SIXTH FRAMEWORK PROGRAMME

PRIORITY 1.6. Sustainable Development, Global Change

and Ecosystem
1.6.2: Sustainable Surface Transport

SPECIFIC TARGETED RESEARCH OR INNOVATION PROJECT

Project acronym: RIPCORD - ISEREST

Project full title: Road Infrastructure Safety Protection – Core-Research and
Development for Road Safety in Europe; Increasing safety and
reliability of secondary roads for a sustainable Surface Transport
Contract no.: 50 61 84
Title: Accident Prediction Models and Road safety Impact
Assessment: Results of the pilot studies
Authors: Rob Eenink, Martine Reurings (SWOV), Rune Elvik (TOI), João
Cardoso, Sofia Wichert (LNEC), Christian Stefan (KfV)
Summary: The results of four pilots that were conducted in Austria,
Portugal, Norway and the Netherlands, are confronted with the
results of a preceding state-of-the-art report.
Status: RD (F: final, D: draft, RD: revised draft)
Distribution: PU (PU: public, AP: all Partners, Specific Partners’ names if only
to some, C: Co-ordinator only)
Document ID: RI-SWOV-WP2-R4-Results
Date: 2007-09-14
Project start: 2005-01-01
Duration: 36 months
Table of Contents
1. Introduction........................................................................................................... 3
2. Conclusions of the state-of-the-art study ........................................................... 4
2.1 Introduction .................................................................................................................. 4
2.2 Methodology for Accident Prediction Models ................................................................ 4
2.3 Urban and Rural road segments and intersections....................................................... 6
2.4 Road safety Impact Assessment .................................................................................. 7
3. Pilots ...................................................................................................................... 9
3.1 Austria ......................................................................................................................... 9
3.2 Norway ........................................................................................................................ 9
3.3 Portugal ......................................................................................................................11
3.4 The Netherlands .........................................................................................................11
3.5 Conclusions ................................................................................................................12
4. Comparison of state-of-the-art study and pilots .............................................. 16
4.1 Coverage ....................................................................................................................16
4.2 Methodology ...............................................................................................................16
4.3 Pilot results .................................................................................................................17
5. Discussion of practical use ............................................................................... 19
References .............................................................................................................. 20
Appendix 1 Quality of methodology used ............................................................ 21
Methodology: Austrian pilot ...............................................................................................22
Methodology: Portuguese pilot ..........................................................................................25
Methodology: Netherlands pilot .........................................................................................27

2
1. Introduction
In order to manage road safety, policy makers and road authorities need to have a
good insight in the variables that explain the accident levels on their roads and
networks. To this end many models are and have been built that (try to) predict the
accident frequency, the so-called Accident Prediction Models (APM) or Road safety
Impact Assessments (RIA). In this study an APM is restricted to road types or
categories and a RIA is a model that aggregates, usually quite simple, APM-models
to the level of a network or area.

There is, however, another type of RIA which is comparable to the Environmental
Impact Assessment, EIA. This is used to assess the safety impact of major road
changes or new roads, on the road itself and (usually) including the adjacent road
network [1]. In some countries specific guidelines for RIA’s have been developed [2].
This is normally restricted to the major roads and therefore these RIA’s are
performed by or under the responsibility of the national road authority. Research on
this type of RIA was not included in the original WP2 plans and is therefore only
briefly dealt with in the state-of-the-art report.

Within workpackage 2 of the Ripcord-Iserest project, task 2.2 consists of the

performance of (4) pilots. The choice of these pilots is based on the outcomes of a
state-of-the-art study (task 2.1) that was reported in June 2005 [3]. In task 2.3 the
results of the pilots are confronted with the findings of the state-of-the-art study. This
report reflects the comparison.

In chapter 2 the conclusions of the state-of-the-art study are given. They regard the
preferred methodology for APMs (2.2), APMs for urban and rural roads (2.3) and RIA
(2.4). The results of the pilots are summarised in chapter 3, an APM for Austria (3.1),
a RIA for Norway (3.2), and APMs for Portugal (3.3) and the Netherlands (3.4).
Conclusions are drawn in 3.5. The comparison of both studies is reported in chapter
4, dealing with the coverage of pilots (4.1), the way in which the preferred
methodology is handled (4.2) and how pilot results are related to what is known in
literature (4.3). At the end the usefulness for practitioners is discussed (chapter 5)
followed by references and appendices.

3
2. Conclusions of the state-of-the-art study
2.1 Introduction
The aim of the study was to give an overview of the state-of-the art of Accident
Prediction Models (APM) and Road safety Impact Assessments (RIA). An APM is
usually a (set of) function(s) that describes how safety depends on explanatory
variables like the amount of traffic, length of the road, road width, number of
crossings etc. A RIA does the same for a network but adds to this the possibility to
calculate quantitatively the effects of different scenarios on mobility, (road safety)
measures etc. In a way a RIA is an aggregate of (quite simple) APMs.

The concept of an APM is widely known and many references could be found. It is
therefore the main topic of the study. A RIA is an interesting new development that
has only been applied a few times because it needs so many good quality data
(chapter 5).

2.2 Methodology for Accident Prediction Models

The basic form of nearly all modern accident prediction models for road sections and
intersections is this:

QMIβ e ∑
β γ i xi
E(λ) = αQMA .

The estimated expected number of accidents, E(λ), is a function of traffic volume, Q

(for road sections only one Q), and a set of risk factors, xi (i = 1, 2, 3, …, n). The
effect of traffic volume on accidents is modelled in terms of an elasticity, that is a
power, β, to which traffic volume is raised. Intersections volumes for the major and
minor roads are included. The effects of various risk factors that influence the
probability of accidents, given exposure, is generally modelled as an exponential
function, that is as e (the base of natural logarithms) raised to a sum of the product of
coefficients, γi, and values of the variables, xi, denoting risk factors.

