the traffic accidents
data/factors to aid prediction, engineering
design/planning and policy analysis. These issues
have not been addressed together in the literature in
the way it is presented in this work. The
computational approach employed in the research is
summarized in Figure1.

Figure 1: Research Approach
A piecewise approach was used to breakdown the main research questionsinto several research questions
thataddresses the various aspects of the data (Figure 2).In view of the complexity, uncertainty, and
heterogeneity, fitting a complex response equation to represent the large number of unknown model coefficients
is difficult. The relationships among the traffic crashes and the various factors could be linear and/or nonlinear.
Thus the high dimensional,multivariate surface (equation 1) could be reduced to a one dimensional space
representing each of the factors (equation 2).
𝑌 𝑋 = 𝑓1 𝑥1 + 𝑓2 𝑥2 + ⋯ + 𝑓𝑛 −1 𝑥 𝑛−1 + 𝑓𝑛 𝑥 𝑛 + 𝑒 𝑛 (1)
𝑦 𝑥1 = 𝑓1 𝑥1 + 𝑒1 … … … . 𝑓𝑜𝑟 𝑠𝑢𝑏𝑓𝑎𝑐𝑡𝑜𝑟 1
𝑦 𝑥2 = 𝑓1 𝑥1 + 𝑒2 … … … . 𝑓𝑜𝑟 𝑠𝑢𝑏𝑓𝑎𝑐𝑡𝑜𝑟 2
…………………………………………………
𝑦 𝑥 𝑛 = 𝑓𝑛 𝑥 𝑛 + 𝑒 𝑛 … … … . 𝑓𝑜𝑟 𝑠𝑢𝑏𝑓𝑎𝑐𝑡𝑜𝑟 𝑛 (2)
Literature
Review
Piecewise Approach: Division into several factors
/ areas and research questions. See Figure 2.
Data Acquisition and Analysis
Primary Research Questions:
What are the nature, peculiarities and trends of the traffic accident data?
Discussions and Conclusions
Equations Curve Fitting /
Plots / Trend Lines

Area of
jurisdic
tion
Road condition – what are the effects of road conditions on accidents?
Type of surface – what type of road surface is more prone to accidents?
Surface condition - does the surface condition affect the accident rate?
Geometrics – is there any impact of road geometrics on accidents?
Intersection type – do more accidents occur at intersections?
Functional class – what class of road is more dangerous?
Major research question: What are the nature, peculiarities and trends of the traffic
accident data?
Agencies – what agencies are responsible for reporting accidents?
Counties - which counties had the most accidents?
Cities - which cities experienced the most accidents?
Rural or urban - which areas experienced more accidents – rural or urban?
Visibility
+ weather
Visibility –which light condition is more critical to traffic accidents?
Weather – which weather-related factors drive accidents?
Road
infrastruc
ture
Crash
related
factors
Human
related
factors
Crash severity - what are the various levels of severity of traffic crashes?
Classes of objects - what classes of objects are associated with accidents?
Vehicle type - which vehicles are more likely to be involved in accidents?
Access control - how does access control relate to accidents?
Manner of collision - are there any similarities in accidents related to the
manner of collision?
Travel direction – does the direction of travel affect the accident rate?
Type of road - what type of road is more critical to traffic accidents?
Road structure - which one is more dangerous- roadway, shoulder, or median?
Type of junction - which type of junction is associated with more accidents?
Other
factors
Time
related
factors
Type -what is the relationship between VMT and the types of accidents?
Licensed drivers – how does the licensed drivers relate to accidents?
Registered vehicles – how does registered vehicles relate to accidents?
Population - what is the relationshipbetween population and
accidents?ccidents?
accidents?
Alcohol/drugs – does alcohol and/or drugs play in a role in accidents?
Visual obstruction - does visual obstruction contribute to traffic
accidents?
Evasive action - do drivers take evasive actions when confronted with
potential crash situations?
Hit and run - are hit and run incidents a major component of accidents?
Year - which year had the most accidents during the study period?
Month - what months are more critical for accidents?
Day – which days of the month did more accidents occur?
Time of day – in which time slots did more traffic accidents occur?
Figure 2: Piece-wise Research Approach

Where x1,x2,…,xn are the subfactors and e1,e2,…,en are the errors. Thus the piecewise approach involves the use
of several, one-dimensional linear/nonlinear equations to accurately and efficiently approximate the complex
nonlinear surface (Ngueveu et al. 2016; Li and Cao 2008).
This approach fully captures the nature, peculiarities and trends underlying the various factors driving traffic
crashes.The research employed six-years (long term in relation to the growth North Dakota’s economy) traffic
accident data from the state of North Dakota.
Analysis and Discussions
North Dakota’sroad traffic accident data from 2008 to 2013was obtained from the Safety Division of North
Dakota Department of Transportation. This large data contain about 107,000 crash reports from 2008 to 2013.
The data comprises of different parameters including their description and coding as per NDDOT
standards.Figure 3 represents a graphical distribution of the total number of crashes within the study period
(2008-2013).
Figure 3: Map of North Dakota Showing the Total Number of Crashes from 2008 to 2013.
Table 1 is a summary of the population, licensed drivers, registered vehicles and the total number of crashesthat
occurred in the state of North Dakota from 2008 to 2013.
Table 1: Summary of data on population, drivers, vehicles and thetotal number of crashes
Year
Populatio
n
Change
%
Licensed
Drivers
Change
%
Registered
Vehicles
Change
%
Crashe
s
Change
%
2008 641,421 475,129 774,346 16387
2009 646,844 0.845 479,921 1.009 728,376 -5.937 17673 7.848
2010 672,591 3.980 487,489 1.577 726,937 -0.198 17076 -3.378
2011 683,932 1.686 496,543 1.857 889,213 22.323 18823 10.231
2012 699,628 2.295 509,195 2.548 844,617 -5.015 18356 -2.481
2013 723,393 3.397 520,083 2.138 804,332 -4.770 18977 3.383
A positive trend was observed in the state’s population from 2008 to 2013. The population of North Dakota
increased each year during the study period. From 2008-09, the population increased by about 0.85% but from
2009-10 the increment was 3.98%;which is the highest for the study period. In 2010-11 and 2011-2012 the
percent increase in population was 1.68% and 2.29%;respectively (Table 2). The increase in population from
2012 to 2013 was 3.39%.
There was a steady positive increase in licensed drivers during the study period. From 2008-09, the percent
change in the number of licensed drivers in North Dakota was about 1%, then the numbers increased to 1.57%,
1.85%, 2.54% for the periods of 2009-2010, 2011-2012 and 2012-2013, respectively. The highest percent
change (2.54%) in the number of licensed drivers was from 2011 to 2012. In 2012-2013 the percent change
slightly dropped from 2.54% to 2.13%.
The number of registered vehicles in North Dakota varied for the entire study period (2008-13). The percent
change in registered vehicles was - 5.93% from 2008 to 2009 as the number of registered vehicles dropped by
about 45,970 vehicles from 2008 to 2009. For 2009-10 the percent change in registered vehicles was only -
0.19% as the numbers of vehicles were more or less the same for the 2009-10 and the difference was of only
1,439 registered vehicles. There was a big jump in the number of registered vehicles between 2010 and 2011;as

