Accident Analysis and Prevention 75 (2015) 164178

Contents lists available at ScienceDirect

Accident Analysis and Prevention

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

Effects on speed and safety of point-to-point speed enforcement

systems: Evaluation on the urban motorway A56 Tangenziale di Napoli
Alfonso Montella *, Lella Liana Imbriani, Vittorio Marzano, Filomena Mauriello
University of Naples Federico II, Department of Civil, Architectural and Environmental Engineering, Via Claudio 21, 80125 Naples, Italy

A R T I C L E I N F O

A B S T R A C T

Article history:
Received 11 August 2014
Received in revised form 27 November 2014
Accepted 28 November 2014
Available online 5 December 2014

In this paper, we evaluated the effects on speed and safety of the point-to-point (P2P) speed enforcement
system activated on the urban motorway A56 in Italy. The P2P speed enforcement is a relatively new
approach to trafc law enforcement that involves the calculation of the average speed over a section. To
evaluate the speed effects, we performed a beforeafter analysis of speed data investigating also effects
on non-compliance to speed limits. To evaluate the safety effects, we carried out an empirical Bayes
observational before-and-after study.
The P2P system led to very positive effects on both speed and safety. As far as the effects on the section
average travel speeds, the system yielded to a reduction in the mean speed, the 85th percentile speed, the
standard deviation of speed, and the proportion of drivers exceeding the speed limits, exceeding
the speed limits more than 10 km/h, and exceeding the speed limits more than 20 km/h. The best results
were the decrease of the speed variability and the reduction of the excessive speeding behaviour. The
decrease in the standard deviation of speed was 26% while the proportion of light and heavy vehicles
exceeding the speed limits more than 20 km/h was reduced respectively by 84 and 77%.
As far as the safety effects, the P2P system yielded to a 32% reduction in the total crashes, with a lower
95% condence limit of the estimate equal to 22%. The greatest crash reductions were in rainy weather
(57%), on wet pavement (51%), on curves (49%), for single vehicle crashes (44%), and for injury crashes
(37%). It is noteworthy that the system produced a statistically signicant reduction of 21% in total
crashes also in the part of the motorway where it was not activated, thus generating a signicant spillover
effect.
The investigation of the effects of the P2P system on speed and safety over time allowed to develop
crash modication functions where the relationship between crash modication factors and speed
parameters (mean speed, 85th percentile speed, and standard deviation of speed) was expressed by a
power function. Crash modication functions show that the effect of speed on safety is greater on curves
and for injury crashes.
Even though the study results show excellent outcomes, we must point out that the crash reduction
effects decreased over time and speed, speed variability, and non-compliance to speed limits signicantly
increased over time. To maintain its effectiveness over time, P2P speed enforcement must be actively
managed, i.e. constantly monitored and supported by appropriate sanctions.
2014 Elsevier Ltd. All rights reserved.

Keywords:
Highway safety
Speeding
Average speed enforcement
Empirical Bayes method
Crash modication factors
Crash modication functions
Safety performance functions

1. Introduction
Drivers speed inconsistent with the road environment is
among the most signicant crash contributing factors (Council
et al., 2010; Hauer, 2009; Montella and Imbriani, 2014; Montella
et al., 2010, 2011; Neuman et al., 2009; OECD, 2006; Yannis et al.,

* Corresponding author. Tel.: +39 817683941; fax: +39 817683946.

E-mail addresses: alfonso.montella@unina.it (A. Montella),
lellaliana.imbriani@unina.it (L.L. Imbriani), vittorio.marzano@unina.it (V. Marzano)
, lomena.mauriello@unina.it (F. Mauriello).
http://dx.doi.org/10.1016/j.aap.2014.11.022
0001-4575/ 2014 Elsevier Ltd. All rights reserved.

2013). Speeding is both driving faster than the posted speed limit
as well as driving too fast for the prevailing weather, light, trafc
and road conditions, but within the speed limits (Montella et al.,
2013; NHTSA, 2012). The relation between speed and safety rests
on two pillars: (1) the relationship between speed and the crash
risk and (2) the relationship between speed and the crash severity.
Higher speeds imply greater driving task difculty and therefore
greater crash risk. At higher speeds, the time to react to changes in
the environment is shorter, the stopping distance is larger, the
manoeuvrability is reduced, and it is more difcult to react in time
and prevent a crash. However, the greater effect of speed is on
the injury consequences of the crashes. The higher the collision

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

speed, the more severe the consequences in terms of injury and

material damage. This is because the energy dissipated in a crash
goes up with the square of collision speed. At a higher impact
speed, more energy is released when colliding with another
vehicle, road user or obstacle. Part of this energy is absorbed by the
vulnerable human body. Hence, higher speeds result in more
severe injury.
Despite understanding that speeding is a high-risk behaviour,
speeding is common and is seen as normal and socially acceptable
by many drivers (Fleiter et al., 2010). The culture of speeding is so
embedded that exceeding the posted speed limit is perceived as
normal. Indeed, there is evidence that many drivers regard
speeding as one of the least serious trafc offences (SARTRE,
2012). Thus, signicant resources are dedicated to reducing
speeding across the road network and new initiatives and
technologies are continually being developed and trialled in an
attempt to enhance speed compliance. These initiatives include
public education campaigns, speed limit reviews, police enforcement, and xed speed cameras. One issue of speed enforcement by
spot speed cameras is that some motorists brake before passing a
speed camera and then speed up to above the speed limit after they
have passed it (De Pauw et al., 2014). Thus, a speed reduction is
obtained only on a short section. This issue is overcome by the
point-to-point (P2P) speed enforcement, named also average
speed enforcement or section speed enforcement (Lynch et al.,
2011; Soole et al., 2012, 2013), which is a relatively new
technological approach to trafc law enforcement that has
increased in use in a number of highly motorized countries in
the last decade. Unlike traditional spotspeed enforcement, which
measures the speed of a vehicle at one point, point-to-point
enforcement involves the calculation of the average speed over a
section and encourages compliance over a greater distance. Pointto-point enforcement involves the installation of a series of
cameras at multiple locations along a road section. The average
speed is calculated by dividing the distance between two camera
sites by the time taken for the vehicle to travel between those two
sites. If the corresponding average speed of a vehicle exceeds the
posted speed limit for that road section, image and offence data are
transmitted to a central processing unit from the local processor
via a communication network. Indeed, with point-to-point
enforcement a sanction is imposed only for an average speed
exceeding the posted speed between the cameras, and not when
the motorist has driven too fast when passing point A or point B.
While there are capabilities for the back-ofce of a system to be
fully automated, almost all current installations involve some
degree of human verication to assess the validity of detected
infringements. Validated offences are subsequently issued with an
infringement notice and data on non-offending vehicles are
typically erased.
To date, there are some evaluations of the system with
encouraging positive ndings. However, the system is not yet
evaluated on a substantial scale (DaCoTA, 2012) and methodological limitations are noted across the majority of the published
evaluations (Soole et al., 2012, 2013). Thus, there is a strong need
for sound scientic evaluations. The P2P system was rst
introduced in the Netherlands where it operated in trial form in
1997 and then as a permanent installation in 2002. Currently, there
are 11 permanent point-to-point speed enforcement locations on
various motorways and rural roads. The number of casualties has
halved on these sections (Olde Karter et al., 2005) and most drivers
obey the speed limit. In the UK, P2P speed enforcement has
increased considerably since its inception via a trial in 1999 in Kent
and is now widely used. The overall result of average speed control
is a conveyor belt type ow, with uniform speeds, little braking and
larger headways (Collins and McConnell, 2008). P2P speed cameras
established on corridors with a history of high crash rates have led

165

to reductions in fatal and serious injury crashes. After the

installation of P2P devices along a section of the A77 in southwest
Scotland in 2005, a 19% reduction in all crashes was observed, with
fatal crashes falling by 46% and serious injury crashes by 37% (Soole
et al., 2012). In Australasia, P2P speed enforcement was introduced
in 2007 in the State of Victoria. Currently, New South Wales has a
total of 21 road segments enforced by P2P speed camera systems
and a trial is being conducted in New Zealand (Soole et al., 2012,
2013). In Italy, the P2P speed enforcement system was introduced
in 2006 and includes a total of 320 P2P speed camera sites which
cover more than 2900 km of the motorway network. In 2012, the
system has also been applied on three national expressways.
Further installations of the system are planned also on regional and
provincial highways. A recent beforeafter study showed that the
vehicle speeds, and subsequently trafc ows, were sensibly
homogenized (Cascetta et al., 2011). To evaluate the safety
effectiveness of the system in an 80 km segment of the Motorway
A1, an empirical Bayes observational beforeafter study was
performed (Montella et al., 2012). The estimate of the total crash
reduction was 31.2%, with a lower 95% condence limit of 24.3%.
The safety effectiveness decreased over time. The crash reduction
was 39.4% in the rst semester after the system activation while it
was 18.7% in the fth semester. Speed data before and after the
system installation were not available.
To obtain a more comprehensive evaluation of the system, this
study investigates both effects on speed as well as effects on
crashes after the P2P system installation on the urban motorway
A56 Tangenziale di Napoli in Italy. The remainder of the paper is
organized as follows. Section 2 describes the P2P system operation,
the geometric data, the trafc data, the speed data, and the crash
data. Section 3 describes the methods used to perform the speed
analysis and the safety evaluations. Section 4 presents effects on
both speed and safety as well as crash modication functions
which take into account the effects of speed on safety, followed by
a discussion that places the results in the context of the highway
engineering practice. The last section presents the conclusions.
2. Study data
2.1. Treatment site data
2.1.1. Point-to-point speed enforcement system
The system is composed of steel gantries at the section entrance
and exit, with one high resolution camera (1600 pixels
1200
pixels) with infrared ash for each lane, mounted on the gantry.
Due to the privacy legislation in Italy, only rearward facing cameras
are used, given that face obscuration is not considered sufcient for
the protection of personal information. Other components include
inductive loop detectors for each lane placed under the road
pavement in two cross-sections slightly downstream from the
cameras, an optical bre network to transfer the data, a central
monitoring and data processing station managed by the police,
and an automatic vehicle identication (AVI) system for
video-based vehicle license plate recognition. Moreover, the
system is equipped with a global position system which allows the
time-synchronization of all the detection points.
Whenever a vehicle crosses over the inductive loop detectors,
the variation of the electromagnetic eld allows the detection of
the vehicle and, simultaneously, the activation of the lane-related
camera. Once activated, the camera records the date and time of
activation and acquires 100 frames, with a frame rate of 25 fps,
which are post-processed by the AVI software for vehicle plate
recognition. When the same vehicle crosses the section exit gantry,
the same operation is performed. As a result, vehicles are classied
in six classes (1 car, moped, caravan; 2 light vehicle with
a trailer; 3 heavy duty vehicle with weight in the range

