Accident Analysis and Prevention 120 (2018) 295–303

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

Effects of modal shares on crash frequencies at aggregate level T

a a,⁎ b
Mehdi Mohammadi , Gholamali Shafabakhsh , Ali Naderan
a
Department of Civil Engineering, Semnan University, Semnan, Iran
b
Department of Civil Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran

A R T I C LE I N FO A B S T R A C T

Keywords: Introduction: Long-range transportation plans often involve proposals for improvements/ changes in different
Safety planning modes of travel. This means that modal share of trips generated at each traffic analysis zone (TAZ) by mode of
Crash generation travel needs to be predicted/ forecasted for safety evaluation purposes. The objective of this research study is to
Crash mode split develop a series of aggregate crash prediction models (ACPMs) that relate with the modal split step of the
Macro model
conventional four-step demand models.
Negative binomial
Method: The models are developed utilizing network and vehicular, socio-economical, trip production/attrac-
tion and trip frequencies by mode at TAZ-level as explanatory variables in a generalized linear regression with
the assumption of a negative binomial error structure. Crash frequencies are split into total crashes (TC) and
severe crashes (SC).
Results: The models prove promising in estimating crash frequencies upon changes in modal shares, which is
essential in safety assessment of alternate transportation demand management (TDM) scenarios. Trips made in
car, bus, and bus Service mode became significant in the estimated TC and trips made in car, taxi, school service,
bus service and moped mode became significant in the estimated SC ACPMs.
Conclusions: The ACPMs may be used from two different points of view. First and most appropriate use is to
consider these as tools to forecast future crash frequencies and develop long-term plans to counteract. In the
second point of view, ACPMs act as the primary planning tool to identify how any increase in a specific mode-
ridership will contribute to crash frequencies. This is of great interest in developing plans that involve increased
use of a specific mode.
Practical application: As modal shares are forecasted in certain years into the future by the modal split step of
demand modeling, crash frequencies could also be forecasted and safety implications of mobility improvement
scenarios (e.g. increased number of trips by bus, car, etc.) would be evaluated.

1. Introduction and background
Transportation planners are often concerned with plans that directly
or indirectly affect modal use in urban trips (Meyer and Miller, 2001),
while a major issue in evaluation/ prioritization step is to estimate
safety implications of proposed policies and plans (Washington et al.,
2006). This requires safety analysis and modeling abilities at macro-
level that target the overall performance of network and enable the
planner to forecast safety figures into the future. On the other hand,
future growth in modal use is a major concern in planning efforts,
especially in developing countries. There is a general understanding
that changed policies and plans will contribute to changing multiple
long-term issues related to sustainability, efficiency and also safety of
transportation in urban environments (Tasic and Porter, 2016). In other
words, planners wish to shift trips from a certain mode of travel to
another (e.g. from private automobiles to public buses). These plans
should be evaluated to see if the modal shift is in the benefit of the trip
makers from a safety point of view.
Evaluating the safety performance of urban transportation network
at an aggregate (macro) level is approached by many researchers. Tasic
et al. (Tasic and Porter, 2016) explored the relationship between access
to multimodal transportation (spatial features, socio-economic char-
acteristics, land use mixture, street network patterns and multimodal
facilities) and safety outcomes in urban environments in census tract
level. Amoh-Gyimah et al. (Amoh-Gyimah et al., 2016) used Annual
Average Daily Traffic (AADT) data, population and housing data and
land-use data to developing macro-level pedestrian and bicycle crashes
model in second level Statistical Area of Melbourne. Li et al. (Li et al.,
2013) using Road network data, Traffic data, and Socio-demographic
data versus California county-level crash data to developing spatial

⁎
Corresponding author.
E-mail addresses: mmahdii@semnan.ac.ir (M. Mohammadi), shafabakhsh@semnan.ac.ir (G. Shafabakhsh), naderan@srbiau.ac.ir (A. Naderan).