a total of 162,276 vehicles were registered in that period. The percent change in the number of registered
vehicles was (22.32%) from 2010 to 2011. The number of registered vehicles declined afterwards. The change
in the number of registered vehicles was -5.01% and -4.76% for year 2011-2012 and 2012-2013,
respectively.The number of accidents in North Dakota has generally increased from year to year (Table 2).
There was about 7.84% increase in the number of crashes from 2008-2009; as 1,286 more crashes were reported
in 2009 compared to 2008. The percent change in crashes from 2009-2010 was - 3.37% due to a decrease in the
number of crashes (about 597 reported in that period). The numbers of crashes increased quite rapidly from
2010 to 2011, as 1,747 more crashes occurred in the one year period (representing an increase of 10.23%). This
rapid change could be the result of a large increase in the number of vehicles (162,276) registered in that period.
After that the numbers of crashes decreasedin 2012 as 467 less crashes were reported as compared to 2011 and
the percent change in crashes was - 2.48% from the previous year. The number of crashes increased in 2013 by
621 from the previous year, despite a decrease of 40,285 in the number of registered vehicles. The percentage
changes in population, the number of licensed drivers, the number of registered vehicles and the number of
crashes during the study period (2008-2013) is also shown in Table 2.
There is a positive trend in the number of crashes from year to year (2008-2013), even though there were slight
increases in 2010 and 2012. A curvilinear equation (3) was fitted to the data.
y = 10.935x3
- 66006x2
+ 1E+08x - 9E+10 (3)
The R² is 0.7843, and so the model could be used to predict the approximate number of crashes with respect to
year.Linear equations(4) and (5) were fitted to the data of population and licensed drivers of North Dakota.The
results show a strong positive trend in the case of population and licensed drivers with R²values of 0.979 and
0.978 respectively. So,these equations could be used to infer meaningful results from the data.
y = 16559x + 620013 (4)
y = 9189.9x + 462562 (5)
A curvilinear equation (6) was fitted to the data of registered vehicles. It showed that the number of registered
vehicles went up and down during different years of the study period (2008-2013). R² is 0.7686 which could be
used to predict the approximate number of registeredvehicles.
y = -12156x3
+ 125064x2
- 348440x + 1E+06 (6)
The following sections addresses the questions raised in Table 1 and discusses the detailed analysis of road
traffic accident data grouped into different categories to ascertain the intricacies of the traffic accident data.
Area and Jurisdiction Factors
Agencies:The statistical analysis shows that city police departments reported the most accidents (54.45%),
followed by county police (25.35%) and highway patrol (19.43%). The rest of the traffic accidents were
reported by campus police and BIA. Military and park rangers reported the lowest number of accidents. Thus a
greater proportion of accidents occurred within the limits of cities.
Counties:Figure 4 is the comparison of crashes versus population in the five major counties of North Dakota. It
is evident that graphs of crashes against population do not followthe same trend in the five major counties
(Figure 4). A total of 58.28% of traffic accidents occurred in the top five counties with respect to the number of
accidents are Cass, Burleigh, Ward, Grand Forks and Williams. Cass County is obviously on top due to its high
population (the highest in the state), as more people may translate into more traffic accidents. Williams County
on the other hand, which has less population (29,563); almost half than the Grand Forks population (69,311);
had a high number of accidents (6.36%).
(a) Cass (b) Burleigh (c) Grand Forks
(d) Ward (e) Williams
Legends

Figure 4: North Dakota Counties (Crashes Vs Population 2008-13)
Cass County:Figure 4(a) shows that the number of crashes in Cass County increased each year from 2008 to
2010 but not in considerable amount. In 2011 the number of crashes decreased substantially by 284. There was a
steep rise (922) in number of crashes from 2012 to 2013. A curvilinear equation (7) was fitted to the crash data.
y = 59.667x3
- 549.64x2
+ 1427.3x + 2250 (7)
R² is 0.9415 which is strong enough to infer reliable results for the number of crashes in Cass County.
Burleigh County:The results from Figure 4(b) depicted that a mixed trend as the number crashes went up and
down throughout the study period (2008-2013). A curvilinear equation (8) was fitted to the data of the number
of crashes.
y = 21.537x3
- 243.32x2
+ 855.43x + 1756.7 (8)
R² is 0.7311, which is quite good and could be used to predict the number of crashes in Burleigh County.
Grand Forks County:The results from Figure 4(c) show aslow and steady decline in the number of crashes from
2008 to 2012; except 2009. In 2013 the number of crashes jumped considerably by 133.A curvilinear equation
(9) was fitted to the number of crashes.
y = 10.843x3
- 105.08x2
+ 285.79x + 1308.7 (9)
R² is 0.949 and it could be used to infer reliable results for number of crashes in Grand Forks County.
Ward County:In Ward County the number of crashes rose gradually from 2008 to 2011 (Figure 4d). The number
of crashes declined noticeably in the last two consecutive years(2012 and 2013). A curvilinear equation (10) was
fitted to data.
y = -13.75x3
+ 98.464x2
- 54.643x + 1540 (10)
R² is 0.8106, which is quite reasonable and the model could be used to predict the number of crashes in Ward
County.
Williams County:Figure 4(e) shows that there is a positive trend and the number of crashes increased
continuously over the study period (2008-2013). A curvilinear equation (11) was fitted to the data of the number
of crashes.
y = -36.574x3
+ 388.35x2
- 922.36x + 1166.7 (11)
R² is 0.9991, so the model could be used confidently to forecast the number of crashes in Williams County. The
Williams County experienced the highest increase in the number of crashes over the study period.
Cities:The following (Figure 5) is the comparison of crashes against population in the five major cities in North
Dakota.Fargo and Bismarck are the two top cities with respect to the number of crashes reported; almost 44% of
the accidents occurred in these two big cities in the state. Bismarck is the worst city with respect to the number
of accidents compared to the population. Its accident rate (total count with respect to population) is 0.198 as
compared to Fargo which has an accident rate of 0.13. After these two cities, about 21.75% accidents occurred
in Grand Forks and Minot. West Fargo (2.94%) is the safest among the big cities in North Dakota. In order to
investigate the trends in the number of crashes from year to year (2008-2013) for all the big cities of North
Dakota, curvilinear equations (12), (13), (14), (15), and (16) were fitted to their crash data.
(a) Fargo (b) Bismarck (c) Grand Forks