166

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

3.512.0 ton; 4 articulated vehicle and multi-trailer vehicle with

weight greater than 12 ton; 5 bus; 6 heavy scooter or
motorcycle) and the individual average travel speeds of all the
vehicles crossing both the entrance and the exit cross-sections are
calculated. Data for vehicles exceeding the posted speed limit, plus
a 5 km/h enforcement tolerance, are sent to the back-ofce
managed by the police. Transmitted data are encrypted to avoid
unauthorised access to the information and then digitally signed
with a unique signature certicate by every PC to avoid
unauthorised manipulation. In addition, all communication is
conducted though secure protocols. Manual checks are performed
on all images of offending vehicles at the back-ofce. These checks
assess the accuracy of the number plate recognition process, as
well as the class of the vehicle. Once veried, an automatic process
identies the address of the offending vehicle and produces an
infringement notice. The only manual step in this process involves
a signature from a police ofcer, which is a legislative requirement.
Vehicles with non-readable plates are out of law and subject to
high fee by the road police. For privacy reasons, images of the
offending vehicle are not included on the infringement notice,
unless explicitly requested by the offender.
The P2P system has very strict accuracy requirements dened
by the Italian Ministry of Transports and Infrastructures and a
homologation process must be conducted prior to the installation.
The factors used to validate the performance of the system include:
the accuracy of the vehicle detection under real trafc conditions
and at various speeds (capture at speeds of at least 260 km/h are
tested in circuit); the accuracy of the speed measurement (speed
data are stated with two decimal points and the calibration
showed an average speed deviation of 0.40%); the accuracy of
vehicle classication under real trafc conditions; the accuracy of
number plate recognition and reading under various light and
weather conditions; and the accuracy of number plate matching
from the entry to exit gantries. In addition, after this process, the
full system operated in test mode, in cooperation with the police,
to certify the correct operation of the system.
In Italy, owner responsibility is associated with cameradetected offences. That is, it is the responsibility of the vehicle
owner to declare the name of the driver if they claim not to have
been in operation of the vehicle at the time of the offence. If
another driver is not declared, the owner is administered with the
subsequent penalties associated with the offence. The issuing of
multiple infringements for the same stretch of road is forbidden by
law in the country. Thus, where multiple infringements are
detected on successive sections of the motorway, only one offence
is pursued. While the system is capable of measuring instantaneous speeds, this function is not employed to minimise the extent
of sudden braking associated with identifying an instantaneous
speed camera. Up to a speeding level of 10 km/h (plus a 5 km/h

[(Fig._1)TD$IG]

enforcement tolerance), offenders get a ne between 41 and

168 Euros. At speeding level between 10 and 40 km/h, the ne is
between 168 and 674 Euros and offenders get three penalty points
on the driver licence. At speeding level between 40 and 60 km/h,
offenders get a driving ban between one and three months and a
ne between 527 and 2108 Euros. At speeding level greater 60 km/
h, offenders get a driving ban between six and twelve months and a
ne between 821 and 3287 Euros.
Before the activation of the P2P system, spot-speed enforcement
was applied. The P2P speed enforcement system was activated on
February 9, 2009 and covered four sections in east carriageway (E1 to
E4) for a total length of 13.0 km and three sections in west
carriageway (W1 to W3) for a total length of 5.1 km. Speed limit
was 80 km/h for light vehicles (weight  3.5 ton) and 70 km/h for
heavy vehicles (weight > 3.5 ton).
2.1.2. Geometric data
The motorway A56 Tangenziale di Napoli is a divided highway
with three lanes in each direction (lane width of 3.75 m), access
control, and interchanges. It is an urban motorway that represents
the ring road of the city of Naples, in southern Italy. The total length
of the motorway is 40.4 km (20.2 km per carriageway). Geometric
characteristics and trafc ow volumes were used to divide the
motorway into 82 homogeneous segments in east carriageway and
85 homogeneous segments in west carriageway. Lengths of
segments vary between 40 and 1387 m. The radii of horizontal
curves vary between 200 and 3000 m. Spiral transitions are not
present. The deection angle varies between 2.20 and 81.10 gon.
The radius of vertical curves varies between 3000 m (crest curve)
and 16,000 m (sag curve). The maximum longitudinal grade is
5.20%.
2.1.3. Trafc data
In the motorway A56, exit points are controlled by toll stations
but entrance is free and entering vehicles are not monitored.
Thus, a specic procedure to estimate trafc volumes along the
motorway was dened. The starting point was represented by a
detailed transport decision support system (DSS) for the entire
metropolitan area of Naples, already developed in previous
studies (Cascetta et al., 2013), able to reproduce four recurrent
system conditions, i.e. both AM/PM and peak/off-peak working
days. Using the partial information given by the available trafc
counts (i.e. counts of the P2P system in some mainstream sections
and total off-ramp ows), an updating of the od matrix of the DSS
was performed for each of the four aforementioned system
conditions, through a generalised least squares-based approach
(Marzano et al., 2009; Cascetta et al., 2013). This allowed obtaining
the most likely od matrix able to t at best the available trafc
counts. The assignment of this matrix to the entire network

Fig. 1. Architecture of the transport decision support system.

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

through the DSS allowed in turn calculating also the most

consistent trafc ows on the unmonitored sections of the A56. In
order to validate this approach, manual trafc counts were carried
out for some of those unmonitored sections (i.e. a hold-out
sample), leading to very satisfactory result. Once calculated the
trafc ows in each section for each of the four time intervals of
the day, an hourly disaggregation within the modelled periods and
an hourly extrapolation for the non-modelled period (e.g.
weekends, night) was performed using specic conversion factors
based on the aforementioned observed trafc counts. This nally
allowed aggregation of such hourly trafc counts in new intervals
within the day, so as to calculate the average annual daily trafcs
in daytime (7:00 AM9:00 PM) and night-time (9:00 PM
7:00 AM). The overall procedure is schematically represented in
Fig. 1.
In the east carriageway, trafc ow increases from section E1 to
section E4. In daytime (7:00 AM9:00 PM), average trafc volume
is about 850 vehicles/lane/hour in section E1 (LOS = B), 1000
vehicles/lane/hour in section E2 (LOS = C), 1300 vehicles/lane/hour
in section E3 (LOS = D), and 1800 vehicles/lane/hour in section E4
(LOS = E). In the west carriageway, trafc ow decreases from
section W1 to section W3. In daytime (7:00 AM9:00 PM), average
trafc volume is about 1500 vehicles/lane/hour in section W1
(LOS = D), 1400 vehicles/lane/hour in section W2 (LOS = D), and
1000 vehicles/lane/hour in section W3 (LOS = C). In night time
(9:00 PM7:00 AM), trafc volumes are less than 300 vehicles/
lane/hour in all the sections.
2.1.4. Crash data
Since the Italian national crash database maintained by the
National Institute of Statistics presents major issues related to the
crash report form, the crash classication, the crash location, and
the crash severity, a new database was developed according to a
framework based on the critical review of the crash databases in
Australasia, the European Union and the United States (Montella
et al., 2013).
Crash data were collected through analysis of police reports and
were integrated with detailed site inspections. Crash data covered
2006 to 2011, with a before period of 3.08 years and an after period
of 2.91 years. Crash count for all treatment sites was 559 in the
before period and 279 in the after period. Crashes at the
interchanges ramps, at rest areas, and at tollbooths were excluded
based on the location descriptions of the police reports.
Crash data were classied according to the alignment (tangent
and curve), collision type (single-vehicle, multi-vehicle rear-end,
multi-vehicle others), crash severity (property damage only,
injuries including fatal injuries), weather conditions (non-rainy
and rainy), pavement conditions (dry and wet), time of the day
(daytime, 7:00 AM9:00 PM, and night-time, 9:00 PM7:00 AM),
and day type (working day and weekend).
2.2. Reference group data
To estimate what the safety experience would have been in the
after period without the installation of the P2P speed enforcement system and to account for the temporal effects on safety of
variation in weather, demography, crash reporting, and so on, the
section between Naples and Candela of the Motorway A16,
located in Campania Region, was used as reference group. In this
motorway the P2P system was not implemented. The motorway
A16 is a divided highway with two lanes in each direction (lane
width of 3.75 m, right shoulder width in the range between 0.50
and 3.50 m, and median width of 2.00 m), access control, and
interchanges. The section between Naples and Candela, with a
total length of 255 km (127.5 km per carriageway), was used.
Geometric characteristics and trafc ow volumes were used to

