Safety Science 43 (2005) 541–557

www.elsevier.com/locate/ssci

Analysis of freeway accident frequencies:

Negative binomial regression versus
artificial neural network
Li-Yen Chang *

Graduate Institute of Transportation and Logistics, National Chia-Yi University,

300 University Road, Chia-Yi 600, Taiwan

Received 27 October 2004; received in revised form 13 April 2005; accepted 17 April 2005

Abstract

The Poisson or negative binomial regression model has been employed to analyze vehicle accident
frequency for many years. However, these models have the pre-defined underlying relationship
between dependent and independent variables. If this assumption is violated, the model could lead
to erroneous estimation of accident likelihood. In contrast, the artificial neural network (ANN),
which does not require any pre-defined underlying relationship between dependent and independent
variables, has been shown to be a powerful tool in dealing with prediction and classification prob-
lems. Thus, this study employs a negative binomial regression model and an ANN model to analyze
1997–1998 accident data for the National Freeway 1 in Taiwan. By comparing the prediction perfor-
mance between the negative binomial regression model and ANN model, this study demonstrates
that ANN is a consistent alternative method for analyzing freeway accident frequency.

2005 Elsevier Ltd. All rights reserved.

Keywords: Accident frequency; Artificial neural network; Negative binomial regression; Freeway

1. Introduction

The impact that traffic accidents have on society is significant. For example, there are
approximately 3000 people are killed and thousands more injured by traffic accidents in

*
Tel.: +886 5 271 7982.
E-mail address: [email protected]

0925-7535/$ - see front matter
2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ssci.2005.04.004
542 L.-Y. Chang / Safety Science 43 (2005) 541–557

Taiwan each year. Individuals injured (or killed) in traffic accidents must deal with pain
and suffering, medical costs, wage loss, higher insurance premium rates, and vehicle repair
costs. For society as a whole, traffic accidents result in enormous costs in terms of lost pro-
ductivity and property damage. Therefore, public agencies have put much effort into pre-
ventive measures, such as illumination and enforcement. However, the annual number of
traffic accidents has not yet significantly decreased. Therefore, there should be further re-
search studies on the risk factors for traffic accidents. This study focuses on the non-behav-
ioral factors of freeway accident risk, specifically highway geometric characteristics and
environmental conditions. Although past statistics indicated that most traffic accidents re-
sulted from drivers
errors (behavioral factors), non-behavioral factors also play an impor-
tant role in traffic safety. Not only can they contribute to certain types of driver errors
(e.g., speeding often occurs at downgrades), but accidents will be likely to occur at the
same location repeatedly if the problem is not mitigated. In addition, with a better under-
standing of non-behavioral factors of freeway accidents, transportation engineers will be
able to design freeways to higher safety standards.
Past research analyzing accident frequencies mainly relied on statistical models such as
linear regression models, Poisson regression or/and negative binomial regression models
because the occurrence of accidents on a highway section can be regarded as a random
event. Another major advantage of applying these statistical models is the ability to iden-
tify a broad range of risk factors that can contribute significantly to accidents. However,
most of the statistical models have their own model assumptions and pre-defined under-
lying relationships between dependent and independent variables. If these assumptions
are violated, the model could lead to erroneous estimation of accident likelihood. Artificial
neural networks (ANN) which do not require any pre-defined underlying relationship be-
tween dependent and independent variables have been widely employed in financial anal-
ysis, decision problems, and pattern recognition. The ANN has been shown to be a
powerful tool, particularly in dealing with prediction and classification problems. There
has also been an increased interest in applying ANN in the field of transportation since
the 1990s, such as driver behavior analysis, pavement maintenance, vehicle detections,
and so on (Dougherty, 1995). However, the applications of ANN to analyze traffic safety
problems have been relatively few. Therefore, this study examines whether ANN can be
used to analyze the relationship between risk factors and accidents. This is done by eval-
uating the prediction performance between the negative binomial regression model and
ANN model. The paper begins with a review of previous literature on modeling accident
frequencies and then presents the methodological approach. A description of the available
data and an assessment of the model estimation results follow this. The paper concludes
with a summary and directions for future research.

2. Literature review

Past research on modeling accident frequencies has been diverse, both empirically and
methodologically. From an empirical standpoint, most research studies (Shankar et al.,
1995; Milton and Mannering, 1998; Carson and Mannering, 2001) have focused on
non-behavioral risk factors of accidents on the freeway or arterial roadways. These
non-behavioral factors included highway geometry (e.g., horizontal and vertical align-
ments, and shoulder width), traffic characteristics (e.g., average annual daily traffic
(AADT) and percentage of trucks) and weather conditions (e.g., rain or snow). The
 L.-Y. Chang / Safety Science 43 (2005) 541–557 543