The volume and risk factors are the explanatory variables of the model and, ideally
speaking, the choice of explanatory variables to be included in an accident prediction
model ought to be based on theory. However, the usual basis for choosing
explanatory variables appears to be simply data availability. They should include
variables that:
• have been found in previous studies to exert a major influence on the number
of accidents;
• can be measured in a valid and reliable way;
• are not very highly correlated with other explanatory variables included.

Observed variation in the number of accidents is nearly always a mixture of

systematic and random variation. It is only the systematic part of the variation that
can be explained by means of accident prediction models. There is systematic
variation in number of accidents whenever the variance exceeds the mean. This is

4
referred to as overdispersion. The amount of overdispersion found in a data set, can
be described in terms of the overdispersion parameter:

Var ( x )
−1
µ= λ
λ

Where λ is the expected value of x, the measured variable.

The success of a model in explaining accidents can be evaluated by comparing the

overdispersion parameter of a fitted model to the overdispersion parameter in the
original data set.

The following criteria have been proposed to help assess if a statistical relationship is
causal:
• internal consistency of the relationship, with respect to, for example, subsets
of data in a study or different specifications of multivariate models;
• invariance with respect to potentially confounding factors, meaning that a
relationship does not vanish when potentially confounding factors are
controlled for;
• plausibility in terms of a known causal mechanism or well-established scientific
law;
• support for counterfactual statements, meaning that the relationship has a
genuine predictive capacity.

Testing predictive performance is essential if one wants to support a causal

interpretation of model estimates. One way of doing so is to use the model to predict
accident counts in future years. Another is to use only half the data set to fit a model
and use the other half of the data set to test its predictive performance.

There are many sources of error in accident prediction models. The most frequently
discussed sources of error include:
• omitted variable bias: possibly the most common form of omitted variable bias
in current accident prediction models is the incompleteness of exposure data;
• bias due to co-linearity among explanatory variables: explanatory variables in
accident prediction models tend to be correlated, sometimes to such a high
degree that inclusion of both or all the correlated variables may lead to
unstable estimates of the coefficients;
• wrong functional form for relationships between variables: two typical
problems are related to the use of average values when estimating the
relationship between traffic volume and accidents. The first occurs when traffic
volume is represented by an average value, like AADT (average annual daily
traffic) rather than the actual traffic volume at the time of each accident. The
second occurs when a single function is used when there is reason to believe
that this relationship varies, depending on circumstances like daytime and
darkness.

Development in the field of accident modelling has been so rapid, that some models
that were considered as state-of-the-art only ten years ago look somewhat primitive

5
today. There is today a danger of moving too far in the direction of mathematical
sophistication and perfect fitting of models. A good model however is rather the
simplest possible model that adequately fits the data, and that contains relationships
that may be presumed to hold in general. Based on the discussion in this section, the
following criteria are proposed for assessing the quality of accident prediction
models:
• The probability distribution of accidents in the original data set should be
tested.
• The structure of residuals should always be tested.
• Separate models should be developed for accidents at different levels of
severity.
• Separate models should be developed for different types of roadway
elements.
• Data on exposure should be decomposed to the maximum extent possible.
• The functional form used to describe the relationship between each
independent variable and accidents should be explicitly chosen.
• Explanatory variables should be entered stepwise into the model.
• The correlations between explanatory variables should be examined.
• The overall goodness-of-fit of the final model should be reported.
• The structure of any systematic variation not explained by a model should be
examined.
• Any model should explicitly identify those variables for which a causal
interpretation is sought.
• Explicit operational criteria for causality should be stated in models seeking
causal interpretation of their findings.
• The possible presence of omitted variable bias should always be discussed.
• The predictive performance of an accident prediction model should be tested.
• Accident prediction model should permit results to be synthesised.

These criteria can be further developed into a quality scoring system for accident
prediction models, designed to assign a numerical quality score to each model. This
quality score will be an important piece of information when synthesising the findings
of several accident prediction models.

2.3 Urban and Rural road segments and intersections

Several forms were used for the accident prediction models for rural road sections:

E( µ ) = ax 1b1 ⋅ K ⋅ x nbn ⋅ e ∑
bn +1x n +1
,
R = a + b1 x1 + K + bn x n ,
E( µ ) = e ∑ ∑
a+ bi x i + c ij x i x j
,

where E(µ) is the expected number of accidents, R is the number of accidents per
1000 vehicles and xi are the explanatory variables. Not surprisingly, the Average
Annual Daily Traffic (AADT) and section lengths are used as explanatory variables in
almost all models. Also the minor access density, the carriageway width and the
shoulder width are used in various models. So it is desirable that accident prediction

6
models to be developed include these explanatory variables. In general the model
coefficients were estimated using generalised linear modelling.

In several papers, models are developed for different accident types and/or for
different road types. In general the disaggregated models are better (and simpler)
than the aggregated models, i.e., they have a better fit. So if data is available for
accident types and road characteristics it is better to develop disaggregated models
instead of aggregated models. At least the accident severity (PDO, injury, fatal)
should be modelled separately.

All the models for rural intersections are of the following form:

e∑
b1 b2 β i xi
E( µ ) = aQmin Qmaj ,

which coincides with the general model given in Chapter 2 of the state-of-the-art
report [3]. Therefore it can be concluded that models for intersections should be of
this form. As expected the AADT on the major and minor road are used as
explanatory variables in all models. Also the “presence of left and right-turn lanes on
the major road” is used in several models. Therefore it is desirable that accident
prediction models for rural intersections include these explanatory variables

The models for urban road sections are generally of the form

E( µ ) = αQ β e ∑
γ i xi
,

as described in the methodology section. Any accident prediction model should

include next to the AADT and section length, the public street approach (and
driveway) density as explanatory variables.