https://doi.org/10.1016/j.aap.2018.08.019
Received 7 May 2017; Received in revised form 20 August 2018; Accepted 21 August 2018
0001-4575/ © 2018 Elsevier Ltd. All rights reserved.
M. Mohammadi et al.
safety macro-model. Ukkusuri et al. (Ukkusuri et al., 2012) investigated
the role of built environment on pedestrian crash frequency at census
tract level. Chang et al. (Chang et al., 2011) analysed the relationship
between Density of Surgeons and deaths from motor vehicle crashes
(MVC) based on general practitioners, urbanicity of the county, and
socioeconomic status of the county for US Counties. Huang et al.
(Huang et al., 2010) tried to detecting county-level variations of crash
risk in Florida by explicitly controlling for exposure variables of daily
vehicle miles traveled and population using county level data. Quddus
(Quddus, 2008a,b) developed a series of relationships between area-
wide different traffic casualties and the contributing factors associated
with ward characteristics in census wards level from the Greater
London metropolitan area. Hadayeghi et al. (Hadayeghi et al., 2003)
and (Hadayeghi et al., 2007) developed macro-level crash prediction
models that estimate crash frequencies in traffic analysis zones (TAZs)
as a function of land use, traffic and network characteristics. Lovegrove
and Sayed (Lovegrove and Sayed, 2006) provided guidelines for prac-
tical application of aggregate crash prediction models (ACPMs). they
have developed their models based on socio-demographic, network and
transportation demand management data. Lord and Persuad (Lord and
Persaud, 2004) developed a tool that would allow the estimation of
crashes on digital or coded urban transportation networks during the
planning process. Chatterjee et al. (Chatterjee et al., 2003) developed
regression models using link-length and daily traffic to forecast future
crash frequencies. Washington et al. (Washington et al., 2006) devel-
oped a series of ACPMs in order to forecast future crash frequencies in
urban TAZs of Arizona. Their models included population by age co-
hort; vehicle-kilometers traveled (VKT); network length by type and
other related variables. The ability of ACPMs to forecast future crash
frequencies at macro level was also reported by Hadayeghi et al.
(Hadayeghi et al., 2006).
Different formulations were used based on the purpose of study and
nature of available data. When the data are over-dispersed the negative
binomial (NB) model structure with a log-link function being the most
favored (Gomes and Vieira, 2013; Ukkusuri et al., 2012; Hadayeghi
et al., 2007; Kim et al., 2006; Qadeer Memon, 2006; Maher and
Summersgill, 1996). when the data are not over-dispersed the poisson
structure (Daniels et al., 2010; Lord et al., 2010, 2005) being the most
favored.
Also some other modeling technique such as zero-inflated poisson
and negative binomial (Malyshkina et al., 2010; Lord et al., 2007, 2005)
and gamma (Daniels et al., 2010; Lord et al., 2005) were used in lit-
erature.
More complex models such as simultaneous estimation of NB
models (de Guevara et al., 2004); Random effects model (Amoh-
Gyimah et al., 2016; Yu et al., 2013); multivariate Poisson model (Ma
and Kockelman, 2006); multivariate NB model (Winkelman, 2003);
multivariate Poisson-lognormal model (Park et al., 2010; El-Basyouny
and Sayed, 2009); full-Bayes hierarchical models with spatial effects
(Vandenbulcke et al., 2014; Huang et al., 2010; Aguero-Valverde et al.,
2006; Quddus, 2008a,b); geographically weighted Poisson regression
models (Li et al., 2013; Hadayeghi et al., 2010); neural network
(Abdelwahab and Abdel-Aty, 2002); bayesian neural network (Xie
et al., 2007), and support vector machine (Li et al., 2008) were also
estimated in previous researches.
Close estimation of model coefficients in spatial and non-spatial NB
models was found by Quddus (Quddus, 2008a,b), and he suggested that
same variables become significant in both models. On the other hand,
simpler models with ordinary least square (OLS) estimation that pro-
vide more understandable results were also used by some researchers,
e.g. linear regression (Clark and Cushing, 2004), linear regression with
a spatial lag parameter (Levine et al., 1995), and log-linear models
(Washington et al., 2006; Wier et al., 2009).
Variables used in previous studies were aggregated at different le-
vels such as state (Noland, 2002), county (Li et al., 2013; Chang et al.,
2011; Huang et al., 2010; Aguero-Valverde et al., 2006), TAZ (Abdel-
296
Accident Analysis and Prevention 120 (2018) 295–303
Fig. 1. Four-step safety modeling (Naderan and Shahi, 2010a,b).
Aty et al., 2013; Pulugurtha et al., 2013; Huang et al., 2013; Wang
et al., 2012; Abdel-Aty et al., 2011; Naderan and Shahi, 2010a,b;
Hadayeghi et al., 2007), ward (Quddus, 2008a,b), census block (Tasic
and Porter, 2016; Ukkusuri et al., 2012; Levine et al., 1995) and even a
grid of 0.1 square miles (Kim et al., 2006). Compared to non-spatial
models, it has been suggested that spatial models provide better esti-
mation results in smaller aggregation units, while they show rather
identical results at higher aggregations (Aguero-Valverde et al., 2006;
Quddus, 2008a,b).
Crash generation models (CGMs) were developed based on results of
trip generation models that related trip frequencies made from/ to a
TAZ by different purposes to crash frequencies in each TAZ (Naderan
and Shahi, 2010a,b). The purpose of those models was to forecast future
crash frequencies upon future changes in frequencies of different trip
purposes (Naderan and Shahi, 2010a,b). These models may be used to
develop four-step safety planning models similar to the process fol-
lowed in four-step demand modeling (Naderan and Shahi, 2010a,b).
Fig. 1 depicts the idea of inter-relating transportation and safety plan-
ning. The concept behind this figure is that each trip generated in a TAZ
by a certain mode of travel (produced in or attracted to a TAZ) will be
subject to a probable risk of crash.
The primary purpose of this paper is to estimate ACPMs that would
predict crash frequencies based on trip frequencies by modes of travel.
This provides a set of policy-related variables: number of trips by mode
and their impact on crash frequencies. As modal shares are forecasted in
certain years into the future by the modal split step of demand mod-
eling, crash frequencies could also be forecasted and safety implications
of mobility improvement scenarios (e.g. increased number of trips by
bus, car, etc.) would be evaluated.
2. Data processing
This study is based on the transportation databases from city of
Mashhad, Iran (MTTO, 2009). City of Mashhad is the second largest city
of Iran with 2.67 million residents (in 2008) and an area of 195 square
kilometers, and is divided into 253 TAZs. In Iran, transportation master
plans done in cities with more than 100,000 population and should be
updated each 5 years. Mashhad transportation master plan studies
updated by conducting a comprehensive household travel survey in
2008 as last time, and zonal trip productions/ attractions per mode are
available. Four-step demand models are also recalibrated to forecast
future trip frequencies.
The data which use in macro level modeling must be accessible in
point type (with precise geographical latitude and longitude) for data
aggregation purpose. On the other hand, these types of data are not
commonly recorded continually in developing countries such as Iran.
M. Mohammadi et al.
The only urban crash dataset that recorded in Iran is Mashhad urban
crashes in 2008 which is used in this research. Also, explanatory data
must be matched with the crash data. Hence, the accessibility to the
2008 trip mode split data is another reason for using the recorded crash