(d) Minot (e) Williston
Figure 5: North Dakota Cities (Crashes Vs Population 2008-13)
Fargo: Figure 5(a) shows that the number of crashes in Fargo went up and down from 2008 to 2012. There was
aconsiderable jump in number of crashes (about 759)from 2012 to 2013. A curvilinear equation (12) was fitted
to data of number of crashes.
y = 47.843x3
- 436.29x2
+ 1145.1x + 1581.7 (12)
R² is 0.977 which is strong enough to infer reliable results for the number of crashes in Fargo.
Bismarck:The results from Figure 5(b) showa mixed trend as the number crashes varied throughout the study
period (2008-2013). A curvilinear equation (13) was fitted to the crash data.
y = 22.954x3
- 257.96x2
+ 916.94x + 1230.7 (13)
R² is 0.8149, which is quite good and the model could be used to predict the number of crashes in Bismarck.
Grand Forks: Figure 5(c) depicts the steady increase in the number of crashes from 2008 to 2011. In 2012 the
number of crashes reduced to 1140 and then it increased in 2013; reaching 1262.A curvilinear equation (14) was
fitted to data.
y = 5.1042x4
- 64.282x3
+ 266.55x2
- 392.35x + 1289.2 (14)
R² is 0.9762, which is good enough to infer reliable results for number of crashes in Grand Forks.
Minot: Like Grand Forks, there was a gradual rise in the number of crashes from 2008 to 2011 (Figure 5d). The
number of crashes declined considerably in 2012 and 2013. A curvilinear equation (15) was fitted to the crash
data.
y = -4.787x3
+ 18.675x2
+ 111.32x + 835.33 (15)
R² is 0.8514, which is quite good and the model could be used to predict the number of crashes in Minot.
Williston: The results from Figure 5(e) show that there is a positive trend and the number of crashes increased
throughout the study period (2008-2013). A curvilinear equation (16) was fitted to the data of the number of
crashes.
y = 21.946x2
+ 15.861x + 229.3 (16)
R² is 0.9375, so the model could be used to forecast the number of crashes in Williston.
Table 5 summarizes the nature of accidents in percentages of all the major cities of North Dakota. Minot is on
the top with respect to fatal accidents (0.216%), whilst Fargo and Williston were the top cities with respect to
accidents involving injuries (26.276%) and property damage only (86.481%). Detailed analysis of each sub
category (fatal, injury, PDO) can be viewed in Table 2 for all the major cities of North Dakota.
Table 2: Nature of accidents in big cities of North Dakota (%)
City
Nature of Accidents (%)
Fatal Injury PDO
Bismarck 0.075 20.906 79.019
Fargo 0.100 26.276 73.624
Grand Forks 0.100 24.704 75.196
Minot 0.216 16.157 83.626
Williston 0.135 13.383 86.481
Rural and Urban: An analysis of the data show that most of the crashes occurred on rural (44%) and on urban
minor arterial (41%) roads. Urban principal arterial (15%) is the safest road.
Legends

Visibility and Weather Related Factors
Visibility and Weather Conditions:The results from Figure 6 (a) show that, almost 60% accidents happened
during daylight. This may be due to lack of concentration by drivers.
(a) (b)
Figure 6: Effects of Light and Weather Conditions on Crashes
About 32% of accidents can be attributed to darkness (road is lighted or unlighted). About 7% of accidents
happened at dawn and dusk.The pie chart Figure 6(b) shows that most accidents occurred on clear days (when
there is no weather event). Clouds, snow, blowing snow and rainare other weather conditions that contributed to
accidents. However all the other factors (aside from clear day) account for less than 40% of traffic incidents in
North Dakota.
Road Infrastructure related Factors:
Table 3 summarizes the road infrastructure related factors involved which had effects on the number of crashes
with respect to top five contributing sub factors and their percentages.
Table 3: Road infrastructure related factors and sub factors
Road Infrastructure related Factors
Road
Condition
Surface
Type
Surface
Condition
Road
Geometrics
Intersection
Type
Functional
class
Sub
Factors
(%) Sub
factors
(%) Sub
factors
(%) Sub
factors
(%) Sub
factors
(%) Sub
factors
(%)
Normal 95.5 Asphalt 68.2 Dry 55.1 Straight
(Level)
79.0 Non-
intersect
ion
73.6 Urban
Minor
Arterial
24.5
Under
Constructi
on
1.47 Concrete 24.5 Ice /
Compact
ed Snow
27.3 Straight
(Grade)
11.7 Four
Way
Intersect
ion
18.0 Urban
Princip
al
Arterial
(Other)
19.7
Loose
Material
Surface
1.23 Gravel/
Scoria
6.27 Wet 8.27 Curve
(Level)
4.61 T
Intersect
ion
7.45 Rural
Princip
al
arterial
(Other)
12.2
Holes/Ruts
/
Bumps/
Washout
0.61 Dirt 0.66 Snow 5.91 Curve
(Grade)
3.49 Five +
Roads
0.51 Rural
Local
Road or
Street
8.26
Debris on
Road
0.36 Concrete
Bridge
Deck
0.20 Slush 1.36 Hillcrest 0.89 Y
Intersect
ion
0.26 Urban
Collect
or
7.10
Road Condition:Table3shows that 95.5% of accidents that occurred in North Dakota in the study period
happened under normal road conditions; only a few other conditions like loose material surfaces (1.233%)and
construction areas (1.472%)contributed to more accidentsthan the other factors.

Surface Type:Table 3 depicts that almost 98.21% of accidents occurred on good surfaces like asphalt and
concrete that again shows the driver - related behavior may be the main contributor to the high traffic accident
rates.Table 4 shows the crash severity with respect to major surface types. It depicts that 71.6% fatal accidents
occurred on asphalt surfaces. Whereas, 16.9% and 10.5% fatal accidents occurred on gravel/scoria and concrete
surfaces respectively.In terms of accidents involving injuries and PDOs, asphalt surfaces are on top, while
concrete and gravel/scoria surfaces are in second and third positions respectively.
Table 4: Crash severity with surface type
Surface Type
Crash Severity (%)
Fatal Injury PDO
Asphalt 71.608 65.0748 69.2416
Concrete 10.490 25.4219 24.3977
Gravel/Scoria 16.923 8.4148 5.5008
Other 0.979 1.0885 0.8601
Surface Condition:Aside from types of road surface, surface conditions are also important. The results from
Table 3 show that almost 55% of crashes were reported when the surface was dry and good for driving, while
ice/ compacted snow, snow and wet surface type contributed to
about 41.75 % of the accidents.
Road Geometrics:In terms of road geometrics,Table 3 depicts that79% accidents occurred on straight and level
sections of the road, which is quite unusual as compared to straight on grade (11.7%), curve on grade (4.6%)
and curve on level (3.5%).
Intersection Type:Intersections are an important aspect ofthe road structure and they could present challenges to
drivers. Table 3reveals that about 18% of accidents occurred at four way intersections and 7.4% happened at T
intersections. The majority of accidents (73.6%)occurred at sections of the road where there were no
intersections.
Functional Class:The results from Table 3 show that the top three dangerous functional classes of road with
respect to accidents were urban minor arterial (24.5%), urban principal arterial (19.8%) and rural principal
arterial (12.3%).
Crash Related Factors
Table 5 summarizes the crash related factors with respect to the top five contributing sub factors and their
percentages during the study period (2008-2013).
Table 5: Crash related factors and sub factors
Crash related Factors
Crash Severity Classes of objects Vehicle Type Access Control Manner of
Collision
Original
Direction of
Travel
Sub
Factors
(%) Sub
factors
(%) Sub
factors
(%) Sub
factors
(%) Sub
factors
(%) Sub
fact
ors
(%)
PDO 79.9 MV in
Transport
57.2 Passeng
er Car
44.0 No
Control
91.1 Non-Coll.
w/Motor
Veh.
31.2 W 24.0
Non-
incapacitat
ing injury
9.70 Deer 16.0 Pickup -
Van -
Utility
43.9 Full
Control
8.28 Rear End 23.9 E 23.6
Possible
Injury
8.09 Parked
Motor
Vehicle
5.67 Truck
Tractor
3.28 Partial
Access
Control
0.61 Angle (Not
Specific)
21.1 N 23.2