167

divide the section into 652 homogeneous segments. Lengths of

segments vary between 62 and 3509 m. The radii of horizontal
curves vary between 245 and 4000 m. Spiral transitions are not
present. The deection angle varies between 4.88 and 109.07 gon.
Trafc data were provided from the motorway management
agency. The annual average daily trafc (AADT) in the period from
2006 to 2011 ranged between 7885 and 18,370 vehicles per day.
Crash data were collected through analysis of police reports and
were integrated with detailed site inspections. The crash database
was developed according to the same framework used for the
treatment site. Crash count was 1873 in the before period and
1445 in the after period.
The criterion for the choice of the reference group is that the
change from the before to after period in the safety of the
reference group is indicative of how safety in the treatment group
would have changed without the treatment (Hauer, 1997).
Specically, this criterion is based on two assumptions: (1) the
factors that affect safety have changed from the before to after
period in the same manner on both the treatment and the
reference group; (2) this change in the factors inuences the
safety of the treatment and the reference group in the same way.
In the absence of treatment, the ratio of the expected number of
crashes before and after would be the same in the treatment and
reference groups. To test this hypothesis, crash data in the before
period were divided into two time series covering years from
2006 to 2008. Each time series is made up of two periods: (1) the
rst period represents the base year (from 2006 to 2007), while
(2) the second period represents the year after (from 2007 to
2008). For each time series, an estimate of the odds ratio (O) was
assessed as follows:
O

K
N=L
M
1 1=L 1=M

(1)

where K is the crash count in the rst period in treatment group, L

is the crash count in the second period in treatment group, M is the
crash count in the rst period in reference group, and N is the crash
count in the second period in reference group.
The average value of the odds ratio was 1.02, close enough to
1 to indicate that the comparison group is suitable for describing
the crash trend when no treatment has been applied to the
treatment group.
3. Methodology
3.1. Speed data analysis
Average travel speeds over the seven study sections in the
motorway A56 were analysed. Speed monitoring was carried out in
four periods: (1) before_2009, twelve days before the implementation of the P2P system (from 28th January to 8th February); (2)
after_2009, seventy-seven days after the implementation of the
P2P system (from 18th February to 5th May); (3) after_2010,
twenty-one days in 2010 (from 12th May to 1st June); and (4)
after_2011, twenty-tree days in 2011 (from 29th March to 20th
April). The total number of observations was 22,556,404. Since
different speed limits were applied to light vehicles (weight  3.5
ton) and heavy vehicles (weight > 3.5 ton), speed data were
disaggregated in relation to the vehicle type. Furthermore, to
investigate the effect of the system in different trafc conditions,
daytime (7:00 AM9:00 PM) and night-time (9:00 PM7:00 AM)
were investigated separately.
Local media advertised the installation of the system only the
day prior to the starting of the enforcement. Therefore, at the time
of the data collection before the system activation (before_2009)
drivers were substantially unaware of the system itself. Moreover,
drivers were already familiar with the P2P gantries, being them

168

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

already installed along the motorway as variable message signs. In

other words, P2P installation works have been very unobtrusive
and basically not detectable by the drivers. The combined effect of
no prior information by the media and the absence of considerable
implementation work ensures that the data collected in the before
scenario are substantially unbiased and thus valuable for
the purposes of the paper.
Speed data were pre-processed testing the normality and
homoscedasticity assumptions. Since different tests of normality
often produce different results (Razali and Wah, 2011), we
veried the normality assumption using the tests Anderson
Darling, JarqueBera, KolmogorovSmirnov, Lilliefors, and ShapiroWilk. To support the formal normality tests, we performed
normal 2015quantile plots (QQ plots). The QQ plot is a plot of
the percentiles of a standard normal distribution against the
corresponding percentiles of the observed data. It shows the
observations on the X axis plotted against the expected normal
score (Z-score) on the Y axis. If the data are normally distributed,
the data points are close to the diagonal line. If the data points
stray from the line in an obvious non-linear fashion, the data are
not normally distributed. All the normality tests showed that
speed data do not t the Gaussian distribution (p < 0.001) and this
result was conrmed by the QQ plots (e.g. Fig. 2). The
homoscedasticity assumption was veried using the Levenes
test, which showed that the homoscedasticity assumption is not
satised (p < 0.001).
Given the non-normality and heteroscedasticity of the speed
data, the speed distributions in the different observation periods
were compared by non-parametric tests. The KruskalWallis
one-way analysis of variance by ranks was performed to test the
signicance of speed changes in the different observation
periods. The KruskalWallis test is a method of testing the
hypothesis that several populations have the same continuous
distribution (Washington et al., 2010). Since the null hypothesis
was rejected and the KruskalWallis test does not identify where
the differences occur or how many differences occur, the two
sample KolmogorovSmirnov and MannWhitney tests were
performed to analyse the specic sample pairs for signicant
differences. The KolmogorovSmirnov test was used to

determine if the samples come from identical distributions

while the MannWhitney test was used to study the relative
positions of the samples. Furthermore, we compared the speed
variances before and after the P2P implementation using the
Levenes test.
Finally, we examined whether the percentage of non-compliance to the speed limits changed after the implementation of the
P2P speed enforcement system, and if this percentage remained
constant over time. At this aim, the Z-test of proportions was
carried out.
3.2. Safety evaluation
In this study, the empirical Bayes (EB) methodology, which
represents the state-of-the-art approach for safety evaluations
(AASHTO, 2010; Gross et al., 2010; Hauer et al., 2002; Montella,
2009; Montella et al., 2012; Persaud and Lyon, 2007), was used.
This methodology is rigorous and properly accounts for regressionto-the-mean, accounts for other changes over time not due to the
treatment being evaluated, overcomes the difculties of using
crash rates in normalizing for trafc volume differences between
the before and after periods, and reduces the level of uncertainty in
the estimates of the safety effect.
In the EB approach, the change in safety for a given crash type at
a site is given by the difference between the expected number of
crashes that would have occurred in the after period without
treatment and the number of reported crashes in the after period
and is assessed as follows:

lp

(2)

where l is the expected number of crashes that would have

occurred in the after period without treatment and p is the number
of reported crashes in the after period.
In estimating l, the effects of regression-to-the-mean
and changes in trafc volume and other factors are explicitly
accounted for using safety performance functions (SPFs), based on
untreated reference sites, relating crashes of different types to
trafc volume and highway features such as horizontal and vertical
alignment. Annual SPF multipliers are calibrated to account for the

[(Fig._2)TD$IG]

Fig. 2. Example of QQ plot, light vehicles in the period after_2010.

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

temporal effects on safety of variation in weather, demography,

crash reporting, and so on.
In the EB procedure, the SPF is used to rst estimate the
number of crashes that would be expected in each year of the
before period at locations with trafc volumes and other
characteristics similar to the one being analysed. The crash
estimates are then combined with the count of crashes in the
before period at the treatment site to obtain a better estimate of
the expected number of crashes before the treatment, produced
as follows:
! 

^
1
EY
b
EBb
(3)


^
k countb
1=k EY
b
where k is the dispersion parameter and is estimated from the SPF
calibration process with the use of a maximum likelihood
^ (Y)b is
procedure, countb is the observed crash frequency, and E
the predicted crash frequency from the SPF.
A factor is then applied to EBb to account for the differences in
duration, trafc volumes and crash trends between the before and
after periods. This factor (R) is the sum of the annual SPF
predictions for the after period divided by the sum of these
predictions for the before period. The result, after applying this
factor, is an estimate of l, the expected number of that would have
occurred in the after period without treatment:
"
#
^
EY
a
l ^
(4)

EBb R
EBb
EY
b

^ (Y)a is the SPF predicted crash frequency in the after

where E
period.
The procedure also produces an estimate of the variance of l,
the expected number of crashes that would have occurred in the
after period without treatment:
"
#2
^
EY
2
b
Varl R


k countb
(5)
^
k EY
b
Disaggregation of total crashes into complementary crash types
gives rise to EB crash estimates with different weights assigned to
the predicted crashes and the crash counts, which, in turn, produce
estimates of complementary crash types whose sum differs from
the estimate of total crashes. As suggested by Hauer et al. (2002),
this discrepancy is corrected by multiplying each disaggregated EB
crash estimate by the ratio (S) of the EB estimate for total crashes to
the sum of the EB estimates for the complementary crash types.
Thus, for each disaggregated crash type, the estimate of l is
multiplied by the factor S and the estimate of the variance of l is
multiplied by S2. The estimate of l is then summed over all sites in
the treatment group of interest (to obtain lsum) and compared
with the count of crashes during the after period in that group
(psum). The variance of l is also summed over all sites in the
treatment group.
The index of effectiveness (u) is estimated as:

psum =lsum
1 S Varl=lsum
2

The standard deviation of u is given by:

v
i
u 2 h
uu
Varp =p2 P Varl=l2
sum
u
sum
sum
St:Devu u
h
i2
t
P
1 Varl=l2sum
The percent reduction in crashes is 100
(1  u).