data in 2008. Also, it should be noted that the results of the models,
made in this study, have been confirmed with existing data.
So Crash data of the same year were also geocoded into the GIS
environment. GIS tools were used to assign crashes represented by
points to the 253 TAZs represented by polygons. Since crashes occur on
network links and nodes, it is believed that the accuracy and precision
of geocoded crash data may result seemingly random errors when GIS
tools assign points to polygons. Previous researches suggest this error is
not significant (Hadayeghi et al., 2003; de Guevara et al., 2004).
Previous researches also clearly indicate that factors affecting cra-
shes vary by crash severity (Quddus, 2008a,b; de Guevara et al., 2004),
thus, crash frequencies are stratified into total and severe (=injury +
fatal) crashes. Trip frequencies by different modes of travel are pro-
posed as explanatory variables in the model. Origin-destination (OD)
trip matrices by mode derived from 2008 household travel survey were
used for calculating trip frequencies by mode per TAZ. First, trip by a
certain mode of travel were filtered. Second, OD matrix was produced
using the pivot-point command of MS Excel. Third, this OD matrix was
then added to its transpose and summed over rows to result trip at-
tractions per TAZ. Again, it was summed over columns to result trip
productions per TAZ. Fourth, productions were added to attractions to
obtain trip generations by mode per TAZ.
Descriptive analysis of data is presented in Table 1. It can be seen
that crash frequencies are very high in Mashhad for a single year
(2008). The minimum of average trip frequencies per TAZ is by bus
services (Bus Service mode) around 145 trips, and the maximum is
8998 trips on foot (Walk mode), then 6196 trips by private automobiles
(Car mode).
Fig. 2 depicts market share of different modes of travel. The major
modes of travel include walk (36%), car (24%), bus (15%) and taxis
(10%). It should be mentioned that taxis in Mashhad do not act as a
door-to-door mode of travel. In fact, taxi is regarded as a para-transit
mode that picks up/ drops off passengers en-route while there are other
passengers on board, and may have predefined route (same as a bus-
line). Minibuses operate in the same manner as taxis. Mopeds include
both bicycles and motor-cycles.
Spatial distributions of trip frequencies by mode are shown in Fig. 3
and those of crash frequencies per TAZ are shown in Fig. 4. Class limits
were determined based on the natural breaks determined by the GIS
software.
3. Modeling approach
VKT had been the most favored variable in crash studies because it
directly relates link volumes to crash frequency (Hadayeghi et al.,
2007). However, VKT for the urban area inherits some uncertainties
regarding omission of streets in lower ranks (i.e. local/ distributor
links) and precision of estimations for link volumes. Some previous
researches suggested use of TAZ population instead of VKT (Wier et al.,
2009; Washington et al., 2006; de Guevara et al., 2004). TAZ popula-
tion is easier to forecast and more accurate. Nevertheless, it has the
disadvantage of neglecting the effect of attracted traffic. In order to
compensate for this, we suggest using trip frequencies by mode gen-
erated in each TAZ. Trips generated in a TAZ include all the trips that
start (originate) from that TAZ and end (are destined to) there.
Therefore, the proposed approach of using trip frequencies generated in
a TAZ dismisses the effect of through traffic, but more precisely account
for the inbound and outbound traffic in TAZ.
Most of previous attempts on aggregate crash modeling were con-
cerned with explaining the cause-effect relation in crash occurrences,
whilst the primary purpose of this research is to investigate how trip
frequencies by different modes in urban areas contribute to crash
297
Accident Analysis and Prevention 120 (2018) 295–303
frequencies at TAZ-level, in order to forecast future crash frequencies.
Because focus here is prediction, and not explanation, models have
fewer restrictions: inference, or effects of isolated variables (estimated
coefficients) not too important, multi co-linearity tolerated, goodness of
fit and predictive-ability most important (Washington et al., 2006,
2003).
The negative binomial (NB) formulation with a log-link function is
used for estimation of ACPMs in this study. In the NB model, the dif-
ference between variance and mean is presented by V[u] = u+ αu2,
where, V[u] is the estimated variance of the crash frequency; u is the
estimated mean crash frequency and α is the NB over-dispersion para-
meter. The Poisson regression model is regarded as a limiting model of
the NB regression model as α approaches zero, which means that the
selection between these two models is dependent on the value of α. The
NB distribution has the general form of Eq. (1).
1/ α yi
Γ ((1/ α ) + yi ) ⎡ 1/ α ⎤ ⎡ λi ⎤
P (yi ) =
⎢ ⎥ ⎢ ⎥
Γ (1/ α ) yi ! ⎣ (1/ α ) + λi ⎦ ⎣ (1/ α ) + λi ⎦ (1)
where Γ(.) is a gamma function. The NB model was developed math-
ematically by assuming that unobserved crash heterogeneity (variation)
across TAZs is gamma distributed, while crashes within sites are
Poisson distributed (Washington et al., 2006). The model form in Eq.
(2) is used where y= predicted crash frequency; Xi= explanatory
variables and bi= model parameters.
n
y = exp ∑ bi Xi
i=1 (2)
The regression coefficients as well as the over-dispersion parameter
were estimated by the maximum likelihood method of the statistical
software SPSS (IBM Corp., 2014), following a forward procedure. In this
procedure, a simple model with only an intercept term is used as a
starting point and the explanatory variables are added to the model one
by one. Keeping these variables in the model was based on whether
they would improve goodness of fit results. Three statistical measures,
(i) t-statistic significant at 5% level, (ii) ratio of deviance over degrees
of freedom between 0.8–1.2, and (iii) the over-dispersion parameter (α)
were used to assess the goodness-of-fit for each model (Washington
et al., 2003) (Cameron et al., 1998).
The significance level for this study was typically chosen set to 5%
(Craparo, 2007) and each explanatory variable should significant at this
level. For a statistical model, Deviance is a goodness-of-fit statistic; in
ordinary least squares sum of squares of residuals used to cases where
model-fitting is achieved by maximum likelihood. Deviance have a si-
milar role in exponential dispersion models and generalized linear
models. Beside this dividing it by its degrees of freedom defined as an
Indicator parameter of under/over-dispersing of data of a model. If
Deviance/df is equal to 1, the data of model is not under-dispersed nor
over-dispersed. The acceptable range of data under/over-dispersing is
0.8 to 1.2. α is the NB over-dispersion parameter. If αis closer to 1, that
the Poisson model should be used and if is greater than 1, NB is more
appropriate.”
4. Modeling results
The regression results for ACPMs per year are presented in Tables
2–5. As shown, the relationship between crashes and selected variables
is highly significant at the 5% significance level. The dispersion para-
meters are significantly different from zero, indicating justification of
NB over the Poisson formulation. Although the primary purpose of
developing ACPMs was forecasting crash frequencies by mode in the
future, the models were also tested for co-linearity and correlation
between independent variables so that an explanatory use of models
should give valid results. Mean deviance (value of deviance over de-
grees of freedom) for the models is between 1.0 and 1.2, indicating an
acceptable goodness of fit.
M. Mohammadi et al. Accident Analysis and Prevention 120 (2018) 295–303