Incapacitat
ing Injury
1.61 Overturn /
Rollover
4.92 Motorcy
cle
1.12 -- -- Sideswipe
(same dir.)
7.69 S 23.2
Fatal 0.66 Ran Off
Roadway
3.03 3+ Axle 1.05 -- -- Right Angle 7.47 SE 1.81
(N = North; S = South; E = East; W = West; SE = Southeast)
Crash Severity:From the data provided in Table 5property damage only accidents were about 80% of the total
crashes in the study period. The most critical accidents are fatal and non-fatal injuries. Fatal accidents were 715
about (0.66%), whilst incapacitating injuries and non- incapacitating injuries were 1.6% and 9.7% respectively.
Classes of objects:In all the crashes reported the top five classes of objects were motor vehicle in transport
(57.2%), deer (16%), parked motor vehicle (5.6%), overturn/ rollover (4.9%) and ran off roadway (3%)
accidents (Table 5).
Vehicle Type (Unit Configuration):Table 5shows that passenger cars (44%) and pickup-van–utility vehicle
(43.9%) were the top two types of vehiclesthat were involved in most crashes in North Dakota. Truck tractors
and motorcycleswere also involved in accidents but in much lesser numbersthan the abovementioned, about
3.3% and 1.1% respectively.
Access Control:Table 5 shows that only 8.3% accidents are related to full control (only ramp entry and exit)
while the rest 91.1% are related to no control (unlimited access).
Manner of Collision:Non collision with motor vehicle in transport was the top manner of collision in most of
the cases (Table 5). The second highest is rear end collision and the third is angular collisions (but there were no
specifics).
Original Direction of Travel:The results (Table 5)show that accidents occurred in all the directions and the
distributionis almost symmetrical in North Dakota.
Type of road:Most accidents occurred on undivided roads (72%) and about 18% happened on divided highway
(Figure 7).
Figure 7: Crash related sub factors with ranking
Roadway structure: The results from Figure 7 show that more accidents happened on the main roadway
(80.4%) as compared to shoulders (15.95%) and median (2.5%).
Type of junction: About 27.6% of crashes occurred at intersections whereas 60.8% of accidents occurred where
there were no junctions or intersections and so the vehicle was travelling in a curve or straight section of road
(Figure 7).
Human related Factors
Alcohol/ Other Drug Involvement:Figure 8(a) shows that only 6.57% accidents were reported as drivers under
the influence of alcohol/ drugs or both. In 85.4% of the cases neither was present.

(a) (b)
(c) (d)
Figure 8: Effects of Human related Factors on Crashes
Visual Obstruction: In 95.3 % of crashes there were no visual obstructions when the accident happened
whereas,rain/snow/frost and fog/smoke/dustcontributed to 1.28% and 0.38% respectively, Figure 8(b).
Evasive Action: A total of 48.5% of drivers did nothing while entering into a potential crash and 41.1% slowed
or stopped the vehicle, while about 9.4% turned left or right, Figure 8(c). Thus the majority of drivers did not
take any evasive actions when confronted with a potential crash.The accidents might have taken them by
surprise.
Hit and Run: Figure 8(d) shows that in almost 92% accidents the parties involved did not leave the scene of the
accident whilstin 8% of the cases, drivers left the accident scene.
Factors Related to Year, Month, Day and Time
Crashes by Year:The numbers of accidents increased in each year except 2010 and 2012 when there were slight
decreases in percentages from previous years.
Crashes by Month:November, December, January and February are the most critical months during the entire
period of 2008-2013. December is the deadliest among all the months.
Crashes by Day of Month:The 2nd
and 11th
day of the month were the most critical, while the rest of the days of
a typical month had close to the same probability of accidents occurring.
Crashes by Time of Day:The most risky time slots were 3:00 pm to 4:00 pm and 5:00 pm to 6:00 pm in the
evening and 7:00 am to 8:00 am in the morning.
Other Factors
Crash Severity - Rate per 100 Million Vehicle Miles Traveled (2008-2013):The following section discusses the
detailed analysis of total crashes and different types of crashes namely fatalities, injuries and property damage
only (PDO), with respect to 100 million vehicle miles traveled in North Dakota. Equations were fitted to see the
positive or negative trends for the different types of crashes with respect to 100 million vehicle miles traveled in
North Dakota.
Fatal Crashes:Figure 9 (a) shows the rate per 100 million VMT for fatal crashes in North Dakota during the
study period (2008-2013). The rate per 100 million VMT for fatal accidents increased from 2008 to 2013. The
rate per 100 million VMT for fatal accidents was 1.46 for year 2009 which was the highest within the study
period (2008-2013). The rate decreased from 1.46 to 1.1 in 2010. Due to an increase in the number of fatalities