findings indicated that number of lanes, narrow shoulder width, vertical grade, horizontal
curves, AADT, amount of snowfall and their interaction can have significant influence
on vehicle accidents. McCarthy (1999) focused on the effectiveness of public policy (e.g.,
traffic regulations, alcohol availability and enforcement) in reducing fatal accidents in indi-
vidual cities in California. The results indicated that fatal accidents were significantly re-
duced due to traffic enforcement, but little effect was found from speed limit and seat belt
use laws.
From a methodological standpoint, the most common approach to analyze accident
frequencies for a specified roadway segment is to apply Poisson or negative binomial
regression models because of the distributional property (i.e., random, discrete and non-
negative) of vehicle accidents (Milton and Mannering, 1998). Although the Poisson regres-
sion model has desirable statistical properties for describing vehicle accidents, it has an
important constraint, which is that the mean and variance of the accident data are con-
strained to be equal. To overcome this constraint, the negative binomial regression model,
which allows this constraint to be relaxed, has been widely employed to analyze vehicle
accidents (Miaou, 1994; Poch and Mannering, 1996; Hadi et al., 1995; Shankar et al.,
1995; McCarthy, 1999; Carson and Mannering, 2001). The findings suggested that most
of vehicle accident data tend to be overdispersed (i.e., the variance will likely be signifi-
cantly greater than the mean) and negative binomial modeling is an appropriate technique
for exploring the frequency of accidents. In addition, zero-inflated Poisson and zero-
inflated negative binomial models were also employed to analyze accident frequencies
by Shankar et al. (1997), Lee and Mannering (2002) and Lee et al. (2002) to deal with
the overdispersion problem potentially caused by the excessive zeroes (i.e., no accidents
being observed) in traffic accident data. The application of zero-altered counting processes
allows modeling roadway section accident frequencies in two states: the zero-accident state
(where no accidents will be ever observed), and the accident state (where accident frequen-
cies follow some known distribution, such as the Poisson or negative binomial distribu-
tion). The findings showed that the zero-altered probability process provides great
flexibility in uncovering processes affecting accident frequencies on roadway sections with
zero accidents and those observed with accidents. In terms of model selection for analyzing
accident frequencies, Miaou (1994) and Lee et al. (2002) recommended that the Poisson
regression model is estimated as an initial model. If the overdispersion of accident data
is found, both negative binomial and zero-inflated count models could be considered.
ANN has also been employed to analyze transportation problems for many years.
According to the review by Dougherty (1995), most studies have focused on modeling dri-
ver behavior, parameter estimation, pavement maintenance, and vehicle detection. More
recent applications in the transportation field using ANN included traffic prediction
(Yin et al., 2002; Zhong et al., 2004), traffic parameters estimation (Tong and Hung,
2002), traffic signal control (Zhang et al., 2001), incident detection (Jin et al., 2002; Yuan
and Cheu, 2003), travel behavior analysis (Subba Rao et al., 1998; Hensher and Ton, 2000;
Vythoulkas and Koutsopoulos, 2003), vehicle emissions (Shiva Nagendra and Khare,
2004) and traffic accident analysis (Mussone et al., 1996; Mussone et al., 1999; Sohn
and Lee, 2003; Abdel-Aty and Pande, 2005). For example, Tong and Hung (2002) em-
ployed a three-layer neural network to estimate vehicle discharge headway. The results
showed that the ANN model could produce reasonably discharge headway estimates
for individual vehicles. Subba Rao et al. (1998) used ANN to model the choice behavior
with respect to access mode for transit. The performance of ANN was found to be
544 L.-Y. Chang / Safety Science 43 (2005) 541–557

superior to multinomial logit model in both calibration and prediction. These studies have
provided general insight into the performance of ANN models. For analyzing traffic safety
problems, there have been relatively few applications using ANN. Mussone et al. (1996)
employed a three-layer neural network to estimate the accident probability using accident
data in Italy. Accident sites, road features, weather and roadbed conditions, human error,
as well as vehicular and environmental factors were used as input variables. ‘‘Careless-
ness’’ and ‘‘excessive speed’’ were found to be the risk factors in defining accident proba-
bility. Mussone et al. (1999) employed ANN modeling approach to analyze the degree of
danger of urban intersections and demonstrated that ANN is a good alternative method
for analyzing the factors contributing to intersection accidents. Abdel-Aty and Pande
(2005) applied a probabilistic neural network (PNN) model to predict crash occurrence
on the Interstate-4 corridor in Orlando. Average and standard deviation of speed around
crash sites were extracted from loop data as input variables. The analysis results showed
that at least 70% of the crashes can be correctly identified by the proposed PNN model.

3. Empirical setting

The study area for this paper is National Freeway 1, which is the most important trans-
portation corridor in Taiwan. National Freeway 1 is a 373 km tolled freeway with 47 inter-
changes and 10 mainline toll plazas. Illumination is provided only at interchange areas,
toll plaza areas and locations with severe geometric changes such as severe downhill or
uphill grades.
To investigate the relationship between vehicle accidents and highway geometry, traffic
characteristics and environment conditions, data from a number of resources were col-
lected. The vehicle accident data were taken from the National Traffic Accident Investiga-
tion Reports provided by the Ministry of Transportation and Communications. The data
were obtained in a computer-ready form, which included coded information on reported
accidents on National Freeway 1. Information on the accidents that occurred in the period
from 1997 to 1998 was extracted for this study. The primary resources of highway geomet-
ric design information and traffic data were obtained from the Taiwan Area National
Freeway Bureau. The highway geometric design information includes number of lanes,
lane width, horizontal curvature, vertical grade and others; while traffic information in-
cludes average daily traffic (ADT) of various vehicle types, peak hour factors, and traffic
distribution over lanes. Weather information was taken from the annual report of clima-
tological data. This report records detailed weather information of cities and towns along
National Freeway 1 including pressure, temperature, humidity, precipitation, wind speed,
and cloudiness.
With these data, the next step is how to divide the study area into manageable roadway
sections. The most common alternatives adopted in previous studies for determining road-
way section length include the use of fixed-length sections or homogeneous sections (in
terms of geometric characteristics). In order to account directly for the effects of highway
geometric characteristics on accident frequencies, homogeneous sections in terms of num-
ber of lanes, horizontal curvature, and vertical grade were considered in this study. A more
detailed discussion on the advantages and disadvantages of these two alternatives can be
found in Shankar et al. (1995). According to this approach, 373 km of freeway were first
disaggregated into 498 sections. Because of the opposite vertical alignment and different
traffic conditions in the northbound and southbound roadway sections, these 498 sections
 L.-Y. Chang / Safety Science 43 (2005) 541–557 545