Table 1
Descriptive Analysis of Variables of TAZs for ACPMs.
Category Variable Definition` Mean* Std. D. Min Max

General TAZ Traffic analysis zone - - 1 253

area m2 TAZ area in km2 1.13 0.12 1.66 19.04
Crashes Total Total No. of Crashes 153.68 135.15 0 1163
By Type Severe No. of Fatal+ Injury Crashes 26.68 24.05 0 250
Network and Vehicular sint No. of signalized intersections 7 5.05 0 27
usint No. of un-signalized intersections 0.45 4.4 0 66
stm Length of total main street in km 500.83 595.87 0 3319.75
stm Total street traversal median 0.4 0.395 0 3.65
sntm Total street non-traversal median 0.42 0.77 0 9.06
tnd Traffic network density** 0.02 0.01 0 0.06
tv Traffic volume (VKT) 78460 50447.1 1 1996288
Socio-Economical per TAZ pop Population per TAZ 10401.63 7796.85 1 42564
dis Distance to Central Business District in meter 6152.2 16130 0 3876
stu No. of students 2728.22 2428.34 0 16009
hbd No. of hospital bed 266.3 272 0 1072
pga Public garden area in hectare 4.33 18.89 0 208
ems No. of employees 2774.58 2264.82 0 13124
emt No. of Employment 2774.59 1367.95 463 7526
shp No. of shops 528 682.64 0 3599
amo Average Motorcycle Ownership 0.019 0.062 0 0.313
aco Average Car Ownership 0.048 0.042 0 0.215
Trip Production/ Attraction per TAZ pw No. of work trip productions 1870.64 1640.82 0 22374
pedu No. of educational trip productions 833.72 1013.7 0 9941
psh No. of shopping trip productions 432.8 387.138 0 2973
pre No. of recreational trip productions 527.77 411.1 0 2805
pper No. of personal trip productions 367.56 360.48 0 3246
pnhb No. of non-home-base trip productions 529.7 641.95 0 5856
aw No. of work trip attractions 1797.26 1551.9 0 8813
aedu No. of educational trip attractions 869.87 927.83 0 8884
ash No. of shopping trip attractions 655.74 2188.9 0 23609
are No. of recreational trip attractions 631.02 948.4 0 7211
aper No. of personal trip attractions 272.5 412.1 0 2763
anhb No. of non-home-base trip attractions 557.45 785.12 0 7633
ptotal Total no. of trip productions 4729.8 3006 108 22834
atotal Total no. of trip attractions 4783.72 5170.68 30 45430
Trips mc No. of trips made by car 6196.68 4544.88 139 32437
Generated mt No. of trips made by taxi*** 2601.27 2376.92 0 17757
by Mode mschs No. of trips made by School Services 832.91 792.71 0 4919
mpick No. of trips made by pickup 245.98 347.87 0 2623
mmb No. of trips made by minibus*** 198.81 339.52 0 3225
mb No. of trips made by bus**** 3876.33 5593.54 0 77798
mbs No. of trips made by bus services***** 145.29 230.13 0 1513
mmo No. of trips made by motorbike 1729.33 1644.97 0 10807
mbk No. of trips made by bike 490.98 613.05 0 5263
mw No. of trips made by walking 8998.20 8592.14 0 58318

*
Number of observations = 253 TAZs, **Traffic network density = (street length*street width)/TAZ Area, ***para-transit, ****transit, *****shuttle.

understudy area. In SC estimation number of un-signalized intersections

and VKT already have positive sign and traffic density have negative
sign; it means that the number of un-signalized intersections and VKT
have positive and traffic density has negative effect on severe crash
frequency in understudy area.
Estimation of NB with Socio-Economical Variables are shown in
Table 3. From all of socio-economical variables population per TAZ,
number of students, average car ownership and average motorcycle
ownership are significant in relation with TC. Also average motorcycle
ownership, distance to central business district and number of hospital
bed are significant in relation with SC. In TC estimation all of four
significant variables already have positive sign; it means that the
variables have positive effect on total crash frequency in understudy
Fig. 2. Modal share of trips (MTTO, 2009).
area. In SC estimation average motorcycle ownership already have
positive sign and distance to central business district and number of
hospital bed has negative sign; it means that average motorcycle
Estimation of NB with Network and Vehicular Variables are shown
ownership have positive and distance to central business district and
in Table 2. From all of network and vehicular variables number of in-
number of hospital bed has negative effect on severe crash frequency in
tersections, number of un-signalized intersections and VKT are sig-
understudy area.
nificant in relation with TC. Also number of un-signalized intersections,
Estimation of NB with trip production/attraction variables are
VKT and traffic density are significant in relation with SC. In TC esti-
shown in Table 4. From all of trip production/attraction variables,
mation all of three significant variables already have positive sign; it
number of shopping trip production, number of non-home-base trip
means that the variables have positive effect on total crash frequency in

298
M. Mohammadi et al. Accident Analysis and Prevention 120 (2018) 295–303

Fig. 3. Spatial distribution of trip frequencies by mode.

production, number of recreational trip production, number of shop-
ping trip attraction and number of non-home-base trip attraction are
significant in relation with TC. Also number of shopping trip produc-
tion, number of non-home-base trip production, number of recreational
trip production, number of shopping trip attraction, number of non-
home-base trip attraction and number of recreational trip attraction are
significant in relation with SC. In TC estimation number of shopping
trip production and number of non-home-base trip attraction already
have positive sign and number of non-home-base trip production,
number of recreational trip production and number of shopping trip
attraction have negative sign; it means that number of shopping trip
production and number of non-home-base trip attraction have positive
and number of non-home-base trip production, number of recreational
trip production and number of shopping trip attraction has negative
effect on total crash frequency in understudy area. In SC estimation
number of recreational trip production and number of non-home-base
trip attraction already have positive sign and number of shopping trip
production, number of non-home-base trip production, number of
shopping trip attraction and number of recreational trip attraction have
negative sign; it means that number of recreational trip production and
number of non-home-base trip attraction have positive and number of
shopping trip production, number of non-home-base trip production,
number of shopping trip attraction and number of recreational trip
attraction have negative effect on severe crash frequency in understudy
area.
Estimation of NB with modal share variables are shown in Table 5.