The model forms for urban intersections are quite similar, namely

QMIβ 2 e ∑
β1 γ i xi
E( µ ) = αQMA .

In most papers separate models were developed for intersections with three arms
and intersections with four arms and/or for different types of control (STOP,
signalised, major/minor priority, roundabouts). This is desirable, because separate
models for different intersection types give a better description of the data then one
model for all intersections together, which includes the intersection type as an
explanatory variable.

The methodology that is used by the majority of studies to estimate the coefficients of
the accident prediction models is generalised linear modelling (GLM).

2.4 Road safety Impact Assessment

In order to make an APM many data of good quality are required. For a RIA the level
of detail per road may be lower but because a network contains much more roads the
total need for good data is much higher, usually too high. There are (at least) two

7
ways to overcome this problem. The first is an extra incentive to gather data, for
instance by introducing a subsidy scheme for it or demanding a RIA as a prerequisite
for funding road safety plans. Another possibility is to link the RIA to a GIS-system
that is already in use by road authorities for other purposes. The VIB of Diepens &
Okkema (see RipCord-Iserest WP 11 and 12) is a good example of such an
instrument. The VIB will be tested in different situations and countries to see if it can
handle different inputs.

There are of course many other ways for road authorities to get an insight in the
safety situation of their networks and suggestions for improvement. For instance in
WP 6 of RipCord-Iserest the safety analysis of networks and black spot management
are dealt with. Another example is the DUMAS project where a framework for the
design and evaluation of cost-effective and successful urban safety initiatives was
developed and tested. However, the advantage of a RIA is that it enables us to see
quantitative effects of different scenarios (mobility, measures, costs). This gives a
better opportunity to improve the cost-effectiveness of road safety programmes.

8
3. Pilots
3.1 Austria
The Austrian pilot of an APM deals with accidents on Austrian motorways [4]. It
turned out to be possible to divide them into 4 classes: injury accidents, fatalities,
severely injured and slightly injured, instead of the previously reported APM just for
fatalities. One of the important advantages of this is the possibility to compare this
APM to the Portuguese and Dutch where APMs are made for (injury) accidents.

Data have been gathered regarding the following variables:

• average annual daily traffic (AADT);
• number of lanes per road sections;
• speed limit;
• amount of heavy goods vehicles (HGV);
• section length.

In the analysis the standard Generalised Linear Model (GLM see [3]) using a
Negative Binomial Distribution was calculated. The coefficients for speed limits and
number of lanes, however, were not significantly different from zero, and therefore
excluded from the model. The result was:

ACC = 2.388 × 10 −4 × AADT 1.048 × Length 0.889 × 0.986 PHGV

ACC = expected number of injury accidents in 5 years

AADT = Annual Average Daily Traffic [vehicles per day]
Length = length of the road section [km]
PHGV = Proportion of Heavy Goods Vehicles [%]

3.2 Norway
This pilot [9] is a road safety impact assessment for Norway, designed to assess the
prospects for improving road safety. The report is to a large extent based on work
done as part of the development of the National Transport Plan for the 2010-2019
planning term.

A broad survey of potentially effective road safety measures has been performed. A
total of 139 road safety measures were surveyed; 45 of these were included in a
formal impact assessment, which also included cost-benefit analyses. The other 94
road safety measures were for various reasons not included in the impact
assessment. Reasons for exclusion include: (1) Effects of the measure are unknown
or too poorly known to support a formal impact assessment; (2) The measure does
not improve road safety; (3) The measure has been fully implemented in Norway; (4)
The measure overlaps another measure; to prevent double counting, only one
measure was included; (5) The measure is analytically intractable.

For the 45 road safety measures included in the impact assessment, use of these
measures during the period until 2020 was considered. Analyses indicate that 39 of
the 45 measures are cost-effective, i.e. their benefits are greater than the costs
according to cost-benefit analyses. Six of these measures were not cost-effective. A

9
preliminary target of halving the number of road accident fatalities and the number of
road users seriously injured has been set in the National Transport Plan for the term
2010-2019. This plan is as yet not finally developed and the road safety target
proposed has not been officially adopted or given political support. It is nevertheless
of interest to examine if such a target can be realised, as previous road safety impact
assessments in Norway have indicated that it is possible to drastically reduce the
number of fatalities and injuries. The preliminary targets in the National Transport
Plan call for a reduction of fatalities from 250 (annual mean 2003-2006) to 125 in
2020. The number of seriously injured road users is to be reduced from 980 (mean
2003-2006) to 490.

The range of options for improving road safety has been described in terms of four
main policy options, all of which apply to the period from 2007 to 2020:
1. Optimal use of road safety measures: All road safety measures are used up to the
point at which marginal benefits equal marginal costs. The surplus of benefits over
costs will then be maximised.
2. “National” optimal use of road safety measures: Not all road safety measures are
under the control of the Norwegian government; in particular new motor vehicle
safety standards are adopted by international bodies. A version of optimal use of
road safety measures confined to those that can be controlled domestically was
therefore developed.
3. Continuing present policies. This option essentially means that road safety
measures continue to be applied as they currently are. There will not be any increase
in police enforcement, nor will new law be introduced (e.g. a law requiring bicycle
helmets to be worn).
4. Strengthening present policies. In this option, those road safety measures that it is
cost-effective to step up, are stepped up. In particular, this implies a drastic increase
of police enforcement.