Table 1
Sample summary of statistics of characteristics of road sections
Minimum Maximum Mean Standard
deviation
Accident frequency (per year) 0 7 0.67 1.00
Section length (km) 0.1 4.2 0.75 0.52
Degree of horizontal curve (angle, in degree, 0 15.68 2.18 2.46
subtended by a 100 m arc, equal to 18,000/(p · radius))
Vertical grade (%)
5.3 5.3 0 1.58
ADT per lane (in 1000
s of vehicles) 12.88 42.13 20.96 4.14
Trucks percentage 0.86 18.45 10.13 4.16
Bus percentage 0.67 11.31 3.53 1.27
Peak hour factor (PHF) 0.77 0.97 0.91 0.04
Number of days with precipitation 56 224 111.1 37.5
Annual precipitation (mm) 1439 5773 2165.6 790.7

were further disaggregated into 996 sections (i.e., northbound and southbound roadway
sections considered separately). During the 1997–1998 study period, there were 1338 acci-
dents resulting in deaths and/or injuries. The summary statistics of these 1992 highway
sections (i.e., each section produces two observations) are presented in Table 1. The ob-
served accident frequency on these freeway sections ranges from 0 to 7, and the average
frequency is 0.67. In order to be able to compare the prediction performance between
the statistical model and ANN model, the collected data were randomly divided into
two subsets, one for training and the other for testing. The number of sections used for
model estimation is 1500, and the number of sections used for testing is 492. A Chi-
squared test shows that the accident frequency distributions between the two sub-samples
are not significantly different.

4. Negative binomial modeling approach to freeway accident frequencies

This study models the number of accidents that occurred on a highway section over a
one-year time period. Because accident frequencies on a highway section are discrete and
non-negative integer values, the Poisson regression technique is a natural first choice for
modeling such data. However, past research has indicated that accident frequency data
are likely to be overdispersed and has suggested using the negative binomial regression
model as an alternative. Deriving the negative binomial regression model can start with
a Poisson model, which is defined by the following equation:
kni i expð
ki Þ
P ðni Þ ¼ ð1Þ
ni !
where P(ni) is the probability of n accidents occurring on a highway section i over a one-
year time period, and ki is the expected accident frequency (i.e., E(ni)) for highway section
i. When applying the Poisson model, the expected accident frequency is assumed to be a
function of explanatory variables such that
ki ¼ expðbXi Þ ð2Þ
where Xi is a vector of explanatory variables that could include the geometry, traffic char-
acteristics, and weather conditions of highway section i that determine accident frequency;
546 L.-Y. Chang / Safety Science 43 (2005) 541–557

and b is a vector of estimable coefficients. With this form of ki, the coefficient vector b can
be estimated by the maximum likelihood method with the likelihood function being
Y exp½
expðbXi Þ½expðbXi Þni
LðbÞ ¼ ð3Þ
i
ni !

To overcome the overdispersion problem, negative binomial regression can be applied

by relaxing the assumption that the mean of accident frequencies equals the variance. To
do this, an error term is added to the expected accident frequency (ki) such that Eq. (2)
becomes
ki ¼ expðbXi þ ei Þ ð4Þ
where exp(ei) is a gamma-distributed error term with mean one and variance a. This gives a
conditional probability.
exp½
ki expðei Þ½ki expðei Þni
P ðni jeÞ ¼ ð5Þ
ni !
Integrating e out of this expression produces the unconditional distribution of ni. The
formulation of this distribution (the negative binomial) is
Cðh þ ni Þ h n
P ðni Þ ¼  u ð1
ui Þ i ð6Þ
½CðhÞ  ni ! i
where ui = h/(h + ki) and h = 1/a, and C (Æ) is a value of gamma distribution. The corre-
sponding likelihood function is
YN
h
ni
Cðh þ ni Þ h ki
Lðki Þ ¼ ð7Þ
i¼1
CðhÞni ! h þ ki h þ ki
where N is the total number of highway sections. This function is maximized to obtain
coefficient estimates for b and a. This model structure allows the mean to differ from
the variance such that,
var½ni  ¼ E½ni ½1 þ aE½ni  ð8Þ
where a is the variance of the gamma-distributed error term and is used as a measure of
dispersion. The choice between the negative binomial model and the Poisson model can be
determined by the statistical significance of the estimated coefficient a. If a is not signifi-
cantly different from zero, the negative binomial model simply reduces to a Poisson model
with var[ni] = E[ni]. If a is significantly different from zero, the negative binomial model is
the correct choice. A more detailed description of negative binomial regression analysis
can be found in Washington et al. (2003).
The estimation results of the negative binomial model of freeway accident frequencies
are presented in Table 2. The model has a reasonable overall statistical fit, as indicated by
the q2 statistic. Fifteen variables are found statistically significant or marginally significant
in determining accident likelihood. It is noteworthy that the dispersion parameter, a, is sig-
nificantly different from zero. This confirms the appropriateness of the negative binomial
model relative to the Poisson model. As shown in Table 2, a number of highway geometric
variables that can significantly influence the accident occurrence were found. The positive
sign of the number of lanes variable indicates that the increase in number of lanes will also
 L.-Y. Chang / Safety Science 43 (2005) 541–557 547