Fig. 4. Spatial distribution of crash frequencies per TAZ.

299
M. Mohammadi et al. Accident Analysis and Prevention 120 (2018) 295–303

Table 2
Estimation of NB – ACPMs with Network and Vehicular Variables.
Total Crashes (TC) Severe Crashes (SC)

Var. Coeff. Std. Err Sig. Var. Coeff. Std. Err Sig.

constant 2.89 0.1312 0.000 constant 1.491 0.1119 0.000

int 0.091840 0.2690 0.001 usint 0.180032 0.2983 0.008
usint 0.129011 0.2610 0.003 vkt 0.114310 0.3317 0.013
vkt 0.093170 0.2501 0.032 td −0.001208 0.0790 0.048

Goodness of fit Goodness of fit

No. of observations 253 No. of observations 253

Degree of freedom (df) 246 Degree of freedom (df) 244
Deviance/df 1.14 Deviance/df 1.16
α 0.436 α 0.367
Log-likelihood (LL) −1423.35 Log-likelihood (LL) −1211.01
ρ2 0.1099 ρ2 0.0975

From all of modal share variables number of trips made by car, number The final models are shown in Eq. (3) and Eq. (4), where the TAZ
of trips made by bus and number of trips made by bus services are index is omitted from all variables for the sake of brevity.
significant in relation with TC. Also number of trips made by car,
TC = Exp(4.302 + .000111car-.000321bus+.001busserv) (3)
number of trips made by taxi, number of trips made by school services,
number of trips made by bus services and number of trips made by SC = Exp(2.567 + .000060car-.000066taxi+.000149 schoolserv
motorbike are significant in relation with SC. In TC estimation number +.000462busserv+.00013moped) (4)
of trips made by car and number of trips made by bus services already
have positive sign; it means that these variables have positive and There were 38,882 crashes in the urban network, while the esti-
number of trips made by bus have negative effect on total crash fre- mated model predicts 39,679 crashes. That is, crash frequencies are
quency in understudy area. In SC estimation number of trips made by over-estimated by +797 or +2.04%, which is a close estimate at ag-
car, number of trips made by school services, number of trips made by gregate level. SC model predicts 6746 crashes that is only +5 crashes or
bus services and number of trips made by motorbike have positive and −0.07% less than actual value of 6751 SC observed in database.
number of trips made by taxi has negative sign; it means that number of
trips made by car, number of trips made by school services, number of
5. Discussion
trips made by bus services and number of trips made by motorbike has
positive and number of trips made by taxi has negative effect on severe
As it can be seen from Tables 2–5, four types of ACPM’s with four
crash frequency in understudy area.
independent variables were estimated. All of these four independent
As it can be seen in Tables 2–5 all types of constructed models have
variables have acceptable results in transportation safety planning
acceptable goodness of fit and have capability for practical usage in
studies. Network and vehicular variables and socio-economical vari-
safety studies. From four type constructed model in this paper, first and
ables were used mostly in previous studies as presented in introduction
second models were familiar in safety studies and have wide use in
and background section. Estimated ACPM’s with network and vehicular
different researches. Third one was firstly introduced in 2010 by
variables is applicable when planners want to change or define new
(Naderan and Shahi, 2010a,b) and led to crash generation concept. The
land use policies. Estimated ACPM’s with socio-economical variables is
simplest definition of this concept is “any trip can lead to one crash”.
applicable when planners want to change or define new policies for
But the last type of models is the main objective of this paper. Using of
transportation facilities supply. Trip production/attraction firstly used
modal share data as variables of model can lead to models which has
by (Naderan and Shahi, 2010a,b) as independent variables of ACPMs.
capability to provide the basic tool for evaluation of travel demand
Estimated ACPM’s with trip production/attraction variables is applic-
management (TDM) scenarios regarding modal shares in urban trans-
able when planners want to change or define new TDM policies. In this
portation planning.
paper another type of variables including modal share data presented as

Table 3
Estimation of NB – ACPMs with Socio-Economical Variables.
Total Crashes (TC) Severe Crashes (SC)

Var. Coeff. Std. Err Sig. Var. Coeff. Std. Err Sig.

constant 2.964 0.6610 0.000 constant 2.213 0.1619 0.000

pop 0.00141 0.00198 0.000 amo 0.00045 0.00037 0.021
stu 0.00007 0.00026 0.015 dis −0.00001 0.000009 0.027
aco 0.00045 0.00077 0.031 hbd −0.00003 0.000024 0.050
amo 0.00014 0.00029 0.044 – – – –

Goodness of fit Goodness of fit

No. of observations 253 No. of observations 253

Degree of freedom (df) 247 Degree of freedom (df) 243
Deviance/df 1.11 Deviance/df 1.16
α 0.61 α 0.48
Log-likelihood (LL) −976.91 Log-likelihood (LL) −614.16
ρ2 0.14 ρ2 0.1211

300
M. Mohammadi et al. Accident Analysis and Prevention 120 (2018) 295–303

Table 4
Estimation of NB – ACPMs with Trip Production/Attraction variables.
Total Crashes (TC) Severe Crashes (SC)

Var. Coeff. Std. Err Sig. Var. Coeff. Std. Err Sig.

constant 4.413 0.0941 0.000 constant 2.314 0.0991 0.000

psh 0.003516 0.001432 0.0039 psh −0.001906 0.000912 0.048
pnhb −0.000781 0.000218 0.0071 pnhb −0.001231 0.000247 0.001
pre −0.001908 0.000750 0.014 pre 0.001338 0.000479 0.009
ash −0.000383 0.000155 0.0038 ash −0.000347 0.000191 0.004
anhb 0.001436 0.000531 0.0043 anhb 0.002538 0.000591 0.000
– – – – are −0.000059 0.000029 0.024