Estimates show that all these policy options can be expected to improve road safety
in Norway. The largest reduction of the number of killed or injured road users is
obtained by implementing policy option 1, optimal use of road safety measures. Full
implementation of this policy option results in a predicted number of fatalities of 138
in 2020. The predicted number of seriously injured road users is 656. These numbers
clearly exceed the targets of, respectively, 125 and 490. It is, however, not realistic to
expect road safety measures to be used optimally. In the first place, some of the road
safety measures that may improve road safety if used optimally, are outside the
power of the Norwegian government. This applies to new motor vehicle safety
standards. In the second place, for some road safety measures, optimal use implies
a drastic increase of efforts. This applies to police enforcement. It is, however,
unlikely that the police will increase traffic law enforcement to the optimal extent. In
the third place, optimal use of road related road safety measures requires a
maximally efficient selection of sites for treatment. Current selection of sites for
treatment is not maximally efficient. It would become so, if sites were selected for
treatment according to traffic volume, but this is not easily accomplished in Norway
due to resource allocation mechanisms favouring regional balancing, rather than
economic efficiency. A more realistic policy is therefore that road safety measures
continue to be used along roughly the same lines as they are today. Such a policy
will not bring about large improvements of road safety in Norway. A conservative
estimate for the number of road accident fatalities in 2020 is about 200. A

10
corresponding estimate for seriously injured road users is about 850. While both
these numbers are lower than the current numbers, they are a long way from
realising the targets set for 2020 (125 road users killed, 490 seriously injured).

It should be stressed that the estimates presented in this report are highly uncertain.
It would therefore not be surprising if actual development turns out to be different
from the one estimated.

3.3 Portugal
Accident prediction models for the national road network links are developed, using
as tentative explanatory variables: carriageway width, AADT, number of lanes, and
shoulder type and width. The network is divided in 7 road classes, ranging from
motorway to minor national roads.

The analysis for motorways is reported in Accident Prediction Models for Portuguese
Motorways [5]. According to the state-of-the-art report [1] GLMs were developed as
simple models and extended models. As an example for comparison the outcome for
the simple model and the preferred (see [3]) Negative Binomial distribution is:

ACC = 8 × 10 −4 × AADT 0.917 × Length 0.932

ACC = expected number of injury accidents in 6 years

AADT = Annual Average Daily Traffic [vehicles per day]
Length = length of the road section [km]

An analysis for rural two-lane roads is reported in Accident Prediction Models for
Portuguese Single Carriageway Roads [10].

3.4 The Netherlands

The pilot in the Netherlands was conducted in Haaglanden, which is a mainly urban
area around the city of The Hague [4]. To make comparisons possible again the
GLM outcomes for the simple model and the Negative Binomial distribution are given:

Urban road segments:

ACC = 3.3 × 10 −4 × AADT 0.318 × Length 0.999

Rural road segments:

ACC = 3.740 × 10 −5 × AADT 0.497 × Length 0.965

ACC = expected number of injury accidents in 3 years

AADT = Annual Average Daily Traffic [vehicles per day]
Length = length of the road section [m]

11
3.5 Conclusions
3.5.1 Comparison of 3 road types: motorways, other rural roads, urban roads
For motorways in Austria and Portugal, and for urban and rural roads in the
Netherlands, 4 different APMs were found. To compare them they are given as
expected values of accidents per km road in 5 years and restricted to max. 3 digits:

Austria Motorways ACC = 2.4 × 10 −4 × AADT 1.05 × Length 0.89 × 0.99 PHGV

Portugal Motorways ACC = 6.7 × 10 −4 × AADT 0.92 × Length 0.93

Netherlands Urban ACC = 0.55 × AADT 0.32 × Length 1.0

Netherlands Rural ACC = 0.047 × AADT 0.50 × Length 0.96

At first glance Portuguese motorways seem to have a much greater risk than
Austrian motorways because of the much higher intercept (6.7 × 10 −4 and 2.4 × 10 −4 ).
The best way to compare them is in a plot of ACC density (ACC per km) against
AADT as is done in fig.1. Please note that the range of AADTs is different for different
APMs.

Figure 1 Comparison of 3 road types in 3 countries

16
Netherlands Urban
Netherlands Rural
14
Austria Motorways
Accidents per kilometre in five years

Portugal Motorways
12

10

0
0 5000 10000 15000 20000 25000 30000 35000 40000
AADT

For a typical AADT of 15,000, segment length of 5 km and PHGV of 10% the
outcomes are for Austria ACC= 22.1 (4.4 accidents per km) and for Portugal ACC =
20.8 (4.2). These are quite comparable. With regards to the direction of change it is
understandable that a longer road segment is safer per km because you expect more
homogeneity in traffic flow. In the Austrian model, however, it seems surprisingly that
risk (ACC/(AADT.km)) increases when the AADT increases. In most literature the

12
opposite is reported as indeed is the case in the Portuguese model. In the Austrian
model, however, an extra explanatory variable, the percentage of heavy goods
vehicles, is included, and this may explain these effects. A brief comparison to the
Dutch situation (see [7]) shows that in the Netherlands the accident density is
comparable to the Austrian and Portuguese level, but at approximately the double
AADT, indicating that risk is much lower at high traffic volumes on motorways.