(6)

(7)

169

4. Safety performance functions

4.1. Model description
Following common practice (Mannering and Bhat, 2014),
generalized linear modelling techniques were used to t the
models, and a negative binomial distribution error structure was
assumed. The selected model form is as follows:
n
X

^
EY
L
ea0 ai
lnAADT
e i1

bi
xi

yri

(8)

^
where EY
is the predicted annual crash frequency, L is the
segment length (m), AADT is the segment average annual daily
trafc (veh/day), ai and bi are the model parameters, xi are the
explanatory variables other than AADT, and yri are the yearly
factors. This model form logically estimates zero crashes if one of
the two exposure variables (AADT or L) is equal to zero. Segment
length is an offset variable. It captures the logical requirement that
if N crashes are expected to occur on 1 km of road, 2N crashes
should be expected to occur on an identical road that is 2 km long.
Explanatory variables related to trafc volume (average annual
daily trafc), horizontal alignment (square of horizontal curvature, length of the tangent preceding the curve, horizontal
curvature of the curve preceding the tangent, and horizontal
curvature of the curve following the tangent), design consistency
(operating speed consistency, consistency in driving dynamics,
and inertial speed consistency), vertical alignment (equivalent
downgrade and equivalent upgrade), roadside context (tunnel
presence), and cross-section (right shoulder width) were
considered (Table 1). The design consistency variables were
assessed using the operating speed predictive models developed
in a previous research (Montella et al., 2014) through an
instrumented vehicle equipped with a GPS continuous speed
tracking from a eld experiment conducted on the same
motorway where the safety performance functions were tted,
i.e. the motorway A16 in Italy.
The model parameters and the dispersion parameter of the
negative binomial distribution were estimated by the maximum
likelihood method using the GENMOD procedure in SAS. First, to
determine which explanatory variables (other than AADT)
signicantly affect crash frequency alone, simple models were
estimated with one independent variable in each. Non-signicant
variables were excluded from further investigation. Second, the
models were developed by the stepwise forward procedure,
adding one explanatory variable at each step. The decision on
whether or not to keep a variable in the model was based on two
criteria. The rst is whether the t-ratio of the variables estimated
coefcient is signicant at the 5% level. The second criterion is
based on the improvement of the goodness-of-t measures of the
model that includes that variable.
4.2. Goodness-of-t measures
Two goodness-of-t measures were used: R2a (Miaou et al.,
1996) and the Akaike information criterion (AIC) (Akaike, 1987).
R2a , a dispersion parameter-based R2, is calculated as follows:


k
(9)
R2a 1 
kmax
where kmax is the largest possible dispersion parameter that is
obtained by having no covariates in the model (by assuming that all
sites have an identical prediction estimate equal to the mean over
all sites) and k is the dispersion parameter for the calibrated model.
This measure is bound between 0 (when no covariate is included)
and 1 (when covariates are perfectly specied).

170

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

Table 1
SPFs: signicant explanatory variables.
Variable

Description

Trafc volume
ln(AADT) [veh/
day]

Natural logarithm of the average annual daily trafc

Horizontal alignment
1/R2 [1/km2]
Square of horizontal curvature
Lt [km]
Length of the tangent preceding the curve
Horizontal curvature of the curve preceding the tangent
1/R1 [1/km]
1/R+1 [1/km]
Horizontal curvature of the curve following the tangent
Design consistency
Absolute value of the operating speed difference in the tangent-to-curve transition = V85tV85c
7:483
 1:290
Gu  0:080
CCR2  14:427
tunnel  4:083
bridge
V 85c 135:490 
R

DV85 [km/h]

V 85t 139:543 1:751
Lt 

Dfr []

V85,iV85,5 [km/
h]

4:983 2:270

 2:507
Gu  0:068
CCR2
R1
R1

where R is the radius of the curve (Km), R1 is the radius of the curve preceding the tangent (Km), R+1 is the radius of the curve following the tangent
(Km), CCR2 is the curvature change ratio of the 2 km preceding the geometric element (gon/km), Lt is the length of the tangent (km), Gu is the
equivalent upgrade (%), bridge is a binary variable equal to 1 if the segment is on a bridge, and tunnel is a binary variable equal to 1 if the segment is on
a tunnel
Difference between side friction assumed with respect to the design speed (fra) and side friction demanded at the 85th-percentile speed (frd)
V2
Df r f ra  f rd 0:6
0:925
0:59  4:85
103
V d 1:51
105
V 2d  85  e
127
R
where Vd is the design speed (km/h) calculated according to the Italian Geometric Design Standards (Italian Ministry of Infrastructures and Transports,
2001), V85 is the 85th percentile of the speed distribution in the curve (km/h), R is the curve radius (m), and e is the curve superelevation
Difference between the operating speed in the curve and the average operating speed along the 5 km preceding the beginning of the curve (weighted to
element length)

Horizontal alignment
1 /R2 [1 /km2]
Square of horizontal curvature
Lt [km]
Length of the tangent preceding the curve
1/R1 [1/km]
Horizontal curvature of the curve preceding the tangent
1/R+1 [1/km]
Horizontal curvature of the curve following the tangent
Vertical alignment
Gd
Equivalent downgrade, obtained by weighing each gradient in relation to the segment length
Gu
Equivalent upgrade, obtained by weighing each gradient in relation to the segment length
Roadside context
Tunnel
Binary variable, equal to 1 if the segment is on a tunnel
Cross-section
RSW [km]

Right shoulder width

The AIC value is calculated as follows:

AIC 2
ML 2
p

(10)

where ML is the maximum log-likelihood of the tted model, and p

is the number of parameters in the model. A smaller AIC value
reects a better model. The rst term in the AIC equation measures
the goodness of t, or bias, when the maximum likelihood
estimates of the parameters are used. The second term measures
the complexity of the model, thus penalizing the model for using
more parameters. The minimisation of the AIC value allows to
reduce over tting, since it permits to choose the best t with the
least complexity.
4.3. Modelling results
4.3.1. Study site
Table 2 presents parameter estimates and goodness-of-t
measures for each crash type. Non-signicant parameters are not
reported. The values of the models range between 31 and 84%. All
the parameters have logical and expected signs. The square
curvature (1/R)2 has a positive sign, that is, the smaller the radius,

the more the expected crash frequency. Furthermore, the

estimated coefcient of the parameter length of the tangent
preceding the curve (Lt) is positive, showing that neighbouring
curves have few predicted crashes than curves more distant to each
other, as reported in the literature (Brenac, 1996; Findley et al.,
2012; Montella et al., 2008). For vertical alignment, the downgrade
parameter (Gd) has a positive sign and is associated with increase
in crash frequency, while the opposite is for the upgrade parameter
(Gu). The variable tunnel is associated with increase in crash
frequency.
4.3.2. Reference site
Table 3 presents parameter estimates, yearly factors and
goodness-of-t measures for each crash type. Non-signicant
parameters are not reported. The R2a values of the models range
between 43 and 93%. All the parameters have logical and expected
signs. The estimated coefcients of the variable related to
curvature and grade have a positive sign. The negative sign of
the estimated coefcient for the deection angle shows that curves
with a larger deection angle (for a given radius) are easier for the
driver to detect in advance and thereby allow safe negotiation of

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

171

Table 2
SPFs in the study site (A56): parameter estimates and goodness of t measures.
Crash type

Dispersion

Constant

ln(AADT)

1 /R2

Lt

Gd

Gu

Tunnel

R2a

AIC

Total
Tangent
Curve
Single-vehicle
MV rear End
MV Others
PDO
Injury
Non-rainy
Rainy
Dry
Wet

0.348 (0.062)
0.202 (0.089)
0.349 (0.079)
0.520 (0.123)
0.469 (0.107)
0.133 (0.117)
0.491 (0.118)
0.290 (0.072)
0.319 (0.062)
2.150 (0.498)
0.285 (0.059)
2.460 (0.452)

16.071 (1.446)
14.405 (1.976)
16.137 (1.846)
10.024 (1.682)
26.677 (2.022)
15.879 (1.830)
15.717 (1.647)
18.715 (1.617)
18.144 (1.362)
12.890 (2.400)
17.020 (1.444)
11.361 (2.373)

0.990 (0.134)
0.830 (0.182)
1.038 (0.169)
0.357 (0.153)
1.867 (0.186)
0.857 (0.166)
0.894 (0.149)
1.184 (0.149)
1.185 (0.124)
0.472 (0.219)
1.065 (0.133)
0.359 (0.216)

0.050 (0.008)

0.024 (0.009)
0.054 (0.009)
0.044 (0.010)
0.052 (0.009)
0.045 (0.009)
0.052 (0.008)
0.041 (0.007)
0.081 (0.013)
0.045 (0.007)
0.091 (0.013)

0.465 (0.174)

0.668 (0.212)

0.480 (0.193)
0.470 (0.175)

0.441 (0.172)

6.928 (2.938)
10.773 (3.581)

11.837 (3.714)
7.072 (3.235)

7.369 (3.145)

8.124 (2.877)

0.398 (0.120)

0.548 (0.159)

0.389 (0.148)

0.410 (0.128)
0.326 (0.117)

0.439 (0.117)

0.66
0.84
0.54
0.52
0.75
0.76
0.58
0.71
0.67
0.36
0.71
0.31

1801
763
1017
1135
1191
856
1162
1467
1742
444
1705
548

0.379 (0.075)

19.106 (1.608)

1.253 (0.150)

0.045 (0.008)

0.468 (0.194)