Goodness of fit Goodness of fit

No. of observations 253 No. of observations 253

Degree of freedom (df) 244 Degree of freedom (df) 242
Deviance/df 1.101 Deviance/df 1.129
α 0.59 α 0.47
Log-likelihood (LL) −971.84 Log-likelihood (LL) −698.38
ρ2 0.0737 ρ2 0.0653

a new set of independent variables for ACPMs. As shown in Table 5 the
estimated models by modal share data for both of TC and SC, presented
acceptable results. Estimated ACPM’s with modal share variables will
be applicable in two different way; (i) when planners want to forecast
future crash frequencies and develop long-term plans to counteract the
increase in crash frequencies. (ii) ACPMs act as the primary planning
tool to identify how any increase in a specific mode-ridership will
contribute to crash frequencies. This is of great interest in developing
plans that involve increased use of a specific mode.
The interpretation of significant explanatory variables in the last
models are given in the following. In general, four most important modes
of travel including car, taxi, bus and moped, show significant influence on
crash frequencies at TAZ-level and are present in the models.
• Car Trips: Trips made by car are highly significant in both estimated
ACPMs and show a positive sign. This means TAZs with more pro-
portions of car trips are associated with more crash frequencies.
More car trips mean more vehicles occupying the streets, which in
turn translates into increased exposure to crashes. From Table 1 it is
concluded that average trip frequencies by automobiles per TAZ is
around 6196 trips. As a working example, if car trips generated in a
TAZ are forecasted to increase by 20% or 1240 trips in the future,
given all other trip modes are held constant; crash frequencies in
that TAZ may be calculated using related ACPM. This results ap-
proximately 14.75% increase in TC (1-exp[.000111*1240]) and
7.72% increase in SC of that TAZ (1-exp[.000060*1240]).
• Bus Trips: Trips made by buses are significant only in TC model and
show a negative sign. This means TAZs with more proportions of bus
trips are associated with less crash frequencies. Table 1 shows that
average trip frequencies by bus per TAZ is around 3876 trips.
Therefore, if bus trips generated in a TAZ are forecasted to increase
by 20% or 775 trips in the future, given all other trip modes are held
constant, TC frequencies in that TAZ decrease approximately
28.24% (1-exp[.000321*775]). This decrease adds up to the many
benefits of public transit modes of travel.
• Bus Services Trips: Trips made by exclusive bus services are sig-
nificant in both estimated ACPMs and show a positive sign. This
mode acts as a door-to-door service mostly for major trip attractors
such as factories, government offices, so on. As seen in Table 1,
average trip frequencies by bus services per TAZ are around 145
trips. Therefore, if trips generated in a TAZ by this mode are fore-
casted to increase by 20% or 30 trips in the future, given all other
trip modes are held constant, TC in that TAZ increase approximately
3.04% (=1-exp[.001*30]) and SC 1.4% (=1-exp[.000462*30]).
• Taxi Trips: Trips made by taxies are significant only in SC model
and show a negative sign. This means TAZs with more proportions
of taxi trips are associated with less crash frequencies. Table 1 shows
that average trip frequencies by taxi per TAZ is around 2601 trips.
Therefore, if taxi trips generated in a TAZ are forecasted to increase
by 20% or ca trips in the future, given all other trip modes are held
constant, TC frequencies in that TAZ decrease approximately 5.24%
(1-exp[.000066*775]).

Table 5
Estimation of NB – ACPMs with Modal Share Variables.
Total Crashes (TC) Severe Crashes (SC)

Var. Coeff. Std. Err Sig. Var. Coeff. Std. Err Sig.

constant 4.302 0.0819 0.000 constant 2.567 0.0823 0.000

mc 0.000111 0.000013 0.000 mc 0.000060 0.000018 0.001
mb −0.000321 0.000012 0.008 mt −0.000066 0.000032 0.038
mbs 0.001 0.0002 0.026 mschs 0.000149 0.000069 0.032
– – – – mbs 0.000462 0.0002 0.019
– – – – mmo 0.000130 0.000028 0.000

Goodness of fit Goodness of fit

No. of observations 253 No. of observations 253

Degree of freedom (df) 244 Degree of freedom (df) 242
Deviance/df 1.12 Deviance/df 1.15
α 0.50 α 0.40
Log-likelihood (LL) −1452.35 Log-likelihood (LL) −1007.55
ρ2 0.0916 ρ2 0.0725