The AADT for urban (3.000 – 40.000) and rural roads (3.000 – 25.000) in the vicinity
of The Hague seems to be rather comparable to motorways in Austria and Portugal.
The city of The Hague has almost 500.000 inhabitants and some of the urban roads
have 2 or 3 lanes per direction of traffic. The influence of segment length is low and
for urban segments negligible. For an AADT of 15000 the accident density (ACC/km)
in 5 years is for urban roads: 11.9 and for rural roads 5.4. At low volumes (AADT of
3000) the accident densities are: Austria 0.8, Portugal 0.9, Netherlands urban 7.1,
and Netherlands rural 2.4. The corresponding risks (ACC/AADT.km) are therefore
much higher for rural and especially urban roads. This is what you would expect, not
because traffic in itself is much safer at high volumes at rural and especially urban
roads, but because road design is adjusted to (expected) high or low volumes. Of
course, one would like to know the effects of different road elements but the data do
not allow incorporating many explanatory variables, such as road design elements.

3.5.2 Comparison of rural two-lane roads in Portugal and Netherlands

After the planned pilots were finished some extra pilots were performed in Portugal
and the Netherlands. An analysis for rural two-lane roads is reported in Accident
Prediction Models for Portuguese Single Carriageway Roads [10]. In the
Netherlands a pilot was done in the province of Gelderland [11]. Portugal has
approximately 9.5 million inhabitants living on 92,000 km2 (100 inh/km2). For
Gelderland these figures are 2 million and 5,000 km2 (400 inh/km2). More details are
given in table 1.

13
 Table 1. Pilot results in Portugal and the Netherlands: rural two-lane roads

Comparison Portugal-Gelderland RipCord APM's rural roads

GLM
Portugal
Lane no av
width sections length volume accidents killed vol/length length acc/length acc/vol AADTpar Lpar
<6 671 7271 49281 18476 950 7 11 2,5 0,37 0,96 0,8
6à7 290 3078 34919 14493 897 11 11 4,7 0,42 0,93 1,02
7 à 7,75 76 843 13661 5167 395 16 11 6,1 0,38 0,91 1,22
>7 58 417 9915 2714 232 24 7 6,5 0,27 0,57 0,88
median 35 418 6042 1316 164 14 12 3,1 0,22 0,26 1,13

Gelderland
single 631 872 12095 2955 14 1,4 3,4 0,24 0,45 0,81
double 133 72 1892 364 26 0,5 5,1 0,19 0,53 0,6
AADT vol5years
single 631 872 7600 12094,64 intersection 44%
double 133 72 14400 1892,16 intersection 57%

To get a better insight the comparison of models is shown in figure 2.

Figure 2 Comparison of two-lane roads in Portugal and the Netherlands

35
Netherlands, single
Crashes on a road segment of 1 km in five years

Netherlands, double
30
Portugal, single, <6m
Portugal, single, 6-7m
25
Portugal, single, 7-7.75m
Portugal, single, > 7.75 m
20
Portugal, double

15

10

0
0 5000 10000 15000 20000 25000 30000 35000 40000
AADT

In most countries rural two-lane roads are the most dangerous, that is, they have the
highest risk (accidents per traffic volume) and highest share of the road safety
problem. It is also the main interest of the RipCord-Iserest project. The road safety
situation is rather different in Portugal and the Netherlands. In 2004 (ERSO-website,
September 2007) the amount of road fatalities per 1 million inhabitants was 125,
whereas in the Netherlands it was 50 (ratio: 2 ½). Therefore one would expect the
risk on rural two-lane roads to be much lower in the Netherlands than in Portugal.

14
From table 1 it is clear that most accidents in Portugal take place on roads without a
median and a road width <7 m: the black and pink lines in figure 1. In Gelderland
these are the ‘enkel’ (single) roads, the dark blue line. For a typical AADT of 15000
Gelderland has about 5 crashes per km in 5 years, for Portugal this is somewhere
between 10 and 15 (ratio 2-3). The situation is therefore what was expected.

The density of inhabitants is rather different between Geldeland (400 inh/km2) and
Portugal (100 inh/km2; ratio 4). The average length of road section is indeed rather
different between Gelderland (single: 1.4 km) and Portugal (< 7m: 11 km). Crashes
on intersections are accounted for in the road section models; this means there are
no separate models for intersections. In the models this is probably the main reason
why ‘length’ is an important explanatory variable with a parameter (Lpar in table 1)
different from “1”. Because the road section is on average much longer in Portugal,
one would expect the Lpar closer to “1”. In fact it is for road widths <7 m somewhere
between 0.8 and 1.02, say 0.9 on average. In Gelderland Lpar is 0.81, indeed what
was expected. For this road type the share of intersection crashes is 44%.

Another remarkable difference is the influence of traffic volume (or intensity AADT;
table 1: vol/length). This influence is expressed by AADTpar (table 1); if this is (far)
below 1 then the influence is high. For Gelderland this is the case (AADTpar = 0.45),
suggesting that traffic behaviour is different at high intensities (p.e. platooning) and/or
road design is different for high volume roads. For Portugal, AADTpar is 0.93-0.96 for
a road width <7m; that means that the influence of intensity (volume) is rather low in
Portugal. If traffic behaviour would explain this difference, there are at least two
possibilities:
- intensity/volume is less varied in Portugal because this is linked to road width.
Indeed the intensity (table 1 vol/length) on <6m wide roads is 7, for 6-7m it is 11, for
7-7.75 m it is 16 and for >7m it is 24. It would be interesting to see the effect of
neglecting road width as an explanatory variable on AADTpar;
- for instance, overtaking behaviour is different in Portugal, that is people accept
smaller gaps to overtake.
If road design is the main explanation it would mean that in Gelderland (Netherlands)
road authorities adjust their roads to volume and expected number of accidents more
than Portuguese road authorities.