Table 2
Negative binomial estimation results
Variable Estimated coefficient t-Statistic
Constant
2.338
7.49
Number of lanes 0.367 5.21
Descent grade
0.064
1.46
Level indicator (1 if
1% 6 grade 6 1%, 0 otherwise)
0.161
1.64
Severe upgrade indicator (1 if grade P 3%, 0 otherwise) 0.353 1.76
Severe horizontal curve indicator
0.538
2.81
(1 if degree of horizontal curve P 6, 0 otherwise)
Military section indicator (1 if section is a military section, 0 otherwise) 0.326 1.73
Interchange indicator (1 if section contains an interchange, 0 otherwise) 0.091 1.05
ADT per lane (in 1000
s of vehicles) 0.028 2.98
High PHF indicator (1 if PHF P 0.95, 0 otherwise)
0.346
3.12
High truck percentage indicator 0.268 2.19
(1 if truck percentage P 30%, 0 otherwise)
Fog zone indicator (1 if section is a fog zone, 0 otherwise)
0.094
1.03
Annual precipitation (mm)
0.00016
2.26
North proportion of freeway indicator 0.218 1.75
(1 if mileage post is between north end and 94.7 km, 0 otherwise)
Section length (km) 0.842 14.66
1998 indicator (1 if accident data were from 1998, 0 otherwise) 0.080 1.06
a (dispersion coefficient) 0.220 3.32
Number of observations 1500
Restricted log-likelihood (constant only)
1775.74
Log-likelihood at convergence
1545.31
q2 0.12

increase accident likelihood. As expected, when the number of lanes increases, the total
amount of lane changing as well as the conflicts between traffic will increase. This result
is consistent with previous findings (Milton and Mannering, 1998; Carson and Mannering,
2001). Vertical and horizontal alignments are other important elements in highway geo-
metric design. Grade can significantly influence vehicle operation speed, particularly for
large trucks and buses. The effect of speed differentials can play an important role in acci-
dent occurrence. The estimated results indicate sections with severe uphill grade (3% or
more) or descent grades have increased likelihood of accident occurrence, while level sec-
tions have reduced likelihood of accidents. It is important to note that the effect of down-
grades on accident likelihood is positive because both the coefficient and the value of
descent grade variable are negative. At downgrades, a greater frequency of accidents is ex-
pected because the speeds on downgrades are high and it is more difficult to control the
vehicle under such conditions. For the effect of horizontal curves, the estimated result
shows reduced accident likelihood for the sections with degree of horizontal curve greater
than six degree. This result may seem counterintuitive, but it is consistent with past find-
ings (Milton and Mannering, 1998). An explanation for this is that drivers are more likely
to drive cautiously at sharp horizontal curves. In addition to vertical and horizontal align-
ments, lane width, median types and shoulder width were also identified by the past studies
to have impact on accident occurrence (Carson and Mannering, 2001). Because most of
the highway sections have the same lane width and the detailed information on shoulder
width and median types was unavailable, this study could not examine their effects on
accident frequencies.
548 L.-Y. Chang / Safety Science 43 (2005) 541–557

In addition to the horizontal and vertical alignments, location-related indicator vari-

ables were also tested to investigate if accidents tend to occur at specific locations. Inter-
change indicator variable is intended to capture the impact of vehicles entering or leaving
freeway mainline on accident risk. The estimated results show an increased likelihood of
accidents at an intersection area. In addition, the estimated results show the increased like-
lihood for accidents in the military sections. The military sections, which provide emer-
gency use for air force aircrafts, are a unique design feature of National Freeway 1.
The design characteristics are being level, straight and with very wide shoulders, and
the two traffic directions are not physically separated. An explanation for the increased
likelihood of accident risk for the military sections is that drivers might easily but uninten-
tionally speed under such design conditions. Because military sections are the only places
where two directions of traffic are not physically separated, this finding implicitly indicates
that physical separation between traffic for a freeway is crucial for preventing severe acci-
dents. Another section location indicator variable shows that accidents are more likely to
occur in the north portion of this freeway. These findings might imply differences of vehi-
cle use and driver behavior associated with level of urbanization because the north propor-
tion of the freeway runs through several metropolitan areas. In addition, this study also
investigated if the mainline toll plazas have any effect on accident occurrence because
mainline toll plazas can significantly interrupt freeway operations. A toll plaza indicator
variable was examined, but the results show that the effect of mainline toll plazas on acci-
dent occurrence is statistically insignificant.
As for the traffic characteristic variables, AADT is typically used to indicate the traffic
conditions. However, these long-term traffic data were not available. In this study, the
interpolated daily traffic volumes for 1997 and 1998 based on traffic surveys conducted
in 1994 and 1999 were used. The positive coefficient of the ADT variable and the high
truck percentage indicator variable imply that conflicts between vehicles and the exposure
to potential risk of accidents increase with increasing number of vehicles and trucks. In
addition, for freeway sections with the peak hour factor higher than 0.95, the accident like-
lihood is reduced. Peak hour factor is used to measure the fluctuation of traffic flow. When
the peak hour factor is greater than 0.95, the traffic condition is congested but relatively
stable. A slow and stable traffic flow is less likely to increase the traffic accident likelihood.
In addition to geometric and traffic factors, environmental factors, such as snow and
rain, were identified by past studies to have significant impacts on accident occurrence.
Among these environmental factors, fog has been taken as a greater contributor to vehicle
accidents because it can significantly reduce drivers
visibility. For example, there was a
major accident caused by fog on this freeway in 1996 involving 99 vehicles, and resulting
in three deaths and 23 injuries. The Taiwan Area National Freeway Bureau has identified
specific freeway sections as fog zones and installed more traffic safety facilities such as illu-
mination, flashing lights, and warning signs. To identify if there is higher accident fre-
quency in fog zones, an indicator variable is selected for these zones. The negative
coefficient indicates that accident frequency tends to decrease in fog zones. An explanation
for this is that fog often occurs at early mornings and less number of vehicles is expected to
be driving at such times compared to non-foggy conditions. In addition, the effect of pre-
cipitation was also investigated and the negative coefficient of the annual precipitation var-
iable shows that an increase in annual precipitation will reduce accidents. A possible
explanation for this is that drivers are more likely to drive at lower speeds and also keep
longer car-following distances under the wet pavement and reduced visibility conditions.
 L.-Y. Chang / Safety Science 43 (2005) 541–557 549