6.638 (3.204)

0.409 (0.130)

0.65

1591

0.635 (0.147)

10.052 (1.431)

0.421 (0.162)

0.060 (0.009)

0.53

956

0.390 (0.079)
0.286 (0.111)

18.274 (1.632)
14.072 (1.772)

1.158 (0.151)
0.720 (0.162)

0.048 (0.008)
0.050 (0.009)

0.570 (0.192)

7.400 (3.316)

0.386 (0.132)
0.361 (0.153)

0.65
0.71

1568
1001

Daytime
7:00AM9:00PM
Night-time
9:00PM7:00AM
Working day
Weekend

Note: MV = multi-vehicle; = variable not signicant at 5% level; standard errors of the parameter estimates are reported in parentheses.

the entry manoeuvre. The signs of the design consistency

parameters show a signicant increase in crash frequency
associated to design inconsistencies. Yearly factors show a
decreasing trend in crash frequency.
5. Results and discussion
5.1. Speed effects
Average speed of light vehicles (weight  3.5 ton) decreased
from 83.4 to 75.2 km/h, i.e. a 10% reduction (Fig. 3 and Table 4). A
greater reduction was observed for the 85th percentile speed (V85)
which decreased from 100.0 to 85.9 km/h, i.e. a 14% reduction.
Night-time speeds were higher than daytime speeds of about
3 km/h but speed reduction was similar. This result is noteworthy
because it shows that the effect of the P2P system is very relevant
not only in free-ow conditions but also during periods with
limitations to freedom to manoeuvre. In fact, in night-time trafc
volumes are less than 300 vehicles/lane/hour over all the sections
whereas in daytime average trafc volumes are up to
1800 vehicles/lane/hour. Nonetheless, average speed reduction
was greater in daytime than in night-time: 9.9 vs. 9.2%.
Effectiveness for heavy vehicles (weight > 3.5 ton) was lower than
for light vehicles: 5 vs. 10% the reduction in the mean speed, and
8 vs. 14% the reduction in the 85th speed.
One of the most important results of the P2P system is an
impressive reduction in the speed variability. Standard deviation of
average speeds of light vehicles over the study sections decreased
from 16.5 to 12.2 km/h (from 13.1 to 10.5 km/h for heavy vehicles),
i.e. a 26% reduction (20% for heavy vehicles). The greater reduction
in speed standard deviation was observed in night-time: 31%.
From before to after the installation of the system, the proportion
of light vehicles exceeding the speed limit (80 km/h) decreased from
56 to 31% (45% decrease) (Table 5). Consistently with previous
studies (Soole et al., 2013), the system was more effective in reducing
excessive speeding behaviour. Indeed, the reduction was from 31 to
9% (72% decrease) for light vehicles exceeding the speed limit more
than 10 km/h and from 14 to 4% (84% decrease) for light vehicles
exceeding the speed limit more than 20 km/h. Decrease in
non-compliance was higher in daytime than in night-time: 48 vs.
39% for the proportion of cars exceeding the speed limit, 76 vs. 66% for
the proportion of cars exceeding the speed limit more than 10 km/h,
and 88 vs. 79% for the proportion of cars exceeding the speed limit

more than 20 km/h. Non-compliance to speed limits of trucks

(70 km/h) was reduced from 58 to 46% (only 16% decrease) even if the
proportion of trucks exceeding the speed limit more than 20 km/h
was reduced from 9 to 2% (77% decrease).
Even though the study results show excellent outcomes, there is
one issue which should not be overlooked. Speed, speed variability
and non-compliance to speed limits signicantly increased over
time. The comparison between the period immediately after the
P2P implementation (after_2009) and the period two years later
(after_2011) shows: (a) increase of 3.1 km/h in the mean speed of
light vehicles; (b) increase of 3.9 km/h in the 85th speed of light
vehicles; (c) increase of 0.7 km/h in the standard deviation of speed
of light vehicles; and (d) increase of non-compliance of light
vehicles to speed limits from 19.4 to 31.1%. Changes over time were
smaller for heavy vehicles.
All results are statistically signicant with a p-value less than
0.001.
5.2. Safety effects in the treatment sites
Table 6 provides the estimates of the index of effectiveness and
the limits of the 95% condence interval of the estimate of the
crash reduction after the installation the P2P speed enforcement
system in the seven sections (length of 18.1 km) where the system
was activated on 9th February 2009. Results are disaggregated
in relation the alignment (tangent and curve), collision type
(single-vehicle, multi-vehicle rear-end, multi-vehicle others),
crash severity (property damage only, injuries including fatal
injuries), weather conditions (non-rainy and rainy), pavement
conditions (dry and wet), time of the day (daytime, 7:00 AM
9:00 PM, and night-time, 9:00 PM7:00 AM), day type (working
day and weekend), and temporal effects (2009, 2010, and 2011).
Results emphasized in boldface indicate statistically signicant
crash reductions (i.e. u < 1) at the 2.5% level.
The crash reduction estimate for the total crashes is 32.0%, with
a lower 95% condence limit of 22.3%.
Crash reduction in the tangents was not statistically signicant
whereas a signicant crash reduction equal to 48.8% was observed
for curves. A greater effectiveness of the P2P system in the curves
(43.4 vs. 28.4% crash reduction) was found also in a previous study
carried out on 80 km of the Motorway A1 (Montella et al., 2012).
The greater safety effect of speed enforcement on curves is
consistent both with the laws of physics, since vehicle stability

172

Table 3
SPFs in the reference site (A16): parameter estimates, yearly factors and goodness of t measures.
Crash type

Dispersion

Constant

ln(AADT)

1 /R2

Def

Gd

Gu

DV85

Dfr

V85,iV85,5

1/R1

1/R+1

RSW

Tunnel Yr09

Total crashes

0.859
(0.054)
0.460
(0.056)
1.034
(0.087)
1.019
(0.067)
0.177
(0.099)
0.839
(0.160)
0.830
(0.059)
0.992
(0.122)
0.667
(0.053)
3.510
(0.253)
0.716
(0.060)
2.580
(0.167)

16.032
(0.835)
17.860
(1.062)
14.604
(1.221)
14.038
(0.893)
28.166
(1.238)
24.956
(1.209)
14.069
(0.860)
24.503
(1.106)
15.581
(0.822)
20.074
(1.259)
15.397
(0.853)
19.490
(1.159)

0.996
(0.089)
1.176
(0.112)
0.840
(0.130)
0.762
(0.096)
1.930
(0.131)
1.715
(0.128)
0.764
(0.092)
1.735
(0.117)
0.936
(0.088)
1.160
(0.134)
0.910
(0.091)
1.189
(0.124)

0.022
(0.011)
0.108
(0.010)
0.118
(0.009)
0.088
(0.009)
0.069
(0.011)
0.095
(0.011)
0.137
(0.007)
0.070
(0.008)
0.147
(0.009)

0.080
(0.017)

0.053
(0.005)

0.95 0.76 0.75 0.46 9121

0.010
(0.002)

1.00 0.79 0.74

0.026
(0.008)
0.030
(0.006)
0.033
(0.009)

0.099
(0.038)

0.234
(0.037)

0.98 0.79 0.82 0.40 4582

2.177
(1.148)

1.08

0.328 0.903

0.82 0.69 0.46 0.93 3269

0.63 0.62 0.59 0.63 3186

0.034
(0.006)

0.97 0.76 0.74

1.08

0.016
(0.006)
0.060
(0.008)
0.023
(0.006)
0.048
(0.008)

0.006
(0.002)

0.126
(0.029)

0.93 0.76 0.83 0.57 8135

1.30

0.041
(0.017)
0.131
(0.026)
0.069
(0.018)
0.095
(0.024)

0.142
(0.009)
0.082
(0.007)

0.121
(0.016)
0.070
(0.017)
0.062
(0.019)
0.090
(0.018)
0.202
(0.022)
0.180
(0.022)
0.118
(0.017)
0.107
(0.016)
0.077
(0.017)
0.244
(0.023)
0.086
(0.017)
0.176
(0.022)

3.710
(0.777)

0.88 0.69 0.89 0.57 7397

1.24

0.935
(0.064)

16.366
(0.884)

0.995
(0.094)

0.041
(0.014)

0.135
(0.017)

0.076
(0.018)

0.033
(0.006)

0.96 0.75 0.76 0.49 7930

1.655
(0.155)

17.139
(1.083)

0.999
(0.115)

0.022
(0.011)

6.283
(1.465)

0.90 0.78 0.71 0.48 3663

0.881
(0.063)
1.444
(0.133)

15.900
(0.877)
18.074
(1.075)

0.948
(0.094)
1.106
(0.114)

0.099
(0.017)

0.005
(0.002)

0.117
(0.017)
0.076
(0.017)

0.070
(0.018)

0.032
(0.006)

0.006
(0.002)

1.01

0.92 0.67 0.81 0.51

Tangent
Curve

MV Others
PDO
Injury
Non-rainy
Rainy
Dry
Wet

Daytime
7:00AM
9:00PM
Night-time
9:00PM
7:00AM
Working day
Weekend

0.005
(0.002)

0.049
(0.019)
0.191
(0.029)
0.122
(0.023)
0.082
(0.018)

5.397
(1.457)

Note: MV = multi-vehicle; = variable not signicant at 5% level; standard errors of the parameter estimates are reported in parentheses.