15
4. Comparison of state-of-the-art study and pilots
4.1 Coverage
From the state-of-the-art report is was clear that much more (scientific) research is
performed on Accident Prediction Models than on Road safety Impact Assessments.
An important reason is that conducting a RIA is time-consuming and expensive.
Improvements are therefore (also) directed towards simplifications and making use of
already existing data, for instance via GIS (geographical information systems or
digital networks). Via the merger of Ripcord and Iserest this type of RIA is tested in
WP11 and demonstrated in WP12 on a regional scale.

These considerations have lead to the following pilots:

• an APM for motorways (Au);
• APMs for distributor roads with a focus on rural roads (as is the Ripcord-
Iserest project focus)(Nl);
• APMs for road links on all road categories of a national network (P);
• a national RIA (N).

The conclusion is that this choice of pilots nicely covers the spectrum of APM/RIA
that was found in literature.

In addition to these (original) pilots, two extra pilots were performed in 2007, both on
rural (non motorway) roads in Portugal and The Netherlands, therefore addressing
the main topic of Ripcord-Iserest.

4.2 Methodology
In Section 2.2 criteria are given for assessing the quality of APMs (for results: see
appendices).

1. The probability distribution of accidents in the original data set should be

tested.
In all 3 pilots the distribution of the data was not tested a priori. However, after
the models were fitted the data were checked for overdispersion. This resulted
in the conclusion that the data were not Poisson distributed and that the
negative binomial distribution described the data better.
2. The structure of the residuals should always be tested.
This was done in all pilots, but in different ways.
3. Separate models should be developed for accidents at different levels of
severity.
In Austria this was done for injuries, slightly injured, severely injured and
fatalities. In Portugal for collisions, injuries and casualties and in the
Netherlands only for accidents with fatalities or severely injured.
4. Separate models should be developed for different types of roadway
elements.
In neither of the pilots the data allowed for this.
5. Data on exposure should be decomposed to the maximum extent possible.
Only in Austria a decomposition of passenger cars and heavy goods vehicles
was performed.

16
 6. The functional form used to describe the relationship between each
independent variable and accidents should be explicitly chosen.
The general model (see [3]) was used.
7. Explanatory variables should be entered stepwise in the model.
This was dealt with differently: Austria: no, Portugal: yes, Netherlands: not
relevant
8. The correlations between explanatory variables should be examined.
This was done in all cases and no correlation was found.
9. The overall goodness-of-fit of the final model should be reported.
This was discussed extensively in all pilot reports
10. The structure of any systematic variation not explained by the model should be
examined.
In one pilot this is not done, in the others still under discussion.
11. Any model should explicitly identify those variables for which a causal
interpretation is sought.
Does not apply to all pilots.
12. Explicit operational criteria for causality should be stated in models seeking
causal interpretation of their findings.
See 11.
13. The possible presence of omitted variable bias should always be discussed.
In neither of the pilots enough data are available to allow for such an
approach.
14. The predictive performance of an accident prediction model should be tested.
This is done in the Austrian and Portuguese pilot.
15. Accident prediction models should permit results to be synthesised.
The standard errors of all coefficients are reported in all pilots, z-values also in
the Austrian pilot

4.3 Pilot results

In all pilots the general form of APM that was found in the state-of-the-art study was
used. Unfortunately not enough good quality data were available for applying many
explanatory variables and this was an important reason why not all criteria (see 2.2)
could be met and not all preferred variables could be incorporated in the APMs.
Nevertheless, the analyses are considered to be of good quality, albeit a judgement
by the researchers and their international colleagues themselves. A more
independent check is advisable.

The literature study showed that the APM outcomes were rather different in different
regions or countries. In our case, the APMs for motorways in Austria and Portugal
are rather comparable. The APMs for rural two-lane roads in Portugal and the
Netherlands show remarkable, but explainable differences. Within the consortium
there is a common understanding and consensus on how to perform an APM-study
so a clear advice to practitioners and their research consultants in the WP2 tasks to
come, is rather likely. This may well lead to a better comparison of APM and a better
exchange of experiences and knowledge, e.g. allowing for a meta-analysis.

The RIA for Norway shows the possible effects of 4 policy scenarios for road safety
programmes. One of these is considered most likely to occur, that is the continuation
of present policies. It is stressed that many uncertainties exist in the assessment.

17
Nevertheless, the conclusion is that this scenario will (probably) not result in the
desired improvements of road safety: the expected number of road fatalities is
estimated at 200, while the target is 125.

18
5. Discussion of practical use
As stated in the work description of RipCord-Iserest [5] the goal of WP2 is to deliver
an advice to users (road authorities) on the best tool for planning and designing. In
September (21st, 22nd) the 1st RipCord-Iserest Conference (Road Infrastructure
Safety in Europe – moving towards a harmonised approach?) was held and some
preliminary ideas on the usefulness for practitioners of an APM was discussed. They
were:
- developing an APM is not an easy task, probably not suited for road authorities with
the exception of the national level;
- a good and detailed APM requires much data of good quality and detail that is
usually not available;
- as a result only a few explanatory variables are included;
- APM could be used to benchmark one’s roads. If the expected amount of accidents
is significantly lower than what is measured in reality, it is likely that there are some
flaws in design. A RSA or RSI could be advisable and of course, if more explanatory
variables are included they might give a hint to what is the problem. To this end a
standard, and preferably simple, tool for testing the significance of the difference
between expected and real accident values should be made available;
- the APM approach is especially of added value when the black spot treatment
suffers from ‘regression to the mean’ effects.

The results of the RIA in Norway where not yet available at the 1st Ripcord-Iserest
conference. It is expected, however, that these results are useful for practitioners, in
this case, national policy makers. Different scenarios are treated and the likeliness of
these to be implemented is discussed as well. This method gives a clear insight in
what extra measures are needed to meet the road safety targets that were set.