in years 2011 and 2012, the rate per 100 million VMT for fatal accidents increased to 1.41 in 2011 and 1.45 in
2012. It dropped from 1.45 to 1.3 in 2013. A curvilinear equation was fitted to the rate per 100 million VMT for
fatal crashes (Eqn. 17).
y = -0.0231x4
+ 0.3135x3
- 1.4415x2
+ 2.5864x - 0.1476 (17)
The R² is 0.572, which is relatively low to deduce meaningful results for the rate per 100 million VMTfor fatal
crashes.
(a) Fatal (b) Injury (c) PDO Rate
Figure 9: Crash Severity - Rate per 100 Million Vehicle Miles Traveled (2008-2013)
Injury Crashes:Figure 9 (b) shows the rate per 100 million VMTfor crashes involving injuries in North Dakota
during the study period (2008-2013). The rate per 100 million VMTfor injury crashes ranges from 37.2 to 40.3
and generally decreased from 2008 to 2013. The highest rate (40.3) was recorded in 2008, whereas the lowest
rate was recorded in 2012. A curvilinear equation was fitted to the rate per 100 million VMT for injury crashes
(Eqn. 18).
y = 0.1288x3
- 1.3663x2
+ 3.6677x + 37.724 (18).
The results show a negative trend in the case of rate of injury crashes per 100 million VMT. R² is 0.768 that
could be used to estimate injury crashes.
Property Damage Only (PDO):Figure 9(c) shows the rate per 100 million VMTfor crashes involving property
damage in North Dakota during the study period (2008-2013). The rate per 100 million VMTfor crashes
involving property damage decreased significantly between 2008 and 2013. The rate consistently dropped each
year from 174.5 to 143.4 from 2008 to 2012. In 2013 the rate again increased from 143.4 to 147.9. A curvilinear
equation was fitted to the rate per 100 million VMT for property damage only crashes (Eqn. 19).
y = 1.2331x3
- 13.5x2
+ 35.909x + 151.26 (19)
The R² is 0.9041, so it could be used to predict the rate per 100 million VMT for property damage only crashes.
The results also show a negative trend in the case of PDO crashes.
Total Crashes: Despite the fact that the VMT in North Dakota increased each year in the study period, the rate
of total crashes per 100 million VMT decreased from 2008 to 2013. The only exception is 2009, where the rate
per 100 million VMT for total crashes increased from the previous year (from 215.4 to 222.5). The rate per 100
million VMT for years 2010 and 2011 were almost the same. In 2012 the rate per 100 million VMT was
dropped to 181.8 and it increased in 2013 and reached 187.9. An equation was fitted to the rate per 100 million
VMT for total crashes (Eqn. 20).
y = 1.3726x3
- 14.982x2
+ 40.079x + 189.2 (20)
The results also show the negative trend in case of total rashes. The R² is 0. 0.8991and the equation could be
used to infer meaningful results.
Crash Severity - Rate per 1,000 Licensed Drivers (2008-2013):The following section discusses the detailed
analysis of the total crashes as well as the different types of crashes - fatalities, injuries and property damage
only (PDO) with respect to 1000 licensed drivers in North Dakota. Equations were fitted to the data.
Fatal Crashes:Figure 10 (a) shows the rate per 1000 licensed drivers involved in fatal crashes in North Dakota
during the study period (2008-2013). The rate per 1000 licensed drivers for fatal accidents increased from 2008
to 2013. The maximum rate per 1000 licensed drivers was 0.28 in 2012. The maximum number of fatal
accidents in 2012 was 144. The lowest rate per 1000 licensed drivers was 0.18 in year 2010 which was the safest

year of the study period. A curvilinear equation was fitted to the rate per 1000 licensed drivers for fatal crashes
(Eqn. 21).
y = -0.0074x3
+ 0.0783x2
- 0.2196x + 0.3554 (21)
The results show a positive trend in the case of fatal crashes. The R² is 0.9715, which is strong enough to deduce
meaningful results for the rate of fatal crashes per 1000 licensed drivers.Injury Crashes:Figure10 (b) shows that
the rate per 1000 licensed drivers for injury crashes in North Dakota during the study period (2008-2013). The
rate per 1000 licensed drivers for injury crashes steadily increased from 2008 to 2013 and the range varied from
6.4 to 7.5. It shows that about 16% increase in the rate of injuries per 1000 licensed drivers from 2008 to 2013.
Figure 10: Crash Severity - Rate per 1,000 Licensed Drivers (2008-2013)
A curvilinear equation (22) was fitted to the rate per 1000 licensed drivers for injury crashes.
y = -0.1275x3
+ 1.3573x2
- 3.7998x + 8.8152 (22)
The results also show the positive trend in the case of rate per 1000 licensed drivers for injury crashes.R² is
0.738 and the equation could be reasonably used to estimate the rate of injury crashes per 1,000 licensed drivers.
Property Damage Only:Figure 10 (c) shows property damage crashes per 1000 licensed drivers for North
Dakota during the study period (2008-2013). The rate per 1000 licensed drivers involved in property damage
accidents went up and down between 2008 and 2013. The rate was 27.9 in 2008, and then it went down to 22.2
in 2009 and went up to 28 in 2010. The maximum rate was 30.4 which was recorded in 2011.A curvilinear
equation was fitted to the rate per 1000 licensed drivers for property damage only crashes (Eqn. 23).
y = 0.4512x4
- 6.7656x3
+ 34.79x2
- 69.305x + 68.747 (23)
The R² is 0.995, so it could be used to predict the rate per 1000 licensed drivers for property damage only
crashes. The results show that there is a positive trend in the case of PDO crashes.
Total Crashes:As the number of licensed drivers in North Dakota increased during the study period, the rate per
1000 licensed drivers for the total crashes also increased. In 2011 the rate per 1000 licensed drivers for the total
crashes was 37.9 which was the highest recorded for the study period and the lowest rate was 34.9 (2008). After
that the rate per 1000 licensed drivers for total crashes decreased from 37.9 to 36 in 2012 and then it went up to
36.4 in 2013. A curvilinear equation was fitted to the rate per 1000 licensed drivers for total crashes (Eqn. 24).
y = -0.5798x3
+ 6.0835x2
- 17.234x + 45.214 (24)
The results show the positive trend in case of total rashes.R² is 0.6242, not good enough to infer meaningful
results.
Crash Severity - Rate per 1,000 Registered Vehicles (2008-2013): The following section discusses the detailed
analysis of total crashes and the different types of crashes - fatalities, injuries and property damage only (PDO)
with respect to 1000 registered vehicles in North Dakota. Equations were fitted to the data to study the trends in
the different types of crashes with respect to 1000 registered vehicles in North Dakota.
Fatal Crashes:Figure 11 (a) shows the rate per 1000 registered vehicles for fatal crashes in North Dakota during
the study period (2008-2013). The rate per 1000 registered vehicles for fatal accidents varied from 2008 to 2013
and ranges from 0.12 to 0.17. The lowest value was recorded in 2008 and the highest value was recorded in
2012. The rate per 1000 registered vehicles dropped from 0.17 in 2012 to 0.16 in 2013 as the numbers of
registered vehicles decreased by 40,285. A curvilinear equation was fitted to the rate per 1000 registered
vehicles for fatal crashes (Eqn. 25).The results show the positive trend in the case of fatal crashes
y = -0.0034x4
+ 0.0479x3
- 0.2289x2
+ 0.4324x - 0.1222 (25).
The R² is 0.9494, and it can be used to infer meaningful results for the rate per 1000 registered vehicles for fatal
crashes.