Yr11

R2a

AIC

0.69 4371

0.84 0.86 0.43 8000

0.49 8088

0.92 0.96 0.66 3356

0.94 0.52 0.49 2982

1.04

0.50 0.44 4414

0.84 0.77 0.50 7672

4175

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

Singlevehicle
MV Rear End

Yr10

[(Fig._3)TD$IG]

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

173

Fig. 3. Speed distribution in the motorway A56 before and after the P2P implementation.

with respect to skidding depends on vehicle speed, as well as with

the AASHTO Strategic Highway Safety Plan, which recommends
speed enforcement as a strategy particularly applicable to reduce
crashes on horizontal curves (Torbic et al., 2004).
Single vehicle crashes showed the greatest crash reductions:
estimate of 43.7% with a lower 95% condence limit of 31.2%.
Indeed, lower speeds reduce probability of skidding, which is one
of the most common events causing single vehicle run-off-theroad crashes. Furthermore, if a motorist travels onto the roadside
the probability of a crash depends to some extent on the speed of

the vehicle and the drivers experience and capabilities (Neuman

et al., 2003).
Consistently with previous studies (Aarts and van Schagen,
2006; Elvik, 2005, 2013; Hauer, 2009; Montella et al., 2012; Nilson,
2004; Soole et al., 2012; Soole et al., 2013; Stefan, 2006), the speed
change associated with the activation of the P2P system produced
greater effects for the most severe crashes. Estimate of the
reduction of injury crashes was 36.8%, with a lower 95% condence
limit of 25.6%, while reduction of property damage only was 21.6%,
with a lower 95% condence limit of the estimate of 3.3%.

174

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

Table 4
Speed data in the motorway A56 before and after the P2P implementation.
Vm (km/h)

D(Vm)

St.dev (km/h)

D(St.dev)

V85 (km/h)

D(V85)

Light vehicles (weight  3.5 ton)

Before_2009
2,023,834
After_2009
12,379,461
After_2010
3,672,345
After_2011
3,928,972

83.4
72.1
74.8
75.2

0.0%
13.5%
10.3%
9.8%

16.5
11.5
11.7
12.2

0.0%
30.4%
29.2%
26.4%

100.0
82.0
85.3
85.9

0.0%
18.0%
14.7%
14.1%

Heavy vehicles (weight > 3.5 ton)

Before_2009
29,806
After_2009
177,361
After_2010
48,021
After_2011
54,535

72.2
65.3
68.5
68.7

0.0%
9.5%
5.0%
4.8%

13.1
9.9
10.0
10.5

0.0%
24.6%
23.4%
19.7%

85.0
74.0
77.6
77.9

0.0%
12.9%
8.7%
8.3%

Light vehicles (weight  3.5 ton), daytime (7:00 AM9:00 PM)

Before_2009
1651,776
82.7
After_2009
10,138,390
71.5
After_2010
2,951,077
74.0
After_2011
3,257,886
74.6

0.0%
13.5%
10.5%
9.9%

16.4
11.4
11.6
12.2

0.0%
30.6%
29.3%
25.6%

99.0
81.0
84.3
85.3

0.0%
18.2%
14.9%
13.8%

Light vehicles (weight  3.5 ton), night-time (9:00 PM7:00 AM)

Before_2009
372,058
86.2
After_2009
2,241,071
74.9
After_2010
721,268
77.8
After_2011
671,086
78.3

0.0%
13.1%
9.8%
9.2%

16.9
11.7
11.8
11.7

0.0%
30.8%
30.5%
31.2%

103.0
85.0
88.7
89.0

0.0%
17.5%
13.9%
13.6%

Heavy vehicles (weight > 3.5 ton), daytime (7:00 AM9:00 PM)
Before_2009
25,673
71.8
After_2009
152,252
64.8
After_2010
40,675
67.8
After_2011
46,842
68.0

0.0%
9.8%
5.5%
5.3%

13.0
9.8
10.0
10.5

0.0%
24.3%
22.9%
19.0%

84.0
73.0
76.9
77.2

0.0%
13.1%
8.4%
8.1%

Heavy vehicles (weight > 3.5 ton), night-time (9:00 PM7:00 AM)
Before_2009
4133
74.3
After_2009
25,109
68.3
After_2010
7346
72.5
After_2011
7693
72.9

0.0%
8.0%
2.4%
1.8%

13.6
9.6
9.2
9.5

0.0%
29.4%
32.5%
30.0%

88.0
77.0
80.5
81.5

0.0%
12.5%
8.6%
7.4%

Number of observations

Weather and pavement conditions showed similar effects. The

P2P system was more effective in the most demanding conditions:
rainy weather and wet pavements. Estimate of the reduction of
rainy crashes was 57.3% (vs. 29.6% for non-rainy crashes), with a
lower 95% condence limit of 34.8%, and estimate of the reduction
wet crashes was 50.9% (vs. 29.4% for dry crashes), with a lower 95%
condence limit of the estimate of 30.1%.
In daytime (7:00 AM9:00 PM) the safety effectiveness was
greater than in night-time (9:00 PM7:00 AM). Estimate of the
reduction of daytime crashes was 33.4%, with a lower 95% condence
limit of 22.4%, while estimate of the reduction of night-time crashes
was 28.5%, with a lower 95% condence limit of the estimate of 10.3%.
Effects on safety are consistent with speed changes (Fig. 1 and
Table 4). Indeed, mean speed reduction (from before the system
implementation to 2011) was greater in daytime than in night-time:
9.9 vs. 9.2% for light vehicles and 5.3 vs. 1.8% for heavy vehicles.
Crash reduction was not signicant during weekends while
during working days it was equal to 37.6%, with a lower 95%
condence limit of the estimate of 26.9%.
Consistently with the results of the previous study in the Italian
motorway A1 (Montella et al., 2012), the safety effectiveness of the
P2P system was statistically signicant but the effectiveness
decreased over time. Crash reduction estimate was 37.3% in the
rst year after the P2P activation, while it was 29.9% in the second
year and 27.9% in the third year. This declining effect is accompanied
by a declining effect on speed reduction (Fig. 1 and Table 4): 13.5,
10.3 and 9.8% respectively in the rst, second, and third year.
5.3. Safety effects in the spillover sites
Table 7 provides the estimates of the index of effectiveness and
the limits of the 95% condence interval of the estimate of the

crash reduction after the installation the P2P speed enforcement

system in the sections of the motorway A56 where the P2P system
was not installed (length of 22.3 km). It is noteworthy that system
produced statistically signicant crash reductions also in the part
of the motorway where it was not activated (the length of the
section where the system was not activated is 22.3 km while the
length of P2P system activation is 18.1 km). Thus, the spillover
effect of the system was signicant. The crash reduction estimate
for the total crashes is 20.8%, with a lower 95% condence limit of
10.6%.
5.4. Crash modication functions
Since our study showed that both the effects on speed and
safety of the P2P system changed over time, we investigated the
relationship among the crash modication factors and the speed
parameters: mean speed, 85th percentile speed, and standard
deviation of speed. The investigation considered CMFs and speeds
in the four analysis periods: (1) before the installation of the P2P
system, (2) year 2009 after the installation of the P2P system, (3)
year 2010, and (4) year 2011. Based on the different safety effects of
the P2P system in the curves and in the tangents, we performed
different analyses for curves and tangents. As a result, we
developed crash modication functions where the relationship
between crash modication factors and speed parameters was
expressed by a power function: CMF = (Speedafter/Speedbefore)a.
The exponents of the power functions were estimated through a
nonlinear regression model using the SPSS Regression procedure.
Exponents of the power functions are reported in Table 8.
A decrease in mean speed, 85th percentile speed, and standard
deviation of speed is associated with a crash reduction. Results
show that the exponent of the power function is greater on curves

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

175

Table 5
Compliance to the speed limits in the motorway A56 before and after the P2P implementation.

DP(S > SL)

P(S > SL + 10)

DP(S > SL + 10)

P(S > SL + 20)

DP(S > SL + 20)

Light vehicles (weight  3.5 ton)

Before_2009
56.4%
After_2009
19.4%
After_2010
27.5%
After_2011
31.1%

0.0%
65.6%
51.2%
44.8%

31.1%
6.3%
8.4%
8.9%

0.0%
79.7%
73.1%
71.5%

14.1%
2.0%
2.3%
2.2%

0.0%
85.8%
83.5%
84.0%

Heavy vehicles (weight > 3.5 ton)

Before_2009
57.7%
After_2009
30.9%
After_2010
45.0%
After_2011
46.0%

0.0%
46.3%
22.0%
20.3%

26.5%
5.5%
9.4%
10.8%

0.0%
79.1%
64.3%
59.1%

8.8%
1.2%
1.7%
1.9%

0.0%
86.8%
80.7%
78.4%

Light vehicles (weight  3.5 ton), daytime (7:00 AM9:00 PM)

Before_2009
57.4%
0.0%
After_2009
18.1%
68.4%
After_2010
25.7%
55.3%
After_2011
29.7%
48.3%

32.6%
5.5%
7.2%
7.9%

0.0%
83.1%
77.9%
75.6%

14.9%
1.6%
1.8%
1.8%

0.0%
89.5%
88.2%
87.8%

Light vehicles (weight  3.5 ton), night-time (9:00 PM7:00 AM)

Before_2009
62.0%
0.0%
After_2009
25.0%
59.7%
After_2010
34.8%
43.9%
After_2011
38.0%
38.7%