19
References
[1] Commision of the European Communities Proposal for a Directive of the
European Parliament and of the Council on Road Infrastructure Safety Management
[SEC (2006) 1231/1232], Brussels 5 October 2006 COM(2006) 569 final

[2] Höhnscheid, K.-J. (2003). Road safety impact assessment. Bergisch Gladbach,
Bundesanstalt für Strassenwesen. [internal report]

[3] Reurings M., Janssen T., Eenink R., Elvik R., Cardoso J., Stefan C. Accident
Prediction Models and Road safety Impact Assessment: a state-of-the-art. RI-SWOV-
WP23-R1-V2-State-of-the-art.

[4] Stefan C. Predictive model of injury accidents on Austrian motorways. KfV. Vienna
July 2006

[5] Wichert S., Cardoso J. Accident Prediction Models for Portuguese Motorways.
LNEC, Lisbon July 2006

[6] Reurings M. and Janssen S.T.M.C Accident prediction models for urban and rural
carriageways. Based on data from The Hague region Haaglanden. SWOV R-2006-
14, Leidschendam July 2006

[7] Commandeur J., Bijleveld F., Braimaister L., Janssen T. De analyse van ongeval-
, weg-, en verkeerskenmerken van de Nederlandse rijkswegen. SWOV (R-2002-19),
Leidschendam, 2002

[8] RiPCORD-iSEREST ANNEX1-“Description of work” BASt, Bergisch Gladbach

January 20th 2004

[9] Elvik R., Prospects for improving road safety in Norway. Institute of Transport
Economics. Oslo June 2007

[10] Wichert S., Cardoso J., Accident Prediction Models for Portuguese Single
Carriageway Roads. LNEC, Lisbon May 2007

[11] Reurings M., Janssen T., De relatie tussen etmaalintensiteit en het aantal
verkeersongevallen op enkelbaans- en dubbelbaanswegen. SWOV, Leidschendam
October 2007

20
Appendix 1 Quality of methodology used

21
Methodology: Austrian pilot
Criteria Description Yes/No Remarks

According to the State-of the Art report and

published literature on Generalized Linear
Models, accident numbers are assumed to
be Poisson or Negative Binomial distributed.
Left out It is important to realize that the distributional
1 form of an APM is determined by the random
intentionally
Testing the probability component of the dependent variable. A
distribution of accidents in fundamental aspect of any regression
the original data set analysis is that the researcher assumes that
the model pulls out the systematic part of the
Y’s, i.e. the dependent variable. What is left
over (termed errors or residuals) is assumed
to have a particular probability distribution. It
is this distribution that determines the link
function
2 Yes
The structure of residuals Poisson and Negative Binomial models were
should always be tested used to test the optimal model fit

In the Austrian pilot, models were developed

for:

• Injury accidents

3 Yes • Fatalities
Separate models should
be developed for
accidents at different
levels of severity

• Severely injured

• Slightly injured
In Austria, data on property damage only
accidents are not collected in a systematic
way and therefore could not be incorporated
into the models.

22
4 No Due to missing data on specific roadway
Separate models should elements such as bridges, tunnels, slip
be developed for different roads, etc. models in this pilot are restricted
types of roadway elements to road sections as the only element
analysed.
Traffic volume on motorways primarily
consists of passenger cars and lorries.
5 Yes Pedestrians and cyclists are not allowed to
Data on exposure should use motorways. Since the pilot uses the
be decomposed to the proportion of Heavy Goods Vehicles (PHGV)
maximum extent possible and Annual Average Daily Traffic as
explanatory variables, decomposition of
exposure is conducted to the highest
possible extent.
The functional form of the pilot matches the
The functional form used model choice of the State-of-the Art report
6 Yes
to describe the which can be regarded as best practice
relationship should be guideline for modelling accident data.
explicitly chosen and
reasons given for the
choice

9 different model forms were used to

examine the stability of the coefficients with
respect to the variables included. The
7 Yes unconstrained model (containing all
Explanatory variables explanatory variables) yielded the smallest
should be entered overdispersion parameter and therefore was
stepwise in the model considered best for fitting the data.
z-values were computed to answer the
question of whether there is a statistically
significant relationship between an
independent and the dependent variable.
The correlations between
8 explanatory variables Yes
Correlation analysis is part of the modelling
should be examined
procedure of the software package GLIM.
The correlation between the explanatory
variables was found to be of no concern.

A measure of goodness-of-fit of the model

(the closeness of the observed to the fitted
values) is the deviance. This summary
measure is always positive and will be small
if the fit is good. For a well-fitting model for
9 (Yes) Poisson-distributed data, the deviance
The overall goodness-of-fit should be close to the residual degrees of
of the final model should freedom (df); equivalently, the mean
be reported deviance the deviance divided by the df)
should be close to 1.
In the pilot study, the mean deviance was
calculated to distinguish the goodness-of-fit
of the 9 models used during the modelling
process. The model with the lowest mean
deviance was chosen for further calculations

23
 (see chapter 2.3).
10 No
The structure of any
systematic variation not
explained by a model
should be examined

11 Any model should No

explicitly identify those
variables for which a
causal interpretation is
sought
The choice of variables in the model is above
12 Explicit operational criteria No
all based on data availability
for causality should be
stated in models seeking
causal interpretation of
their findings

13 No
The possible presence of
omitted variable bias
should always be
discussed

Considering injury accidents, the pilot

overestimates the observed values in most
14 Yes
The predictive cases. Similar results apply to fatal and
performance of an slightly injured, respectively
accident prediction model
should be tested

15 Yes Standard errors and z-values are reported

Accident prediction
models should permit
results to be synthesised

24
Methodology: Portuguese pilot

1. The probability distribution of accidents in the original data set should be

tested.
The distribution of accidents was not formally tested. This was because the
number of accidents consists of count data and the benchmark model for
count data is the Poisson distribution.
Once the model was fitted to the data it was observed that the variance was
larger than the mean, meaning that overdispersion occurred. A standard
generalisation of the Poisson distribution is then the Negative Binomial
distribution which allows a dispersion parameter.