301
M. Mohammadi et al.
• School Services Trips: Trips made by students by school services are
significant only in SC model and show a positive sign. A conclusion
may be that TAZs with more proportions of school trips suffer SC
more than other TAZs. In Table 1, average trip frequencies by school
services per TAZ are around 832 trips. Therefore, if trips generated
in a TAZ by this mode are forecasted to increase by 20% or 167 trips
in the future, given all other trip modes are held constant, SC in that
TAZ increase approximately 2.52% (=1-exp[.000149*167]).
• Moped Trips: Trips made by moped are significant only in SC model
and show a positive sign. mopeds are one of the most vulnerable
road users and if involved in a crash, they are likely to suffer in-
juries. This is why they show a positive sign in the SC model.
According to Table 1, average moped trip frequencies per TAZ are
around 1729 trips. Therefore, if moped trips generated in a TAZ are
forecasted to increase by 20% or 346 trips in the future, given all
other trip modes are held constant, SC frequencies in that TAZ in-
crease approximately 4.60% (=1-exp[.000130*346]).
Knowledge of these changes in crash frequencies is essential in order
to develop long-term safety plans and establish detailed studies that
result counter measures of specific types. It would be expected that trip
frequencies by other modes would also change (decrease or increase),
and they would have a combined effect on crash frequencies.
6. Conclusions
ACPMs are estimated with different types of explanatory variables
in previous studies. Traffic network, vehicular and socio-economical
variables, are widely used in the literature. Besides, trip production/
attraction variables are different set of explanatory variables which
suggested for ACPMs estimations. Each of mentioned variables has
significant application in transportation safety planning as declared in
discussion section. So in this paper by using of Mashhad transportation
databases, three types of ACPMs estimated for Mashhad and all of es-
timated models presented acceptable results.
However the primary purpose of this paper was to investigate how
travel demand by specific modes of travel in urban areas contributes to
crash frequencies at TAZ-level. To this end, a series of ACPMs were
developed using trip frequencies by mode generated per TAZ as the
primary variables affecting crash occurrence, was the primary purpose
of this study. These models provide the basic tool for evaluation of
travel demand management (TDM) scenarios regarding modal shares in
urban transportation planning. A certain TDM scenario may be devel-
oped to reduce modal share of a specific mode and the related change in
crash frequency could be predicted by the appropriate ACPM. These
models could easily be used for forecasting future crash frequencies,
based on future forecasts of trip frequencies.
The ACPMs may be used from two different points of view. First and
most appropriate use is to consider these as tools to forecast future
crash frequencies and develop long-term plans to counteract. In this
way, the models may also be used to identify which group of TAZs
(rather than a single TAZ) will experience a rapid increase in crash
frequencies and develop focused detailed studies and specific treat-
ments for the problematic areas. This involves a microscopic (dis-
aggregate) view and may include a road safety audit (RSA) to identify
symptoms and appropriate countermeasures.
In the second point of view, ACPMs act as the primary planning tool
to identify how any increase in a specific mode-ridership will contribute
to crash frequencies. This is of great interest in developing plans that
involve increased use of a specific mode. For example, a new transit
plan may be proposed that will increase trips by buses by 20% and thus
decrease car trips by somewhat the same amount. ACPMs are able to
roughly predict changes in crash frequency at an aggregate level and
provide for the safety evaluation of that proposed plan. On the other
hand, cautious should be exercised while interpreting model coeffi-
cients: negative or positive signs maybe forced due to an unseen
302
Accident Analysis and Prevention 120 (2018) 295–303
(unaccounted for) variable that was omitted from the model.
Acknowledgment
This paper is part of an ongoing research on PhD dissertation at
Faculty of Civil Engineering, Semnan University, Semnan, Iran. This
research did not receive any specific grant from funding agencies in the
public, commercial, or not-for-profit sectors.
References
Abdel-Aty, M.A., Siddiqui, C., Huang, H., 2011. Integrating trip and roadway character-
istics in managing safety at traffic analysis zones. Transp. Res. Rec. 2213, 20–28.
Abdel-Aty, M., et al., 2013. Geographical unit based analysis in the context of transpor-
tation safety planning. Transp. Res. Part A Policy Pract. 49, 62–75.
Abdelwahab, H.T., Abdel-Aty, M.A., 2002. Artificial neural networks and logit models for
traffic safety analysis of toll plazas. Transp. Res. Rec. 1784, 115–125.
Aguero-Valverde, Jonathan, Jovanis Paul, P., 2006. Spatial analysis of fatal and injury
crashes in Pennsylvania. Accid. Anal. Prev. 38, 618–625 2006.
Amoh-Gyimah, Richard, Saberi, Meead, Sarvi, Majid, 2016. Macroscopic modeling of
pedestrian and bicycle crashes: a cross-comparison of estimation methods. Accid.
Anal. Prev. 93, 147–159.
Cameron, A., Colin, Trivedi, Pravin, K., 1998. Regression Analysis of Count Data.
Cambridge University Press, New York.
Chang, D.C., et al., 2011. Density of surgeons is significantly associated with reduced risk
of deaths from motor vehicle crashes in US counties. J. Am. Coll. Surg. 212 (5),
862–866.
Chatterjee, A., et al., 2003. Tools for Assesing Safety Impact of Long-range Transportation
Plans in Urban Areass. Office of Metropolitan Planning and Programs, Federal
Highway Administration, US Department of Transportation.
Clark, David E., Cushing, Brad M., 2004. Rural and urban traffic fatalities, vehicle miles,
and population density. Accid. Anal. Prev. 36, 967–972.
Craparo, Robert M. 2007. Significance level, In Salkind, Neil J. Encyclopedia of
Measurement and Statistics. s.l. : CA: SAGE Publications. pp. 889–891., 2007. ISBN 1-
412-91611-9.
Daniels, Stijn, et al., 2010. Explaining variation in safety performance of roundabouts.
Accid. Anal. Prev. 42, 393–402.
de Guevara, Felipe, Ladrón, Washington, Simon P., Oh, Jutaek, 2004. Forecasting Crashes
at the Planning Level: A Simultaneous Negative Binomial Crash Model Applied in
Tucson, Arizona. Transportation Research Record 1897. TRB, National Research
Council, Washington, D.C.
El-Basyouny, K., Sayed, T., 2009. Collision prediction models using multivariate poisson-
lognormal regression. Accid. Anal. Prev. 41 (4), 820–828.
Gomes, Sandra, Vieira, 2013. The influence of the infrastructure characteristics in urban
road accidents occurrence. Accid. Anal. Prev. 60, 289–297.