However, detailed data on rain intensity and time of raining are unavailable. To have a
better understanding of the effect of rain on accidents, a more detailed study is suggested.
Overall, the environmental variables do not appear to be risk factors for vehicle accidents.
To gain a better understanding of the marginal effects of the variables on accident fre-
quency, elasticities were computed. In general, the direct elasticity is defined as
oki xij
Ekxiji ¼  ð9Þ
oxij ki
where E represents the elasticity, xij is the value of variable j being considered for highway
section i, and ki is the mean of accident frequency on highway section i. Applying this to
Eq. (4) gives
Ekxiji ¼ bj xij ð10Þ
where bj is the coefficient corresponding to variable j.
The elasticity defined in Eq. (10) is used to measure the effect that 1% change in an acci-
dent covariate will have on accident frequency. Thus, it is only valid when the accident
covariates are continuous variables, and it is not applicable for indicator variables (i.e.,
variables that take on values of zero or one). For indicator variables, a ‘‘pseudo-elasticity’’
can be used to give an approximate elasticity of the variable. The pseudo-elasticity gives
the incremental change in accident frequency caused by the indicator variable in the model.
In this case, the pseudo-elasticity is defined as
expðbj Þ
1
Ekxiji ¼ ð11Þ
expðbj Þ
The elasticities for the independent variables are shown in Table 3. None of them are
elastic (i.e., absolute value of elasticity greater than one). The values in the table can be
readily interpreted. For example, the elasticity for ADT is 0.58. This means that a 1% in-
crease in ADT will result in a 0.58% increase in accidents. It is also important to note that
elasticity estimates can only be applied to examine the effect of a small change of indepen-
dent variable (e.g., 1–5% increase of ADT) on the expected accident frequency.

Table 3
Elasticity estimates of key variables
Variables Elasticity
Number of lanes 0.81
Descent grade 0.03
Level indicator
0.17
Severe upgrade indicator 0.30
Severe horizontal curve indicator
0.71
Interchange indicator 0.09
Military section indicator 0.28
ADT per lane (in 1000
s of vehicles) 0.58
High PHF indicator
0.41
High percentage of truck indicator 0.24
Fog zone indicator
0.10
Annual precipitation (mm)
0.35
550 L.-Y. Chang / Safety Science 43 (2005) 541–557

5. Artificial neural network approach to freeway accident frequencies

The structure of the artificial neural network used for this study is a three-layer neural
network, as shown in Fig. 1. The basic elements are the artificial neurons, and each neuron
is interconnected with all the neurons in the next layer. The first layer is the input layer,
where the data are presented to the neural network. The values of the input variables
can be either a numerical value (generally normalized) or a binary code (e.g., gender).
The intermediate layer is the hidden layer. The function of the hidden layer is to compute
the complicated pattern associations. A single hidden layer has been found to be satisfac-
tory in most applications, while the number of neurons in the hidden layer is generally
determined through experimentation. The third layer is the output layer, representing
the network response to the corresponding input (i.e., accident frequencies). The neural
network can then be trained through a training algorithm. Currently, there are a number
of training algorithms available for artificial neural network models, and the back-prop-
agation rule, which is one of the most widely used training algorithms, is adopted in this
study. The principle of this rule is to minimize the total output error described in Eq. (12).
1 X N X k
2
MSE ¼ ðtij
aij Þ ð12Þ
N  K i¼1 j¼1

where MSE is the mean squares error, t is the target output value, a is the model output
value, K is the number of output neurons and N is the number of testing data. More
detailed description on the back-propagation rule can be found in Hagan et al. (1996).

P1 Year

P2 Location

a1
Geometric variables
P3

a5
Traffic variables
P4

Environmental variables

Pn

Input layer Hidden layer Output layer

(15 neurons) (10 neurons) (5 neurons)

Fig. 1. The structure of the ANN model.

 L.-Y. Chang / Safety Science 43 (2005) 541–557 551

To estimate the ANN model, there are a number of software packages ready to perform
the back-propagation algorithm, and MATLAB was chosen for this study. Here, the input
layer contains 15 neurons which are the input variables representing the potential risk fac-
tors for accidents. Currently, there is no systematic way to select the number of neurons in
the input layer. Thus 15 neurons, the statistically significant variables found in the negative
binomial regression model, are used. The other reason for using these 15 input variables is
to allow the comparisons of the model prediction performance and the marginal effects of
each independent variable between the negative binomial regression and ANN models.
Table 4 shows the definition of input variables. Because the effectiveness of the back prop-
agation training algorithm depends on the number of neurons in the hidden layer, various
numbers of neurons (ranging from 1 to 29) in the hidden layer were tested. The ‘‘optimal’’
number of neurons in the hidden layer was found to be 10. In this study, five neurons in
the output layer were taken to represent the accident frequencies for the freeway sections
because relatively few highway sections had more than three accidents. Treating freeway
sections with accident frequencies greater than three as a group can reduce the complexity
of the network. By minimizing the MSE calculated on the testing data, the best solution
was found after 800 learning cycles. The change of MSE in relation to learning cycles is
shown in Fig. 2. The MSE of the best solution for the training and for the testing data
are 0.097 and 0.112, respectively.
To evaluate the marginal effects of input neurons (variables) on output neurons for an
ANN model, sensitivity analysis is commonly applied (Mussone et al., 1999; Tong and
Hung, 2002). The sensitivity analysis conducted by this study is to examine the effect of
a particular independent variable change on accident frequency distribution by holding
all other variables fixed. Table 5 shows the analysis results. For example, of the level high-
way sections, about 95.4% had no accident occurrence, 2.8% had one accident, 1.2% had
two accidents, 0% had three accidents, and 0.6% had four or more accidents. When the
highway section is not level (i.e., the grade is greater than 1% or less than
1%), the dis-
tribution of accident frequencies of highway sections are 95.2%, 2.4%, 0.6%, 1.8%, and 0%
for zero, one, two, three and four or more accidents, respectively. Thus the average