2. The structure of the residuals should always be tested.

The residuals obtained by the fitted models were used for diagnostic checks,
namely in visual analysis.

3. Separate models should be developed for accidents at different levels of

severity.
The dependent variable consisted on the number of accidents for all the fitted
models considered.

4. Separate models should be developed for different types of roadway

elements.
The available data at the time of the first draft of WP2 report consisted only on
sections of the Portuguese motorway network. In the mean time more data
has been made available consisting on road sections with one and two lanes
each as well as intersections. These are presently being analysed.

5. Data on exposure should be decomposed to the maximum extent possible.

The data available on the Portuguese road network does not specify the type
of vehicles contributing to the traffic.

6. The functional form used to describe the relationship between each

independent variable and accidents should be explicitly chosen and reasons
given for the choice.
The functional forms used were based on the state-of-the-art report which
concluded that there is a basic form for nearly all modern accident prediction
models.

7. Explanatory variables should be entered stepwise in the model.

The explanatory variables for the Portuguese motorway network consisted of
AADT, Length and Number of Lanes. They were entered stepwise in the
model.

8. The correlations between explanatory variables should be examined.

The AADT and Length were not correlated.

9. The overall goodness-of-fit of the final model should be reported.

25
 The goodness-of-fit of the models was calculated using the Freeman-Tukey
R 2 and the Elvik index.

10. The structure of any systematic variation not explained by the model should be
examined.
The systematic variation in the number of accidents was not calculated. It is
now under investigation.

11. Any model should explicitly identify those variables for which a causal
interpretation is sought.
Not applicable in the Portuguese case.

12. Explicit operational criteria for causality should be stated in models seeking
causal interpretation of their findings.
Not applicable in the Portuguese case.

13. The possible presence of omitted variable bias should always be discussed.
It is hoped that with the new available data (by having more explanatory
variables) this issue would be better tackled.

14. The predictive performance of an accident prediction model should be tested.

It was tested in some models which include the explanatory variable Lanes.
The model was applied on simulated data.

15. Accident prediction models should permit results to be synthesised.

The coefficient estimates and their corresponding standard errors were
reported for all the accident prediction models considered.

26
Methodology: Netherlands pilot
In Section 2.12 of the state-of-the-art report about accident prediction models en road
safety impact assessments a total of 15 criteria are given for assessing the quality of
APMs. Each of these criteria will be discussed for the pilot study carried out in the
Netherlands.

1. The probability distribution of accidents in the original data set should be

tested. The distribution of the data was not tested a priori. However, after the
models were fitted the data were checked for overdispersion. This resulted in
the conclusion that the data were not Poisson distributed and that the negative
binomial distribution described the data better.
2. The structure of the residuals should always be tested. For each of the fitted
models the residuals were tested with the help of several plots. The tests were
used to determine if there was heteroscedasticity and whether or not the
residuals were normally distributed. Especially for the models based on the
negative binomial distribution there was no reason to believe that the residuals
were not normally distributed or not homoscedastic.
3. Separate models should be developed for accidents at different levels of
severity. In the Dutch pilots only KSI accidents were considered.
4. Separate models should be developed for different types of roadway
elements. It was only possible to develop models for road segments and not
for intersections. For intersections information is needed about the minor and
major traffic flows and this was not available. It was also not known if the
considered road sections were tunnels or bridges.
5. Data on exposure should be decomposed to the maximum extent possible.
Only the AADT for motor vehicles was available.
6. The functional form used to describe the relationship between each
independent variable and accidents should be explicitly chosen. The length
and AADT are the only explanatory variables and the way they are included in
the model in the same way as in the literature. Only the so called extended
model was slightly different. The model form was indeed checked with the
data.
7. Explanatory variables should be entered stepwise in the model. Does not
apply to the Dutch pilots.
8. The correlations between explanatory variables should be examined. The
length and AADT were not correlated.
9. The overall goodness-of-fit of the final model should be reported. The
goodness-of-fit of the models was discussed very extensively. However, in the
SotA it is proposed that the variation has to be decomposed in (not) explained
systematic variation and random variation.
10. The structure of any systematic variation not explained by the model should be
examined. See 9.
11. Any model should explicitly identify those variables for which a causal
interpretation is sought. Does not apply to the Dutch pilots.
12. Explicit operational criteria for causality should be stated in models seeking
causal interpretation of their findings. Does not apply to the Dutch pilots.
13. The possible presence of omitted variable bias should always be discussed. In
the Dutch pilots there is a small discussion about the model form. It is
explained that this form does not indicate that roads with high AADT are saver,

27
 but that roads with high AADT are designed to handle this high AADT in a
save manner. It is preferred to develop separate models for different road
types, because then this problem should be eliminated.
14. The predictive performance of an accident prediction model should be tested.
This is not done.
15. Accident prediction models should permit results to be synthesised. The
standard errors of all coefficients are reported.

There are some criteria which do not apply to the Dutch pilots. These are in particular
the criteria involving explanatory variables. We wanted to developed separate models
for different road types instead of entering a lot of explanatory variables into the
model. Due to the limited data that was available this was not yet possible. At the
moment we are working on APMs for a few provinces in the Netherlands and it
seems that we can divide the data into different categories.

28