39.0%
9.9%
13.1%
13.4%

0.0%
74.6%
66.4%
65.6%

20.5%
4.0%
4.6%
4.3%

0.0%
80.7%
77.4%
78.9%

Heavy vehicles (weight > 3.5 ton), daytime (7:00 AM9:00 PM)
Before_2009
53.2%
0.0%
After_2009
24.5%
53.9%
After_2010
41.8%
21.4%
After_2011
42.9%
19.5%

22.3%
4.0%
8.2%
9.4%

0.0%
82.1%
63.1%
57.8%

7.2%
0.8%
1.4%
1.7%

0.0%
89.0%
80.5%
76.7%

Heavy vehicles (weight > 3.5 ton), night-time (9:00 PM7:00 AM)
Before_2009
62.7%
0.0%
After_2009
43.2%
31.1%
After_2010
62.4%
0.5%
After_2011
64.8%
3.4%

34.4%
9.8%
16.1%
19.4%

0.0%
71.5%
53.2%
43.7%

12.1%
2.4%
3.3%
3.3%

0.0%
80.1%
72.3%
72.5%

P(S > SL)

Note: DP = 1  Pafter/Pbefore.

and for injury crashes. These results are compared with previous
studies in Table 9. It can be observed that results are similar to the
ones obtained by Elvik (2013) that performed a re-parameterisation of the power model of the relationship between speed and

crashes. For PDO crashes, our study estimated a coefcient of the

power function equal to 2.30 while Elvik estimated a coefcient
equal to 1.91. Elviks exponent for fatal crashes (4.23) was similar to
our exponent for injury and fatal crashes (4.68).

Table 6
Effectiveness of the P2P system in the treatment sites.
Reported
crashes before
(countb)

Expected
crashes
(l)

Reported
crashes
(p)

Index of
effectiveness (u)

Standard
deviation
(s )

Crash reduction

95% Condence interval

lower limit
(%)

Estimate
(%)

95% Condence interval

upper limit
(%)

Total crashes

559

409

279

0.68

0.05

22.3

32.0

41.6

Alignment
Tangent
Curve

214
345

156
253

149
130

0.95
0.51

0.10
0.05

14.8
38.6

4.9
48.8

24.5
59.1

Collision type
Single-vehicle
MV rear end
MV others

213
217
129

196
135
78

111
107
61

0.56
0.79
0.78

0.06
0.09
0.12

31.2
3.0
0.7

43.7
20.9
22.4

56.3
38.8
44.7

Crash severity
PDO
Injury

196
363

128
281

101
178

0.78
0.63

0.09
0.06

3.3
25.6

21.6
36.8

39.9
48.0

Weather conditions
Non-rainy
509
Rainy
50

374
35

264
15

0.70
0.43

0.05
0.11

19.3
34.8

29.6
57.3

40.0
79.8

176

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

Table 6 (Continued)
Reported
crashes before
(countb)

Expected
crashes
(l)

Reported
crashes
(p)

Index of
effectiveness (u )

Crash reduction

Standard
deviation
(s )

95% Condence interval

lower limit
(%)

Estimate
(%)

95% Condence interval

upper limit
(%)

Pavement conditions
Dry
496
Wet
63

361
49

255
24

0.71
0.49

0.05
0.11

18.0
30.1

29.4
50.9

40.0
71.7

Time
Daytime
Night-time

401
158

294
116

196
83

0.67
0.71

0.06
0.09

22.2
10.3

33.4
28.5

44.7
46.7

Day type
Working day
Weekend

402
157

300
109

188
91

0.62
0.83

0.05
0.11

26.9
4.0

37.6
16.9

48.2
37.8

Temporal effects
2009
559
2010
559
2011
559

148
134
127

93
94
92

0.63
0.70
0.72

0.07
0.08
0.08

23.6
14.7
12.0

37.3
29.9
27.9

51.0
45.2
43.7

Table 7
Effectiveness of the P2P system in the spillover sites.
Reported crashes before
(countb)

Expected
crashes
(l)

Reported
crashes
(p)

Index of
effectiveness (u)

Standard
deviation
(s )

Crash reduction

95% Condence
interval lower limit
(%)

Estimate
(%)

95% Condence interval upper limit

(%)

Total crashes

616

453

359

0.79

0.05

10.6

20.8

31.0

Alignment
Tangent
Curve

239
377

173
280

159
200

0.91
0.71

0.09
0.06

9.4
16.5

8.7
28.6

26.8
40.7

Crash type
Single-vehicle
MV rear end
MV others

200
298
118

191
188
73

91
187
81

0.47
0.99
1.10

0.06
0.09
0.15

41.3
16.8
39.9

52.7
0.9
9.7

64.0
18.6
20.5

Crash severity
PDO
Injury

239
377

157
295

139
220

0.88
0.74

0.09
0.06

5.8
13.5

12.0
25.7

29.8
37.9

Weather conditions
Non-rainy
573
Rainy
43

416
37

341
18

0.82
0.49

0.06
0.12

7.4
27.5

18.2
51.1

29.1
74.6

Pavement conditions
Dry
552
Wet
64

399
54

335
24

0.84
0.45

0.06
0.1

5.0
36.8

16.2
55.5

27.5
74.1

Time
Daytime
Night-time

469
147

342
111

290
69

0.85
0.62

0.06
0.09

3.0
21.5

15.3
38.3

27.5
55.1

Day type
Working day
Weekend

438
178

328
125

266
93

0.81
0.74

0.06
0.09

6.9
7.9

19.0
26.0

31.2
44.1

Temporal effects
2009
616
2010
616
2011
616

164
148
141

123
106
130

0.75
0.71
0.92

0.07
0.07
0.09

10.6
13.9
9.5

25.0
28.6
7.8

39.5
43.2
25.2

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

177

Table 8
Estimates of the exponent of the power function.
Vm_after/Vm_before

V85_after/V85_before

St.dev(Vafter)/St.dev(Vbefore)

Curves
Total crashes
PDO crashes
Injury crashes

6.85 (R2 = 0.99)

5.50 (R2 = 0.66)
7.42 (R2 = 0.96)

4.16 (R2 = 0.97)

3.35 (R2 = 0.98)
4.46 (R2 = 0.88)

1.77 (R2 = 0.93)

1.45 (R2 = 0.97)
1.93 (R2 = 0.86)

Tangents
Total crashes
PDO crashes
Injury crashes

0.49 (R2 = 0.60)

0.28 (R2 = 0.23)
1.60 (R2 = 0.56)

0.31 (R2 = 0.55)

0.48 (R2 = 0.08)
0.98 (R2 = 0.59)

0.13 (R2 = 0.50)

0.15 (R2 = 0.03)
0.42 (R2 = 0.59)

Curves + tangents
Total crashes
PDO crashes
Injury crashes

3.84 (R2 = 0.98)

2.30 (R2 = 0.19)
4.68 (R2 = 0.97)

2.43 (R2 = 0.99)

1.19 (R2 = 0.31)
3.10 (R2 = 0.98)

1.03 (R2 = 0.95)

0.67 (R2 = 0.47)
1.24 (R2 = 0.81)

Table 9
Estimates of the exponent of the power function: comparison with previous studies.

Total crashes
PDO crashes
Injury crashes
Fatal crashes
Injury + fatal crashes

Present study

Elvik, 2013

Nilsson, 2004

3.84
2.30

4.68

1.91
2.12
4.23

4.00
2.00

6. Conclusions
The installation of the point-to-point speed enforcement
system on the urban motorway A56 led to very positive effects
on both speed and safety.
As far as the effects on the section average travel speeds, the P2P
system yielded to a considerable reduction in in all the calculated
statistics: the mean speed, the 85th percentile speed, the standard
deviation of speed, and the proportion of drivers exceeding the
speed limits, exceeding the speed limits more than 10 km/h, and
exceeding the speed limits more than 20 km/h. The best results
were the decrease of the speed variability and the reduction of the
excessive speeding behaviour. The decrease in the standard
deviation of speed was 26% while the proportion of light and
heavy vehicles exceeding the speed limits more than 20 km/h was
reduced respectively by 84 and 77%.
As far as the safety effects, the P2P system yielded to a 32%
reduction in the total crashes, with a lower 95% condence limit of
the estimate equal to 22%. The greatest crash reductions were
produced in rainy weather (57%), on wet pavement (51%), on
curves (49%), for single vehicle crashes (44%), and for injury crashes
(37%). It is noteworthy that the P2P system produced a statistically
signicant reduction of 21% in total crashes also in the part of the
motorway where it was not activated (22.3 km vs. 18.1 km), thus
generating a signicant spillover effect.
The investigation of the effects of the P2P system on speed and
safety over time allowed to develop crash modication functions
where the relationship between crash modication factors and
speed parameters (mean speed, 85th percentile speed, and
standard deviation of speed) was expressed by a power function.
Results show that the effect of speed on safety is greater on curves
and for injury crashes. These crash modication functions allow to
quantify the safety effect of the speed changes even though they
are affected by some limitations. Estimates of the speed values in
the different observation periods are based on accurate measurements of the average travel speeds of each vehicle crossing the
enforced sections before and after the system activation. However,
the study is based only on one motorway and the estimates of the