Injury Crashes:Figure 11 (b) shows the rate per 1000 registered vehicles for crashes involving injuries in North
Dakota during the study period (2008-2013). The rate per 1000 registered vehicles for injury crashes ranges
from 3.9 to 4.8 and generally increased during the study period (2008-2013). The rate per 1000 registered
vehicles increased from 3.9 to 4.6 from 2008 to 2010 and then it dropped to 4.0 in 2011. The rate went up to 4.4
in 2012 and then again increased to
Figure 11: Crash Severity - Rate per 1,000 Registered Vehicles (2008-2013)
4.8 in 2013. A curvilinear equation was fitted to the rate per 1000 registered vehicles for injury crashes (Eqn.
26).
y = 0.055x3
- 0.5658x2
+ 1.7807x + 2.6877 (26)
R² is 0.8192, which is good enough to infer results from the model. The results also show a positive trend in the
case of rate per 1000 registered vehicles for injury crashes.
Property Damage Only:Figure 11 (c) shows the rate per 1000 registered vehicles for property damage only
crashes in North Dakota during the study period (2008-2013). The rate per 1000 registered vehicles for property
damage crashes was not consistent in the study period (2008-2013). The maximum rate per 1000 registered
vehicles of 19.7 was recorded in 2009 and the minimum of 17.1 was recorded in 2012. A curvilinear equation
was fitted to the rate per 1000 registered vehicles for property damage only crashes (Eqn. 27).
y = 0.3057x3
- 3.2377x2
+ 9.8118x + 10.37 (27)
The R² is 0.9307, so it could be used to predict the rate per 1000 registered vehicles for property damage only
crashes. The results show that there is mixed trend in the case of PDO crashes as the rate varied during different
years of the study period (2008-2013).
Total Crashes: Even though there was an increase in the number of vehicles registered in North Dakota each
year, the rate of total crashes per 1000 registered vehicles did not increase consistently from 2008 to 2013. A
maximum rate of 24.2 was recorded in 2009, which was surprisingly quite high as the highest numbers of
vehicles (889,213) were registered in 2011. The rate per 1000 registered vehicles for years 2008, 2011 and 2012
were almost same. In 2013 the rate per 1000 registered vehicles increased to 23.6. A curvilinear equation was
fitted to the rate per 1000 registered vehicles for total crashes (Eqn. 28).
y = 0.3626x3
- 3.8223x2
+ 11.664x + 13.065 (28)
The results show that there is a mixed trend in the case of total crashes.R² is 0.9139, so it could be used to
predict the rate per 1000 registered vehicles for total crashes.
Crash Severity - Rate per 1,000 Population (2008-2013):This section discusses the detailed analysis of the total
crashes and the different types of crashes - fatalities, injuries and property damage only (PDO) with respect to
1000 population in North Dakota. Equations were fitted to the data to study the trend.
Fatal Crashes:Figure12 (a) shows the rate per 1000 population for fatal crashes in North Dakota during the
study period (2008-2013). The rate per 1000 population for fatal accidents increased from 2008 to 2013; but not
consistently. The maximum rate per 1000 population was 0.21 in 2012, where the maximum number of fatal
accidents was 144. The lowest rate per 1000 population was 0.13 in year 2010. A curvilinear equation was fitted
to the rate per 1000 population for fatal crashes (Eqn. 29).
y = -0.0037x4
+ 0.0499x3
- 0.2241x2
+ 0.3974x - 0.0666 (29)
The results show the positive trend in fatal crashes. The R² is 0.8123, which could be used to infer meaningful
results for the rate per 1000 population for fatal crashes.

Figure 12: Crash Severity - Rate per 1,000 Population (2008-2013)
Injury Crashes:Figure 12 (b) shows that the rate per 1000 population for crashes involving injuries in North
Dakota during the study period (2008-2013). The rate per 1000 population for injury crashes increased from 4.7
to 5.4 in 2008 to 2013. It shows that about 12.6% rate per 1000 population for injury crashes increased from
2008 to 2013. Linear equation was fitted to the rate per 1000 population for injury crashes (Eqn. 30).
y = 0.1346x + 4.6467 (30)
The R² is 0.9348, which could be used confidently to infer meaningful results. The results show asteady positive
trend in case of rate per 1000 population for injury crashes.
Property Damage Only:Figure 12 (c) shows property damage crashes per 1000 population for North Dakota
during the study period (2008-2013). The rate per 1000 population involved in property damage accidents
varied from 2008 to 2013. It did not exhibit consistency and went up and down each year. The maximum rate
was 22.2 which was recorded in 2009. The rate was 20.3 in 2010, and then it went up to 22.0 in 2011 and went
down to 20.6 for the years 2012 and 2013. A curvilinear equation was fitted to the rate per 1000 population for
property damage only crashes (Eqn. 31).
y = -0.0538x4
+ 0.7897x3
- 4.0771x2
+ 8.4726x + 15.668 (31)
The R² is 0.2651, so it could not be used to predict the rate per 1000 registered vehicles for property damage
only crashes. The results show that there is a mixedtrend in the case of PDO crashes.
Total Crashes:The population in North Dakota increased each year during the study period. However, the
number of total crashes per 1000 population did not increase substantially from 2008 to 2013. In 2011 the total
number of crashes per 1000 population was 27.5; which was the highest recorded for the study period and the
lowest rate was 25.3 (2010). After 2011, the rate per 1000 population for total crashes decreased from 27.5 to
26.2 in 2012 and remained the same in 2013. A curvilinear equation was fitted to the rate per 1000 population
for total crashes (Eqn. 32).
y = 0.023x3
- 0.3544x2
+ 1.583x + 24.515 (32)
The results show that there is a varied trend in case of rate per 1000 population for total rashes.
R² is also quite low (0.1488) to infer meaningful results.
Conclusions
Traffic crashes and its associated fatalities, injuries
and property damages remain a long-standing
problem in road transportation; despite several
advances in technology, highway design and
engineering and safety policy initiatives.Various
flavors of generalized linear models - negative
binomial and Poisson are the predominant
approaches used to characterize traffic crash data. Of
late machine learning and other models have been
used. Generally these models were used without prior
exploratory analysis of the traffic data; even though
traffic accidents are driven by several discrete
variables and understanding them could throw more
light on research results. In view of its pure vastness,
diversity, complexity, heterogeneity, uncertainty, and
other characteristics of traffic crash data,aggregating
the data for use during modeling could lead to wrong
conclusions. Research is needed to ascertain the
nature, peculiarities and trends in the various factors
underlying traffic accidents. In this work a piecewise
approach was used to break down the research
question into several sub-questions (organized around