Table 4
Definition of the input variables
Variable Definition Binary/numerical code
X1 Number of lanes Numerical value
X2 Vertical alignment 1 Numerical value
X3 Vertical alignment 2 1 for section with grade between
1% and 1%, 0 for others
X4 Vertical alignment 3 1 for section with grade P 3%, 0 for others
X5 Horizontal alignment 1 1 for section with degree of horizontal curve P 6, 0 for others
X6 Military section 1 for section within a military section; 0 for others
X7 Interchange 1 for section containing an interchange; 0 for others
X8 ADT per lane Numerical value
X9 High truck percentage 1 for section with truck percentage P 30%, 0 for others
X10 High PHF 1 for PHF P 0.95, 0 for others
X11 Fog zone 1 for section located in fog zone, 0 for others
X12 Annual precipitation Numerical value
X13 Section location 1 for section located in northern Taiwan, 0 for others
X14 Section length Numerical value
X15 Year of 1998 1 for accident data from 1998, 0 for others
552 L.-Y. Chang / Safety Science 43 (2005) 541–557

0.14

0.12

0.1

0.08
MSE

0.06 Train

Test
0.04

0.02

0
100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500
Learning cycles

Fig. 2. The training curve for the ANN model.

accident frequency for level highway sections is about 19% lower than that in the highway
sections with grades. The shift of accident frequency distribution indicates that highway
sections with grades have greater likelihood of having an accident. This finding is consis-
tent with the elasticity analysis from the negative binomial regression model. However, it is
important to note that not all of the sensitivity analysis results are consistent with those
from the negative binomial regression model. For example, the average accident frequency
is 0.22 for the highway sections with an interchange and 0.06 for those without an inter-
change in the sensitivity analysis. The presence of an interchange in a highway section in-
creases the accident likelihood by 270%, while the elasticity analysis shows only a 9%
increase in the accident likelihood. It should also be noted that for the continuous inde-
pendent variables such as ADT per lane, only a certain range of changes was analyzed.
Although the results of sensitivity analysis and elasticity are not quite similar, the sensitiv-
ity analysis still provides valuable insight into the relationship between risk factors and
accident frequency.

6. Comparisons of the prediction performance of neural networks and negative binomial

regression models

In order to examine the performance of the ANN model, the comparison of model pre-
diction performance between the negative binomial regression model and ANN model is
examined. Tables 6 and 7 show the comparison results. For the negative binomial regres-
sion model, the overall model prediction accuracy for the training data is about 58.3%,
while that for the testing data is about 60.8%. For the ANN model, the overall model pre-
diction performances for the training data and the testing data are 64% and 61.4%, respec-
tively. The proposed ANN model performs slightly better than the negative binomial
regression model in analyzing the training data. In predicting the accident frequency on
 L.-Y. Chang / Safety Science 43 (2005) 541–557 553

Table 5
Sensitivity analysis of ANN model
Variable N = 0 (%) N = 1 (%) N = 2 (%) N = 3 (%) N P 4 (%)
1 Number of lanes = 2 97.30 2.70 0.00 0.00 0.00
Number of lanes = 3 92.31 3.55 1.78 1.18 1.18
Number of lanes = 4 96.52 1.99 1.00 0.50 0.00
2 Level indicator = 0 95.15 2.42 0.61 1.82 0.00
Level indicator = 1 95.41 2.75 1.22 0.00 0.61
3 Severe upgrade indicator = 0 95.14 2.75 1.06 0.63 0.42
Severe upgrade indicator = 1 100.00 0.00 0.00 0.00 0.00
4 Severe horizontal curve indicator = 0 95.31 2.77 0.85 0.64 0.43
Severe horizontal curve indicator = 1 95.65 0.00 4.35 0.00 0.00
5 Interchange indicator = 0 96.81 1.72 1.23 0.00 0.25
Interchange indicator = 1 88.10 7.14 0.00 3.57 1.19
6 Military section indicator = 0 95.67 2.47 1.03 0.62 0.21
Military section indicator = 1 71.43 14.29 0.00 0.00 14.29
7 ADT per lane < 15,000 97.30 2.70 0.00 0.00 0.00
ADT per lane < 20,000 92.31 3.55 1.78 1.18 1.18
ADT per lane < 25,000 96.52 1.99 1.00 0.50 0.00
ADT per lane P 25,000 97.65 2.35 0.00 0.00 0.00
8 High PHF indicator = 0 94.72 2.88 1.20 0.72 0.48
High PHF indicator = 1 98.67 1.33 0.00 0.00 0.00
9 High percentage of truck indicator = 0 89.09 7.27 1.82 1.82 0.00
High percentage of truck indicator = 1 95.77 2.82 0.00 0.00 1.41
10 Fog zone indicator = 0 94.63 3.32 0.77 0.77 0.51
Fog zone indicator = 1 98.02 0.00 1.98 0.00 0.00
11 Annual precipitation < 2000 97.65 1.57 0.78 0.00 0.00
Annual precipitation < 3000 92.23 3.63 1.55 1.55 1.04
Annual precipitation P 3000 95.45 4.55 0.00 0.00 0.00