speed parameters in the three-years before period were based only

on twelve days of speed monitoring. More data, including both
other motorways and other road types as well as longer
observation periods, would allow to strengthen the study results.
Furthermore, we developed different crash modication functions
for tangents and curves based on the assumption that average
travel speeds of individual vehicles in the curves and in the
tangents were equal to the average travel speeds of the section
between the two gantries of the P2P system. This hypothesis
should be tested by additional speed monitoring in the different
curves and tangents under investigation.
Even though the study results show excellent outcomes, we
must point out that crash reduction effects decreased over time
and speed, speed variability and non-compliance to speed limits
signicantly increased over time. To maintain its effectiveness over
time, P2P speed enforcement must be actively managed, i.e.,
constantly monitored and supported by appropriate sanctions.
References
AASHTO, 2010. Highway Safety Manual, 1st ed. AASHTO, Washington D.C.
Aarts, L., van Schagen, I., 2006. Driving speed and the risk of road crashes: a review.
Accid. Anal. Prev. 38 (2), 215224.
Akaike, H., 1987. Factor analysis and AIC. Psycometrika 52, 317332.
Brenac, R., 1996. Safety at curves and road geometry standards in some European
countries. Transport. Res. Rec. 1523, 99106.
Cascetta, E., Punzo, V., Montanino, M., 2011. Empirical evidence of speed
management effects on trafc ow at freeway bottleneck. Presented at 90th
Annual Meeting of the Transportation Research Board, Washington D.C..
Cascetta, E., Papola, A., Marzano, V., Simonelli, F., Vitiello, I., 2013. Quasi-dynamic
estimation of od ows from trafc counts: formulation, statistical validation
and performance analysis on real data. Transport. Res. Part B 55, 171187.
Collins, G., McConnell, D., 2008. Speed harmonisation with average speed
enforcement. Trafc Eng. Control 49 (1), 69.
Council, F.M., Reurings, M., Srinivasan, R., Masten, S., Carter, D., 2010. Development
of a Speeding-Related Crash Typology. Report FHWA-HRT-10-024.
DaCoTA, 2012. Speed Enforcement. Deliverable 4.8 of the EC FP7 project DaCoTA.
De Pauw, E., Daniels, S., Brijs, T., Hermans, E., Wets, G., 2014. Behavioural effects of
xed speed cameras on motorways: overall improved speed compliance or
kangaroo jumps? Accid. Anal. Prev. 73, 132140.
Elvik, R., 2005. Speed and road safety: synthesis of evidence from evaluation
studies. Transport. Res. Rec. 1908, 5969.
Elvik, R., 2013. A re-parameterisation of the Power Model of the relationship
between the speed of trafc and the number of accidents and accident victims.
Accid. Anal. Prev. 50, 854860.
Findley, D.J., Hummer, J.E., Rasdorf, W., Zegeer, C.V., Fowler, T.J., 2012. Modeling the
impact of spatial relationships on orizontal curve safety. Accid. Anal. Prev. 45,
296304.
Fleiter, J., Lennon, A., Watson, B., 2010. How do other people inuence your driving
speed? Exploring the who and the how of social inuences on speeding from a
qualitative perspective. Transport. Res. Part F: Trafc Psychol. Behav.13 (1), 4962.
Gross, F., Persaud, B., Lyon, C., 2010. A Guide to Developing Quality Crash
Modication Factors. Report FHWA-SA-10-032, Washington, D.C.
Hauer, E., 1997. Observational Before? After Studies in Road Safety: Estimating the
Effect of Highway and Trafc Engineering Measures on Road Safety. Pergamon
Press (Elsevier Science), Oxford.
Hauer, E., 2009. Speed and safety. Transport. Res. Rec. 2103, 1017.

178

A. Montella et al. / Accident Analysis and Prevention 75 (2015) 164178

Hauer, E., Harwood, D.W., Council, F.M., Grifth, M.S., 2002. Estimating safety by the
empirical Bayes method: a tutorial. Transport.Res. Rec. 1784, 126131.
Italian Ministry of Infrastructures and Transports, 2001. Guidelines for the Design of
Road Infrastructures: D.M. n. 6792, 5/11/2001.
Lynch, M., White, M., Napier, R., 2011. Investigation into the use of point-to-point
speed cameras. NZ Transport Agency Research Report 465, Wellington.
Mannering, F., Bhat, C.R., 2014. Analytic methods in accident research:
methodological frontier and future directions. Anal. Methods Accid. Res. 1, 122.
Marzano, V., Papola, A., Simonelli, F., 2009. Limits and perspectives of effective od
matrix correction using trafc counts. Transport. Res. Part C 17 (2), 120132.
Miaou, S.P., Lu, A., Lum, H., 1996. Pitfalls of using R2 to evaluate goodness of t of
accident prediction models. Transport. Res. Rec. 1542, 613.
Montella, A., 2009. Safety evaluation of curve delineation improvements. Transport.
Res. Rec. 2103, 6979.
Montella, A., Imbriani, L.L., 2014. Safety performance functions incorporating design
consistency variables. Accid. Anal. Prev. 74, 133144.
Montella, A., Colantuoni, L., Lamberti, R., 2008. Crash prediction models for rural
motorways. Transport. Res. Rec. 2083, 180189.
Montella, A., Aria, M., DAmbrosio, A., Galante, F., Mauriello, F., Pernetti, M., 2010.
Perceptual measures to inuence operating speeds and reduce crashes at rural
intersections: driving simulator experiment. Transport. Res. Rec. 2149, 1120.
Montella, A., Aria, M., DAmbrosio, A., Galante, F., Mauriello, F., Pernetti, M., 2011.
Simulator evaluation of drivers speed, deceleration and lateral position at rural
intersections in relation to different perceptual cues. Accid. Anal. Prev. 43 (6),
20722084.
Montella, A., Persaud, B., DApuzzo, M., Imbriani, L.L., 2012. Safety evaluation of an
automated section speed enforcement system. Transport. Res. Rec. 2281, 1625.
Montella, A., Andreassen, D., Tarko, A., Turner, S., Mauriello, F., Imbriani, L.L.,
Romero, M., 2013. Crash databases in Australasia, the European Union, and the
United States: review and prospects for improvement. Transport. Res. Rec. 2386,
128136.
Montella, A., Pariota, L., Galante, F., Imbriani, L.L., Mauriello, F., 2014. Prediction of
drivers speed behaviour on rural motorways based on an instrumented vehicle
study. Transport. Res. Rec. 2434, 5262.
Neuman, T.R., Pferer, R., Slack, K.L., Council, F., McGee, H., Prothe, L., Eccles, K., 2003.
NCHRP Report 500: Guidance for Implementation of the AASHTO Strategic
Highway Safety Plan, Vol. 6: A Guide for Addressing Run-Off-Road Collisions.
Transportation Research Board, Washington, D.C.

Neuman, T.R., Slack, K.L., Hardy, K.K., Bond, V.L., Potts, I., Alberson, B., Lerner, N.,
2009. NCHRP Report 500: Guidance for Implementation of the AASHTO
Strategic Highway Safety Plan, Vol. 23: A Guide for Reducing Speeding-Related
Crashes. Transportation Research Board, Washington, D.C.
Nilsson, G., 2004. Trafc Safety Dimensions and the Power Model to Describe the
Effect of Speed on Safety. Bulletin 221. Lund Institute of Technology Department
of Technology and Society, Trafc Engineering, Lund.
NHTSA, 2012. MMUCC Guideline: Model Minimum Uniform Crash Criteria, 4th ed.
Report DOT HS 811 631.
OECD, 2006. Speed Management. Joint OECD/ECMT Transport Research Centre, Paris.
Olde Karter, M.J.T., van Beek, P., Stemerding, M.P., Havermans, P.F., 2005. Reducing
speed limits on highways: Dutch experiences and impact on air pollution,
noise-level, trafc safety and trafc ow. Proceedings of the ETC Conference
Strassbourg, Strassbourg.
Persaud, B., Lyon, C., 2007. Empirical Bayes beforeafter studies: lessons learned
from two decades of experience and future directions. Accid. Anal. Prev. 39,
546555.
Razali, N., Wah, Y.B., 2011. Power comparisons of ShapiroWilk, Kolmogorov
Smirnov, Lilliefors and AndersonDarling tests. J. Stat. Model. Anal. 2 (1), 2133.
SARTRE Consortium, 2012. European Road Users Risk Perception and Mobility: The
SARTRE 4 Survey. Cestac & Delhomme Editors, France.
Soole, D.W., Watson, B.C., Fleiter, J.J., 2012. Point-to-Point Speed Enforcement.
Austroads Report AP-R415-12, Sydney.
Soole, D.W., Watson, B.C., Fleiter, J.J., 2013. Effects of average speed enforcement on
speed compliance and crashes: a review of the literature. Accid. Anal. Prev. 54,
4656.
Stefan, C., 2006. Section Control Automatic Speed Enforcement in the
Kaisermhlen Tunnel (Vienna, A22 Motorway). Austrian Road Safety Board
(KvF), Vienna.
Torbic, D.J., Harwood, D.W., Gilmore, D.K., Pferer, R., Neuman, T.R., Slack, K.L., Hardy,
K.K., 2004. NCHRP Report 500: Guidance for Implementation of the AASHTO
Strategic Highway Safety Plan. Vol. 7: A Guide for Redducing Collisions on
Horizontal Curves. Transportation Research Board, Washington, D.C.
Washington, S.P., Karlaftis, M.G., Mannering, F.L., 2010. Statistical and Econometric
Methods for Transportation Data Analysis, 2nd ed. Chapmann & Hall/CRC.
Yannis, G., Louca, G., Vardaki, S., Kanelladis, G., 2013. Why do drivers exceed speed
limits. Eur. Transport Res. Rev. 5, 165177.