seven major factors) to enhance the understanding of
the nature, peculiarities and trends in the data.
It was evident that each county, city, and other sub-
areas have their own trends and peculiarities. There
was a positive trend statewide in the number of
crashes from year to year even though the number of
registered vehicles went up and down during the
period of the study (2008-2013). City Police
departments documented most of the crashes
(54.45% compared with 25.35% by the counties), and
William’s County (in the oil producing Bakken
region of North Dakota) experienced the highest
increase in the number of crashes. Even though the
most accidents occurred in the two largest cities –
Fargo and Bismarck- Minot experienced the highest
number of traffic fatalities; whilst Fargo and
Williston saw the most accidents involving injuries
and property damage only, respectively. The irony is
thata high proportion of the accidents (60%) occurred
during the day, when the weather was clear. Almost
all the accidents (95.5%) occurred under normal road
conditions and asphalt paved roads saw the highest
number of accidents (68.2%). Asphalt roads
accounted for the most accidents - fatalities (72%),
injuries (65%) and property damages only (69%) and
is the most dangerous road surface to travel in the
state. The highest number of accidents occurred on
straight and level sections of the roadway (79%);
where there were no intersections (73.6%). Property
damages only accidents account for the highest
percentage of accidents in the state (80%) and almost
all the accidents occurred on sections of the main
roadway (80%) where there is unlimited (no control)
access (91%). Undivided roads are the most
dangerous in the state and accounted for 72% of
accidents. A high percentage of the accidents did not
involve drunk driving (85.4%) and even though there
were no visual obstructions (95 %); a greater
proportion of the drivers did not take any evasive
actions (48.6%). The months of November,
December, January and February saw the most
accidents and the most risky days were 2nd
and 11th
days of the month. It is evident that the approach
used in this research brought a lot of salient
information about the accident data to light. Even
though this research was limited to North Dakota, it
could be extended to other states to enhance it
societal benefits. Further work is needed to fully
address the uncertainty and heterogeneity associated
with the traffic accident data.
Acknowledgements
This research was fully performed by the
Computational and Sustainable Infrastructure
Research Laboratory and was not funded by any
public or private entity. The data used in the research
was provided by the Safety Office of North
Department of Transportation; for which we are very
grateful.
References
[1.] Abdel-Aty, M. A., and Abdelwahab, H. T.
(2004). “Predicting Injury Severity Levels in
Traffic Crashes: A Modeling Comparison.”
Journal of Transportation Engineering, 130(2),
204-201.
[2.] Abdelwahab, H.T., Abdel-Aty, M.A. (2001).
“Development of Artificial Neural Network
Models to Predict Driver Injury Severity in
Traffic Accidents at Signalized Intersections.”
Transportation Research Record, 1746, 6–13.
[3.] Anastasopoulos, P., Mannering, F., Shankar, V.,
and Haddock, J. (2012). “A Study of Factors
Affecting Highway Accident Rates Using the
Random-parameters Tobit Model.” Accident
Analysis and Prevention, 45 (1), 628–633.
[4.] Anderson, T. K. (2009). “Kernel Density
Estimation and K-Means Clustering to Profile
Road Accident Hotspots.” Accident Analysis and
Prevention, 41, 359-364.
[5.] Barua,S., El-Basyouny, K., and Islam, M.
(2015). “Effects of Spatial Correlation in
Random Parameters Collision Count-data
Models.” Analytic Methods in Accident
Research, 5–6, 28–42.
[6.] Barua,S., El-Basyouny, K., and Islam, M.
(2016). “Multivariate Random Parameters
Collision Count-data Models with Spatial
Heterogeneity.” Analytic Methods in Accident
Research, 9, 1-15.
[7.] Bham, G. H., Bhanu, S. J., and Manepalli, U. R.
R. (2012). “ Multinomial Logistic Regression of
Single Vehicle and Multi Vehicle Collisions on
Urban U. S. Highways in Arkansas.” Journal of
Transportation Engineering, 138(6), 786-797.
[8.] Bhat, C. R., Born, K., Sidharthan, R., and Bhat,
P. C. (2014). “A Count Data Model with
Endogenous Covariates: Formulation and
Application to Roadway Crash Frequency at
Intersections.” Analytic Methods in Accident
Research, 1, 53–71.
[9.] CDC, Centers for Disease Control and
Prevention. (2010),WISQARS (Web-based
Injury Statistics Query and Reporting System),
US Department of Health and Human Services,
CDC, Atlanta, GA
(http://www.cdc.gov/injury/wisqars. Accessed
on August 23rd, 2013)
[10.] Cerwick, D. M., Gkritkza, Shaheed, M. S. and
Hans, Z. (2014). “A Comparison of the Mixed
Logit and Latent Class Methods for Crash
Severity Analysis.” Analytic Methods in
Accident Research, 3-4, 11-27.
[11.] Chin, H. C. and Quddus, M. A. (2003).
“Applying the Random Effect Negative
Binomial model to Examine Traffic Accident

Occurrence at Signalized Intersections.”
Accident Analysis and Prevention, 35, 253-259.
[12.] Conner, K. A., and Smith, G. A. (2014). “The
impact of aggressive driving-related injuries in
Ohio, 2004–2009.” Journal of Safety Research,
51, 23-31.
[13.] Coruh, E., Bilgic, A., and Tortum, A. (2015).
“Accident Analysis with Aggregated Data: The
Random Parameters Negative Binomial Panel
Count Data Model.” Analytic Methods in
Accident Research, 7, 37-49.
[14.] Delorme, R. and Lassarre, S. (2014). “A new
theory of complexity for safety research. The
case of the long-lasting gap in road safety
outcomes between France and Great Britain.”
Safety Science, 70, 488-503.
[15.] De Ona, J., Lopez, G., Mujalli, R., and Calvo,
F.J. (2013). “Analysis of Traffic Accidents on
Rural Highways Using Latent Class Clustering
and Bayesian Networks.” Accident Analysis and
[16.] Effati, M., Rajabi, M. A., Hakimpour F., and
Shabani, S. (2015). “Prediction of Crash Severity
on Two-Lane, Two-Way Roads Based on Fuzzy
Classification and Regression Tree Using
Geospatial Analysis.” Journal of Computing
Civil Engineering, 29(6), 04014099.
[17.] El-Basyouny, K., Burua, S., and Islam, T. M.
(2014). “Investigation of Time and Weather
Effects on Crash Types Using Full Bayesian
Multivariate Poisson Lognormal Models.”
[18.] Geedipally, S. R., Lord, D., and Dhavala, S. S.
(2012). “The Negative Binomial-Lindley
Generalized Linear Model: Characteristics and
Applications Using Crash Data.” Accident
Analysis and Prevention, 45, 258-265.
[19.] Gkritza, K., Souleyrette, R. R., Baird, M. J., and
Danielson, B. (2014). “Empirical Bayes
Approach for Estimating Urban Deer-Vehicle
Crashes Using Police and Maintenance
Records.” Journal of Transportation Engineering,
140(2), 04013002.
[20.] Hadayeghi, A., Shalaby, A. S., Persaud, B. N.
(2010). “Development of Planning Level
Transportation Safety Tools Using
Geographically Weighted Poisson Regression.”
[21.] Kim, D-G., Washington, S and Oh, J. (2006),
Journal of Transportation Engineering, 132(4),
282-292.
[22.] Lao, Y., Wu Y-J, Corey, J. and Wang, Y. (2011).
“Modeling Animal-Vehicle Collisions Diagonal
Inflated Bivariate Poisson Regression.” Accident
[23.] Li, Z., Liu, P., Wang, W., and Xu, C. (2012).
“Using Support Vector Machine Models for
Crash Injury Severity Analysis.” Accident
[24.] Li, X., and Cao, Y. (2008). “Projection-Based
Piecewise-Linear Response Surface Modeling
for Strongly Nonlinear VLSI Performance
Variations.” 9th
International Symposium on
Quality Electronic Design, IEEE Computer
Society, 108-113, March 17th
– 19th
, San Jose,
CA.
[25.] Lord, D. and Geedipally, S. R. (2011). “The
Negative Binomial-Lindley Distribution as a
Tool for Analyzing Crash Data Characterized by
Large Amounts of Zeros.” Accident Analysis
and Prevention, 43, 1738-1742.
[26.] Ma, J., Kockelman, K. M. and Damien, P.
(2008). “A Multivariate Poisson-lognormal
Regression Model for Prediction of Crash
Counts by Severity, Using Bayesian Methods.”
[27.] Mahmud, S. M. S., Ferreira, L., and Tavassoli,
A. (2015). “Traditional Approaches to Traffic
Safety Evaluation (TSE): Application Challenges
and Future Directions.” Bridging the East and
the West, 11th
Asia Pacific Transportation
Development Conference and 29th
ICTPA
Annual Conference, ASCE, May 27-29, 242-
262.
[28.] Malyshkina, N., and Mannering, F. (2010).
“Zero-state Markov Switching Count-data
Models: An Empirical Assessment.” Accident
Analysis and Prevention, 42 (1), 122–130.
[29.] Mannering, F. L. and Bhat, C. R. (2014).
“Analytic Methods in Accident Research:
Methodological Frontier and Future Directions.”
Analytic Methods in Accident Research, 1, 1-22.
[30.] Mannering, F. L., Shankar, V., and Bhat, C. R.
(2016). “Unobserved Heterogeneity and the
Statistical Analysis of Highway Accident Data.”
Analytic Methods in Accident Research, 11, 1-
16.
[31.] Mohammadi, M. A., Samaranayake, V. A. and
Bham, G. H. (2014). “Crash Frequency modeling
Using Negative Binomial Models: An
Application of Generalized Estimating Equation
to Longitudinal Data.” Analytic Methods in
Accident Research, 2, 52-69.
[32.] Motella, A., Aria, M., D’Ambrosio, A., and
Mauriello, F. (2011). “Data Mining Techniques
for Exploratory Analysis of Pedestrian Crashes.”
Transportation Research Record. 2237, 107-116.
[33.] Mothafer, G. I. M. A., Yamamoto, T, and
Shankar V. N. (2016). “Evaluating Crash Type
Covariances and Roadway Geometric Marginal
Effects Using the Multivariate Mixture Model.”
26.
[34.] NDDOT, North Dakota Department of
Transportation (2014). “North Dakota Crash
Summary 2014.” Report Prepared by NDDOT.
[35.] Ngueveu, S. U., Artigues, C., and Lopez, P.
(2016). “Scheduling Under a Non-reversible