Table 6
Prediction performance for the negative binomial regression model
Training data Testing data
Observed Predicted Correctly Observed Predicted Correctly
frequency frequency predicted frequency frequency predicted
N=0 862 1354 833 (97%) 295 452 286 (97%)
N=1 389 105 32 (8%) 127 26 10 (8%)
N=2 154 0 0 (0%) 46 0 0 (0%)
N=3 58 0 0 (0%) 18 0 0 (0%)
NP4 37 41 9 (24%) 6 14 3 (50%)
The overall prediction accuracy is 58.3% for training data and 60.8% for testing data.

a highway section, the proposed ANN model performs better for the highway sections
with one or more accidents, while the negative binomial regression model performs slightly
better for the sections with zero accidents.
554 L.-Y. Chang / Safety Science 43 (2005) 541–557

Table 7
Prediction performance for the ANN model
Training data Testing data
Observed Predicted Correctly Observed Predicted Correctly
frequency frequency predicted frequency frequency predicted
N=0 862 1142 780 (90%) 295 469 290 (98%)
N=1 389 312 149 (38%) 127 13 6 (5%)
N=2 154 21 13 (8%) 46 5 1 (2%)
N=3 58 4 4 (7%) 18 3 3 (17%)
NP4 37 21 14 (38%) 6 2 2 (33%)
The overall prediction accuracy is 64% for training data and 61.4% for testing data.

7. Discussion

In this particular application, the negative binomial regression model and ANN model
provide similar results in terms of prediction performance on the training data and testing
data. This demonstrates that the ANN model is an appropriate methodology for analyzing
traffic accidents. Although it is difficult to distinguish which modeling approach is better
according to the analysis results of this study, there are some aspects might be of great
interest for future research.
In past research, the Poisson or negative binomial model was commonly employed for
traffic accident analysis because of the nature of random, discrete, and non-negative char-
acteristics of vehicle accidents and their capacity of identifying effectively a broad range of
risk factors for accidents. In addition, the elasticity of each risk factor can be mathemat-
ically defined and ready to compute. These analysis results not only can be easily inter-
preted, but also provide clear and valuable information for traffic engineers to perform
mitigation. In constrast, the advantage of the ANN model is that it requires no assump-
tions of underlying relationship between risk factors and traffic accidents. In this applica-
tion, if the underlying relationship between risk factors and traffic accidents does not
follow a gamma distribution, the relationship estimated by the negative binomial regres-
sion could be erroneous. Another advantage of the ANN model is that it can effectively
handle interrelation problems between independent variables. When a serious correlation
exists between independent variables, the variability of estimated coefficients will be in-
flated and interpretation of relationship between independent variables and dependent
variable will also be difficult. But when the ANN model is applied, the correlation prob-
lems between independent variables would not be a great concern. Compared to the com-
monly applied regression models in traffic accident analysis, this is an important advantage
of employing ANN models because an accident is rarely due to a single risk factor, but is
rather the outcome of a series of factors.
Despite these advantages, the ANN model has its own drawbacks. Firstly, developing
an ANN model is very time-consuming. The time required to develop an ANN model de-
pends on the size of training data and network structure. As discussed earlier, there is no
general rule in determining the network structure and it can only be done by experimen-
tation. Therefore, it always takes a great deal of time to determine the model structure,
including the network structure (number of hidden layers and number of neurons in the
hidden layer), transfer functions, and so on. Once a network is specified, it usually takes
hours to complete an experiment especially when the size of training data is large because a
 L.-Y. Chang / Safety Science 43 (2005) 541–557 555

training algorithm usually needs to go through several hundred of iterations to obtain an

‘‘optimal’’ weighting for the network. Secondly, unlike the elasticity analysis for most sta-
tistical models, the sensitivity analysis for the ANN model cannot be mathematically de-
fined. Therefore, the sensitivity analysis for continuous variables is difficult to perform.
For example, to analyze the relationship between accidents and precipitation, the ANN
model can only examine the distribution change of accident frequency against a certain
amount of annual precipitation. In addition, the selection of input neurons is also a critical
issue in developing an ANN model. Because the input neurons are usually regarded as the
input variables, only the variables with cause–effect relationship should be selected. Hav-
ing more input neurons in the input layer can significantly increase the computing time,
but does not guarantee a better prediction performance of the network. As stated above,
various combinations of input variables were tested during model development. The per-
formance of the proposed model was better than that of all tested models. The selection of
input neurons can significantly influence the model performance and so selection should be
done with caution.

8. Conclusions

A negative binomial regression and an ANN model were proposed to establish the
empirical relationship between traffic accidents and highway geometric variables, traffic
characteristics and environmental factors. The results of this study can eventually be em-
ployed to identify locations of high accident frequency for the most important transpor-
tation corridor in Taiwan. This study also demonstrated that ANN is a consistent
alternative for analyzing freeway accident frequency by comparing the prediction per-
formance with negative binomial regression analysis. This represents an important
methodological step in studying traffic accident frequency. The results obtained here, by
exploring a broad range of variables including highway geometry, traffic and environmen-
tal characteristics, provide valuable insight into the underlying relationship between risk
factors and vehicle accidents. In terms of future work, an application of the methodolo-
gical approaches used in this paper to different roadway types, such as interchange ramps,
would be worthwhile. The accident database revealed that quite a number of accidents oc-
curred at interchange ramps. Further exploration could provide a better understanding of
the characteristics for accidents occurred at interchange ramps and safer designs for the
freeway systems. In addition, the overall prediction performance of the proposed ANN
model is approximately 60%. Future work might focus on how to improve the prediction
performance of ANN models. First of all, because the structure of ANN can significantly
influence the prediction performance, it would be worthwhile for future studies to develop
a new ANN model for predicting accident rate (i.e., the neurons in the output layer will be
reduced to one) instead of accident frequency and to check if the resulting model can give a
better prediction. Secondly, training an ANN model using a different ratio of testing to
training data or the same ratio of testing to training data for a different number of obser-
vations could result in different results. Further investigation on the effects of the ratio of
testing to training data on the model performance would also be a good direction. Finally,
it would also be interesting for future studies to employ different training algorithms such
as the PNN to explore the factors that affect accident frequency and to see if the prediction
performance could be improved.
556 L.-Y. Chang / Safety Science 43 (2005) 541–557

Acknowledgements

The author would like to thank the Editor and referee for their constructive comments
and the National Science Council of Taiwan for financially supporting this study under
Contract No. NSC91–2211–E–415–002.