Hadayeghi, A., Shalaby, A.S., Persaud, B., 2003. Macro-Level Accident Prediction Models
for Evaluating Safety of Urban Transportation Systems. Transportation Research
Record 1840. TRB, National Research Council, Washington D.C.
Hadayeghi, Alireza, et al., 2006. Temporal transferability and updating of zonal-level
accident prediction models. Accid. Anal. Prev. 38, 579–589.
Hadayeghi, A., Shalaby, A.S., Pesuad, B., 2007. safety prediction models: a proactive tool
for safety evaluation in Urban transportation planning applications. Transportation
Research Record, TRB. National Research Council, Washington D.C.
Hadayeghi, A., Shalaby, A.S., Persaud, B.N., 2010. Development of planning level
transportation safety tools using geographically Weighted Poisson Regression. Accid.
Anal. Prev. 42 (2), 676–688.
Huang, HelaiA., Mohamed, Abdel-Aty, Lotfi Ali, D., 2010. County-level crash risk analysis
in Florida, bayesian spatial modeling. Transportation Research Record: Journal of the
Transportation Research Board (2148), 27–37 Transportation Research Board of the
National Academies, Washington, D.C.,.
Huang, H., Xu, P., Abdel-Aty, M.A., 2013. Transportation safety planning: spatial analysis
approach. Presented in Transportation ResEarch Board 92nd Annual Meeting.
IBM Corp, 2014. IBM SPSS statistics 23. s.l. IBM SPSS Statistics for Windows, Version
22.0. IBM Corp., Armonk, NY Released 2013. www.vsni.co.uk.
Kim, K., Brunner, I.M., Yamashita, E.Y., 2006. The influence of land use, population,
employment and economic activity on accidents. Transportation Research Record
1953. TRB, National research Council, Washington DC, pp. 56–64.
Levine, N., Kim, K.E., Nitz, L.H., 1995. Spatial analysis of Honolulu motor vehicle crashes:
II. Zonal generators. Accid. Anal. Prev. 27 (5), 675–685.
Li, X., et al., 2008. Predicting motor vehicle crashes using support vector machine models.
Accid. Anal. Prev. 40 (4), 1611–1618.
Li, Zhibin, et al., 2013. Using geographically weighted poisson regression for county-level
crash modeling in California. Saf. Sci. 58, 89–97.
Lord, D., Persaud, B.N., 2004. Estimating the safety performance of urban road trans-
portation networks. Accid. Anal. Prev. 36 (5), 609–620.
Lord, Dominique, Washingtonb, SimonP., Ivanc, JohnN., 2005. Poisson, Poisson-gamma
and zero-inflated regression models of motor vehicle crashes: balancing statistical fit
and theory. Accid. Anal. Prev. 37 (1), 35–46.
Dominique, Lord, Washington, Simon, Ivan, JohnN., 2007. Further notes on the appli-
cation of zero-inflated models in highway safety. Accid. Anal. Prev. 39 (1), 53–57.
Lord, D., Geedipally, S.R., Guikema, S.D., 2010. Extension of the application of conway-
maxwell-poisson models: analyzing traffic crash data exhibiting underdispersion.
M. Mohammadi et al.
Risk Anal. 2010 30 (8), 1268–1276.
Lovegrove, G.R., Sayed, T., 2006. Using Macro-level Collision Prediction Models Inroad
Safety Planning Application. Transportation research record 1950, TRB, National
Research Council, Washington D.C.
Ma, J., Kockelman, K.M., 2006. Bayesian multivariate Poisson regression for models of
injury count by severity. Transp. Res. Rec. 1950, 24–34.
Maher, Michael J., Summersgill, Ian., 1996. A comprehensive methodology for the fitting
of predictive accident models. Accid. Anal. Prev. 28 (3), pp281–296.
Malyshkina, Nataliya V., Fred, L., Mannering, 2010. Zero-state Markov switching count-
data models: an empirical assessment. Accid. Anal. Prev. 42, 122–130.
Meyer, Miller, 2001. Urban Transportation Planning: A Decision Oriented Approach s.l.
Prentice-Hall.
MTTO, 2009. Comprehensive Transportation Studies of Mashhad City, Update. Mashhad
Transportation and Traffic Organization (MTTO), Mashhad, Iran.
Naderan, A., Shahi, J., 2010a. Aggregate crash prediction models: introducing crash
generation concept. Accid. Anal. Prev. 42 (1), 339–346.
Naderan, Ali, Shahi, Jalil., 2010b. Crash generation models: forecasting crashes in urban
areas. Transp. Res. Rec.: J. Transp. Res. Board (2148), 101–106 Transportation
Research Board of the National Academies, Washington, D.C., 2010.
Noland, Robert B., 2002. Traffic Fatalities and Injuries: the Effect of Changes in
Infrastructure and Other Trends. s.l. Center for transport studies, Dept. of civil and
environmental engineering, Imperial college of science, technology and medicine.
Park, E.S., Park, J., Lomax, T.J., 2010. A fully Bayesian multivariate approach to before-
after safety evaluation. Accid. Anal. Prev. 42 (4), 1118–1127.
Pulugurtha, S.S., Duddu, V.R., Kotagiri, Y., 2013. Traffic analysis zone level crash esti-
mation models based on land use characteristics. Accid. Anal. Prev. 50, 678–687.
Qadeer Memon, A., 2006. Road Accident Prediction Models Developed From a National
Database: Poisson and Negative Binomial Regressions. Transportation Research
Board, TRB.
Quddus, M.A., 2008a. Modelling area-wide count outcomes with spatial correlation and
303
Accident Analysis and Prevention 120 (2018) 295–303
heterogeneity: an analysis of London crash data. Accid. Anal. Prev. 40 (4),
1486–1497.
Quddus, Mohammed A., 2008b. Modelling area-wide count outcomes with spatial cor-
relation and heterogeneity: an analysis of London crash data. Accid. Anal. Prev. 40
(2008), 1486–1497.
Tasic, Ivana, Porter, Richard J., 2016. Modeling spatial relationships between multimodal
transportation infrastructure and traffic safety outcomes in urban environments. Saf.
Sci. 82 (2016), 325–337 2016.
Ukkusuri, S., et al., 2012. The role of built environment on pedestrian crash frequency.
Saf. Sci. 50 (4), 1141–1151.
Vandenbulcke, Grégory, Thomas, Isabelle, Panis Luc Int, 2014. Predicting cycling acci-
dent risk in Brussels: a spatial case–control approach. Accid. Anal. Prev. 62, 341–357.
Wang, X., et al., 2012. Macrolevel model development for safety assessment of road
network structures. Transp. Res. Rec.: J. Transp. Res. Board 2280, 100–109.
Washington, Simon P., Karlaftis, Matthew G., Mannering, Fred L., 2003. Statistical and
Econometirc Method Methods for Transportation Data Analysis. s.l. Chapman & Hall/
CRC 2003.
Washington, S., et al., 2006. Incorporating Safety Into Long-range Transportation
Planning. s.l. : NCHRP Report 546. TRB, National Cooperative Highway Research
Program, Washington D.C.
Wier, Megan, et al., 2009. An area-level model of vehicle-pedestrian injury collisions with
implications for land use and transportation planning. Accid. Anal. Prev. 41 (1),
137–145.
Winkelman, R., 2003. Econometric Analysis of Count Data, fourth ed. Springer, New York
2003.
Xie, Y., Lord, D., Zhang, Y., 2007. Predicting motor vehicle collisions using Bayesian
neural networks: an empirical analysis. Accid. Anal. Prev. 39 (5), 922–933.
Yu, R.A., Abdel-Aty, M., Ahmed, M., 2013. Bayesian random effect models incorporating
real-time weather and traffic data to investigate mountainous freeway hazardous
factors. Accid. Anal. Prev. 50, 371–376.