Energy Source: An Application of Piecewise
Linear Bounding of Non-linear Demand/Cost
Functions.” Discrete Applied Mathematics, 208,
98-113.
[36.] Pakgohar, A., Tabrizi, R.S., Khalilli, M.,
Esmaeili, A. (2010). “The role of human factor
in Incidence and Severity of Road Crashes Based
on the CART and LR regression: a Data Mining
Approach.”Procedia Computer Science 3, 764–
769.
[37.] Polus, A., and Cohen, M. (2012). “A Non-
Canonical Poisson Regression Model for the
Prediction of Crashes on Low-Volume Rural
Roads.” IATSS Research, 35, 98-103.
[38.] Powers, S., Gerlach, R., and Stamey, J. (2010).
“Bayesian Variable Selection for Poisson
Regression with Underreported Responses.”
Computational Statistics and Data Analysis, 54,
3289-3299.
[39.] Pirdavani A., Bellemans T., Brijs, T. and Wets,
G. 2014. “Application of Geographically
Weighted Regression Technique in Spatial
Analysis of Fatal and Injury Crashes.” Journal of
Transportation Engineering, 140(8), 04014032.
[40.] Prieto, F., Gomez-Deniz, E., and Sarabia, J. M.
(2014). “Modeling Road Accident Blackspots
Data with the Discrete Generalized Pareto
Distribution.” Accident Analysis and Prevention,
71, 38-49.
[41.] Quddus, M. (2008). “Time Series Count Data
Models: An Empirical Application to Traffic
Accidents, 40, 1732-7241.
[42.] Rhee, K-A., Kim, J-K., Lee, Y-i., and Ulfarsson,
G. F. (2016). “Spatial Regression Analysis of
Traffic Crashes in Seoul.” Accident Analysis and
[43.] Shirazi, M., Lord, D., Dhavala, S. S., and
Geedipally, S. R. (2016). “A Semiparametric
Negative Binomial Generalized Linear Model for
Modeling Over-Dispersed Count Data with a
Heavy Tail: Characteristics and Applications to
Crash Data.” Accident Analysis and Prevention,
91, 10-18.
[44.] Vangala, P., Lord, D., and Geedipally, S. R.
(2015). “Exploring the Application of the
Negative Binomial-Generalized Exponential
Model for Analyzing Traffic Crash Data with
Excess Zeros.” Analytic Methods in Accident
Research, 7, 29-36.
[45.] Wang, Y. and Kockelman, K. M. (2013). “A
Poisson-lognormal Conditional-autoregressive
Model for Multivariate Spatial Analysis of
Pedestrian Crash Counts Across
Neighborhoods.” Accident Analysis and
[46.] Wu, Q., Chen, F., Zhang, G., Liu, X. C., Wang,
H. and Bogus, S. M. (2014). “Mixed Logit
Model-Based Driver Injury Severity
Investigations in Single- and Multi-Vehicle
Crashes on Rural Two-Lane Highways.”
[47.] Wu, H., Gao, L., and Zhang, Z. (2014).
“Analysis of Crash Data Using Quantile
Regression for Counts.” Journal of
Transportation Engineering, 140(4), 04013025.
[48.] Yasmin, S., Eluru, N., Bhat, C., and Tay, R.
(2014). “A Latent Segmentation Based
Generalized Ordered Logit Model to Examine
Factors Influencing Driver Injury Severity.”
Analytic Methods in Accident Research, 1, 23–
38.
[49.] Ye, F. and Lord, D. (2014). “Comparing Three
Commonly Used Crash Severity Models on
Sample Size Requirements: Multinomial, Logit,
Ordered Probit and Mixed Logit Models.”
85.
[50.] Zou, Y., Zhang, Y., and Lord, D. (2013).
“Application of Finite Mixture of Negative
Binomial Regression Models with Varying
Weight Parameters for Vehicle Crash Data
Analysis.” Accident Analysis and Prevention,
50, 1042-1051.
[51.] Zou, Y., Zhang, Y., and Lord, D. (2014).
“Analyzing Different Functional Forms of the
Varying Weight Parameter for Finite Mixture of
Negative Binomial Regression Models.”
52.

An Exploratory Computational Piecewise Approach to Characterizing and Analyzing Traffic Accident Data

More Related Content

What's hot (8)

Similar to An Exploratory Computational Piecewise Approach to Characterizing and Analyzing Traffic Accident Data (20)

More from International Journal of Modern Research in Engineering and Technology (20)

Recently uploaded (20)

An Exploratory Computational Piecewise Approach to Characterizing and Analyzing Traffic Accident Data