References

Abdel-Aty, M., Pande, A., 2005. Identifying crash propensity using traffic speed conditions. Journal of Safety
Research 36 (1), 97–108.
Carson, J., Mannering, F., 2001. The effect of ice warning signs on ice-accident frequency and severity. Accident
Analysis and Prevention 33 (1), 99–109.
Dougherty, M., 1995. A review of neural networks applied to transport. Transportation Research Part C 3 (4),
247–260.
Hadi, M., Aruldhas, J., Chow, L.-F., Wattleworth, J., 1995. Estimating safety effects of cross-section design for
various highway types using negative binomial regression. Transportation Research Record 1500, 169–177.
Hagan, M.T., Demuth, H.B., Beale, M., 1996. Neural Network Design. Thomson Learning Publishing Inc.
Hensher, D.A., Ton, T.T., 2000. A comparison of the predictive potential of artificial neural networks and nested
logit models for commuter mode choice. Transportation Research Part E 36 (3), 155–172.
Jin, X., Cheu, R.L., Dipti, S., 2002. Development and adaptation of constructive probabilistic neural network in
freeway incident detection. Transportation Research Part C 10 (2), 121–147.
Lee, J., Mannering, F., 2002. Impact of roadside features on the frequency and severity of run-off-roadway
accidents: an empirical analysis. Accident Analysis and Prevention 34 (2), 149–161.
Lee, A.H., Stevenson, M.R., Wang, K., Yau, K., 2002. Modeling young driver motor vehicle crashes: data with
extra zero. Accident Analysis and Prevention 34 (4), 515–521.
McCarthy, P.S., 1999. Public policy and highway safety: a city-wide perspective. Regional Science and Urban
Economics 29 (3), 231–244.
Miaou, S.P., 1994. The relationship between truck accidents and geometric design of road sections: Poisson
versus negative binomial regressions. Accident Analysis and Prevention 26 (4), 471–482.
Milton, J., Mannering, F., 1998. The relationship among highway geometrics, traffic-related elements and motor-
vehicle accident frequencies. Transportation 25, 395–413.
Mussone, L., Rinelli, S., Reitani, G., 1996. Estimating the accident probability of a vehicular flow by means of an
artificial neural network. Environment and Planning B 23, 667–675.
Mussone, L., Ferrari, A., Oneta, M., 1999. An analysis of urban collisions using an artificial intelligence model.
Accident Analysis and Prevention 31 (6), 705–718.
Poch, M., Mannering, F., 1996. Negative binomial analysis of intersection—accident frequencies. Journal of
Transportation Engineering 122 (2), 105–113.
Shankar, V.N., Mannering, F., Barfield, W., 1995. Effect of roadway geometrics and environmental factors on
rural freeway accident frequencies. Accident Analysis and Prevention 27 (3), 371–389.
Shankar, V.N., Milton, J., Mannering, F., 1997. Modeling accident frequencies as zero-altered probability
processes: an empirical inquiry. Accident Analysis and Prevention 29 (6), 829–837.
Shiva Nagendra, S.M., Khare, M., 2004. Artificial neural network based line source models for vehicular exhaust
emission predictions of an urban roadway. Transportation Research Part D 9 (3), 199–208.
Sohn, S., Lee, S., 2003. Data fusion, ensemble and clustering to improve the classification accuracy for the
severity of road traffic accident in Korea. Safety Science 41 (1), 1–14.
Subba Rao, P.V., Sikdar, P.K., Krishna Rao, K.V., 1998. Another insight into artificial neural networks through
behavioral analysis of access mode choice. Computers, Environment and Urban Systems 22 (5), 485–496.
Tong, H.Y., Hung, W.T., 2002. Neural network modeling of vehicle discharge headway at signalized intersection:
model descriptions and the results. Transportation Research Part A 36 (1), 17–40.
Vythoulkas, P.C., Koutsopoulos, H.N., 2003. Modeling discrete choice behavior using concepts from fuzzy set
theory, approximate reasoning and neural networks. Transportation Research Part C 11 (1), 51–73.
Washington, S., Karlaftis, M.G., Mannering, F.L., 2003. Statistical and Econometric Methods for Transpor-
tation Data Analysis. Chapman & Hall/CRC Press.
 L.-Y. Chang / Safety Science 43 (2005) 541–557 557

Yin, H., Wong, S.C., Xu, J., Wong, C.K., 2002. Urban traffic flow prediction using fuzzy-neural network.
Transportation Research Part C 10 (2), 85–98.
Yuan, F., Cheu, R.L., 2003. Incident detection using support vector machines. Transportation Research Part C
11 (3-4), 309–328.
Zhang, H.M., Ritchie, S.G., Jayakrishnan, R., 2001. Coordinated traffic-responsive ramp control via nonlinear
state feedback. Transportation Research Part C 9 (5), 337–352.
Zhong, M., Lingras, P., Sharma, S., 2004. Estimation of missing traffic counts using factor, genetic, neural and
regression techniques. Transportation Research Part C 12 (2), 139–166.